1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where 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/ParamAttrsList.h"
43 #include "llvm/Analysis/ConstantFolding.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
63 using namespace llvm::PatternMatch;
65 STATISTIC(NumCombined , "Number of insts combined");
66 STATISTIC(NumConstProp, "Number of constant folds");
67 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
68 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
69 STATISTIC(NumSunkInst , "Number of instructions sunk");
72 class VISIBILITY_HIDDEN InstCombiner
73 : public FunctionPass,
74 public InstVisitor<InstCombiner, Instruction*> {
75 // Worklist of all of the instructions that need to be simplified.
76 std::vector<Instruction*> Worklist;
77 DenseMap<Instruction*, unsigned> WorklistMap;
79 bool MustPreserveLCSSA;
81 static char ID; // Pass identification, replacement for typeid
82 InstCombiner() : FunctionPass((intptr_t)&ID) {}
84 /// AddToWorkList - Add the specified instruction to the worklist if it
85 /// isn't already in it.
86 void AddToWorkList(Instruction *I) {
87 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
88 Worklist.push_back(I);
91 // RemoveFromWorkList - remove I from the worklist if it exists.
92 void RemoveFromWorkList(Instruction *I) {
93 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
94 if (It == WorklistMap.end()) return; // Not in worklist.
96 // Don't bother moving everything down, just null out the slot.
97 Worklist[It->second] = 0;
99 WorklistMap.erase(It);
102 Instruction *RemoveOneFromWorkList() {
103 Instruction *I = Worklist.back();
105 WorklistMap.erase(I);
110 /// AddUsersToWorkList - When an instruction is simplified, add all users of
111 /// the instruction to the work lists because they might get more simplified
114 void AddUsersToWorkList(Value &I) {
115 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
117 AddToWorkList(cast<Instruction>(*UI));
120 /// AddUsesToWorkList - When an instruction is simplified, add operands to
121 /// the work lists because they might get more simplified now.
123 void AddUsesToWorkList(Instruction &I) {
124 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
125 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
129 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
130 /// dead. Add all of its operands to the worklist, turning them into
131 /// undef's to reduce the number of uses of those instructions.
133 /// Return the specified operand before it is turned into an undef.
135 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
136 Value *R = I.getOperand(op);
138 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
139 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
141 // Set the operand to undef to drop the use.
142 I.setOperand(i, UndefValue::get(Op->getType()));
149 virtual bool runOnFunction(Function &F);
151 bool DoOneIteration(Function &F, unsigned ItNum);
153 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
154 AU.addRequired<TargetData>();
155 AU.addPreservedID(LCSSAID);
156 AU.setPreservesCFG();
159 TargetData &getTargetData() const { return *TD; }
161 // Visitation implementation - Implement instruction combining for different
162 // instruction types. The semantics are as follows:
164 // null - No change was made
165 // I - Change was made, I is still valid, I may be dead though
166 // otherwise - Change was made, replace I with returned instruction
168 Instruction *visitAdd(BinaryOperator &I);
169 Instruction *visitSub(BinaryOperator &I);
170 Instruction *visitMul(BinaryOperator &I);
171 Instruction *visitURem(BinaryOperator &I);
172 Instruction *visitSRem(BinaryOperator &I);
173 Instruction *visitFRem(BinaryOperator &I);
174 Instruction *commonRemTransforms(BinaryOperator &I);
175 Instruction *commonIRemTransforms(BinaryOperator &I);
176 Instruction *commonDivTransforms(BinaryOperator &I);
177 Instruction *commonIDivTransforms(BinaryOperator &I);
178 Instruction *visitUDiv(BinaryOperator &I);
179 Instruction *visitSDiv(BinaryOperator &I);
180 Instruction *visitFDiv(BinaryOperator &I);
181 Instruction *visitAnd(BinaryOperator &I);
182 Instruction *visitOr (BinaryOperator &I);
183 Instruction *visitXor(BinaryOperator &I);
184 Instruction *visitShl(BinaryOperator &I);
185 Instruction *visitAShr(BinaryOperator &I);
186 Instruction *visitLShr(BinaryOperator &I);
187 Instruction *commonShiftTransforms(BinaryOperator &I);
188 Instruction *visitFCmpInst(FCmpInst &I);
189 Instruction *visitICmpInst(ICmpInst &I);
190 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
191 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
194 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
195 ConstantInt *DivRHS);
197 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
198 ICmpInst::Predicate Cond, Instruction &I);
199 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
201 Instruction *commonCastTransforms(CastInst &CI);
202 Instruction *commonIntCastTransforms(CastInst &CI);
203 Instruction *commonPointerCastTransforms(CastInst &CI);
204 Instruction *visitTrunc(TruncInst &CI);
205 Instruction *visitZExt(ZExtInst &CI);
206 Instruction *visitSExt(SExtInst &CI);
207 Instruction *visitFPTrunc(FPTruncInst &CI);
208 Instruction *visitFPExt(CastInst &CI);
209 Instruction *visitFPToUI(CastInst &CI);
210 Instruction *visitFPToSI(CastInst &CI);
211 Instruction *visitUIToFP(CastInst &CI);
212 Instruction *visitSIToFP(CastInst &CI);
213 Instruction *visitPtrToInt(CastInst &CI);
214 Instruction *visitIntToPtr(IntToPtrInst &CI);
215 Instruction *visitBitCast(BitCastInst &CI);
216 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
218 Instruction *visitSelectInst(SelectInst &CI);
219 Instruction *visitCallInst(CallInst &CI);
220 Instruction *visitInvokeInst(InvokeInst &II);
221 Instruction *visitPHINode(PHINode &PN);
222 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
223 Instruction *visitAllocationInst(AllocationInst &AI);
224 Instruction *visitFreeInst(FreeInst &FI);
225 Instruction *visitLoadInst(LoadInst &LI);
226 Instruction *visitStoreInst(StoreInst &SI);
227 Instruction *visitBranchInst(BranchInst &BI);
228 Instruction *visitSwitchInst(SwitchInst &SI);
229 Instruction *visitInsertElementInst(InsertElementInst &IE);
230 Instruction *visitExtractElementInst(ExtractElementInst &EI);
231 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
233 // visitInstruction - Specify what to return for unhandled instructions...
234 Instruction *visitInstruction(Instruction &I) { return 0; }
237 Instruction *visitCallSite(CallSite CS);
238 bool transformConstExprCastCall(CallSite CS);
239 Instruction *transformCallThroughTrampoline(CallSite CS);
242 // InsertNewInstBefore - insert an instruction New before instruction Old
243 // in the program. Add the new instruction to the worklist.
245 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
246 assert(New && New->getParent() == 0 &&
247 "New instruction already inserted into a basic block!");
248 BasicBlock *BB = Old.getParent();
249 BB->getInstList().insert(&Old, New); // Insert inst
254 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
255 /// This also adds the cast to the worklist. Finally, this returns the
257 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
259 if (V->getType() == Ty) return V;
261 if (Constant *CV = dyn_cast<Constant>(V))
262 return ConstantExpr::getCast(opc, CV, Ty);
264 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
269 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
270 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
274 // ReplaceInstUsesWith - This method is to be used when an instruction is
275 // found to be dead, replacable with another preexisting expression. Here
276 // we add all uses of I to the worklist, replace all uses of I with the new
277 // value, then return I, so that the inst combiner will know that I was
280 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
281 AddUsersToWorkList(I); // Add all modified instrs to worklist
283 I.replaceAllUsesWith(V);
286 // If we are replacing the instruction with itself, this must be in a
287 // segment of unreachable code, so just clobber the instruction.
288 I.replaceAllUsesWith(UndefValue::get(I.getType()));
293 // UpdateValueUsesWith - This method is to be used when an value is
294 // found to be replacable with another preexisting expression or was
295 // updated. Here we add all uses of I to the worklist, replace all uses of
296 // I with the new value (unless the instruction was just updated), then
297 // return true, so that the inst combiner will know that I was modified.
299 bool UpdateValueUsesWith(Value *Old, Value *New) {
300 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
302 Old->replaceAllUsesWith(New);
303 if (Instruction *I = dyn_cast<Instruction>(Old))
305 if (Instruction *I = dyn_cast<Instruction>(New))
310 // EraseInstFromFunction - When dealing with an instruction that has side
311 // effects or produces a void value, we can't rely on DCE to delete the
312 // instruction. Instead, visit methods should return the value returned by
314 Instruction *EraseInstFromFunction(Instruction &I) {
315 assert(I.use_empty() && "Cannot erase instruction that is used!");
316 AddUsesToWorkList(I);
317 RemoveFromWorkList(&I);
319 return 0; // Don't do anything with FI
323 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
324 /// InsertBefore instruction. This is specialized a bit to avoid inserting
325 /// casts that are known to not do anything...
327 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
328 Value *V, const Type *DestTy,
329 Instruction *InsertBefore);
331 /// SimplifyCommutative - This performs a few simplifications for
332 /// commutative operators.
333 bool SimplifyCommutative(BinaryOperator &I);
335 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
336 /// most-complex to least-complex order.
337 bool SimplifyCompare(CmpInst &I);
339 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
340 /// on the demanded bits.
341 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
342 APInt& KnownZero, APInt& KnownOne,
345 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
346 uint64_t &UndefElts, unsigned Depth = 0);
348 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
349 // PHI node as operand #0, see if we can fold the instruction into the PHI
350 // (which is only possible if all operands to the PHI are constants).
351 Instruction *FoldOpIntoPhi(Instruction &I);
353 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
354 // operator and they all are only used by the PHI, PHI together their
355 // inputs, and do the operation once, to the result of the PHI.
356 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
357 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
360 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
361 ConstantInt *AndRHS, BinaryOperator &TheAnd);
363 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
364 bool isSub, Instruction &I);
365 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
366 bool isSigned, bool Inside, Instruction &IB);
367 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
368 Instruction *MatchBSwap(BinaryOperator &I);
369 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
370 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
373 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
376 char InstCombiner::ID = 0;
377 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
380 // getComplexity: Assign a complexity or rank value to LLVM Values...
381 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
382 static unsigned getComplexity(Value *V) {
383 if (isa<Instruction>(V)) {
384 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
388 if (isa<Argument>(V)) return 3;
389 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
392 // isOnlyUse - Return true if this instruction will be deleted if we stop using
394 static bool isOnlyUse(Value *V) {
395 return V->hasOneUse() || isa<Constant>(V);
398 // getPromotedType - Return the specified type promoted as it would be to pass
399 // though a va_arg area...
400 static const Type *getPromotedType(const Type *Ty) {
401 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
402 if (ITy->getBitWidth() < 32)
403 return Type::Int32Ty;
408 /// getBitCastOperand - If the specified operand is a CastInst or a constant
409 /// expression bitcast, return the operand value, otherwise return null.
410 static Value *getBitCastOperand(Value *V) {
411 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
412 return I->getOperand(0);
413 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
414 if (CE->getOpcode() == Instruction::BitCast)
415 return CE->getOperand(0);
419 /// This function is a wrapper around CastInst::isEliminableCastPair. It
420 /// simply extracts arguments and returns what that function returns.
421 static Instruction::CastOps
422 isEliminableCastPair(
423 const CastInst *CI, ///< The first cast instruction
424 unsigned opcode, ///< The opcode of the second cast instruction
425 const Type *DstTy, ///< The target type for the second cast instruction
426 TargetData *TD ///< The target data for pointer size
429 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
430 const Type *MidTy = CI->getType(); // B from above
432 // Get the opcodes of the two Cast instructions
433 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
434 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
436 return Instruction::CastOps(
437 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
438 DstTy, TD->getIntPtrType()));
441 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
442 /// in any code being generated. It does not require codegen if V is simple
443 /// enough or if the cast can be folded into other casts.
444 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
445 const Type *Ty, TargetData *TD) {
446 if (V->getType() == Ty || isa<Constant>(V)) return false;
448 // If this is another cast that can be eliminated, it isn't codegen either.
449 if (const CastInst *CI = dyn_cast<CastInst>(V))
450 if (isEliminableCastPair(CI, opcode, Ty, TD))
455 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
456 /// InsertBefore instruction. This is specialized a bit to avoid inserting
457 /// casts that are known to not do anything...
459 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
460 Value *V, const Type *DestTy,
461 Instruction *InsertBefore) {
462 if (V->getType() == DestTy) return V;
463 if (Constant *C = dyn_cast<Constant>(V))
464 return ConstantExpr::getCast(opcode, C, DestTy);
466 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
469 // SimplifyCommutative - This performs a few simplifications for commutative
472 // 1. Order operands such that they are listed from right (least complex) to
473 // left (most complex). This puts constants before unary operators before
476 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
477 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
479 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
480 bool Changed = false;
481 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
482 Changed = !I.swapOperands();
484 if (!I.isAssociative()) return Changed;
485 Instruction::BinaryOps Opcode = I.getOpcode();
486 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
487 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
488 if (isa<Constant>(I.getOperand(1))) {
489 Constant *Folded = ConstantExpr::get(I.getOpcode(),
490 cast<Constant>(I.getOperand(1)),
491 cast<Constant>(Op->getOperand(1)));
492 I.setOperand(0, Op->getOperand(0));
493 I.setOperand(1, Folded);
495 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
496 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
497 isOnlyUse(Op) && isOnlyUse(Op1)) {
498 Constant *C1 = cast<Constant>(Op->getOperand(1));
499 Constant *C2 = cast<Constant>(Op1->getOperand(1));
501 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
502 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
503 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
507 I.setOperand(0, New);
508 I.setOperand(1, Folded);
515 /// SimplifyCompare - For a CmpInst this function just orders the operands
516 /// so that theyare listed from right (least complex) to left (most complex).
517 /// This puts constants before unary operators before binary operators.
518 bool InstCombiner::SimplifyCompare(CmpInst &I) {
519 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
522 // Compare instructions are not associative so there's nothing else we can do.
526 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
527 // if the LHS is a constant zero (which is the 'negate' form).
529 static inline Value *dyn_castNegVal(Value *V) {
530 if (BinaryOperator::isNeg(V))
531 return BinaryOperator::getNegArgument(V);
533 // Constants can be considered to be negated values if they can be folded.
534 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
535 return ConstantExpr::getNeg(C);
539 static inline Value *dyn_castNotVal(Value *V) {
540 if (BinaryOperator::isNot(V))
541 return BinaryOperator::getNotArgument(V);
543 // Constants can be considered to be not'ed values...
544 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
545 return ConstantInt::get(~C->getValue());
549 // dyn_castFoldableMul - If this value is a multiply that can be folded into
550 // other computations (because it has a constant operand), return the
551 // non-constant operand of the multiply, and set CST to point to the multiplier.
552 // Otherwise, return null.
554 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
555 if (V->hasOneUse() && V->getType()->isInteger())
556 if (Instruction *I = dyn_cast<Instruction>(V)) {
557 if (I->getOpcode() == Instruction::Mul)
558 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
559 return I->getOperand(0);
560 if (I->getOpcode() == Instruction::Shl)
561 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
562 // The multiplier is really 1 << CST.
563 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
564 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
565 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
566 return I->getOperand(0);
572 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
573 /// expression, return it.
574 static User *dyn_castGetElementPtr(Value *V) {
575 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
576 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
577 if (CE->getOpcode() == Instruction::GetElementPtr)
578 return cast<User>(V);
582 /// AddOne - Add one to a ConstantInt
583 static ConstantInt *AddOne(ConstantInt *C) {
584 APInt Val(C->getValue());
585 return ConstantInt::get(++Val);
587 /// SubOne - Subtract one from a ConstantInt
588 static ConstantInt *SubOne(ConstantInt *C) {
589 APInt Val(C->getValue());
590 return ConstantInt::get(--Val);
592 /// Add - Add two ConstantInts together
593 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
594 return ConstantInt::get(C1->getValue() + C2->getValue());
596 /// And - Bitwise AND two ConstantInts together
597 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
598 return ConstantInt::get(C1->getValue() & C2->getValue());
600 /// Subtract - Subtract one ConstantInt from another
601 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
602 return ConstantInt::get(C1->getValue() - C2->getValue());
604 /// Multiply - Multiply two ConstantInts together
605 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
606 return ConstantInt::get(C1->getValue() * C2->getValue());
608 /// MultiplyOverflows - True if the multiply can not be expressed in an int
610 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
611 uint32_t W = C1->getBitWidth();
612 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
621 APInt MulExt = LHSExt * RHSExt;
624 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
625 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
626 return MulExt.slt(Min) || MulExt.sgt(Max);
628 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
631 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
632 /// known to be either zero or one and return them in the KnownZero/KnownOne
633 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
635 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
636 /// we cannot optimize based on the assumption that it is zero without changing
637 /// it to be an explicit zero. If we don't change it to zero, other code could
638 /// optimized based on the contradictory assumption that it is non-zero.
639 /// Because instcombine aggressively folds operations with undef args anyway,
640 /// this won't lose us code quality.
641 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
642 APInt& KnownOne, unsigned Depth = 0) {
643 assert(V && "No Value?");
644 assert(Depth <= 6 && "Limit Search Depth");
645 uint32_t BitWidth = Mask.getBitWidth();
646 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
647 KnownZero.getBitWidth() == BitWidth &&
648 KnownOne.getBitWidth() == BitWidth &&
649 "V, Mask, KnownOne and KnownZero should have same BitWidth");
650 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
651 // We know all of the bits for a constant!
652 KnownOne = CI->getValue() & Mask;
653 KnownZero = ~KnownOne & Mask;
657 if (Depth == 6 || Mask == 0)
658 return; // Limit search depth.
660 Instruction *I = dyn_cast<Instruction>(V);
663 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
664 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
666 switch (I->getOpcode()) {
667 case Instruction::And: {
668 // If either the LHS or the RHS are Zero, the result is zero.
669 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
670 APInt Mask2(Mask & ~KnownZero);
671 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
672 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
673 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
675 // Output known-1 bits are only known if set in both the LHS & RHS.
676 KnownOne &= KnownOne2;
677 // Output known-0 are known to be clear if zero in either the LHS | RHS.
678 KnownZero |= KnownZero2;
681 case Instruction::Or: {
682 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
683 APInt Mask2(Mask & ~KnownOne);
684 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
685 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
686 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
688 // Output known-0 bits are only known if clear in both the LHS & RHS.
689 KnownZero &= KnownZero2;
690 // Output known-1 are known to be set if set in either the LHS | RHS.
691 KnownOne |= KnownOne2;
694 case Instruction::Xor: {
695 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
696 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
697 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
698 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
700 // Output known-0 bits are known if clear or set in both the LHS & RHS.
701 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
702 // Output known-1 are known to be set if set in only one of the LHS, RHS.
703 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
704 KnownZero = KnownZeroOut;
707 case Instruction::Select:
708 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
709 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
710 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
711 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
713 // Only known if known in both the LHS and RHS.
714 KnownOne &= KnownOne2;
715 KnownZero &= KnownZero2;
717 case Instruction::FPTrunc:
718 case Instruction::FPExt:
719 case Instruction::FPToUI:
720 case Instruction::FPToSI:
721 case Instruction::SIToFP:
722 case Instruction::PtrToInt:
723 case Instruction::UIToFP:
724 case Instruction::IntToPtr:
725 return; // Can't work with floating point or pointers
726 case Instruction::Trunc: {
727 // All these have integer operands
728 uint32_t SrcBitWidth =
729 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
731 MaskIn.zext(SrcBitWidth);
732 KnownZero.zext(SrcBitWidth);
733 KnownOne.zext(SrcBitWidth);
734 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
735 KnownZero.trunc(BitWidth);
736 KnownOne.trunc(BitWidth);
739 case Instruction::BitCast: {
740 const Type *SrcTy = I->getOperand(0)->getType();
741 if (SrcTy->isInteger()) {
742 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
747 case Instruction::ZExt: {
748 // Compute the bits in the result that are not present in the input.
749 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
750 uint32_t SrcBitWidth = SrcTy->getBitWidth();
753 MaskIn.trunc(SrcBitWidth);
754 KnownZero.trunc(SrcBitWidth);
755 KnownOne.trunc(SrcBitWidth);
756 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
757 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
758 // The top bits are known to be zero.
759 KnownZero.zext(BitWidth);
760 KnownOne.zext(BitWidth);
761 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
764 case Instruction::SExt: {
765 // Compute the bits in the result that are not present in the input.
766 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
767 uint32_t SrcBitWidth = SrcTy->getBitWidth();
770 MaskIn.trunc(SrcBitWidth);
771 KnownZero.trunc(SrcBitWidth);
772 KnownOne.trunc(SrcBitWidth);
773 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
774 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
775 KnownZero.zext(BitWidth);
776 KnownOne.zext(BitWidth);
778 // If the sign bit of the input is known set or clear, then we know the
779 // top bits of the result.
780 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
781 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
782 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
783 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
786 case Instruction::Shl:
787 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
788 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
789 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
790 APInt Mask2(Mask.lshr(ShiftAmt));
791 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
792 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
793 KnownZero <<= ShiftAmt;
794 KnownOne <<= ShiftAmt;
795 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
799 case Instruction::LShr:
800 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
801 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
802 // Compute the new bits that are at the top now.
803 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
805 // Unsigned shift right.
806 APInt Mask2(Mask.shl(ShiftAmt));
807 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
808 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
809 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
810 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
811 // high bits known zero.
812 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
816 case Instruction::AShr:
817 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
818 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
819 // Compute the new bits that are at the top now.
820 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
822 // Signed shift right.
823 APInt Mask2(Mask.shl(ShiftAmt));
824 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
825 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
826 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
827 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
829 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
830 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
831 KnownZero |= HighBits;
832 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
833 KnownOne |= HighBits;
837 case Instruction::SRem:
838 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
839 APInt RA = Rem->getValue();
840 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
841 APInt LowBits = RA.isStrictlyPositive() ? ((RA - 1) | RA) : ~RA;
842 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
843 ComputeMaskedBits(I->getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
845 // The sign of a remainder is equal to the sign of the first
846 // operand (zero being positive).
847 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
848 KnownZero2 |= ~LowBits;
849 else if (KnownOne2[BitWidth-1])
850 KnownOne2 |= ~LowBits;
852 KnownZero |= KnownZero2 & Mask;
853 KnownOne |= KnownOne2 & Mask;
855 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
859 case Instruction::URem:
860 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
861 APInt RA = Rem->getValue();
862 if (RA.isStrictlyPositive() && RA.isPowerOf2()) {
863 APInt LowBits = (RA - 1) | RA;
864 APInt Mask2 = LowBits & Mask;
865 KnownZero |= ~LowBits & Mask;
866 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
867 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
870 // Since the result is less than or equal to RHS, any leading zero bits
871 // in RHS must also exist in the result.
872 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
873 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2, Depth+1);
875 uint32_t Leaders = KnownZero2.countLeadingOnes();
876 KnownZero |= APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
877 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
883 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
884 /// this predicate to simplify operations downstream. Mask is known to be zero
885 /// for bits that V cannot have.
886 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
887 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
888 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
889 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
890 return (KnownZero & Mask) == Mask;
893 /// ShrinkDemandedConstant - Check to see if the specified operand of the
894 /// specified instruction is a constant integer. If so, check to see if there
895 /// are any bits set in the constant that are not demanded. If so, shrink the
896 /// constant and return true.
897 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
899 assert(I && "No instruction?");
900 assert(OpNo < I->getNumOperands() && "Operand index too large");
902 // If the operand is not a constant integer, nothing to do.
903 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
904 if (!OpC) return false;
906 // If there are no bits set that aren't demanded, nothing to do.
907 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
908 if ((~Demanded & OpC->getValue()) == 0)
911 // This instruction is producing bits that are not demanded. Shrink the RHS.
912 Demanded &= OpC->getValue();
913 I->setOperand(OpNo, ConstantInt::get(Demanded));
917 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
918 // set of known zero and one bits, compute the maximum and minimum values that
919 // could have the specified known zero and known one bits, returning them in
921 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
922 const APInt& KnownZero,
923 const APInt& KnownOne,
924 APInt& Min, APInt& Max) {
925 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
926 assert(KnownZero.getBitWidth() == BitWidth &&
927 KnownOne.getBitWidth() == BitWidth &&
928 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
929 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
930 APInt UnknownBits = ~(KnownZero|KnownOne);
932 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
933 // bit if it is unknown.
935 Max = KnownOne|UnknownBits;
937 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
939 Max.clear(BitWidth-1);
943 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
944 // a set of known zero and one bits, compute the maximum and minimum values that
945 // could have the specified known zero and known one bits, returning them in
947 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
948 const APInt &KnownZero,
949 const APInt &KnownOne,
950 APInt &Min, APInt &Max) {
951 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
952 assert(KnownZero.getBitWidth() == BitWidth &&
953 KnownOne.getBitWidth() == BitWidth &&
954 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
955 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
956 APInt UnknownBits = ~(KnownZero|KnownOne);
958 // The minimum value is when the unknown bits are all zeros.
960 // The maximum value is when the unknown bits are all ones.
961 Max = KnownOne|UnknownBits;
964 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
965 /// value based on the demanded bits. When this function is called, it is known
966 /// that only the bits set in DemandedMask of the result of V are ever used
967 /// downstream. Consequently, depending on the mask and V, it may be possible
968 /// to replace V with a constant or one of its operands. In such cases, this
969 /// function does the replacement and returns true. In all other cases, it
970 /// returns false after analyzing the expression and setting KnownOne and known
971 /// to be one in the expression. KnownZero contains all the bits that are known
972 /// to be zero in the expression. These are provided to potentially allow the
973 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
974 /// the expression. KnownOne and KnownZero always follow the invariant that
975 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
976 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
977 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
978 /// and KnownOne must all be the same.
979 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
980 APInt& KnownZero, APInt& KnownOne,
982 assert(V != 0 && "Null pointer of Value???");
983 assert(Depth <= 6 && "Limit Search Depth");
984 uint32_t BitWidth = DemandedMask.getBitWidth();
985 const IntegerType *VTy = cast<IntegerType>(V->getType());
986 assert(VTy->getBitWidth() == BitWidth &&
987 KnownZero.getBitWidth() == BitWidth &&
988 KnownOne.getBitWidth() == BitWidth &&
989 "Value *V, DemandedMask, KnownZero and KnownOne \
990 must have same BitWidth");
991 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
992 // We know all of the bits for a constant!
993 KnownOne = CI->getValue() & DemandedMask;
994 KnownZero = ~KnownOne & DemandedMask;
1000 if (!V->hasOneUse()) { // Other users may use these bits.
1001 if (Depth != 0) { // Not at the root.
1002 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
1003 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
1006 // If this is the root being simplified, allow it to have multiple uses,
1007 // just set the DemandedMask to all bits.
1008 DemandedMask = APInt::getAllOnesValue(BitWidth);
1009 } else if (DemandedMask == 0) { // Not demanding any bits from V.
1010 if (V != UndefValue::get(VTy))
1011 return UpdateValueUsesWith(V, UndefValue::get(VTy));
1013 } else if (Depth == 6) { // Limit search depth.
1017 Instruction *I = dyn_cast<Instruction>(V);
1018 if (!I) return false; // Only analyze instructions.
1020 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1021 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
1022 switch (I->getOpcode()) {
1024 case Instruction::And:
1025 // If either the LHS or the RHS are Zero, the result is zero.
1026 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1027 RHSKnownZero, RHSKnownOne, Depth+1))
1029 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1030 "Bits known to be one AND zero?");
1032 // If something is known zero on the RHS, the bits aren't demanded on the
1034 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
1035 LHSKnownZero, LHSKnownOne, Depth+1))
1037 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1038 "Bits known to be one AND zero?");
1040 // If all of the demanded bits are known 1 on one side, return the other.
1041 // These bits cannot contribute to the result of the 'and'.
1042 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1043 (DemandedMask & ~LHSKnownZero))
1044 return UpdateValueUsesWith(I, I->getOperand(0));
1045 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1046 (DemandedMask & ~RHSKnownZero))
1047 return UpdateValueUsesWith(I, I->getOperand(1));
1049 // If all of the demanded bits in the inputs are known zeros, return zero.
1050 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1051 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1053 // If the RHS is a constant, see if we can simplify it.
1054 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1055 return UpdateValueUsesWith(I, I);
1057 // Output known-1 bits are only known if set in both the LHS & RHS.
1058 RHSKnownOne &= LHSKnownOne;
1059 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1060 RHSKnownZero |= LHSKnownZero;
1062 case Instruction::Or:
1063 // If either the LHS or the RHS are One, the result is One.
1064 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1065 RHSKnownZero, RHSKnownOne, Depth+1))
1067 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1068 "Bits known to be one AND zero?");
1069 // If something is known one on the RHS, the bits aren't demanded on the
1071 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1072 LHSKnownZero, LHSKnownOne, Depth+1))
1074 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1075 "Bits known to be one AND zero?");
1077 // If all of the demanded bits are known zero on one side, return the other.
1078 // These bits cannot contribute to the result of the 'or'.
1079 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1080 (DemandedMask & ~LHSKnownOne))
1081 return UpdateValueUsesWith(I, I->getOperand(0));
1082 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1083 (DemandedMask & ~RHSKnownOne))
1084 return UpdateValueUsesWith(I, I->getOperand(1));
1086 // If all of the potentially set bits on one side are known to be set on
1087 // the other side, just use the 'other' side.
1088 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1089 (DemandedMask & (~RHSKnownZero)))
1090 return UpdateValueUsesWith(I, I->getOperand(0));
1091 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1092 (DemandedMask & (~LHSKnownZero)))
1093 return UpdateValueUsesWith(I, I->getOperand(1));
1095 // If the RHS is a constant, see if we can simplify it.
1096 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1097 return UpdateValueUsesWith(I, I);
1099 // Output known-0 bits are only known if clear in both the LHS & RHS.
1100 RHSKnownZero &= LHSKnownZero;
1101 // Output known-1 are known to be set if set in either the LHS | RHS.
1102 RHSKnownOne |= LHSKnownOne;
1104 case Instruction::Xor: {
1105 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1106 RHSKnownZero, RHSKnownOne, Depth+1))
1108 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1109 "Bits known to be one AND zero?");
1110 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1111 LHSKnownZero, LHSKnownOne, Depth+1))
1113 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1114 "Bits known to be one AND zero?");
1116 // If all of the demanded bits are known zero on one side, return the other.
1117 // These bits cannot contribute to the result of the 'xor'.
1118 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1119 return UpdateValueUsesWith(I, I->getOperand(0));
1120 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1121 return UpdateValueUsesWith(I, I->getOperand(1));
1123 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1124 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1125 (RHSKnownOne & LHSKnownOne);
1126 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1127 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1128 (RHSKnownOne & LHSKnownZero);
1130 // If all of the demanded bits are known to be zero on one side or the
1131 // other, turn this into an *inclusive* or.
1132 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1133 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1135 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1137 InsertNewInstBefore(Or, *I);
1138 return UpdateValueUsesWith(I, Or);
1141 // If all of the demanded bits on one side are known, and all of the set
1142 // bits on that side are also known to be set on the other side, turn this
1143 // into an AND, as we know the bits will be cleared.
1144 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1145 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1147 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1148 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1150 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1151 InsertNewInstBefore(And, *I);
1152 return UpdateValueUsesWith(I, And);
1156 // If the RHS is a constant, see if we can simplify it.
1157 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1158 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1159 return UpdateValueUsesWith(I, I);
1161 RHSKnownZero = KnownZeroOut;
1162 RHSKnownOne = KnownOneOut;
1165 case Instruction::Select:
1166 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1167 RHSKnownZero, RHSKnownOne, Depth+1))
1169 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1170 LHSKnownZero, LHSKnownOne, Depth+1))
1172 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1173 "Bits known to be one AND zero?");
1174 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1175 "Bits known to be one AND zero?");
1177 // If the operands are constants, see if we can simplify them.
1178 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1179 return UpdateValueUsesWith(I, I);
1180 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1181 return UpdateValueUsesWith(I, I);
1183 // Only known if known in both the LHS and RHS.
1184 RHSKnownOne &= LHSKnownOne;
1185 RHSKnownZero &= LHSKnownZero;
1187 case Instruction::Trunc: {
1189 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1190 DemandedMask.zext(truncBf);
1191 RHSKnownZero.zext(truncBf);
1192 RHSKnownOne.zext(truncBf);
1193 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1194 RHSKnownZero, RHSKnownOne, Depth+1))
1196 DemandedMask.trunc(BitWidth);
1197 RHSKnownZero.trunc(BitWidth);
1198 RHSKnownOne.trunc(BitWidth);
1199 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1200 "Bits known to be one AND zero?");
1203 case Instruction::BitCast:
1204 if (!I->getOperand(0)->getType()->isInteger())
1207 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1208 RHSKnownZero, RHSKnownOne, Depth+1))
1210 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1211 "Bits known to be one AND zero?");
1213 case Instruction::ZExt: {
1214 // Compute the bits in the result that are not present in the input.
1215 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1216 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1218 DemandedMask.trunc(SrcBitWidth);
1219 RHSKnownZero.trunc(SrcBitWidth);
1220 RHSKnownOne.trunc(SrcBitWidth);
1221 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1222 RHSKnownZero, RHSKnownOne, Depth+1))
1224 DemandedMask.zext(BitWidth);
1225 RHSKnownZero.zext(BitWidth);
1226 RHSKnownOne.zext(BitWidth);
1227 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1228 "Bits known to be one AND zero?");
1229 // The top bits are known to be zero.
1230 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1233 case Instruction::SExt: {
1234 // Compute the bits in the result that are not present in the input.
1235 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1236 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1238 APInt InputDemandedBits = DemandedMask &
1239 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1241 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1242 // If any of the sign extended bits are demanded, we know that the sign
1244 if ((NewBits & DemandedMask) != 0)
1245 InputDemandedBits.set(SrcBitWidth-1);
1247 InputDemandedBits.trunc(SrcBitWidth);
1248 RHSKnownZero.trunc(SrcBitWidth);
1249 RHSKnownOne.trunc(SrcBitWidth);
1250 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1251 RHSKnownZero, RHSKnownOne, Depth+1))
1253 InputDemandedBits.zext(BitWidth);
1254 RHSKnownZero.zext(BitWidth);
1255 RHSKnownOne.zext(BitWidth);
1256 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1257 "Bits known to be one AND zero?");
1259 // If the sign bit of the input is known set or clear, then we know the
1260 // top bits of the result.
1262 // If the input sign bit is known zero, or if the NewBits are not demanded
1263 // convert this into a zero extension.
1264 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1266 // Convert to ZExt cast
1267 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1268 return UpdateValueUsesWith(I, NewCast);
1269 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1270 RHSKnownOne |= NewBits;
1274 case Instruction::Add: {
1275 // Figure out what the input bits are. If the top bits of the and result
1276 // are not demanded, then the add doesn't demand them from its input
1278 uint32_t NLZ = DemandedMask.countLeadingZeros();
1280 // If there is a constant on the RHS, there are a variety of xformations
1282 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1283 // If null, this should be simplified elsewhere. Some of the xforms here
1284 // won't work if the RHS is zero.
1288 // If the top bit of the output is demanded, demand everything from the
1289 // input. Otherwise, we demand all the input bits except NLZ top bits.
1290 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1292 // Find information about known zero/one bits in the input.
1293 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1294 LHSKnownZero, LHSKnownOne, Depth+1))
1297 // If the RHS of the add has bits set that can't affect the input, reduce
1299 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1300 return UpdateValueUsesWith(I, I);
1302 // Avoid excess work.
1303 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1306 // Turn it into OR if input bits are zero.
1307 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1309 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1311 InsertNewInstBefore(Or, *I);
1312 return UpdateValueUsesWith(I, Or);
1315 // We can say something about the output known-zero and known-one bits,
1316 // depending on potential carries from the input constant and the
1317 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1318 // bits set and the RHS constant is 0x01001, then we know we have a known
1319 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1321 // To compute this, we first compute the potential carry bits. These are
1322 // the bits which may be modified. I'm not aware of a better way to do
1324 const APInt& RHSVal = RHS->getValue();
1325 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1327 // Now that we know which bits have carries, compute the known-1/0 sets.
1329 // Bits are known one if they are known zero in one operand and one in the
1330 // other, and there is no input carry.
1331 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1332 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1334 // Bits are known zero if they are known zero in both operands and there
1335 // is no input carry.
1336 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1338 // If the high-bits of this ADD are not demanded, then it does not demand
1339 // the high bits of its LHS or RHS.
1340 if (DemandedMask[BitWidth-1] == 0) {
1341 // Right fill the mask of bits for this ADD to demand the most
1342 // significant bit and all those below it.
1343 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1344 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1345 LHSKnownZero, LHSKnownOne, Depth+1))
1347 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1348 LHSKnownZero, LHSKnownOne, Depth+1))
1354 case Instruction::Sub:
1355 // If the high-bits of this SUB are not demanded, then it does not demand
1356 // the high bits of its LHS or RHS.
1357 if (DemandedMask[BitWidth-1] == 0) {
1358 // Right fill the mask of bits for this SUB to demand the most
1359 // significant bit and all those below it.
1360 uint32_t NLZ = DemandedMask.countLeadingZeros();
1361 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1362 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1363 LHSKnownZero, LHSKnownOne, Depth+1))
1365 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1366 LHSKnownZero, LHSKnownOne, Depth+1))
1370 case Instruction::Shl:
1371 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1372 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1373 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1374 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1375 RHSKnownZero, RHSKnownOne, Depth+1))
1377 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1378 "Bits known to be one AND zero?");
1379 RHSKnownZero <<= ShiftAmt;
1380 RHSKnownOne <<= ShiftAmt;
1381 // low bits known zero.
1383 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1386 case Instruction::LShr:
1387 // For a logical shift right
1388 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1389 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1391 // Unsigned shift right.
1392 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1393 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1394 RHSKnownZero, RHSKnownOne, Depth+1))
1396 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1397 "Bits known to be one AND zero?");
1398 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1399 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1401 // Compute the new bits that are at the top now.
1402 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1403 RHSKnownZero |= HighBits; // high bits known zero.
1407 case Instruction::AShr:
1408 // If this is an arithmetic shift right and only the low-bit is set, we can
1409 // always convert this into a logical shr, even if the shift amount is
1410 // variable. The low bit of the shift cannot be an input sign bit unless
1411 // the shift amount is >= the size of the datatype, which is undefined.
1412 if (DemandedMask == 1) {
1413 // Perform the logical shift right.
1414 Value *NewVal = BinaryOperator::createLShr(
1415 I->getOperand(0), I->getOperand(1), I->getName());
1416 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1417 return UpdateValueUsesWith(I, NewVal);
1420 // If the sign bit is the only bit demanded by this ashr, then there is no
1421 // need to do it, the shift doesn't change the high bit.
1422 if (DemandedMask.isSignBit())
1423 return UpdateValueUsesWith(I, I->getOperand(0));
1425 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1426 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1428 // Signed shift right.
1429 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1430 // If any of the "high bits" are demanded, we should set the sign bit as
1432 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1433 DemandedMaskIn.set(BitWidth-1);
1434 if (SimplifyDemandedBits(I->getOperand(0),
1436 RHSKnownZero, RHSKnownOne, Depth+1))
1438 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1439 "Bits known to be one AND zero?");
1440 // Compute the new bits that are at the top now.
1441 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1442 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1443 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1445 // Handle the sign bits.
1446 APInt SignBit(APInt::getSignBit(BitWidth));
1447 // Adjust to where it is now in the mask.
1448 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1450 // If the input sign bit is known to be zero, or if none of the top bits
1451 // are demanded, turn this into an unsigned shift right.
1452 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1453 (HighBits & ~DemandedMask) == HighBits) {
1454 // Perform the logical shift right.
1455 Value *NewVal = BinaryOperator::createLShr(
1456 I->getOperand(0), SA, I->getName());
1457 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1458 return UpdateValueUsesWith(I, NewVal);
1459 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1460 RHSKnownOne |= HighBits;
1464 case Instruction::SRem:
1465 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1466 APInt RA = Rem->getValue();
1467 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1468 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) | RA : ~RA;
1469 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1470 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1471 LHSKnownZero, LHSKnownOne, Depth+1))
1474 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1475 LHSKnownZero |= ~LowBits;
1476 else if (LHSKnownOne[BitWidth-1])
1477 LHSKnownOne |= ~LowBits;
1479 KnownZero |= LHSKnownZero & DemandedMask;
1480 KnownOne |= LHSKnownOne & DemandedMask;
1482 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1486 case Instruction::URem:
1487 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1488 APInt RA = Rem->getValue();
1489 if (RA.isPowerOf2()) {
1490 APInt LowBits = (RA - 1) | RA;
1491 APInt Mask2 = LowBits & DemandedMask;
1492 KnownZero |= ~LowBits & DemandedMask;
1493 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1494 KnownZero, KnownOne, Depth+1))
1497 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1500 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1501 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1502 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1503 KnownZero2, KnownOne2, Depth+1))
1506 uint32_t Leaders = KnownZero2.countLeadingOnes();
1507 KnownZero |= APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1512 // If the client is only demanding bits that we know, return the known
1514 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1515 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1520 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1521 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1522 /// actually used by the caller. This method analyzes which elements of the
1523 /// operand are undef and returns that information in UndefElts.
1525 /// If the information about demanded elements can be used to simplify the
1526 /// operation, the operation is simplified, then the resultant value is
1527 /// returned. This returns null if no change was made.
1528 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1529 uint64_t &UndefElts,
1531 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1532 assert(VWidth <= 64 && "Vector too wide to analyze!");
1533 uint64_t EltMask = ~0ULL >> (64-VWidth);
1534 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1535 "Invalid DemandedElts!");
1537 if (isa<UndefValue>(V)) {
1538 // If the entire vector is undefined, just return this info.
1539 UndefElts = EltMask;
1541 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1542 UndefElts = EltMask;
1543 return UndefValue::get(V->getType());
1547 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1548 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1549 Constant *Undef = UndefValue::get(EltTy);
1551 std::vector<Constant*> Elts;
1552 for (unsigned i = 0; i != VWidth; ++i)
1553 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1554 Elts.push_back(Undef);
1555 UndefElts |= (1ULL << i);
1556 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1557 Elts.push_back(Undef);
1558 UndefElts |= (1ULL << i);
1559 } else { // Otherwise, defined.
1560 Elts.push_back(CP->getOperand(i));
1563 // If we changed the constant, return it.
1564 Constant *NewCP = ConstantVector::get(Elts);
1565 return NewCP != CP ? NewCP : 0;
1566 } else if (isa<ConstantAggregateZero>(V)) {
1567 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1569 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1570 Constant *Zero = Constant::getNullValue(EltTy);
1571 Constant *Undef = UndefValue::get(EltTy);
1572 std::vector<Constant*> Elts;
1573 for (unsigned i = 0; i != VWidth; ++i)
1574 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1575 UndefElts = DemandedElts ^ EltMask;
1576 return ConstantVector::get(Elts);
1579 if (!V->hasOneUse()) { // Other users may use these bits.
1580 if (Depth != 0) { // Not at the root.
1581 // TODO: Just compute the UndefElts information recursively.
1585 } else if (Depth == 10) { // Limit search depth.
1589 Instruction *I = dyn_cast<Instruction>(V);
1590 if (!I) return false; // Only analyze instructions.
1592 bool MadeChange = false;
1593 uint64_t UndefElts2;
1595 switch (I->getOpcode()) {
1598 case Instruction::InsertElement: {
1599 // If this is a variable index, we don't know which element it overwrites.
1600 // demand exactly the same input as we produce.
1601 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1603 // Note that we can't propagate undef elt info, because we don't know
1604 // which elt is getting updated.
1605 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1606 UndefElts2, Depth+1);
1607 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1611 // If this is inserting an element that isn't demanded, remove this
1613 unsigned IdxNo = Idx->getZExtValue();
1614 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1615 return AddSoonDeadInstToWorklist(*I, 0);
1617 // Otherwise, the element inserted overwrites whatever was there, so the
1618 // input demanded set is simpler than the output set.
1619 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1620 DemandedElts & ~(1ULL << IdxNo),
1621 UndefElts, Depth+1);
1622 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1624 // The inserted element is defined.
1625 UndefElts |= 1ULL << IdxNo;
1628 case Instruction::BitCast: {
1629 // Vector->vector casts only.
1630 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1632 unsigned InVWidth = VTy->getNumElements();
1633 uint64_t InputDemandedElts = 0;
1636 if (VWidth == InVWidth) {
1637 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1638 // elements as are demanded of us.
1640 InputDemandedElts = DemandedElts;
1641 } else if (VWidth > InVWidth) {
1645 // If there are more elements in the result than there are in the source,
1646 // then an input element is live if any of the corresponding output
1647 // elements are live.
1648 Ratio = VWidth/InVWidth;
1649 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1650 if (DemandedElts & (1ULL << OutIdx))
1651 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1657 // If there are more elements in the source than there are in the result,
1658 // then an input element is live if the corresponding output element is
1660 Ratio = InVWidth/VWidth;
1661 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1662 if (DemandedElts & (1ULL << InIdx/Ratio))
1663 InputDemandedElts |= 1ULL << InIdx;
1666 // div/rem demand all inputs, because they don't want divide by zero.
1667 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1668 UndefElts2, Depth+1);
1670 I->setOperand(0, TmpV);
1674 UndefElts = UndefElts2;
1675 if (VWidth > InVWidth) {
1676 assert(0 && "Unimp");
1677 // If there are more elements in the result than there are in the source,
1678 // then an output element is undef if the corresponding input element is
1680 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1681 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1682 UndefElts |= 1ULL << OutIdx;
1683 } else if (VWidth < InVWidth) {
1684 assert(0 && "Unimp");
1685 // If there are more elements in the source than there are in the result,
1686 // then a result element is undef if all of the corresponding input
1687 // elements are undef.
1688 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1689 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1690 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1691 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1695 case Instruction::And:
1696 case Instruction::Or:
1697 case Instruction::Xor:
1698 case Instruction::Add:
1699 case Instruction::Sub:
1700 case Instruction::Mul:
1701 // div/rem demand all inputs, because they don't want divide by zero.
1702 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1703 UndefElts, Depth+1);
1704 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1705 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1706 UndefElts2, Depth+1);
1707 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1709 // Output elements are undefined if both are undefined. Consider things
1710 // like undef&0. The result is known zero, not undef.
1711 UndefElts &= UndefElts2;
1714 case Instruction::Call: {
1715 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1717 switch (II->getIntrinsicID()) {
1720 // Binary vector operations that work column-wise. A dest element is a
1721 // function of the corresponding input elements from the two inputs.
1722 case Intrinsic::x86_sse_sub_ss:
1723 case Intrinsic::x86_sse_mul_ss:
1724 case Intrinsic::x86_sse_min_ss:
1725 case Intrinsic::x86_sse_max_ss:
1726 case Intrinsic::x86_sse2_sub_sd:
1727 case Intrinsic::x86_sse2_mul_sd:
1728 case Intrinsic::x86_sse2_min_sd:
1729 case Intrinsic::x86_sse2_max_sd:
1730 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1731 UndefElts, Depth+1);
1732 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1733 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1734 UndefElts2, Depth+1);
1735 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1737 // If only the low elt is demanded and this is a scalarizable intrinsic,
1738 // scalarize it now.
1739 if (DemandedElts == 1) {
1740 switch (II->getIntrinsicID()) {
1742 case Intrinsic::x86_sse_sub_ss:
1743 case Intrinsic::x86_sse_mul_ss:
1744 case Intrinsic::x86_sse2_sub_sd:
1745 case Intrinsic::x86_sse2_mul_sd:
1746 // TODO: Lower MIN/MAX/ABS/etc
1747 Value *LHS = II->getOperand(1);
1748 Value *RHS = II->getOperand(2);
1749 // Extract the element as scalars.
1750 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1751 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1753 switch (II->getIntrinsicID()) {
1754 default: assert(0 && "Case stmts out of sync!");
1755 case Intrinsic::x86_sse_sub_ss:
1756 case Intrinsic::x86_sse2_sub_sd:
1757 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1758 II->getName()), *II);
1760 case Intrinsic::x86_sse_mul_ss:
1761 case Intrinsic::x86_sse2_mul_sd:
1762 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1763 II->getName()), *II);
1768 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1770 InsertNewInstBefore(New, *II);
1771 AddSoonDeadInstToWorklist(*II, 0);
1776 // Output elements are undefined if both are undefined. Consider things
1777 // like undef&0. The result is known zero, not undef.
1778 UndefElts &= UndefElts2;
1784 return MadeChange ? I : 0;
1787 /// @returns true if the specified compare predicate is
1788 /// true when both operands are equal...
1789 /// @brief Determine if the icmp Predicate is true when both operands are equal
1790 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1791 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1792 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1793 pred == ICmpInst::ICMP_SLE;
1796 /// @returns true if the specified compare instruction is
1797 /// true when both operands are equal...
1798 /// @brief Determine if the ICmpInst returns true when both operands are equal
1799 static bool isTrueWhenEqual(ICmpInst &ICI) {
1800 return isTrueWhenEqual(ICI.getPredicate());
1803 /// AssociativeOpt - Perform an optimization on an associative operator. This
1804 /// function is designed to check a chain of associative operators for a
1805 /// potential to apply a certain optimization. Since the optimization may be
1806 /// applicable if the expression was reassociated, this checks the chain, then
1807 /// reassociates the expression as necessary to expose the optimization
1808 /// opportunity. This makes use of a special Functor, which must define
1809 /// 'shouldApply' and 'apply' methods.
1811 template<typename Functor>
1812 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1813 unsigned Opcode = Root.getOpcode();
1814 Value *LHS = Root.getOperand(0);
1816 // Quick check, see if the immediate LHS matches...
1817 if (F.shouldApply(LHS))
1818 return F.apply(Root);
1820 // Otherwise, if the LHS is not of the same opcode as the root, return.
1821 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1822 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1823 // Should we apply this transform to the RHS?
1824 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1826 // If not to the RHS, check to see if we should apply to the LHS...
1827 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1828 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1832 // If the functor wants to apply the optimization to the RHS of LHSI,
1833 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1835 BasicBlock *BB = Root.getParent();
1837 // Now all of the instructions are in the current basic block, go ahead
1838 // and perform the reassociation.
1839 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1841 // First move the selected RHS to the LHS of the root...
1842 Root.setOperand(0, LHSI->getOperand(1));
1844 // Make what used to be the LHS of the root be the user of the root...
1845 Value *ExtraOperand = TmpLHSI->getOperand(1);
1846 if (&Root == TmpLHSI) {
1847 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1850 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1851 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1852 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1853 BasicBlock::iterator ARI = &Root; ++ARI;
1854 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1857 // Now propagate the ExtraOperand down the chain of instructions until we
1859 while (TmpLHSI != LHSI) {
1860 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1861 // Move the instruction to immediately before the chain we are
1862 // constructing to avoid breaking dominance properties.
1863 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1864 BB->getInstList().insert(ARI, NextLHSI);
1867 Value *NextOp = NextLHSI->getOperand(1);
1868 NextLHSI->setOperand(1, ExtraOperand);
1870 ExtraOperand = NextOp;
1873 // Now that the instructions are reassociated, have the functor perform
1874 // the transformation...
1875 return F.apply(Root);
1878 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1884 // AddRHS - Implements: X + X --> X << 1
1887 AddRHS(Value *rhs) : RHS(rhs) {}
1888 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1889 Instruction *apply(BinaryOperator &Add) const {
1890 return BinaryOperator::createShl(Add.getOperand(0),
1891 ConstantInt::get(Add.getType(), 1));
1895 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1897 struct AddMaskingAnd {
1899 AddMaskingAnd(Constant *c) : C2(c) {}
1900 bool shouldApply(Value *LHS) const {
1902 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1903 ConstantExpr::getAnd(C1, C2)->isNullValue();
1905 Instruction *apply(BinaryOperator &Add) const {
1906 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1910 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1912 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1913 if (Constant *SOC = dyn_cast<Constant>(SO))
1914 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1916 return IC->InsertNewInstBefore(CastInst::create(
1917 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1920 // Figure out if the constant is the left or the right argument.
1921 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1922 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1924 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1926 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1927 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1930 Value *Op0 = SO, *Op1 = ConstOperand;
1932 std::swap(Op0, Op1);
1934 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1935 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1936 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1937 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1938 SO->getName()+".cmp");
1940 assert(0 && "Unknown binary instruction type!");
1943 return IC->InsertNewInstBefore(New, I);
1946 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1947 // constant as the other operand, try to fold the binary operator into the
1948 // select arguments. This also works for Cast instructions, which obviously do
1949 // not have a second operand.
1950 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1952 // Don't modify shared select instructions
1953 if (!SI->hasOneUse()) return 0;
1954 Value *TV = SI->getOperand(1);
1955 Value *FV = SI->getOperand(2);
1957 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1958 // Bool selects with constant operands can be folded to logical ops.
1959 if (SI->getType() == Type::Int1Ty) return 0;
1961 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1962 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1964 return new SelectInst(SI->getCondition(), SelectTrueVal,
1971 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1972 /// node as operand #0, see if we can fold the instruction into the PHI (which
1973 /// is only possible if all operands to the PHI are constants).
1974 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1975 PHINode *PN = cast<PHINode>(I.getOperand(0));
1976 unsigned NumPHIValues = PN->getNumIncomingValues();
1977 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1979 // Check to see if all of the operands of the PHI are constants. If there is
1980 // one non-constant value, remember the BB it is. If there is more than one
1981 // or if *it* is a PHI, bail out.
1982 BasicBlock *NonConstBB = 0;
1983 for (unsigned i = 0; i != NumPHIValues; ++i)
1984 if (!isa<Constant>(PN->getIncomingValue(i))) {
1985 if (NonConstBB) return 0; // More than one non-const value.
1986 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1987 NonConstBB = PN->getIncomingBlock(i);
1989 // If the incoming non-constant value is in I's block, we have an infinite
1991 if (NonConstBB == I.getParent())
1995 // If there is exactly one non-constant value, we can insert a copy of the
1996 // operation in that block. However, if this is a critical edge, we would be
1997 // inserting the computation one some other paths (e.g. inside a loop). Only
1998 // do this if the pred block is unconditionally branching into the phi block.
2000 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2001 if (!BI || !BI->isUnconditional()) return 0;
2004 // Okay, we can do the transformation: create the new PHI node.
2005 PHINode *NewPN = new PHINode(I.getType(), "");
2006 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2007 InsertNewInstBefore(NewPN, *PN);
2008 NewPN->takeName(PN);
2010 // Next, add all of the operands to the PHI.
2011 if (I.getNumOperands() == 2) {
2012 Constant *C = cast<Constant>(I.getOperand(1));
2013 for (unsigned i = 0; i != NumPHIValues; ++i) {
2015 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2016 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2017 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2019 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2021 assert(PN->getIncomingBlock(i) == NonConstBB);
2022 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2023 InV = BinaryOperator::create(BO->getOpcode(),
2024 PN->getIncomingValue(i), C, "phitmp",
2025 NonConstBB->getTerminator());
2026 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2027 InV = CmpInst::create(CI->getOpcode(),
2029 PN->getIncomingValue(i), C, "phitmp",
2030 NonConstBB->getTerminator());
2032 assert(0 && "Unknown binop!");
2034 AddToWorkList(cast<Instruction>(InV));
2036 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2039 CastInst *CI = cast<CastInst>(&I);
2040 const Type *RetTy = CI->getType();
2041 for (unsigned i = 0; i != NumPHIValues; ++i) {
2043 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2044 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2046 assert(PN->getIncomingBlock(i) == NonConstBB);
2047 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
2048 I.getType(), "phitmp",
2049 NonConstBB->getTerminator());
2050 AddToWorkList(cast<Instruction>(InV));
2052 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2055 return ReplaceInstUsesWith(I, NewPN);
2059 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
2060 /// value is never equal to -0.0.
2062 /// Note that this function will need to be revisited when we support nondefault
2065 static bool CannotBeNegativeZero(const Value *V) {
2066 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2067 return !CFP->getValueAPF().isNegZero();
2069 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2070 if (const Instruction *I = dyn_cast<Instruction>(V)) {
2071 if (I->getOpcode() == Instruction::Add &&
2072 isa<ConstantFP>(I->getOperand(1)) &&
2073 cast<ConstantFP>(I->getOperand(1))->isNullValue())
2076 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2077 if (II->getIntrinsicID() == Intrinsic::sqrt)
2078 return CannotBeNegativeZero(II->getOperand(1));
2080 if (const CallInst *CI = dyn_cast<CallInst>(I))
2081 if (const Function *F = CI->getCalledFunction()) {
2082 if (F->isDeclaration()) {
2083 switch (F->getNameLen()) {
2084 case 3: // abs(x) != -0.0
2085 if (!strcmp(F->getNameStart(), "abs")) return true;
2087 case 4: // abs[lf](x) != -0.0
2088 if (!strcmp(F->getNameStart(), "absf")) return true;
2089 if (!strcmp(F->getNameStart(), "absl")) return true;
2100 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2101 bool Changed = SimplifyCommutative(I);
2102 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2104 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2105 // X + undef -> undef
2106 if (isa<UndefValue>(RHS))
2107 return ReplaceInstUsesWith(I, RHS);
2110 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2111 if (RHSC->isNullValue())
2112 return ReplaceInstUsesWith(I, LHS);
2113 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2114 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2115 (I.getType())->getValueAPF()))
2116 return ReplaceInstUsesWith(I, LHS);
2119 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2120 // X + (signbit) --> X ^ signbit
2121 const APInt& Val = CI->getValue();
2122 uint32_t BitWidth = Val.getBitWidth();
2123 if (Val == APInt::getSignBit(BitWidth))
2124 return BinaryOperator::createXor(LHS, RHS);
2126 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2127 // (X & 254)+1 -> (X&254)|1
2128 if (!isa<VectorType>(I.getType())) {
2129 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2130 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2131 KnownZero, KnownOne))
2136 if (isa<PHINode>(LHS))
2137 if (Instruction *NV = FoldOpIntoPhi(I))
2140 ConstantInt *XorRHS = 0;
2142 if (isa<ConstantInt>(RHSC) &&
2143 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2144 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2145 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2147 uint32_t Size = TySizeBits / 2;
2148 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2149 APInt CFF80Val(-C0080Val);
2151 if (TySizeBits > Size) {
2152 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2153 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2154 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2155 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2156 // This is a sign extend if the top bits are known zero.
2157 if (!MaskedValueIsZero(XorLHS,
2158 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2159 Size = 0; // Not a sign ext, but can't be any others either.
2164 C0080Val = APIntOps::lshr(C0080Val, Size);
2165 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2166 } while (Size >= 1);
2168 // FIXME: This shouldn't be necessary. When the backends can handle types
2169 // with funny bit widths then this whole cascade of if statements should
2170 // be removed. It is just here to get the size of the "middle" type back
2171 // up to something that the back ends can handle.
2172 const Type *MiddleType = 0;
2175 case 32: MiddleType = Type::Int32Ty; break;
2176 case 16: MiddleType = Type::Int16Ty; break;
2177 case 8: MiddleType = Type::Int8Ty; break;
2180 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2181 InsertNewInstBefore(NewTrunc, I);
2182 return new SExtInst(NewTrunc, I.getType(), I.getName());
2188 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2189 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2191 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2192 if (RHSI->getOpcode() == Instruction::Sub)
2193 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2194 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2196 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2197 if (LHSI->getOpcode() == Instruction::Sub)
2198 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2199 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2204 // -A + -B --> -(A + B)
2205 if (Value *LHSV = dyn_castNegVal(LHS)) {
2206 if (LHS->getType()->isIntOrIntVector()) {
2207 if (Value *RHSV = dyn_castNegVal(RHS)) {
2208 Instruction *NewAdd = BinaryOperator::createAdd(LHSV, RHSV, "sum");
2209 InsertNewInstBefore(NewAdd, I);
2210 return BinaryOperator::createNeg(NewAdd);
2214 return BinaryOperator::createSub(RHS, LHSV);
2218 if (!isa<Constant>(RHS))
2219 if (Value *V = dyn_castNegVal(RHS))
2220 return BinaryOperator::createSub(LHS, V);
2224 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2225 if (X == RHS) // X*C + X --> X * (C+1)
2226 return BinaryOperator::createMul(RHS, AddOne(C2));
2228 // X*C1 + X*C2 --> X * (C1+C2)
2230 if (X == dyn_castFoldableMul(RHS, C1))
2231 return BinaryOperator::createMul(X, Add(C1, C2));
2234 // X + X*C --> X * (C+1)
2235 if (dyn_castFoldableMul(RHS, C2) == LHS)
2236 return BinaryOperator::createMul(LHS, AddOne(C2));
2238 // X + ~X --> -1 since ~X = -X-1
2239 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2240 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2243 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2244 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2245 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2248 // W*X + Y*Z --> W * (X+Z) iff W == Y
2249 if (I.getType()->isIntOrIntVector()) {
2250 Value *W, *X, *Y, *Z;
2251 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2252 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2256 } else if (Y == X) {
2258 } else if (X == Z) {
2265 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, Z,
2266 LHS->getName()), I);
2267 return BinaryOperator::createMul(W, NewAdd);
2272 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2274 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2275 return BinaryOperator::createSub(SubOne(CRHS), X);
2277 // (X & FF00) + xx00 -> (X+xx00) & FF00
2278 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2279 Constant *Anded = And(CRHS, C2);
2280 if (Anded == CRHS) {
2281 // See if all bits from the first bit set in the Add RHS up are included
2282 // in the mask. First, get the rightmost bit.
2283 const APInt& AddRHSV = CRHS->getValue();
2285 // Form a mask of all bits from the lowest bit added through the top.
2286 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2288 // See if the and mask includes all of these bits.
2289 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2291 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2292 // Okay, the xform is safe. Insert the new add pronto.
2293 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2294 LHS->getName()), I);
2295 return BinaryOperator::createAnd(NewAdd, C2);
2300 // Try to fold constant add into select arguments.
2301 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2302 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2306 // add (cast *A to intptrtype) B ->
2307 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2309 CastInst *CI = dyn_cast<CastInst>(LHS);
2312 CI = dyn_cast<CastInst>(RHS);
2315 if (CI && CI->getType()->isSized() &&
2316 (CI->getType()->getPrimitiveSizeInBits() ==
2317 TD->getIntPtrType()->getPrimitiveSizeInBits())
2318 && isa<PointerType>(CI->getOperand(0)->getType())) {
2320 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2321 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2322 PointerType::get(Type::Int8Ty, AS), I);
2323 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2324 return new PtrToIntInst(I2, CI->getType());
2328 // add (select X 0 (sub n A)) A --> select X A n
2330 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2333 SI = dyn_cast<SelectInst>(RHS);
2336 if (SI && SI->hasOneUse()) {
2337 Value *TV = SI->getTrueValue();
2338 Value *FV = SI->getFalseValue();
2341 // Can we fold the add into the argument of the select?
2342 // We check both true and false select arguments for a matching subtract.
2343 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2344 A == Other) // Fold the add into the true select value.
2345 return new SelectInst(SI->getCondition(), N, A);
2346 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2347 A == Other) // Fold the add into the false select value.
2348 return new SelectInst(SI->getCondition(), A, N);
2352 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2353 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2354 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2355 return ReplaceInstUsesWith(I, LHS);
2357 return Changed ? &I : 0;
2360 // isSignBit - Return true if the value represented by the constant only has the
2361 // highest order bit set.
2362 static bool isSignBit(ConstantInt *CI) {
2363 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2364 return CI->getValue() == APInt::getSignBit(NumBits);
2367 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2368 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2370 if (Op0 == Op1) // sub X, X -> 0
2371 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2373 // If this is a 'B = x-(-A)', change to B = x+A...
2374 if (Value *V = dyn_castNegVal(Op1))
2375 return BinaryOperator::createAdd(Op0, V);
2377 if (isa<UndefValue>(Op0))
2378 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2379 if (isa<UndefValue>(Op1))
2380 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2382 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2383 // Replace (-1 - A) with (~A)...
2384 if (C->isAllOnesValue())
2385 return BinaryOperator::createNot(Op1);
2387 // C - ~X == X + (1+C)
2389 if (match(Op1, m_Not(m_Value(X))))
2390 return BinaryOperator::createAdd(X, AddOne(C));
2392 // -(X >>u 31) -> (X >>s 31)
2393 // -(X >>s 31) -> (X >>u 31)
2395 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2396 if (SI->getOpcode() == Instruction::LShr) {
2397 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2398 // Check to see if we are shifting out everything but the sign bit.
2399 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2400 SI->getType()->getPrimitiveSizeInBits()-1) {
2401 // Ok, the transformation is safe. Insert AShr.
2402 return BinaryOperator::create(Instruction::AShr,
2403 SI->getOperand(0), CU, SI->getName());
2407 else if (SI->getOpcode() == Instruction::AShr) {
2408 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2409 // Check to see if we are shifting out everything but the sign bit.
2410 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2411 SI->getType()->getPrimitiveSizeInBits()-1) {
2412 // Ok, the transformation is safe. Insert LShr.
2413 return BinaryOperator::createLShr(
2414 SI->getOperand(0), CU, SI->getName());
2421 // Try to fold constant sub into select arguments.
2422 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2423 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2426 if (isa<PHINode>(Op0))
2427 if (Instruction *NV = FoldOpIntoPhi(I))
2431 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2432 if (Op1I->getOpcode() == Instruction::Add &&
2433 !Op0->getType()->isFPOrFPVector()) {
2434 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2435 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2436 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2437 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2438 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2439 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2440 // C1-(X+C2) --> (C1-C2)-X
2441 return BinaryOperator::createSub(Subtract(CI1, CI2),
2442 Op1I->getOperand(0));
2446 if (Op1I->hasOneUse()) {
2447 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2448 // is not used by anyone else...
2450 if (Op1I->getOpcode() == Instruction::Sub &&
2451 !Op1I->getType()->isFPOrFPVector()) {
2452 // Swap the two operands of the subexpr...
2453 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2454 Op1I->setOperand(0, IIOp1);
2455 Op1I->setOperand(1, IIOp0);
2457 // Create the new top level add instruction...
2458 return BinaryOperator::createAdd(Op0, Op1);
2461 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2463 if (Op1I->getOpcode() == Instruction::And &&
2464 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2465 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2468 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2469 return BinaryOperator::createAnd(Op0, NewNot);
2472 // 0 - (X sdiv C) -> (X sdiv -C)
2473 if (Op1I->getOpcode() == Instruction::SDiv)
2474 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2476 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2477 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2478 ConstantExpr::getNeg(DivRHS));
2480 // X - X*C --> X * (1-C)
2481 ConstantInt *C2 = 0;
2482 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2483 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2484 return BinaryOperator::createMul(Op0, CP1);
2487 // X - ((X / Y) * Y) --> X % Y
2488 if (Op1I->getOpcode() == Instruction::Mul)
2489 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2490 if (Op0 == I->getOperand(0) &&
2491 Op1I->getOperand(1) == I->getOperand(1)) {
2492 if (I->getOpcode() == Instruction::SDiv)
2493 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2494 if (I->getOpcode() == Instruction::UDiv)
2495 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2500 if (!Op0->getType()->isFPOrFPVector())
2501 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2502 if (Op0I->getOpcode() == Instruction::Add) {
2503 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2504 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2505 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2506 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2507 } else if (Op0I->getOpcode() == Instruction::Sub) {
2508 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2509 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2514 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2515 if (X == Op1) // X*C - X --> X * (C-1)
2516 return BinaryOperator::createMul(Op1, SubOne(C1));
2518 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2519 if (X == dyn_castFoldableMul(Op1, C2))
2520 return BinaryOperator::createMul(X, Subtract(C1, C2));
2525 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2526 /// comparison only checks the sign bit. If it only checks the sign bit, set
2527 /// TrueIfSigned if the result of the comparison is true when the input value is
2529 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2530 bool &TrueIfSigned) {
2532 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2533 TrueIfSigned = true;
2534 return RHS->isZero();
2535 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2536 TrueIfSigned = true;
2537 return RHS->isAllOnesValue();
2538 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2539 TrueIfSigned = false;
2540 return RHS->isAllOnesValue();
2541 case ICmpInst::ICMP_UGT:
2542 // True if LHS u> RHS and RHS == high-bit-mask - 1
2543 TrueIfSigned = true;
2544 return RHS->getValue() ==
2545 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2546 case ICmpInst::ICMP_UGE:
2547 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2548 TrueIfSigned = true;
2549 return RHS->getValue() ==
2550 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2556 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2557 bool Changed = SimplifyCommutative(I);
2558 Value *Op0 = I.getOperand(0);
2560 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2561 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2563 // Simplify mul instructions with a constant RHS...
2564 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2565 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2567 // ((X << C1)*C2) == (X * (C2 << C1))
2568 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2569 if (SI->getOpcode() == Instruction::Shl)
2570 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2571 return BinaryOperator::createMul(SI->getOperand(0),
2572 ConstantExpr::getShl(CI, ShOp));
2575 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2576 if (CI->equalsInt(1)) // X * 1 == X
2577 return ReplaceInstUsesWith(I, Op0);
2578 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2579 return BinaryOperator::createNeg(Op0, I.getName());
2581 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2582 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2583 return BinaryOperator::createShl(Op0,
2584 ConstantInt::get(Op0->getType(), Val.logBase2()));
2586 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2587 if (Op1F->isNullValue())
2588 return ReplaceInstUsesWith(I, Op1);
2590 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2591 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2592 // We need a better interface for long double here.
2593 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2594 if (Op1F->isExactlyValue(1.0))
2595 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2598 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2599 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2600 isa<ConstantInt>(Op0I->getOperand(1))) {
2601 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2602 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2604 InsertNewInstBefore(Add, I);
2605 Value *C1C2 = ConstantExpr::getMul(Op1,
2606 cast<Constant>(Op0I->getOperand(1)));
2607 return BinaryOperator::createAdd(Add, C1C2);
2611 // Try to fold constant mul into select arguments.
2612 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2613 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2616 if (isa<PHINode>(Op0))
2617 if (Instruction *NV = FoldOpIntoPhi(I))
2621 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2622 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2623 return BinaryOperator::createMul(Op0v, Op1v);
2625 // If one of the operands of the multiply is a cast from a boolean value, then
2626 // we know the bool is either zero or one, so this is a 'masking' multiply.
2627 // See if we can simplify things based on how the boolean was originally
2629 CastInst *BoolCast = 0;
2630 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2631 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2634 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2635 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2638 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2639 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2640 const Type *SCOpTy = SCIOp0->getType();
2643 // If the icmp is true iff the sign bit of X is set, then convert this
2644 // multiply into a shift/and combination.
2645 if (isa<ConstantInt>(SCIOp1) &&
2646 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2648 // Shift the X value right to turn it into "all signbits".
2649 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2650 SCOpTy->getPrimitiveSizeInBits()-1);
2652 InsertNewInstBefore(
2653 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2654 BoolCast->getOperand(0)->getName()+
2657 // If the multiply type is not the same as the source type, sign extend
2658 // or truncate to the multiply type.
2659 if (I.getType() != V->getType()) {
2660 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2661 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2662 Instruction::CastOps opcode =
2663 (SrcBits == DstBits ? Instruction::BitCast :
2664 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2665 V = InsertCastBefore(opcode, V, I.getType(), I);
2668 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2669 return BinaryOperator::createAnd(V, OtherOp);
2674 return Changed ? &I : 0;
2677 /// This function implements the transforms on div instructions that work
2678 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2679 /// used by the visitors to those instructions.
2680 /// @brief Transforms common to all three div instructions
2681 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2682 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2684 // undef / X -> 0 for integer.
2685 // undef / X -> undef for FP (the undef could be a snan).
2686 if (isa<UndefValue>(Op0)) {
2687 if (Op0->getType()->isFPOrFPVector())
2688 return ReplaceInstUsesWith(I, Op0);
2689 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2692 // X / undef -> undef
2693 if (isa<UndefValue>(Op1))
2694 return ReplaceInstUsesWith(I, Op1);
2696 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2697 // This does not apply for fdiv.
2698 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2699 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2700 // the same basic block, then we replace the select with Y, and the
2701 // condition of the select with false (if the cond value is in the same BB).
2702 // If the select has uses other than the div, this allows them to be
2703 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2704 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2705 if (ST->isNullValue()) {
2706 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2707 if (CondI && CondI->getParent() == I.getParent())
2708 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2709 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2710 I.setOperand(1, SI->getOperand(2));
2712 UpdateValueUsesWith(SI, SI->getOperand(2));
2716 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2717 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2718 if (ST->isNullValue()) {
2719 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2720 if (CondI && CondI->getParent() == I.getParent())
2721 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2722 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2723 I.setOperand(1, SI->getOperand(1));
2725 UpdateValueUsesWith(SI, SI->getOperand(1));
2733 /// This function implements the transforms common to both integer division
2734 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2735 /// division instructions.
2736 /// @brief Common integer divide transforms
2737 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2738 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2740 if (Instruction *Common = commonDivTransforms(I))
2743 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2745 if (RHS->equalsInt(1))
2746 return ReplaceInstUsesWith(I, Op0);
2748 // (X / C1) / C2 -> X / (C1*C2)
2749 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2750 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2751 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2752 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2753 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2755 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2756 Multiply(RHS, LHSRHS));
2759 if (!RHS->isZero()) { // avoid X udiv 0
2760 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2761 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2763 if (isa<PHINode>(Op0))
2764 if (Instruction *NV = FoldOpIntoPhi(I))
2769 // 0 / X == 0, we don't need to preserve faults!
2770 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2771 if (LHS->equalsInt(0))
2772 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2777 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2778 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2780 // Handle the integer div common cases
2781 if (Instruction *Common = commonIDivTransforms(I))
2784 // X udiv C^2 -> X >> C
2785 // Check to see if this is an unsigned division with an exact power of 2,
2786 // if so, convert to a right shift.
2787 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2788 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2789 return BinaryOperator::createLShr(Op0,
2790 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2793 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2794 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2795 if (RHSI->getOpcode() == Instruction::Shl &&
2796 isa<ConstantInt>(RHSI->getOperand(0))) {
2797 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2798 if (C1.isPowerOf2()) {
2799 Value *N = RHSI->getOperand(1);
2800 const Type *NTy = N->getType();
2801 if (uint32_t C2 = C1.logBase2()) {
2802 Constant *C2V = ConstantInt::get(NTy, C2);
2803 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2805 return BinaryOperator::createLShr(Op0, N);
2810 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2811 // where C1&C2 are powers of two.
2812 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2813 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2814 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2815 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2816 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2817 // Compute the shift amounts
2818 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2819 // Construct the "on true" case of the select
2820 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2821 Instruction *TSI = BinaryOperator::createLShr(
2822 Op0, TC, SI->getName()+".t");
2823 TSI = InsertNewInstBefore(TSI, I);
2825 // Construct the "on false" case of the select
2826 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2827 Instruction *FSI = BinaryOperator::createLShr(
2828 Op0, FC, SI->getName()+".f");
2829 FSI = InsertNewInstBefore(FSI, I);
2831 // construct the select instruction and return it.
2832 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2838 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2839 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2841 // Handle the integer div common cases
2842 if (Instruction *Common = commonIDivTransforms(I))
2845 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2847 if (RHS->isAllOnesValue())
2848 return BinaryOperator::createNeg(Op0);
2851 if (Value *LHSNeg = dyn_castNegVal(Op0))
2852 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2855 // If the sign bits of both operands are zero (i.e. we can prove they are
2856 // unsigned inputs), turn this into a udiv.
2857 if (I.getType()->isInteger()) {
2858 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2859 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2860 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2861 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2868 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2869 return commonDivTransforms(I);
2872 /// This function implements the transforms on rem instructions that work
2873 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2874 /// is used by the visitors to those instructions.
2875 /// @brief Transforms common to all three rem instructions
2876 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2877 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2879 // 0 % X == 0 for integer, we don't need to preserve faults!
2880 if (Constant *LHS = dyn_cast<Constant>(Op0))
2881 if (LHS->isNullValue())
2882 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2884 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2885 if (I.getType()->isFPOrFPVector())
2886 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2887 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2889 if (isa<UndefValue>(Op1))
2890 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2892 // Handle cases involving: rem X, (select Cond, Y, Z)
2893 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2894 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2895 // the same basic block, then we replace the select with Y, and the
2896 // condition of the select with false (if the cond value is in the same
2897 // BB). If the select has uses other than the div, this allows them to be
2899 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2900 if (ST->isNullValue()) {
2901 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2902 if (CondI && CondI->getParent() == I.getParent())
2903 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2904 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2905 I.setOperand(1, SI->getOperand(2));
2907 UpdateValueUsesWith(SI, SI->getOperand(2));
2910 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2911 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2912 if (ST->isNullValue()) {
2913 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2914 if (CondI && CondI->getParent() == I.getParent())
2915 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2916 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2917 I.setOperand(1, SI->getOperand(1));
2919 UpdateValueUsesWith(SI, SI->getOperand(1));
2927 /// This function implements the transforms common to both integer remainder
2928 /// instructions (urem and srem). It is called by the visitors to those integer
2929 /// remainder instructions.
2930 /// @brief Common integer remainder transforms
2931 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2932 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2934 if (Instruction *common = commonRemTransforms(I))
2937 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2938 // X % 0 == undef, we don't need to preserve faults!
2939 if (RHS->equalsInt(0))
2940 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2942 if (RHS->equalsInt(1)) // X % 1 == 0
2943 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2945 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2946 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2947 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2949 } else if (isa<PHINode>(Op0I)) {
2950 if (Instruction *NV = FoldOpIntoPhi(I))
2954 // See if we can fold away this rem instruction.
2955 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
2956 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2957 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2958 KnownZero, KnownOne))
2966 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2967 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2969 if (Instruction *common = commonIRemTransforms(I))
2972 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2973 // X urem C^2 -> X and C
2974 // Check to see if this is an unsigned remainder with an exact power of 2,
2975 // if so, convert to a bitwise and.
2976 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2977 if (C->getValue().isPowerOf2())
2978 return BinaryOperator::createAnd(Op0, SubOne(C));
2981 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2982 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2983 if (RHSI->getOpcode() == Instruction::Shl &&
2984 isa<ConstantInt>(RHSI->getOperand(0))) {
2985 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2986 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2987 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2989 return BinaryOperator::createAnd(Op0, Add);
2994 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2995 // where C1&C2 are powers of two.
2996 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2997 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2998 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2999 // STO == 0 and SFO == 0 handled above.
3000 if ((STO->getValue().isPowerOf2()) &&
3001 (SFO->getValue().isPowerOf2())) {
3002 Value *TrueAnd = InsertNewInstBefore(
3003 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3004 Value *FalseAnd = InsertNewInstBefore(
3005 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3006 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
3014 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3015 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3017 // Handle the integer rem common cases
3018 if (Instruction *common = commonIRemTransforms(I))
3021 if (Value *RHSNeg = dyn_castNegVal(Op1))
3022 if (!isa<ConstantInt>(RHSNeg) ||
3023 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
3025 AddUsesToWorkList(I);
3026 I.setOperand(1, RHSNeg);
3030 // If the sign bits of both operands are zero (i.e. we can prove they are
3031 // unsigned inputs), turn this into a urem.
3032 if (I.getType()->isInteger()) {
3033 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3034 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3035 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3036 return BinaryOperator::createURem(Op0, Op1, I.getName());
3043 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3044 return commonRemTransforms(I);
3047 // isMaxValueMinusOne - return true if this is Max-1
3048 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
3049 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3051 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
3052 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
3055 // isMinValuePlusOne - return true if this is Min+1
3056 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
3058 return C->getValue() == 1; // unsigned
3060 // Calculate 1111111111000000000000
3061 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3062 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
3065 // isOneBitSet - Return true if there is exactly one bit set in the specified
3067 static bool isOneBitSet(const ConstantInt *CI) {
3068 return CI->getValue().isPowerOf2();
3071 // isHighOnes - Return true if the constant is of the form 1+0+.
3072 // This is the same as lowones(~X).
3073 static bool isHighOnes(const ConstantInt *CI) {
3074 return (~CI->getValue() + 1).isPowerOf2();
3077 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3078 /// are carefully arranged to allow folding of expressions such as:
3080 /// (A < B) | (A > B) --> (A != B)
3082 /// Note that this is only valid if the first and second predicates have the
3083 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3085 /// Three bits are used to represent the condition, as follows:
3090 /// <=> Value Definition
3091 /// 000 0 Always false
3098 /// 111 7 Always true
3100 static unsigned getICmpCode(const ICmpInst *ICI) {
3101 switch (ICI->getPredicate()) {
3103 case ICmpInst::ICMP_UGT: return 1; // 001
3104 case ICmpInst::ICMP_SGT: return 1; // 001
3105 case ICmpInst::ICMP_EQ: return 2; // 010
3106 case ICmpInst::ICMP_UGE: return 3; // 011
3107 case ICmpInst::ICMP_SGE: return 3; // 011
3108 case ICmpInst::ICMP_ULT: return 4; // 100
3109 case ICmpInst::ICMP_SLT: return 4; // 100
3110 case ICmpInst::ICMP_NE: return 5; // 101
3111 case ICmpInst::ICMP_ULE: return 6; // 110
3112 case ICmpInst::ICMP_SLE: return 6; // 110
3115 assert(0 && "Invalid ICmp predicate!");
3120 /// getICmpValue - This is the complement of getICmpCode, which turns an
3121 /// opcode and two operands into either a constant true or false, or a brand
3122 /// new ICmp instruction. The sign is passed in to determine which kind
3123 /// of predicate to use in new icmp instructions.
3124 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3126 default: assert(0 && "Illegal ICmp code!");
3127 case 0: return ConstantInt::getFalse();
3130 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3132 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3133 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3136 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3138 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3141 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3143 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3144 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3147 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3149 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3150 case 7: return ConstantInt::getTrue();
3154 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3155 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3156 (ICmpInst::isSignedPredicate(p1) &&
3157 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3158 (ICmpInst::isSignedPredicate(p2) &&
3159 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3163 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3164 struct FoldICmpLogical {
3167 ICmpInst::Predicate pred;
3168 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3169 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3170 pred(ICI->getPredicate()) {}
3171 bool shouldApply(Value *V) const {
3172 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3173 if (PredicatesFoldable(pred, ICI->getPredicate()))
3174 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3175 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3178 Instruction *apply(Instruction &Log) const {
3179 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3180 if (ICI->getOperand(0) != LHS) {
3181 assert(ICI->getOperand(1) == LHS);
3182 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3185 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3186 unsigned LHSCode = getICmpCode(ICI);
3187 unsigned RHSCode = getICmpCode(RHSICI);
3189 switch (Log.getOpcode()) {
3190 case Instruction::And: Code = LHSCode & RHSCode; break;
3191 case Instruction::Or: Code = LHSCode | RHSCode; break;
3192 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3193 default: assert(0 && "Illegal logical opcode!"); return 0;
3196 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3197 ICmpInst::isSignedPredicate(ICI->getPredicate());
3199 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3200 if (Instruction *I = dyn_cast<Instruction>(RV))
3202 // Otherwise, it's a constant boolean value...
3203 return IC.ReplaceInstUsesWith(Log, RV);
3206 } // end anonymous namespace
3208 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3209 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3210 // guaranteed to be a binary operator.
3211 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3213 ConstantInt *AndRHS,
3214 BinaryOperator &TheAnd) {
3215 Value *X = Op->getOperand(0);
3216 Constant *Together = 0;
3218 Together = And(AndRHS, OpRHS);
3220 switch (Op->getOpcode()) {
3221 case Instruction::Xor:
3222 if (Op->hasOneUse()) {
3223 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3224 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3225 InsertNewInstBefore(And, TheAnd);
3227 return BinaryOperator::createXor(And, Together);
3230 case Instruction::Or:
3231 if (Together == AndRHS) // (X | C) & C --> C
3232 return ReplaceInstUsesWith(TheAnd, AndRHS);
3234 if (Op->hasOneUse() && Together != OpRHS) {
3235 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3236 Instruction *Or = BinaryOperator::createOr(X, Together);
3237 InsertNewInstBefore(Or, TheAnd);
3239 return BinaryOperator::createAnd(Or, AndRHS);
3242 case Instruction::Add:
3243 if (Op->hasOneUse()) {
3244 // Adding a one to a single bit bit-field should be turned into an XOR
3245 // of the bit. First thing to check is to see if this AND is with a
3246 // single bit constant.
3247 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3249 // If there is only one bit set...
3250 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3251 // Ok, at this point, we know that we are masking the result of the
3252 // ADD down to exactly one bit. If the constant we are adding has
3253 // no bits set below this bit, then we can eliminate the ADD.
3254 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3256 // Check to see if any bits below the one bit set in AndRHSV are set.
3257 if ((AddRHS & (AndRHSV-1)) == 0) {
3258 // If not, the only thing that can effect the output of the AND is
3259 // the bit specified by AndRHSV. If that bit is set, the effect of
3260 // the XOR is to toggle the bit. If it is clear, then the ADD has
3262 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3263 TheAnd.setOperand(0, X);
3266 // Pull the XOR out of the AND.
3267 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3268 InsertNewInstBefore(NewAnd, TheAnd);
3269 NewAnd->takeName(Op);
3270 return BinaryOperator::createXor(NewAnd, AndRHS);
3277 case Instruction::Shl: {
3278 // We know that the AND will not produce any of the bits shifted in, so if
3279 // the anded constant includes them, clear them now!
3281 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3282 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3283 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3284 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3286 if (CI->getValue() == ShlMask) {
3287 // Masking out bits that the shift already masks
3288 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3289 } else if (CI != AndRHS) { // Reducing bits set in and.
3290 TheAnd.setOperand(1, CI);
3295 case Instruction::LShr:
3297 // We know that the AND will not produce any of the bits shifted in, so if
3298 // the anded constant includes them, clear them now! This only applies to
3299 // unsigned shifts, because a signed shr may bring in set bits!
3301 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3302 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3303 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3304 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3306 if (CI->getValue() == ShrMask) {
3307 // Masking out bits that the shift already masks.
3308 return ReplaceInstUsesWith(TheAnd, Op);
3309 } else if (CI != AndRHS) {
3310 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3315 case Instruction::AShr:
3317 // See if this is shifting in some sign extension, then masking it out
3319 if (Op->hasOneUse()) {
3320 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3321 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3322 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3323 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3324 if (C == AndRHS) { // Masking out bits shifted in.
3325 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3326 // Make the argument unsigned.
3327 Value *ShVal = Op->getOperand(0);
3328 ShVal = InsertNewInstBefore(
3329 BinaryOperator::createLShr(ShVal, OpRHS,
3330 Op->getName()), TheAnd);
3331 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3340 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3341 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3342 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3343 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3344 /// insert new instructions.
3345 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3346 bool isSigned, bool Inside,
3348 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3349 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3350 "Lo is not <= Hi in range emission code!");
3353 if (Lo == Hi) // Trivially false.
3354 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3356 // V >= Min && V < Hi --> V < Hi
3357 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3358 ICmpInst::Predicate pred = (isSigned ?
3359 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3360 return new ICmpInst(pred, V, Hi);
3363 // Emit V-Lo <u Hi-Lo
3364 Constant *NegLo = ConstantExpr::getNeg(Lo);
3365 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3366 InsertNewInstBefore(Add, IB);
3367 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3368 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3371 if (Lo == Hi) // Trivially true.
3372 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3374 // V < Min || V >= Hi -> V > Hi-1
3375 Hi = SubOne(cast<ConstantInt>(Hi));
3376 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3377 ICmpInst::Predicate pred = (isSigned ?
3378 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3379 return new ICmpInst(pred, V, Hi);
3382 // Emit V-Lo >u Hi-1-Lo
3383 // Note that Hi has already had one subtracted from it, above.
3384 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3385 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3386 InsertNewInstBefore(Add, IB);
3387 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3388 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3391 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3392 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3393 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3394 // not, since all 1s are not contiguous.
3395 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3396 const APInt& V = Val->getValue();
3397 uint32_t BitWidth = Val->getType()->getBitWidth();
3398 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3400 // look for the first zero bit after the run of ones
3401 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3402 // look for the first non-zero bit
3403 ME = V.getActiveBits();
3407 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3408 /// where isSub determines whether the operator is a sub. If we can fold one of
3409 /// the following xforms:
3411 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3412 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3413 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3415 /// return (A +/- B).
3417 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3418 ConstantInt *Mask, bool isSub,
3420 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3421 if (!LHSI || LHSI->getNumOperands() != 2 ||
3422 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3424 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3426 switch (LHSI->getOpcode()) {
3428 case Instruction::And:
3429 if (And(N, Mask) == Mask) {
3430 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3431 if ((Mask->getValue().countLeadingZeros() +
3432 Mask->getValue().countPopulation()) ==
3433 Mask->getValue().getBitWidth())
3436 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3437 // part, we don't need any explicit masks to take them out of A. If that
3438 // is all N is, ignore it.
3439 uint32_t MB = 0, ME = 0;
3440 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3441 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3442 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3443 if (MaskedValueIsZero(RHS, Mask))
3448 case Instruction::Or:
3449 case Instruction::Xor:
3450 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3451 if ((Mask->getValue().countLeadingZeros() +
3452 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3453 && And(N, Mask)->isZero())
3460 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3462 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3463 return InsertNewInstBefore(New, I);
3466 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3467 bool Changed = SimplifyCommutative(I);
3468 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3470 if (isa<UndefValue>(Op1)) // X & undef -> 0
3471 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3475 return ReplaceInstUsesWith(I, Op1);
3477 // See if we can simplify any instructions used by the instruction whose sole
3478 // purpose is to compute bits we don't care about.
3479 if (!isa<VectorType>(I.getType())) {
3480 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3481 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3482 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3483 KnownZero, KnownOne))
3486 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3487 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3488 return ReplaceInstUsesWith(I, I.getOperand(0));
3489 } else if (isa<ConstantAggregateZero>(Op1)) {
3490 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3494 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3495 const APInt& AndRHSMask = AndRHS->getValue();
3496 APInt NotAndRHS(~AndRHSMask);
3498 // Optimize a variety of ((val OP C1) & C2) combinations...
3499 if (isa<BinaryOperator>(Op0)) {
3500 Instruction *Op0I = cast<Instruction>(Op0);
3501 Value *Op0LHS = Op0I->getOperand(0);
3502 Value *Op0RHS = Op0I->getOperand(1);
3503 switch (Op0I->getOpcode()) {
3504 case Instruction::Xor:
3505 case Instruction::Or:
3506 // If the mask is only needed on one incoming arm, push it up.
3507 if (Op0I->hasOneUse()) {
3508 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3509 // Not masking anything out for the LHS, move to RHS.
3510 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3511 Op0RHS->getName()+".masked");
3512 InsertNewInstBefore(NewRHS, I);
3513 return BinaryOperator::create(
3514 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3516 if (!isa<Constant>(Op0RHS) &&
3517 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3518 // Not masking anything out for the RHS, move to LHS.
3519 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3520 Op0LHS->getName()+".masked");
3521 InsertNewInstBefore(NewLHS, I);
3522 return BinaryOperator::create(
3523 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3528 case Instruction::Add:
3529 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3530 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3531 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3532 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3533 return BinaryOperator::createAnd(V, AndRHS);
3534 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3535 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3538 case Instruction::Sub:
3539 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3540 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3541 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3542 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3543 return BinaryOperator::createAnd(V, AndRHS);
3547 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3548 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3550 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3551 // If this is an integer truncation or change from signed-to-unsigned, and
3552 // if the source is an and/or with immediate, transform it. This
3553 // frequently occurs for bitfield accesses.
3554 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3555 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3556 CastOp->getNumOperands() == 2)
3557 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3558 if (CastOp->getOpcode() == Instruction::And) {
3559 // Change: and (cast (and X, C1) to T), C2
3560 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3561 // This will fold the two constants together, which may allow
3562 // other simplifications.
3563 Instruction *NewCast = CastInst::createTruncOrBitCast(
3564 CastOp->getOperand(0), I.getType(),
3565 CastOp->getName()+".shrunk");
3566 NewCast = InsertNewInstBefore(NewCast, I);
3567 // trunc_or_bitcast(C1)&C2
3568 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3569 C3 = ConstantExpr::getAnd(C3, AndRHS);
3570 return BinaryOperator::createAnd(NewCast, C3);
3571 } else if (CastOp->getOpcode() == Instruction::Or) {
3572 // Change: and (cast (or X, C1) to T), C2
3573 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3574 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3575 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3576 return ReplaceInstUsesWith(I, AndRHS);
3582 // Try to fold constant and into select arguments.
3583 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3584 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3586 if (isa<PHINode>(Op0))
3587 if (Instruction *NV = FoldOpIntoPhi(I))
3591 Value *Op0NotVal = dyn_castNotVal(Op0);
3592 Value *Op1NotVal = dyn_castNotVal(Op1);
3594 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3595 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3597 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3598 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3599 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3600 I.getName()+".demorgan");
3601 InsertNewInstBefore(Or, I);
3602 return BinaryOperator::createNot(Or);
3606 Value *A = 0, *B = 0, *C = 0, *D = 0;
3607 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3608 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3609 return ReplaceInstUsesWith(I, Op1);
3611 // (A|B) & ~(A&B) -> A^B
3612 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3613 if ((A == C && B == D) || (A == D && B == C))
3614 return BinaryOperator::createXor(A, B);
3618 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3619 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3620 return ReplaceInstUsesWith(I, Op0);
3622 // ~(A&B) & (A|B) -> A^B
3623 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3624 if ((A == C && B == D) || (A == D && B == C))
3625 return BinaryOperator::createXor(A, B);
3629 if (Op0->hasOneUse() &&
3630 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3631 if (A == Op1) { // (A^B)&A -> A&(A^B)
3632 I.swapOperands(); // Simplify below
3633 std::swap(Op0, Op1);
3634 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3635 cast<BinaryOperator>(Op0)->swapOperands();
3636 I.swapOperands(); // Simplify below
3637 std::swap(Op0, Op1);
3640 if (Op1->hasOneUse() &&
3641 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3642 if (B == Op0) { // B&(A^B) -> B&(B^A)
3643 cast<BinaryOperator>(Op1)->swapOperands();
3646 if (A == Op0) { // A&(A^B) -> A & ~B
3647 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3648 InsertNewInstBefore(NotB, I);
3649 return BinaryOperator::createAnd(A, NotB);
3654 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3655 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3656 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3659 Value *LHSVal, *RHSVal;
3660 ConstantInt *LHSCst, *RHSCst;
3661 ICmpInst::Predicate LHSCC, RHSCC;
3662 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3663 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3664 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3665 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3666 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3667 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3668 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3669 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3671 // Don't try to fold ICMP_SLT + ICMP_ULT.
3672 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3673 ICmpInst::isSignedPredicate(LHSCC) ==
3674 ICmpInst::isSignedPredicate(RHSCC))) {
3675 // Ensure that the larger constant is on the RHS.
3676 ICmpInst::Predicate GT;
3677 if (ICmpInst::isSignedPredicate(LHSCC) ||
3678 (ICmpInst::isEquality(LHSCC) &&
3679 ICmpInst::isSignedPredicate(RHSCC)))
3680 GT = ICmpInst::ICMP_SGT;
3682 GT = ICmpInst::ICMP_UGT;
3684 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3685 ICmpInst *LHS = cast<ICmpInst>(Op0);
3686 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3687 std::swap(LHS, RHS);
3688 std::swap(LHSCst, RHSCst);
3689 std::swap(LHSCC, RHSCC);
3692 // At this point, we know we have have two icmp instructions
3693 // comparing a value against two constants and and'ing the result
3694 // together. Because of the above check, we know that we only have
3695 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3696 // (from the FoldICmpLogical check above), that the two constants
3697 // are not equal and that the larger constant is on the RHS
3698 assert(LHSCst != RHSCst && "Compares not folded above?");
3701 default: assert(0 && "Unknown integer condition code!");
3702 case ICmpInst::ICMP_EQ:
3704 default: assert(0 && "Unknown integer condition code!");
3705 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3706 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3707 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3708 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3709 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3710 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3711 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3712 return ReplaceInstUsesWith(I, LHS);
3714 case ICmpInst::ICMP_NE:
3716 default: assert(0 && "Unknown integer condition code!");
3717 case ICmpInst::ICMP_ULT:
3718 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3719 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3720 break; // (X != 13 & X u< 15) -> no change
3721 case ICmpInst::ICMP_SLT:
3722 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3723 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3724 break; // (X != 13 & X s< 15) -> no change
3725 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3726 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3727 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3728 return ReplaceInstUsesWith(I, RHS);
3729 case ICmpInst::ICMP_NE:
3730 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3731 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3732 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3733 LHSVal->getName()+".off");
3734 InsertNewInstBefore(Add, I);
3735 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3736 ConstantInt::get(Add->getType(), 1));
3738 break; // (X != 13 & X != 15) -> no change
3741 case ICmpInst::ICMP_ULT:
3743 default: assert(0 && "Unknown integer condition code!");
3744 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3745 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3746 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3747 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3749 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3750 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3751 return ReplaceInstUsesWith(I, LHS);
3752 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3756 case ICmpInst::ICMP_SLT:
3758 default: assert(0 && "Unknown integer condition code!");
3759 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3760 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3761 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3762 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3764 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3765 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3766 return ReplaceInstUsesWith(I, LHS);
3767 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3771 case ICmpInst::ICMP_UGT:
3773 default: assert(0 && "Unknown integer condition code!");
3774 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3775 return ReplaceInstUsesWith(I, LHS);
3776 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3777 return ReplaceInstUsesWith(I, RHS);
3778 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3780 case ICmpInst::ICMP_NE:
3781 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3782 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3783 break; // (X u> 13 & X != 15) -> no change
3784 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3785 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3787 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3791 case ICmpInst::ICMP_SGT:
3793 default: assert(0 && "Unknown integer condition code!");
3794 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3795 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3796 return ReplaceInstUsesWith(I, RHS);
3797 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3799 case ICmpInst::ICMP_NE:
3800 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3801 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3802 break; // (X s> 13 & X != 15) -> no change
3803 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3804 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3806 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3814 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3815 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3816 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3817 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3818 const Type *SrcTy = Op0C->getOperand(0)->getType();
3819 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3820 // Only do this if the casts both really cause code to be generated.
3821 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3823 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3825 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3826 Op1C->getOperand(0),
3828 InsertNewInstBefore(NewOp, I);
3829 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3833 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3834 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3835 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3836 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3837 SI0->getOperand(1) == SI1->getOperand(1) &&
3838 (SI0->hasOneUse() || SI1->hasOneUse())) {
3839 Instruction *NewOp =
3840 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3842 SI0->getName()), I);
3843 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3844 SI1->getOperand(1));
3848 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3849 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3850 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3851 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3852 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3853 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3854 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3855 // If either of the constants are nans, then the whole thing returns
3857 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3858 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3859 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3860 RHS->getOperand(0));
3865 return Changed ? &I : 0;
3868 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3869 /// in the result. If it does, and if the specified byte hasn't been filled in
3870 /// yet, fill it in and return false.
3871 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3872 Instruction *I = dyn_cast<Instruction>(V);
3873 if (I == 0) return true;
3875 // If this is an or instruction, it is an inner node of the bswap.
3876 if (I->getOpcode() == Instruction::Or)
3877 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3878 CollectBSwapParts(I->getOperand(1), ByteValues);
3880 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3881 // If this is a shift by a constant int, and it is "24", then its operand
3882 // defines a byte. We only handle unsigned types here.
3883 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3884 // Not shifting the entire input by N-1 bytes?
3885 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3886 8*(ByteValues.size()-1))
3890 if (I->getOpcode() == Instruction::Shl) {
3891 // X << 24 defines the top byte with the lowest of the input bytes.
3892 DestNo = ByteValues.size()-1;
3894 // X >>u 24 defines the low byte with the highest of the input bytes.
3898 // If the destination byte value is already defined, the values are or'd
3899 // together, which isn't a bswap (unless it's an or of the same bits).
3900 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3902 ByteValues[DestNo] = I->getOperand(0);
3906 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3908 Value *Shift = 0, *ShiftLHS = 0;
3909 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3910 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3911 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3913 Instruction *SI = cast<Instruction>(Shift);
3915 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3916 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3917 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3920 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3922 if (AndAmt->getValue().getActiveBits() > 64)
3924 uint64_t AndAmtVal = AndAmt->getZExtValue();
3925 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3926 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3928 // Unknown mask for bswap.
3929 if (DestByte == ByteValues.size()) return true;
3931 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3933 if (SI->getOpcode() == Instruction::Shl)
3934 SrcByte = DestByte - ShiftBytes;
3936 SrcByte = DestByte + ShiftBytes;
3938 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3939 if (SrcByte != ByteValues.size()-DestByte-1)
3942 // If the destination byte value is already defined, the values are or'd
3943 // together, which isn't a bswap (unless it's an or of the same bits).
3944 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3946 ByteValues[DestByte] = SI->getOperand(0);
3950 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3951 /// If so, insert the new bswap intrinsic and return it.
3952 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3953 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3954 if (!ITy || ITy->getBitWidth() % 16)
3955 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3957 /// ByteValues - For each byte of the result, we keep track of which value
3958 /// defines each byte.
3959 SmallVector<Value*, 8> ByteValues;
3960 ByteValues.resize(ITy->getBitWidth()/8);
3962 // Try to find all the pieces corresponding to the bswap.
3963 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3964 CollectBSwapParts(I.getOperand(1), ByteValues))
3967 // Check to see if all of the bytes come from the same value.
3968 Value *V = ByteValues[0];
3969 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3971 // Check to make sure that all of the bytes come from the same value.
3972 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3973 if (ByteValues[i] != V)
3975 const Type *Tys[] = { ITy };
3976 Module *M = I.getParent()->getParent()->getParent();
3977 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3978 return new CallInst(F, V);
3982 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3983 bool Changed = SimplifyCommutative(I);
3984 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3986 if (isa<UndefValue>(Op1)) // X | undef -> -1
3987 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3991 return ReplaceInstUsesWith(I, Op0);
3993 // See if we can simplify any instructions used by the instruction whose sole
3994 // purpose is to compute bits we don't care about.
3995 if (!isa<VectorType>(I.getType())) {
3996 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3997 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3998 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3999 KnownZero, KnownOne))
4001 } else if (isa<ConstantAggregateZero>(Op1)) {
4002 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4003 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4004 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4005 return ReplaceInstUsesWith(I, I.getOperand(1));
4011 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4012 ConstantInt *C1 = 0; Value *X = 0;
4013 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4014 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4015 Instruction *Or = BinaryOperator::createOr(X, RHS);
4016 InsertNewInstBefore(Or, I);
4018 return BinaryOperator::createAnd(Or,
4019 ConstantInt::get(RHS->getValue() | C1->getValue()));
4022 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4023 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4024 Instruction *Or = BinaryOperator::createOr(X, RHS);
4025 InsertNewInstBefore(Or, I);
4027 return BinaryOperator::createXor(Or,
4028 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4031 // Try to fold constant and into select arguments.
4032 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4033 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4035 if (isa<PHINode>(Op0))
4036 if (Instruction *NV = FoldOpIntoPhi(I))
4040 Value *A = 0, *B = 0;
4041 ConstantInt *C1 = 0, *C2 = 0;
4043 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4044 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4045 return ReplaceInstUsesWith(I, Op1);
4046 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4047 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4048 return ReplaceInstUsesWith(I, Op0);
4050 // (A | B) | C and A | (B | C) -> bswap if possible.
4051 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4052 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4053 match(Op1, m_Or(m_Value(), m_Value())) ||
4054 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4055 match(Op1, m_Shift(m_Value(), m_Value())))) {
4056 if (Instruction *BSwap = MatchBSwap(I))
4060 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4061 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4062 MaskedValueIsZero(Op1, C1->getValue())) {
4063 Instruction *NOr = BinaryOperator::createOr(A, Op1);
4064 InsertNewInstBefore(NOr, I);
4066 return BinaryOperator::createXor(NOr, C1);
4069 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4070 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4071 MaskedValueIsZero(Op0, C1->getValue())) {
4072 Instruction *NOr = BinaryOperator::createOr(A, Op0);
4073 InsertNewInstBefore(NOr, I);
4075 return BinaryOperator::createXor(NOr, C1);
4079 Value *C = 0, *D = 0;
4080 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4081 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4082 Value *V1 = 0, *V2 = 0, *V3 = 0;
4083 C1 = dyn_cast<ConstantInt>(C);
4084 C2 = dyn_cast<ConstantInt>(D);
4085 if (C1 && C2) { // (A & C1)|(B & C2)
4086 // If we have: ((V + N) & C1) | (V & C2)
4087 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4088 // replace with V+N.
4089 if (C1->getValue() == ~C2->getValue()) {
4090 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4091 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4092 // Add commutes, try both ways.
4093 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4094 return ReplaceInstUsesWith(I, A);
4095 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4096 return ReplaceInstUsesWith(I, A);
4098 // Or commutes, try both ways.
4099 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4100 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4101 // Add commutes, try both ways.
4102 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4103 return ReplaceInstUsesWith(I, B);
4104 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4105 return ReplaceInstUsesWith(I, B);
4108 V1 = 0; V2 = 0; V3 = 0;
4111 // Check to see if we have any common things being and'ed. If so, find the
4112 // terms for V1 & (V2|V3).
4113 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4114 if (A == B) // (A & C)|(A & D) == A & (C|D)
4115 V1 = A, V2 = C, V3 = D;
4116 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4117 V1 = A, V2 = B, V3 = C;
4118 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4119 V1 = C, V2 = A, V3 = D;
4120 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4121 V1 = C, V2 = A, V3 = B;
4125 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
4126 return BinaryOperator::createAnd(V1, Or);
4131 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4132 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4133 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4134 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4135 SI0->getOperand(1) == SI1->getOperand(1) &&
4136 (SI0->hasOneUse() || SI1->hasOneUse())) {
4137 Instruction *NewOp =
4138 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4140 SI0->getName()), I);
4141 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4142 SI1->getOperand(1));
4146 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4147 if (A == Op1) // ~A | A == -1
4148 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4152 // Note, A is still live here!
4153 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4155 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4157 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4158 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4159 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4160 I.getName()+".demorgan"), I);
4161 return BinaryOperator::createNot(And);
4165 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4166 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4167 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4170 Value *LHSVal, *RHSVal;
4171 ConstantInt *LHSCst, *RHSCst;
4172 ICmpInst::Predicate LHSCC, RHSCC;
4173 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4174 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4175 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4176 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4177 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4178 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4179 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4180 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4181 // We can't fold (ugt x, C) | (sgt x, C2).
4182 PredicatesFoldable(LHSCC, RHSCC)) {
4183 // Ensure that the larger constant is on the RHS.
4184 ICmpInst *LHS = cast<ICmpInst>(Op0);
4186 if (ICmpInst::isSignedPredicate(LHSCC))
4187 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4189 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4192 std::swap(LHS, RHS);
4193 std::swap(LHSCst, RHSCst);
4194 std::swap(LHSCC, RHSCC);
4197 // At this point, we know we have have two icmp instructions
4198 // comparing a value against two constants and or'ing the result
4199 // together. Because of the above check, we know that we only have
4200 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4201 // FoldICmpLogical check above), that the two constants are not
4203 assert(LHSCst != RHSCst && "Compares not folded above?");
4206 default: assert(0 && "Unknown integer condition code!");
4207 case ICmpInst::ICMP_EQ:
4209 default: assert(0 && "Unknown integer condition code!");
4210 case ICmpInst::ICMP_EQ:
4211 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4212 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4213 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4214 LHSVal->getName()+".off");
4215 InsertNewInstBefore(Add, I);
4216 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4217 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4219 break; // (X == 13 | X == 15) -> no change
4220 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4221 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4223 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4224 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4225 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4226 return ReplaceInstUsesWith(I, RHS);
4229 case ICmpInst::ICMP_NE:
4231 default: assert(0 && "Unknown integer condition code!");
4232 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4233 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4234 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4235 return ReplaceInstUsesWith(I, LHS);
4236 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4237 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4238 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4239 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4242 case ICmpInst::ICMP_ULT:
4244 default: assert(0 && "Unknown integer condition code!");
4245 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4247 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4248 // If RHSCst is [us]MAXINT, it is always false. Not handling
4249 // this can cause overflow.
4250 if (RHSCst->isMaxValue(false))
4251 return ReplaceInstUsesWith(I, LHS);
4252 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4254 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4256 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4257 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4258 return ReplaceInstUsesWith(I, RHS);
4259 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4263 case ICmpInst::ICMP_SLT:
4265 default: assert(0 && "Unknown integer condition code!");
4266 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4268 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4269 // If RHSCst is [us]MAXINT, it is always false. Not handling
4270 // this can cause overflow.
4271 if (RHSCst->isMaxValue(true))
4272 return ReplaceInstUsesWith(I, LHS);
4273 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4275 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4277 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4278 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4279 return ReplaceInstUsesWith(I, RHS);
4280 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4284 case ICmpInst::ICMP_UGT:
4286 default: assert(0 && "Unknown integer condition code!");
4287 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4288 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4289 return ReplaceInstUsesWith(I, LHS);
4290 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4292 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4293 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4294 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4295 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4299 case ICmpInst::ICMP_SGT:
4301 default: assert(0 && "Unknown integer condition code!");
4302 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4303 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4304 return ReplaceInstUsesWith(I, LHS);
4305 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4307 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4308 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4309 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4310 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4318 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4319 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4320 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4321 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4322 const Type *SrcTy = Op0C->getOperand(0)->getType();
4323 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4324 // Only do this if the casts both really cause code to be generated.
4325 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4327 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4329 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4330 Op1C->getOperand(0),
4332 InsertNewInstBefore(NewOp, I);
4333 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4339 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4340 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4341 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4342 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4343 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4344 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4345 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4346 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4347 // If either of the constants are nans, then the whole thing returns
4349 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4350 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4352 // Otherwise, no need to compare the two constants, compare the
4354 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4355 RHS->getOperand(0));
4360 return Changed ? &I : 0;
4363 // XorSelf - Implements: X ^ X --> 0
4366 XorSelf(Value *rhs) : RHS(rhs) {}
4367 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4368 Instruction *apply(BinaryOperator &Xor) const {
4374 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4375 bool Changed = SimplifyCommutative(I);
4376 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4378 if (isa<UndefValue>(Op1))
4379 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4381 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4382 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4383 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4384 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4387 // See if we can simplify any instructions used by the instruction whose sole
4388 // purpose is to compute bits we don't care about.
4389 if (!isa<VectorType>(I.getType())) {
4390 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4391 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4392 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4393 KnownZero, KnownOne))
4395 } else if (isa<ConstantAggregateZero>(Op1)) {
4396 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4399 // Is this a ~ operation?
4400 if (Value *NotOp = dyn_castNotVal(&I)) {
4401 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4402 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4403 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4404 if (Op0I->getOpcode() == Instruction::And ||
4405 Op0I->getOpcode() == Instruction::Or) {
4406 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4407 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4409 BinaryOperator::createNot(Op0I->getOperand(1),
4410 Op0I->getOperand(1)->getName()+".not");
4411 InsertNewInstBefore(NotY, I);
4412 if (Op0I->getOpcode() == Instruction::And)
4413 return BinaryOperator::createOr(Op0NotVal, NotY);
4415 return BinaryOperator::createAnd(Op0NotVal, NotY);
4422 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4423 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4424 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4425 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4426 return new ICmpInst(ICI->getInversePredicate(),
4427 ICI->getOperand(0), ICI->getOperand(1));
4429 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4430 return new FCmpInst(FCI->getInversePredicate(),
4431 FCI->getOperand(0), FCI->getOperand(1));
4434 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4435 // ~(c-X) == X-c-1 == X+(-c-1)
4436 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4437 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4438 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4439 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4440 ConstantInt::get(I.getType(), 1));
4441 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4444 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4445 if (Op0I->getOpcode() == Instruction::Add) {
4446 // ~(X-c) --> (-c-1)-X
4447 if (RHS->isAllOnesValue()) {
4448 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4449 return BinaryOperator::createSub(
4450 ConstantExpr::getSub(NegOp0CI,
4451 ConstantInt::get(I.getType(), 1)),
4452 Op0I->getOperand(0));
4453 } else if (RHS->getValue().isSignBit()) {
4454 // (X + C) ^ signbit -> (X + C + signbit)
4455 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4456 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4459 } else if (Op0I->getOpcode() == Instruction::Or) {
4460 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4461 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4462 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4463 // Anything in both C1 and C2 is known to be zero, remove it from
4465 Constant *CommonBits = And(Op0CI, RHS);
4466 NewRHS = ConstantExpr::getAnd(NewRHS,
4467 ConstantExpr::getNot(CommonBits));
4468 AddToWorkList(Op0I);
4469 I.setOperand(0, Op0I->getOperand(0));
4470 I.setOperand(1, NewRHS);
4477 // Try to fold constant and into select arguments.
4478 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4479 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4481 if (isa<PHINode>(Op0))
4482 if (Instruction *NV = FoldOpIntoPhi(I))
4486 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4488 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4490 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4492 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4495 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4498 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4499 if (A == Op0) { // B^(B|A) == (A|B)^B
4500 Op1I->swapOperands();
4502 std::swap(Op0, Op1);
4503 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4504 I.swapOperands(); // Simplified below.
4505 std::swap(Op0, Op1);
4507 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4508 if (Op0 == A) // A^(A^B) == B
4509 return ReplaceInstUsesWith(I, B);
4510 else if (Op0 == B) // A^(B^A) == B
4511 return ReplaceInstUsesWith(I, A);
4512 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4513 if (A == Op0) { // A^(A&B) -> A^(B&A)
4514 Op1I->swapOperands();
4517 if (B == Op0) { // A^(B&A) -> (B&A)^A
4518 I.swapOperands(); // Simplified below.
4519 std::swap(Op0, Op1);
4524 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4527 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4528 if (A == Op1) // (B|A)^B == (A|B)^B
4530 if (B == Op1) { // (A|B)^B == A & ~B
4532 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4533 return BinaryOperator::createAnd(A, NotB);
4535 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4536 if (Op1 == A) // (A^B)^A == B
4537 return ReplaceInstUsesWith(I, B);
4538 else if (Op1 == B) // (B^A)^A == B
4539 return ReplaceInstUsesWith(I, A);
4540 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4541 if (A == Op1) // (A&B)^A -> (B&A)^A
4543 if (B == Op1 && // (B&A)^A == ~B & A
4544 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4546 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4547 return BinaryOperator::createAnd(N, Op1);
4552 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4553 if (Op0I && Op1I && Op0I->isShift() &&
4554 Op0I->getOpcode() == Op1I->getOpcode() &&
4555 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4556 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4557 Instruction *NewOp =
4558 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4559 Op1I->getOperand(0),
4560 Op0I->getName()), I);
4561 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4562 Op1I->getOperand(1));
4566 Value *A, *B, *C, *D;
4567 // (A & B)^(A | B) -> A ^ B
4568 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4569 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4570 if ((A == C && B == D) || (A == D && B == C))
4571 return BinaryOperator::createXor(A, B);
4573 // (A | B)^(A & B) -> A ^ B
4574 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4575 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4576 if ((A == C && B == D) || (A == D && B == C))
4577 return BinaryOperator::createXor(A, B);
4581 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4582 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4583 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4584 // (X & Y)^(X & Y) -> (Y^Z) & X
4585 Value *X = 0, *Y = 0, *Z = 0;
4587 X = A, Y = B, Z = D;
4589 X = A, Y = B, Z = C;
4591 X = B, Y = A, Z = D;
4593 X = B, Y = A, Z = C;
4596 Instruction *NewOp =
4597 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4598 return BinaryOperator::createAnd(NewOp, X);
4603 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4604 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4605 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4608 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4609 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4610 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4611 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4612 const Type *SrcTy = Op0C->getOperand(0)->getType();
4613 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4614 // Only do this if the casts both really cause code to be generated.
4615 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4617 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4619 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4620 Op1C->getOperand(0),
4622 InsertNewInstBefore(NewOp, I);
4623 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4627 return Changed ? &I : 0;
4630 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4631 /// overflowed for this type.
4632 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4633 ConstantInt *In2, bool IsSigned = false) {
4634 Result = cast<ConstantInt>(Add(In1, In2));
4637 if (In2->getValue().isNegative())
4638 return Result->getValue().sgt(In1->getValue());
4640 return Result->getValue().slt(In1->getValue());
4642 return Result->getValue().ult(In1->getValue());
4645 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4646 /// code necessary to compute the offset from the base pointer (without adding
4647 /// in the base pointer). Return the result as a signed integer of intptr size.
4648 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4649 TargetData &TD = IC.getTargetData();
4650 gep_type_iterator GTI = gep_type_begin(GEP);
4651 const Type *IntPtrTy = TD.getIntPtrType();
4652 Value *Result = Constant::getNullValue(IntPtrTy);
4654 // Build a mask for high order bits.
4655 unsigned IntPtrWidth = TD.getPointerSize()*8;
4656 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4658 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4659 Value *Op = GEP->getOperand(i);
4660 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4661 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4662 if (OpC->isZero()) continue;
4664 // Handle a struct index, which adds its field offset to the pointer.
4665 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4666 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4668 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4669 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4671 Result = IC.InsertNewInstBefore(
4672 BinaryOperator::createAdd(Result,
4673 ConstantInt::get(IntPtrTy, Size),
4674 GEP->getName()+".offs"), I);
4678 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4679 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4680 Scale = ConstantExpr::getMul(OC, Scale);
4681 if (Constant *RC = dyn_cast<Constant>(Result))
4682 Result = ConstantExpr::getAdd(RC, Scale);
4684 // Emit an add instruction.
4685 Result = IC.InsertNewInstBefore(
4686 BinaryOperator::createAdd(Result, Scale,
4687 GEP->getName()+".offs"), I);
4691 // Convert to correct type.
4692 if (Op->getType() != IntPtrTy) {
4693 if (Constant *OpC = dyn_cast<Constant>(Op))
4694 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4696 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4697 Op->getName()+".c"), I);
4700 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4701 if (Constant *OpC = dyn_cast<Constant>(Op))
4702 Op = ConstantExpr::getMul(OpC, Scale);
4703 else // We'll let instcombine(mul) convert this to a shl if possible.
4704 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4705 GEP->getName()+".idx"), I);
4708 // Emit an add instruction.
4709 if (isa<Constant>(Op) && isa<Constant>(Result))
4710 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4711 cast<Constant>(Result));
4713 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4714 GEP->getName()+".offs"), I);
4719 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4720 /// else. At this point we know that the GEP is on the LHS of the comparison.
4721 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4722 ICmpInst::Predicate Cond,
4724 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4726 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4727 if (isa<PointerType>(CI->getOperand(0)->getType()))
4728 RHS = CI->getOperand(0);
4730 Value *PtrBase = GEPLHS->getOperand(0);
4731 if (PtrBase == RHS) {
4732 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4733 // This transformation is valid because we know pointers can't overflow.
4734 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4735 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4736 Constant::getNullValue(Offset->getType()));
4737 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4738 // If the base pointers are different, but the indices are the same, just
4739 // compare the base pointer.
4740 if (PtrBase != GEPRHS->getOperand(0)) {
4741 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4742 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4743 GEPRHS->getOperand(0)->getType();
4745 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4746 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4747 IndicesTheSame = false;
4751 // If all indices are the same, just compare the base pointers.
4753 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4754 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4756 // Otherwise, the base pointers are different and the indices are
4757 // different, bail out.
4761 // If one of the GEPs has all zero indices, recurse.
4762 bool AllZeros = true;
4763 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4764 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4765 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4770 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4771 ICmpInst::getSwappedPredicate(Cond), I);
4773 // If the other GEP has all zero indices, recurse.
4775 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4776 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4777 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4782 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4784 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4785 // If the GEPs only differ by one index, compare it.
4786 unsigned NumDifferences = 0; // Keep track of # differences.
4787 unsigned DiffOperand = 0; // The operand that differs.
4788 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4789 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4790 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4791 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4792 // Irreconcilable differences.
4796 if (NumDifferences++) break;
4801 if (NumDifferences == 0) // SAME GEP?
4802 return ReplaceInstUsesWith(I, // No comparison is needed here.
4803 ConstantInt::get(Type::Int1Ty,
4804 isTrueWhenEqual(Cond)));
4806 else if (NumDifferences == 1) {
4807 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4808 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4809 // Make sure we do a signed comparison here.
4810 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4814 // Only lower this if the icmp is the only user of the GEP or if we expect
4815 // the result to fold to a constant!
4816 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4817 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4818 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4819 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4820 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4821 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4827 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4828 bool Changed = SimplifyCompare(I);
4829 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4831 // Fold trivial predicates.
4832 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4833 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4834 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4835 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4837 // Simplify 'fcmp pred X, X'
4839 switch (I.getPredicate()) {
4840 default: assert(0 && "Unknown predicate!");
4841 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4842 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4843 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4844 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4845 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4846 case FCmpInst::FCMP_OLT: // True if ordered and less than
4847 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4848 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4850 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4851 case FCmpInst::FCMP_ULT: // True if unordered or less than
4852 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4853 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4854 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4855 I.setPredicate(FCmpInst::FCMP_UNO);
4856 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4859 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4860 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4861 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4862 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4863 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4864 I.setPredicate(FCmpInst::FCMP_ORD);
4865 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4870 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4871 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4873 // Handle fcmp with constant RHS
4874 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4875 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4876 switch (LHSI->getOpcode()) {
4877 case Instruction::PHI:
4878 if (Instruction *NV = FoldOpIntoPhi(I))
4881 case Instruction::Select:
4882 // If either operand of the select is a constant, we can fold the
4883 // comparison into the select arms, which will cause one to be
4884 // constant folded and the select turned into a bitwise or.
4885 Value *Op1 = 0, *Op2 = 0;
4886 if (LHSI->hasOneUse()) {
4887 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4888 // Fold the known value into the constant operand.
4889 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4890 // Insert a new FCmp of the other select operand.
4891 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4892 LHSI->getOperand(2), RHSC,
4894 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4895 // Fold the known value into the constant operand.
4896 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4897 // Insert a new FCmp of the other select operand.
4898 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4899 LHSI->getOperand(1), RHSC,
4905 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4910 return Changed ? &I : 0;
4913 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4914 bool Changed = SimplifyCompare(I);
4915 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4916 const Type *Ty = Op0->getType();
4920 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4921 isTrueWhenEqual(I)));
4923 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4924 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4926 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4927 // addresses never equal each other! We already know that Op0 != Op1.
4928 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4929 isa<ConstantPointerNull>(Op0)) &&
4930 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4931 isa<ConstantPointerNull>(Op1)))
4932 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4933 !isTrueWhenEqual(I)));
4935 // icmp's with boolean values can always be turned into bitwise operations
4936 if (Ty == Type::Int1Ty) {
4937 switch (I.getPredicate()) {
4938 default: assert(0 && "Invalid icmp instruction!");
4939 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4940 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4941 InsertNewInstBefore(Xor, I);
4942 return BinaryOperator::createNot(Xor);
4944 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4945 return BinaryOperator::createXor(Op0, Op1);
4947 case ICmpInst::ICMP_UGT:
4948 case ICmpInst::ICMP_SGT:
4949 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4951 case ICmpInst::ICMP_ULT:
4952 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4953 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4954 InsertNewInstBefore(Not, I);
4955 return BinaryOperator::createAnd(Not, Op1);
4957 case ICmpInst::ICMP_UGE:
4958 case ICmpInst::ICMP_SGE:
4959 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4961 case ICmpInst::ICMP_ULE:
4962 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4963 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4964 InsertNewInstBefore(Not, I);
4965 return BinaryOperator::createOr(Not, Op1);
4970 // See if we are doing a comparison between a constant and an instruction that
4971 // can be folded into the comparison.
4972 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4975 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
4976 if (I.isEquality() && CI->isNullValue() &&
4977 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
4978 // (icmp cond A B) if cond is equality
4979 return new ICmpInst(I.getPredicate(), A, B);
4982 switch (I.getPredicate()) {
4984 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4985 if (CI->isMinValue(false))
4986 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4987 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4988 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4989 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4990 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4991 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4992 if (CI->isMinValue(true))
4993 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4994 ConstantInt::getAllOnesValue(Op0->getType()));
4998 case ICmpInst::ICMP_SLT:
4999 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5000 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5001 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5002 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5003 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5004 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5007 case ICmpInst::ICMP_UGT:
5008 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5009 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5010 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5011 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5012 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5013 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5015 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5016 if (CI->isMaxValue(true))
5017 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5018 ConstantInt::getNullValue(Op0->getType()));
5021 case ICmpInst::ICMP_SGT:
5022 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5023 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5024 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5025 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5026 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5027 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5030 case ICmpInst::ICMP_ULE:
5031 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5032 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5033 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5034 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5035 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5036 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5039 case ICmpInst::ICMP_SLE:
5040 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5041 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5042 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5043 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5044 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5045 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5048 case ICmpInst::ICMP_UGE:
5049 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5050 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5051 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5052 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5053 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5054 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5057 case ICmpInst::ICMP_SGE:
5058 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5059 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5060 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5061 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5062 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5063 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5067 // If we still have a icmp le or icmp ge instruction, turn it into the
5068 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5069 // already been handled above, this requires little checking.
5071 switch (I.getPredicate()) {
5073 case ICmpInst::ICMP_ULE:
5074 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5075 case ICmpInst::ICMP_SLE:
5076 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5077 case ICmpInst::ICMP_UGE:
5078 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5079 case ICmpInst::ICMP_SGE:
5080 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5083 // See if we can fold the comparison based on bits known to be zero or one
5084 // in the input. If this comparison is a normal comparison, it demands all
5085 // bits, if it is a sign bit comparison, it only demands the sign bit.
5088 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5090 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5091 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5092 if (SimplifyDemandedBits(Op0,
5093 isSignBit ? APInt::getSignBit(BitWidth)
5094 : APInt::getAllOnesValue(BitWidth),
5095 KnownZero, KnownOne, 0))
5098 // Given the known and unknown bits, compute a range that the LHS could be
5100 if ((KnownOne | KnownZero) != 0) {
5101 // Compute the Min, Max and RHS values based on the known bits. For the
5102 // EQ and NE we use unsigned values.
5103 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5104 const APInt& RHSVal = CI->getValue();
5105 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5106 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5109 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5112 switch (I.getPredicate()) { // LE/GE have been folded already.
5113 default: assert(0 && "Unknown icmp opcode!");
5114 case ICmpInst::ICMP_EQ:
5115 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5116 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5118 case ICmpInst::ICMP_NE:
5119 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5120 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5122 case ICmpInst::ICMP_ULT:
5123 if (Max.ult(RHSVal))
5124 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5125 if (Min.uge(RHSVal))
5126 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5128 case ICmpInst::ICMP_UGT:
5129 if (Min.ugt(RHSVal))
5130 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5131 if (Max.ule(RHSVal))
5132 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5134 case ICmpInst::ICMP_SLT:
5135 if (Max.slt(RHSVal))
5136 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5137 if (Min.sgt(RHSVal))
5138 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5140 case ICmpInst::ICMP_SGT:
5141 if (Min.sgt(RHSVal))
5142 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5143 if (Max.sle(RHSVal))
5144 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5149 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5150 // instruction, see if that instruction also has constants so that the
5151 // instruction can be folded into the icmp
5152 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5153 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5157 // Handle icmp with constant (but not simple integer constant) RHS
5158 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5159 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5160 switch (LHSI->getOpcode()) {
5161 case Instruction::GetElementPtr:
5162 if (RHSC->isNullValue()) {
5163 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5164 bool isAllZeros = true;
5165 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5166 if (!isa<Constant>(LHSI->getOperand(i)) ||
5167 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5172 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5173 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5177 case Instruction::PHI:
5178 if (Instruction *NV = FoldOpIntoPhi(I))
5181 case Instruction::Select: {
5182 // If either operand of the select is a constant, we can fold the
5183 // comparison into the select arms, which will cause one to be
5184 // constant folded and the select turned into a bitwise or.
5185 Value *Op1 = 0, *Op2 = 0;
5186 if (LHSI->hasOneUse()) {
5187 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5188 // Fold the known value into the constant operand.
5189 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5190 // Insert a new ICmp of the other select operand.
5191 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5192 LHSI->getOperand(2), RHSC,
5194 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5195 // Fold the known value into the constant operand.
5196 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5197 // Insert a new ICmp of the other select operand.
5198 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5199 LHSI->getOperand(1), RHSC,
5205 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5208 case Instruction::Malloc:
5209 // If we have (malloc != null), and if the malloc has a single use, we
5210 // can assume it is successful and remove the malloc.
5211 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5212 AddToWorkList(LHSI);
5213 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5214 !isTrueWhenEqual(I)));
5220 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5221 if (User *GEP = dyn_castGetElementPtr(Op0))
5222 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5224 if (User *GEP = dyn_castGetElementPtr(Op1))
5225 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5226 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5229 // Test to see if the operands of the icmp are casted versions of other
5230 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5232 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5233 if (isa<PointerType>(Op0->getType()) &&
5234 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5235 // We keep moving the cast from the left operand over to the right
5236 // operand, where it can often be eliminated completely.
5237 Op0 = CI->getOperand(0);
5239 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5240 // so eliminate it as well.
5241 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5242 Op1 = CI2->getOperand(0);
5244 // If Op1 is a constant, we can fold the cast into the constant.
5245 if (Op0->getType() != Op1->getType()) {
5246 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5247 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5249 // Otherwise, cast the RHS right before the icmp
5250 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5253 return new ICmpInst(I.getPredicate(), Op0, Op1);
5257 if (isa<CastInst>(Op0)) {
5258 // Handle the special case of: icmp (cast bool to X), <cst>
5259 // This comes up when you have code like
5262 // For generality, we handle any zero-extension of any operand comparison
5263 // with a constant or another cast from the same type.
5264 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5265 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5269 if (I.isEquality()) {
5270 Value *A, *B, *C, *D;
5271 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5272 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5273 Value *OtherVal = A == Op1 ? B : A;
5274 return new ICmpInst(I.getPredicate(), OtherVal,
5275 Constant::getNullValue(A->getType()));
5278 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5279 // A^c1 == C^c2 --> A == C^(c1^c2)
5280 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5281 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5282 if (Op1->hasOneUse()) {
5283 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5284 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5285 return new ICmpInst(I.getPredicate(), A,
5286 InsertNewInstBefore(Xor, I));
5289 // A^B == A^D -> B == D
5290 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5291 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5292 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5293 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5297 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5298 (A == Op0 || B == Op0)) {
5299 // A == (A^B) -> B == 0
5300 Value *OtherVal = A == Op0 ? B : A;
5301 return new ICmpInst(I.getPredicate(), OtherVal,
5302 Constant::getNullValue(A->getType()));
5304 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5305 // (A-B) == A -> B == 0
5306 return new ICmpInst(I.getPredicate(), B,
5307 Constant::getNullValue(B->getType()));
5309 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5310 // A == (A-B) -> B == 0
5311 return new ICmpInst(I.getPredicate(), B,
5312 Constant::getNullValue(B->getType()));
5315 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5316 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5317 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5318 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5319 Value *X = 0, *Y = 0, *Z = 0;
5322 X = B; Y = D; Z = A;
5323 } else if (A == D) {
5324 X = B; Y = C; Z = A;
5325 } else if (B == C) {
5326 X = A; Y = D; Z = B;
5327 } else if (B == D) {
5328 X = A; Y = C; Z = B;
5331 if (X) { // Build (X^Y) & Z
5332 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5333 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5334 I.setOperand(0, Op1);
5335 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5340 return Changed ? &I : 0;
5344 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5345 /// and CmpRHS are both known to be integer constants.
5346 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5347 ConstantInt *DivRHS) {
5348 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5349 const APInt &CmpRHSV = CmpRHS->getValue();
5351 // FIXME: If the operand types don't match the type of the divide
5352 // then don't attempt this transform. The code below doesn't have the
5353 // logic to deal with a signed divide and an unsigned compare (and
5354 // vice versa). This is because (x /s C1) <s C2 produces different
5355 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5356 // (x /u C1) <u C2. Simply casting the operands and result won't
5357 // work. :( The if statement below tests that condition and bails
5359 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5360 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5362 if (DivRHS->isZero())
5363 return 0; // The ProdOV computation fails on divide by zero.
5365 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5366 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5367 // C2 (CI). By solving for X we can turn this into a range check
5368 // instead of computing a divide.
5369 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5371 // Determine if the product overflows by seeing if the product is
5372 // not equal to the divide. Make sure we do the same kind of divide
5373 // as in the LHS instruction that we're folding.
5374 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5375 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5377 // Get the ICmp opcode
5378 ICmpInst::Predicate Pred = ICI.getPredicate();
5380 // Figure out the interval that is being checked. For example, a comparison
5381 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5382 // Compute this interval based on the constants involved and the signedness of
5383 // the compare/divide. This computes a half-open interval, keeping track of
5384 // whether either value in the interval overflows. After analysis each
5385 // overflow variable is set to 0 if it's corresponding bound variable is valid
5386 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5387 int LoOverflow = 0, HiOverflow = 0;
5388 ConstantInt *LoBound = 0, *HiBound = 0;
5391 if (!DivIsSigned) { // udiv
5392 // e.g. X/5 op 3 --> [15, 20)
5394 HiOverflow = LoOverflow = ProdOV;
5396 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5397 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5398 if (CmpRHSV == 0) { // (X / pos) op 0
5399 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5400 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5402 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5403 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5404 HiOverflow = LoOverflow = ProdOV;
5406 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5407 } else { // (X / pos) op neg
5408 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5409 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5410 LoOverflow = AddWithOverflow(LoBound, Prod,
5411 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5412 HiBound = AddOne(Prod);
5413 HiOverflow = ProdOV ? -1 : 0;
5415 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5416 if (CmpRHSV == 0) { // (X / neg) op 0
5417 // e.g. X/-5 op 0 --> [-4, 5)
5418 LoBound = AddOne(DivRHS);
5419 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5420 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5421 HiOverflow = 1; // [INTMIN+1, overflow)
5422 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5424 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5425 // e.g. X/-5 op 3 --> [-19, -14)
5426 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5428 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5429 HiBound = AddOne(Prod);
5430 } else { // (X / neg) op neg
5431 // e.g. X/-5 op -3 --> [15, 20)
5433 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5434 HiBound = Subtract(Prod, DivRHS);
5437 // Dividing by a negative swaps the condition. LT <-> GT
5438 Pred = ICmpInst::getSwappedPredicate(Pred);
5441 Value *X = DivI->getOperand(0);
5443 default: assert(0 && "Unhandled icmp opcode!");
5444 case ICmpInst::ICMP_EQ:
5445 if (LoOverflow && HiOverflow)
5446 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5447 else if (HiOverflow)
5448 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5449 ICmpInst::ICMP_UGE, X, LoBound);
5450 else if (LoOverflow)
5451 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5452 ICmpInst::ICMP_ULT, X, HiBound);
5454 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5455 case ICmpInst::ICMP_NE:
5456 if (LoOverflow && HiOverflow)
5457 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5458 else if (HiOverflow)
5459 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5460 ICmpInst::ICMP_ULT, X, LoBound);
5461 else if (LoOverflow)
5462 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5463 ICmpInst::ICMP_UGE, X, HiBound);
5465 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5466 case ICmpInst::ICMP_ULT:
5467 case ICmpInst::ICMP_SLT:
5468 if (LoOverflow == +1) // Low bound is greater than input range.
5469 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5470 if (LoOverflow == -1) // Low bound is less than input range.
5471 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5472 return new ICmpInst(Pred, X, LoBound);
5473 case ICmpInst::ICMP_UGT:
5474 case ICmpInst::ICMP_SGT:
5475 if (HiOverflow == +1) // High bound greater than input range.
5476 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5477 else if (HiOverflow == -1) // High bound less than input range.
5478 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5479 if (Pred == ICmpInst::ICMP_UGT)
5480 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5482 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5487 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5489 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5492 const APInt &RHSV = RHS->getValue();
5494 switch (LHSI->getOpcode()) {
5495 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5496 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5497 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5499 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5500 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5501 Value *CompareVal = LHSI->getOperand(0);
5503 // If the sign bit of the XorCST is not set, there is no change to
5504 // the operation, just stop using the Xor.
5505 if (!XorCST->getValue().isNegative()) {
5506 ICI.setOperand(0, CompareVal);
5507 AddToWorkList(LHSI);
5511 // Was the old condition true if the operand is positive?
5512 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5514 // If so, the new one isn't.
5515 isTrueIfPositive ^= true;
5517 if (isTrueIfPositive)
5518 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5520 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5524 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5525 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5526 LHSI->getOperand(0)->hasOneUse()) {
5527 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5529 // If the LHS is an AND of a truncating cast, we can widen the
5530 // and/compare to be the input width without changing the value
5531 // produced, eliminating a cast.
5532 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5533 // We can do this transformation if either the AND constant does not
5534 // have its sign bit set or if it is an equality comparison.
5535 // Extending a relational comparison when we're checking the sign
5536 // bit would not work.
5537 if (Cast->hasOneUse() &&
5538 (ICI.isEquality() ||
5539 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5541 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5542 APInt NewCST = AndCST->getValue();
5543 NewCST.zext(BitWidth);
5545 NewCI.zext(BitWidth);
5546 Instruction *NewAnd =
5547 BinaryOperator::createAnd(Cast->getOperand(0),
5548 ConstantInt::get(NewCST),LHSI->getName());
5549 InsertNewInstBefore(NewAnd, ICI);
5550 return new ICmpInst(ICI.getPredicate(), NewAnd,
5551 ConstantInt::get(NewCI));
5555 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5556 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5557 // happens a LOT in code produced by the C front-end, for bitfield
5559 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5560 if (Shift && !Shift->isShift())
5564 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5565 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5566 const Type *AndTy = AndCST->getType(); // Type of the and.
5568 // We can fold this as long as we can't shift unknown bits
5569 // into the mask. This can only happen with signed shift
5570 // rights, as they sign-extend.
5572 bool CanFold = Shift->isLogicalShift();
5574 // To test for the bad case of the signed shr, see if any
5575 // of the bits shifted in could be tested after the mask.
5576 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5577 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5579 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5580 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5581 AndCST->getValue()) == 0)
5587 if (Shift->getOpcode() == Instruction::Shl)
5588 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5590 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5592 // Check to see if we are shifting out any of the bits being
5594 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5595 // If we shifted bits out, the fold is not going to work out.
5596 // As a special case, check to see if this means that the
5597 // result is always true or false now.
5598 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5599 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5600 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5601 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5603 ICI.setOperand(1, NewCst);
5604 Constant *NewAndCST;
5605 if (Shift->getOpcode() == Instruction::Shl)
5606 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5608 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5609 LHSI->setOperand(1, NewAndCST);
5610 LHSI->setOperand(0, Shift->getOperand(0));
5611 AddToWorkList(Shift); // Shift is dead.
5612 AddUsesToWorkList(ICI);
5618 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5619 // preferable because it allows the C<<Y expression to be hoisted out
5620 // of a loop if Y is invariant and X is not.
5621 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5622 ICI.isEquality() && !Shift->isArithmeticShift() &&
5623 isa<Instruction>(Shift->getOperand(0))) {
5626 if (Shift->getOpcode() == Instruction::LShr) {
5627 NS = BinaryOperator::createShl(AndCST,
5628 Shift->getOperand(1), "tmp");
5630 // Insert a logical shift.
5631 NS = BinaryOperator::createLShr(AndCST,
5632 Shift->getOperand(1), "tmp");
5634 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5636 // Compute X & (C << Y).
5637 Instruction *NewAnd =
5638 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5639 InsertNewInstBefore(NewAnd, ICI);
5641 ICI.setOperand(0, NewAnd);
5647 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5648 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5651 uint32_t TypeBits = RHSV.getBitWidth();
5653 // Check that the shift amount is in range. If not, don't perform
5654 // undefined shifts. When the shift is visited it will be
5656 if (ShAmt->uge(TypeBits))
5659 if (ICI.isEquality()) {
5660 // If we are comparing against bits always shifted out, the
5661 // comparison cannot succeed.
5663 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5664 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5665 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5666 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5667 return ReplaceInstUsesWith(ICI, Cst);
5670 if (LHSI->hasOneUse()) {
5671 // Otherwise strength reduce the shift into an and.
5672 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5674 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5677 BinaryOperator::createAnd(LHSI->getOperand(0),
5678 Mask, LHSI->getName()+".mask");
5679 Value *And = InsertNewInstBefore(AndI, ICI);
5680 return new ICmpInst(ICI.getPredicate(), And,
5681 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5685 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5686 bool TrueIfSigned = false;
5687 if (LHSI->hasOneUse() &&
5688 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5689 // (X << 31) <s 0 --> (X&1) != 0
5690 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5691 (TypeBits-ShAmt->getZExtValue()-1));
5693 BinaryOperator::createAnd(LHSI->getOperand(0),
5694 Mask, LHSI->getName()+".mask");
5695 Value *And = InsertNewInstBefore(AndI, ICI);
5697 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5698 And, Constant::getNullValue(And->getType()));
5703 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5704 case Instruction::AShr: {
5705 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5708 if (ICI.isEquality()) {
5709 // Check that the shift amount is in range. If not, don't perform
5710 // undefined shifts. When the shift is visited it will be
5712 uint32_t TypeBits = RHSV.getBitWidth();
5713 if (ShAmt->uge(TypeBits))
5715 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5717 // If we are comparing against bits always shifted out, the
5718 // comparison cannot succeed.
5719 APInt Comp = RHSV << ShAmtVal;
5720 if (LHSI->getOpcode() == Instruction::LShr)
5721 Comp = Comp.lshr(ShAmtVal);
5723 Comp = Comp.ashr(ShAmtVal);
5725 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5726 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5727 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5728 return ReplaceInstUsesWith(ICI, Cst);
5731 if (LHSI->hasOneUse() || RHSV == 0) {
5732 // Otherwise strength reduce the shift into an and.
5733 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5734 Constant *Mask = ConstantInt::get(Val);
5737 BinaryOperator::createAnd(LHSI->getOperand(0),
5738 Mask, LHSI->getName()+".mask");
5739 Value *And = InsertNewInstBefore(AndI, ICI);
5740 return new ICmpInst(ICI.getPredicate(), And,
5741 ConstantExpr::getShl(RHS, ShAmt));
5747 case Instruction::SDiv:
5748 case Instruction::UDiv:
5749 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5750 // Fold this div into the comparison, producing a range check.
5751 // Determine, based on the divide type, what the range is being
5752 // checked. If there is an overflow on the low or high side, remember
5753 // it, otherwise compute the range [low, hi) bounding the new value.
5754 // See: InsertRangeTest above for the kinds of replacements possible.
5755 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5756 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5761 case Instruction::Add:
5762 // Fold: icmp pred (add, X, C1), C2
5764 if (!ICI.isEquality()) {
5765 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5767 const APInt &LHSV = LHSC->getValue();
5769 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
5772 if (ICI.isSignedPredicate()) {
5773 if (CR.getLower().isSignBit()) {
5774 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
5775 ConstantInt::get(CR.getUpper()));
5776 } else if (CR.getUpper().isSignBit()) {
5777 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
5778 ConstantInt::get(CR.getLower()));
5781 if (CR.getLower().isMinValue()) {
5782 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
5783 ConstantInt::get(CR.getUpper()));
5784 } else if (CR.getUpper().isMinValue()) {
5785 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
5786 ConstantInt::get(CR.getLower()));
5793 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5794 if (ICI.isEquality()) {
5795 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5797 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5798 // the second operand is a constant, simplify a bit.
5799 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5800 switch (BO->getOpcode()) {
5801 case Instruction::SRem:
5802 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5803 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5804 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5805 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5806 Instruction *NewRem =
5807 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5809 InsertNewInstBefore(NewRem, ICI);
5810 return new ICmpInst(ICI.getPredicate(), NewRem,
5811 Constant::getNullValue(BO->getType()));
5815 case Instruction::Add:
5816 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5817 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5818 if (BO->hasOneUse())
5819 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5820 Subtract(RHS, BOp1C));
5821 } else if (RHSV == 0) {
5822 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5823 // efficiently invertible, or if the add has just this one use.
5824 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5826 if (Value *NegVal = dyn_castNegVal(BOp1))
5827 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5828 else if (Value *NegVal = dyn_castNegVal(BOp0))
5829 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5830 else if (BO->hasOneUse()) {
5831 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5832 InsertNewInstBefore(Neg, ICI);
5834 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5838 case Instruction::Xor:
5839 // For the xor case, we can xor two constants together, eliminating
5840 // the explicit xor.
5841 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5842 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5843 ConstantExpr::getXor(RHS, BOC));
5846 case Instruction::Sub:
5847 // Replace (([sub|xor] A, B) != 0) with (A != B)
5849 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5853 case Instruction::Or:
5854 // If bits are being or'd in that are not present in the constant we
5855 // are comparing against, then the comparison could never succeed!
5856 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5857 Constant *NotCI = ConstantExpr::getNot(RHS);
5858 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5859 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5864 case Instruction::And:
5865 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5866 // If bits are being compared against that are and'd out, then the
5867 // comparison can never succeed!
5868 if ((RHSV & ~BOC->getValue()) != 0)
5869 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5872 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5873 if (RHS == BOC && RHSV.isPowerOf2())
5874 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5875 ICmpInst::ICMP_NE, LHSI,
5876 Constant::getNullValue(RHS->getType()));
5878 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5879 if (isSignBit(BOC)) {
5880 Value *X = BO->getOperand(0);
5881 Constant *Zero = Constant::getNullValue(X->getType());
5882 ICmpInst::Predicate pred = isICMP_NE ?
5883 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5884 return new ICmpInst(pred, X, Zero);
5887 // ((X & ~7) == 0) --> X < 8
5888 if (RHSV == 0 && isHighOnes(BOC)) {
5889 Value *X = BO->getOperand(0);
5890 Constant *NegX = ConstantExpr::getNeg(BOC);
5891 ICmpInst::Predicate pred = isICMP_NE ?
5892 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5893 return new ICmpInst(pred, X, NegX);
5898 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5899 // Handle icmp {eq|ne} <intrinsic>, intcst.
5900 if (II->getIntrinsicID() == Intrinsic::bswap) {
5902 ICI.setOperand(0, II->getOperand(1));
5903 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5907 } else { // Not a ICMP_EQ/ICMP_NE
5908 // If the LHS is a cast from an integral value of the same size,
5909 // then since we know the RHS is a constant, try to simlify.
5910 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5911 Value *CastOp = Cast->getOperand(0);
5912 const Type *SrcTy = CastOp->getType();
5913 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5914 if (SrcTy->isInteger() &&
5915 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5916 // If this is an unsigned comparison, try to make the comparison use
5917 // smaller constant values.
5918 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5919 // X u< 128 => X s> -1
5920 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5921 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5922 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5923 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5924 // X u> 127 => X s< 0
5925 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5926 Constant::getNullValue(SrcTy));
5934 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5935 /// We only handle extending casts so far.
5937 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5938 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5939 Value *LHSCIOp = LHSCI->getOperand(0);
5940 const Type *SrcTy = LHSCIOp->getType();
5941 const Type *DestTy = LHSCI->getType();
5944 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5945 // integer type is the same size as the pointer type.
5946 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5947 getTargetData().getPointerSizeInBits() ==
5948 cast<IntegerType>(DestTy)->getBitWidth()) {
5950 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5951 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5952 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5953 RHSOp = RHSC->getOperand(0);
5954 // If the pointer types don't match, insert a bitcast.
5955 if (LHSCIOp->getType() != RHSOp->getType())
5956 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
5960 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5963 // The code below only handles extension cast instructions, so far.
5965 if (LHSCI->getOpcode() != Instruction::ZExt &&
5966 LHSCI->getOpcode() != Instruction::SExt)
5969 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5970 bool isSignedCmp = ICI.isSignedPredicate();
5972 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5973 // Not an extension from the same type?
5974 RHSCIOp = CI->getOperand(0);
5975 if (RHSCIOp->getType() != LHSCIOp->getType())
5978 // If the signedness of the two casts doesn't agree (i.e. one is a sext
5979 // and the other is a zext), then we can't handle this.
5980 if (CI->getOpcode() != LHSCI->getOpcode())
5983 // Deal with equality cases early.
5984 if (ICI.isEquality())
5985 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5987 // A signed comparison of sign extended values simplifies into a
5988 // signed comparison.
5989 if (isSignedCmp && isSignedExt)
5990 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5992 // The other three cases all fold into an unsigned comparison.
5993 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
5996 // If we aren't dealing with a constant on the RHS, exit early
5997 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6001 // Compute the constant that would happen if we truncated to SrcTy then
6002 // reextended to DestTy.
6003 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6004 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6006 // If the re-extended constant didn't change...
6008 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6009 // For example, we might have:
6010 // %A = sext short %X to uint
6011 // %B = icmp ugt uint %A, 1330
6012 // It is incorrect to transform this into
6013 // %B = icmp ugt short %X, 1330
6014 // because %A may have negative value.
6016 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6017 // OR operation is EQ/NE.
6018 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6019 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6024 // The re-extended constant changed so the constant cannot be represented
6025 // in the shorter type. Consequently, we cannot emit a simple comparison.
6027 // First, handle some easy cases. We know the result cannot be equal at this
6028 // point so handle the ICI.isEquality() cases
6029 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6030 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6031 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6032 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6034 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6035 // should have been folded away previously and not enter in here.
6038 // We're performing a signed comparison.
6039 if (cast<ConstantInt>(CI)->getValue().isNegative())
6040 Result = ConstantInt::getFalse(); // X < (small) --> false
6042 Result = ConstantInt::getTrue(); // X < (large) --> true
6044 // We're performing an unsigned comparison.
6046 // We're performing an unsigned comp with a sign extended value.
6047 // This is true if the input is >= 0. [aka >s -1]
6048 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6049 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6050 NegOne, ICI.getName()), ICI);
6052 // Unsigned extend & unsigned compare -> always true.
6053 Result = ConstantInt::getTrue();
6057 // Finally, return the value computed.
6058 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6059 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6060 return ReplaceInstUsesWith(ICI, Result);
6062 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6063 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6064 "ICmp should be folded!");
6065 if (Constant *CI = dyn_cast<Constant>(Result))
6066 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6068 return BinaryOperator::createNot(Result);
6072 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6073 return commonShiftTransforms(I);
6076 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6077 return commonShiftTransforms(I);
6080 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6081 if (Instruction *R = commonShiftTransforms(I))
6084 Value *Op0 = I.getOperand(0);
6086 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6087 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6088 if (CSI->isAllOnesValue())
6089 return ReplaceInstUsesWith(I, CSI);
6091 // See if we can turn a signed shr into an unsigned shr.
6092 if (MaskedValueIsZero(Op0,
6093 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6094 return BinaryOperator::createLShr(Op0, I.getOperand(1));
6099 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6100 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6101 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6103 // shl X, 0 == X and shr X, 0 == X
6104 // shl 0, X == 0 and shr 0, X == 0
6105 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6106 Op0 == Constant::getNullValue(Op0->getType()))
6107 return ReplaceInstUsesWith(I, Op0);
6109 if (isa<UndefValue>(Op0)) {
6110 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6111 return ReplaceInstUsesWith(I, Op0);
6112 else // undef << X -> 0, undef >>u X -> 0
6113 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6115 if (isa<UndefValue>(Op1)) {
6116 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6117 return ReplaceInstUsesWith(I, Op0);
6118 else // X << undef, X >>u undef -> 0
6119 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6122 // Try to fold constant and into select arguments.
6123 if (isa<Constant>(Op0))
6124 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6125 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6128 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6129 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6134 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6135 BinaryOperator &I) {
6136 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6138 // See if we can simplify any instructions used by the instruction whose sole
6139 // purpose is to compute bits we don't care about.
6140 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6141 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6142 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6143 KnownZero, KnownOne))
6146 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6147 // of a signed value.
6149 if (Op1->uge(TypeBits)) {
6150 if (I.getOpcode() != Instruction::AShr)
6151 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6153 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6158 // ((X*C1) << C2) == (X * (C1 << C2))
6159 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6160 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6161 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6162 return BinaryOperator::createMul(BO->getOperand(0),
6163 ConstantExpr::getShl(BOOp, Op1));
6165 // Try to fold constant and into select arguments.
6166 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6167 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6169 if (isa<PHINode>(Op0))
6170 if (Instruction *NV = FoldOpIntoPhi(I))
6173 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6174 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6175 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6176 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6177 // place. Don't try to do this transformation in this case. Also, we
6178 // require that the input operand is a shift-by-constant so that we have
6179 // confidence that the shifts will get folded together. We could do this
6180 // xform in more cases, but it is unlikely to be profitable.
6181 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6182 isa<ConstantInt>(TrOp->getOperand(1))) {
6183 // Okay, we'll do this xform. Make the shift of shift.
6184 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6185 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6187 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6189 // For logical shifts, the truncation has the effect of making the high
6190 // part of the register be zeros. Emulate this by inserting an AND to
6191 // clear the top bits as needed. This 'and' will usually be zapped by
6192 // other xforms later if dead.
6193 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6194 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6195 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6197 // The mask we constructed says what the trunc would do if occurring
6198 // between the shifts. We want to know the effect *after* the second
6199 // shift. We know that it is a logical shift by a constant, so adjust the
6200 // mask as appropriate.
6201 if (I.getOpcode() == Instruction::Shl)
6202 MaskV <<= Op1->getZExtValue();
6204 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6205 MaskV = MaskV.lshr(Op1->getZExtValue());
6208 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6210 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6212 // Return the value truncated to the interesting size.
6213 return new TruncInst(And, I.getType());
6217 if (Op0->hasOneUse()) {
6218 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6219 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6222 switch (Op0BO->getOpcode()) {
6224 case Instruction::Add:
6225 case Instruction::And:
6226 case Instruction::Or:
6227 case Instruction::Xor: {
6228 // These operators commute.
6229 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6230 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6231 match(Op0BO->getOperand(1),
6232 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6233 Instruction *YS = BinaryOperator::createShl(
6234 Op0BO->getOperand(0), Op1,
6236 InsertNewInstBefore(YS, I); // (Y << C)
6238 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6239 Op0BO->getOperand(1)->getName());
6240 InsertNewInstBefore(X, I); // (X + (Y << C))
6241 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6242 return BinaryOperator::createAnd(X, ConstantInt::get(
6243 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6246 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6247 Value *Op0BOOp1 = Op0BO->getOperand(1);
6248 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6250 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6251 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6253 Instruction *YS = BinaryOperator::createShl(
6254 Op0BO->getOperand(0), Op1,
6256 InsertNewInstBefore(YS, I); // (Y << C)
6258 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6259 V1->getName()+".mask");
6260 InsertNewInstBefore(XM, I); // X & (CC << C)
6262 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6267 case Instruction::Sub: {
6268 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6269 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6270 match(Op0BO->getOperand(0),
6271 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6272 Instruction *YS = BinaryOperator::createShl(
6273 Op0BO->getOperand(1), Op1,
6275 InsertNewInstBefore(YS, I); // (Y << C)
6277 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6278 Op0BO->getOperand(0)->getName());
6279 InsertNewInstBefore(X, I); // (X + (Y << C))
6280 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6281 return BinaryOperator::createAnd(X, ConstantInt::get(
6282 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6285 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6286 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6287 match(Op0BO->getOperand(0),
6288 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6289 m_ConstantInt(CC))) && V2 == Op1 &&
6290 cast<BinaryOperator>(Op0BO->getOperand(0))
6291 ->getOperand(0)->hasOneUse()) {
6292 Instruction *YS = BinaryOperator::createShl(
6293 Op0BO->getOperand(1), Op1,
6295 InsertNewInstBefore(YS, I); // (Y << C)
6297 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6298 V1->getName()+".mask");
6299 InsertNewInstBefore(XM, I); // X & (CC << C)
6301 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6309 // If the operand is an bitwise operator with a constant RHS, and the
6310 // shift is the only use, we can pull it out of the shift.
6311 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6312 bool isValid = true; // Valid only for And, Or, Xor
6313 bool highBitSet = false; // Transform if high bit of constant set?
6315 switch (Op0BO->getOpcode()) {
6316 default: isValid = false; break; // Do not perform transform!
6317 case Instruction::Add:
6318 isValid = isLeftShift;
6320 case Instruction::Or:
6321 case Instruction::Xor:
6324 case Instruction::And:
6329 // If this is a signed shift right, and the high bit is modified
6330 // by the logical operation, do not perform the transformation.
6331 // The highBitSet boolean indicates the value of the high bit of
6332 // the constant which would cause it to be modified for this
6335 if (isValid && I.getOpcode() == Instruction::AShr)
6336 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6339 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6341 Instruction *NewShift =
6342 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6343 InsertNewInstBefore(NewShift, I);
6344 NewShift->takeName(Op0BO);
6346 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6353 // Find out if this is a shift of a shift by a constant.
6354 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6355 if (ShiftOp && !ShiftOp->isShift())
6358 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6359 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6360 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6361 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6362 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6363 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6364 Value *X = ShiftOp->getOperand(0);
6366 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6367 if (AmtSum > TypeBits)
6370 const IntegerType *Ty = cast<IntegerType>(I.getType());
6372 // Check for (X << c1) << c2 and (X >> c1) >> c2
6373 if (I.getOpcode() == ShiftOp->getOpcode()) {
6374 return BinaryOperator::create(I.getOpcode(), X,
6375 ConstantInt::get(Ty, AmtSum));
6376 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6377 I.getOpcode() == Instruction::AShr) {
6378 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6379 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6380 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6381 I.getOpcode() == Instruction::LShr) {
6382 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6383 Instruction *Shift =
6384 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6385 InsertNewInstBefore(Shift, I);
6387 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6388 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6391 // Okay, if we get here, one shift must be left, and the other shift must be
6392 // right. See if the amounts are equal.
6393 if (ShiftAmt1 == ShiftAmt2) {
6394 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6395 if (I.getOpcode() == Instruction::Shl) {
6396 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6397 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6399 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6400 if (I.getOpcode() == Instruction::LShr) {
6401 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6402 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6404 // We can simplify ((X << C) >>s C) into a trunc + sext.
6405 // NOTE: we could do this for any C, but that would make 'unusual' integer
6406 // types. For now, just stick to ones well-supported by the code
6408 const Type *SExtType = 0;
6409 switch (Ty->getBitWidth() - ShiftAmt1) {
6416 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6421 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6422 InsertNewInstBefore(NewTrunc, I);
6423 return new SExtInst(NewTrunc, Ty);
6425 // Otherwise, we can't handle it yet.
6426 } else if (ShiftAmt1 < ShiftAmt2) {
6427 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6429 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6430 if (I.getOpcode() == Instruction::Shl) {
6431 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6432 ShiftOp->getOpcode() == Instruction::AShr);
6433 Instruction *Shift =
6434 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6435 InsertNewInstBefore(Shift, I);
6437 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6438 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6441 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6442 if (I.getOpcode() == Instruction::LShr) {
6443 assert(ShiftOp->getOpcode() == Instruction::Shl);
6444 Instruction *Shift =
6445 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6446 InsertNewInstBefore(Shift, I);
6448 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6449 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6452 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6454 assert(ShiftAmt2 < ShiftAmt1);
6455 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6457 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6458 if (I.getOpcode() == Instruction::Shl) {
6459 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6460 ShiftOp->getOpcode() == Instruction::AShr);
6461 Instruction *Shift =
6462 BinaryOperator::create(ShiftOp->getOpcode(), X,
6463 ConstantInt::get(Ty, ShiftDiff));
6464 InsertNewInstBefore(Shift, I);
6466 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6467 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6470 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6471 if (I.getOpcode() == Instruction::LShr) {
6472 assert(ShiftOp->getOpcode() == Instruction::Shl);
6473 Instruction *Shift =
6474 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6475 InsertNewInstBefore(Shift, I);
6477 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6478 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6481 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6488 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6489 /// expression. If so, decompose it, returning some value X, such that Val is
6492 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6494 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6495 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6496 Offset = CI->getZExtValue();
6498 return ConstantInt::get(Type::Int32Ty, 0);
6499 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6500 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6501 if (I->getOpcode() == Instruction::Shl) {
6502 // This is a value scaled by '1 << the shift amt'.
6503 Scale = 1U << RHS->getZExtValue();
6505 return I->getOperand(0);
6506 } else if (I->getOpcode() == Instruction::Mul) {
6507 // This value is scaled by 'RHS'.
6508 Scale = RHS->getZExtValue();
6510 return I->getOperand(0);
6511 } else if (I->getOpcode() == Instruction::Add) {
6512 // We have X+C. Check to see if we really have (X*C2)+C1,
6513 // where C1 is divisible by C2.
6516 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6517 Offset += RHS->getZExtValue();
6524 // Otherwise, we can't look past this.
6531 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6532 /// try to eliminate the cast by moving the type information into the alloc.
6533 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6534 AllocationInst &AI) {
6535 const PointerType *PTy = cast<PointerType>(CI.getType());
6537 // Remove any uses of AI that are dead.
6538 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6540 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6541 Instruction *User = cast<Instruction>(*UI++);
6542 if (isInstructionTriviallyDead(User)) {
6543 while (UI != E && *UI == User)
6544 ++UI; // If this instruction uses AI more than once, don't break UI.
6547 DOUT << "IC: DCE: " << *User;
6548 EraseInstFromFunction(*User);
6552 // Get the type really allocated and the type casted to.
6553 const Type *AllocElTy = AI.getAllocatedType();
6554 const Type *CastElTy = PTy->getElementType();
6555 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6557 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6558 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6559 if (CastElTyAlign < AllocElTyAlign) return 0;
6561 // If the allocation has multiple uses, only promote it if we are strictly
6562 // increasing the alignment of the resultant allocation. If we keep it the
6563 // same, we open the door to infinite loops of various kinds.
6564 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6566 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6567 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6568 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6570 // See if we can satisfy the modulus by pulling a scale out of the array
6572 unsigned ArraySizeScale;
6574 Value *NumElements = // See if the array size is a decomposable linear expr.
6575 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6577 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6579 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6580 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6582 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6587 // If the allocation size is constant, form a constant mul expression
6588 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6589 if (isa<ConstantInt>(NumElements))
6590 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6591 // otherwise multiply the amount and the number of elements
6592 else if (Scale != 1) {
6593 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6594 Amt = InsertNewInstBefore(Tmp, AI);
6598 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6599 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6600 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6601 Amt = InsertNewInstBefore(Tmp, AI);
6604 AllocationInst *New;
6605 if (isa<MallocInst>(AI))
6606 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6608 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6609 InsertNewInstBefore(New, AI);
6612 // If the allocation has multiple uses, insert a cast and change all things
6613 // that used it to use the new cast. This will also hack on CI, but it will
6615 if (!AI.hasOneUse()) {
6616 AddUsesToWorkList(AI);
6617 // New is the allocation instruction, pointer typed. AI is the original
6618 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6619 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6620 InsertNewInstBefore(NewCast, AI);
6621 AI.replaceAllUsesWith(NewCast);
6623 return ReplaceInstUsesWith(CI, New);
6626 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6627 /// and return it as type Ty without inserting any new casts and without
6628 /// changing the computed value. This is used by code that tries to decide
6629 /// whether promoting or shrinking integer operations to wider or smaller types
6630 /// will allow us to eliminate a truncate or extend.
6632 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6633 /// extension operation if Ty is larger.
6634 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6635 unsigned CastOpc, int &NumCastsRemoved) {
6636 // We can always evaluate constants in another type.
6637 if (isa<ConstantInt>(V))
6640 Instruction *I = dyn_cast<Instruction>(V);
6641 if (!I) return false;
6643 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6645 // If this is an extension or truncate, we can often eliminate it.
6646 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6647 // If this is a cast from the destination type, we can trivially eliminate
6648 // it, and this will remove a cast overall.
6649 if (I->getOperand(0)->getType() == Ty) {
6650 // If the first operand is itself a cast, and is eliminable, do not count
6651 // this as an eliminable cast. We would prefer to eliminate those two
6653 if (!isa<CastInst>(I->getOperand(0)))
6659 // We can't extend or shrink something that has multiple uses: doing so would
6660 // require duplicating the instruction in general, which isn't profitable.
6661 if (!I->hasOneUse()) return false;
6663 switch (I->getOpcode()) {
6664 case Instruction::Add:
6665 case Instruction::Sub:
6666 case Instruction::And:
6667 case Instruction::Or:
6668 case Instruction::Xor:
6669 // These operators can all arbitrarily be extended or truncated.
6670 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6672 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6675 case Instruction::Mul:
6676 // A multiply can be truncated by truncating its operands.
6677 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6678 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6680 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6683 case Instruction::Shl:
6684 // If we are truncating the result of this SHL, and if it's a shift of a
6685 // constant amount, we can always perform a SHL in a smaller type.
6686 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6687 uint32_t BitWidth = Ty->getBitWidth();
6688 if (BitWidth < OrigTy->getBitWidth() &&
6689 CI->getLimitedValue(BitWidth) < BitWidth)
6690 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6694 case Instruction::LShr:
6695 // If this is a truncate of a logical shr, we can truncate it to a smaller
6696 // lshr iff we know that the bits we would otherwise be shifting in are
6698 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6699 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6700 uint32_t BitWidth = Ty->getBitWidth();
6701 if (BitWidth < OrigBitWidth &&
6702 MaskedValueIsZero(I->getOperand(0),
6703 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6704 CI->getLimitedValue(BitWidth) < BitWidth) {
6705 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6710 case Instruction::ZExt:
6711 case Instruction::SExt:
6712 case Instruction::Trunc:
6713 // If this is the same kind of case as our original (e.g. zext+zext), we
6714 // can safely replace it. Note that replacing it does not reduce the number
6715 // of casts in the input.
6716 if (I->getOpcode() == CastOpc)
6721 // TODO: Can handle more cases here.
6728 /// EvaluateInDifferentType - Given an expression that
6729 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6730 /// evaluate the expression.
6731 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6733 if (Constant *C = dyn_cast<Constant>(V))
6734 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6736 // Otherwise, it must be an instruction.
6737 Instruction *I = cast<Instruction>(V);
6738 Instruction *Res = 0;
6739 switch (I->getOpcode()) {
6740 case Instruction::Add:
6741 case Instruction::Sub:
6742 case Instruction::Mul:
6743 case Instruction::And:
6744 case Instruction::Or:
6745 case Instruction::Xor:
6746 case Instruction::AShr:
6747 case Instruction::LShr:
6748 case Instruction::Shl: {
6749 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6750 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6751 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6752 LHS, RHS, I->getName());
6755 case Instruction::Trunc:
6756 case Instruction::ZExt:
6757 case Instruction::SExt:
6758 // If the source type of the cast is the type we're trying for then we can
6759 // just return the source. There's no need to insert it because it is not
6761 if (I->getOperand(0)->getType() == Ty)
6762 return I->getOperand(0);
6764 // Otherwise, must be the same type of case, so just reinsert a new one.
6765 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6769 // TODO: Can handle more cases here.
6770 assert(0 && "Unreachable!");
6774 return InsertNewInstBefore(Res, *I);
6777 /// @brief Implement the transforms common to all CastInst visitors.
6778 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6779 Value *Src = CI.getOperand(0);
6781 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6782 // eliminate it now.
6783 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6784 if (Instruction::CastOps opc =
6785 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6786 // The first cast (CSrc) is eliminable so we need to fix up or replace
6787 // the second cast (CI). CSrc will then have a good chance of being dead.
6788 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6792 // If we are casting a select then fold the cast into the select
6793 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6794 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6797 // If we are casting a PHI then fold the cast into the PHI
6798 if (isa<PHINode>(Src))
6799 if (Instruction *NV = FoldOpIntoPhi(CI))
6805 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6806 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6807 Value *Src = CI.getOperand(0);
6809 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6810 // If casting the result of a getelementptr instruction with no offset, turn
6811 // this into a cast of the original pointer!
6812 if (GEP->hasAllZeroIndices()) {
6813 // Changing the cast operand is usually not a good idea but it is safe
6814 // here because the pointer operand is being replaced with another
6815 // pointer operand so the opcode doesn't need to change.
6817 CI.setOperand(0, GEP->getOperand(0));
6821 // If the GEP has a single use, and the base pointer is a bitcast, and the
6822 // GEP computes a constant offset, see if we can convert these three
6823 // instructions into fewer. This typically happens with unions and other
6824 // non-type-safe code.
6825 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6826 if (GEP->hasAllConstantIndices()) {
6827 // We are guaranteed to get a constant from EmitGEPOffset.
6828 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6829 int64_t Offset = OffsetV->getSExtValue();
6831 // Get the base pointer input of the bitcast, and the type it points to.
6832 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6833 const Type *GEPIdxTy =
6834 cast<PointerType>(OrigBase->getType())->getElementType();
6835 if (GEPIdxTy->isSized()) {
6836 SmallVector<Value*, 8> NewIndices;
6838 // Start with the index over the outer type. Note that the type size
6839 // might be zero (even if the offset isn't zero) if the indexed type
6840 // is something like [0 x {int, int}]
6841 const Type *IntPtrTy = TD->getIntPtrType();
6842 int64_t FirstIdx = 0;
6843 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6844 FirstIdx = Offset/TySize;
6847 // Handle silly modulus not returning values values [0..TySize).
6851 assert(Offset >= 0);
6853 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6856 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6858 // Index into the types. If we fail, set OrigBase to null.
6860 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6861 const StructLayout *SL = TD->getStructLayout(STy);
6862 if (Offset < (int64_t)SL->getSizeInBytes()) {
6863 unsigned Elt = SL->getElementContainingOffset(Offset);
6864 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6866 Offset -= SL->getElementOffset(Elt);
6867 GEPIdxTy = STy->getElementType(Elt);
6869 // Otherwise, we can't index into this, bail out.
6873 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6874 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6875 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6876 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6879 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6881 GEPIdxTy = STy->getElementType();
6883 // Otherwise, we can't index into this, bail out.
6889 // If we were able to index down into an element, create the GEP
6890 // and bitcast the result. This eliminates one bitcast, potentially
6892 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6894 NewIndices.end(), "");
6895 InsertNewInstBefore(NGEP, CI);
6896 NGEP->takeName(GEP);
6898 if (isa<BitCastInst>(CI))
6899 return new BitCastInst(NGEP, CI.getType());
6900 assert(isa<PtrToIntInst>(CI));
6901 return new PtrToIntInst(NGEP, CI.getType());
6908 return commonCastTransforms(CI);
6913 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6914 /// integer types. This function implements the common transforms for all those
6916 /// @brief Implement the transforms common to CastInst with integer operands
6917 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6918 if (Instruction *Result = commonCastTransforms(CI))
6921 Value *Src = CI.getOperand(0);
6922 const Type *SrcTy = Src->getType();
6923 const Type *DestTy = CI.getType();
6924 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6925 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6927 // See if we can simplify any instructions used by the LHS whose sole
6928 // purpose is to compute bits we don't care about.
6929 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6930 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6931 KnownZero, KnownOne))
6934 // If the source isn't an instruction or has more than one use then we
6935 // can't do anything more.
6936 Instruction *SrcI = dyn_cast<Instruction>(Src);
6937 if (!SrcI || !Src->hasOneUse())
6940 // Attempt to propagate the cast into the instruction for int->int casts.
6941 int NumCastsRemoved = 0;
6942 if (!isa<BitCastInst>(CI) &&
6943 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6944 CI.getOpcode(), NumCastsRemoved)) {
6945 // If this cast is a truncate, evaluting in a different type always
6946 // eliminates the cast, so it is always a win. If this is a zero-extension,
6947 // we need to do an AND to maintain the clear top-part of the computation,
6948 // so we require that the input have eliminated at least one cast. If this
6949 // is a sign extension, we insert two new casts (to do the extension) so we
6950 // require that two casts have been eliminated.
6952 switch (CI.getOpcode()) {
6954 // All the others use floating point so we shouldn't actually
6955 // get here because of the check above.
6956 assert(0 && "Unknown cast type");
6957 case Instruction::Trunc:
6960 case Instruction::ZExt:
6961 DoXForm = NumCastsRemoved >= 1;
6963 case Instruction::SExt:
6964 DoXForm = NumCastsRemoved >= 2;
6969 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6970 CI.getOpcode() == Instruction::SExt);
6971 assert(Res->getType() == DestTy);
6972 switch (CI.getOpcode()) {
6973 default: assert(0 && "Unknown cast type!");
6974 case Instruction::Trunc:
6975 case Instruction::BitCast:
6976 // Just replace this cast with the result.
6977 return ReplaceInstUsesWith(CI, Res);
6978 case Instruction::ZExt: {
6979 // We need to emit an AND to clear the high bits.
6980 assert(SrcBitSize < DestBitSize && "Not a zext?");
6981 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6983 return BinaryOperator::createAnd(Res, C);
6985 case Instruction::SExt:
6986 // We need to emit a cast to truncate, then a cast to sext.
6987 return CastInst::create(Instruction::SExt,
6988 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6994 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6995 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6997 switch (SrcI->getOpcode()) {
6998 case Instruction::Add:
6999 case Instruction::Mul:
7000 case Instruction::And:
7001 case Instruction::Or:
7002 case Instruction::Xor:
7003 // If we are discarding information, rewrite.
7004 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7005 // Don't insert two casts if they cannot be eliminated. We allow
7006 // two casts to be inserted if the sizes are the same. This could
7007 // only be converting signedness, which is a noop.
7008 if (DestBitSize == SrcBitSize ||
7009 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7010 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7011 Instruction::CastOps opcode = CI.getOpcode();
7012 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7013 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7014 return BinaryOperator::create(
7015 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7019 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7020 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7021 SrcI->getOpcode() == Instruction::Xor &&
7022 Op1 == ConstantInt::getTrue() &&
7023 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7024 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7025 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
7028 case Instruction::SDiv:
7029 case Instruction::UDiv:
7030 case Instruction::SRem:
7031 case Instruction::URem:
7032 // If we are just changing the sign, rewrite.
7033 if (DestBitSize == SrcBitSize) {
7034 // Don't insert two casts if they cannot be eliminated. We allow
7035 // two casts to be inserted if the sizes are the same. This could
7036 // only be converting signedness, which is a noop.
7037 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7038 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7039 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7041 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7043 return BinaryOperator::create(
7044 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7049 case Instruction::Shl:
7050 // Allow changing the sign of the source operand. Do not allow
7051 // changing the size of the shift, UNLESS the shift amount is a
7052 // constant. We must not change variable sized shifts to a smaller
7053 // size, because it is undefined to shift more bits out than exist
7055 if (DestBitSize == SrcBitSize ||
7056 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7057 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7058 Instruction::BitCast : Instruction::Trunc);
7059 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7060 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7061 return BinaryOperator::createShl(Op0c, Op1c);
7064 case Instruction::AShr:
7065 // If this is a signed shr, and if all bits shifted in are about to be
7066 // truncated off, turn it into an unsigned shr to allow greater
7068 if (DestBitSize < SrcBitSize &&
7069 isa<ConstantInt>(Op1)) {
7070 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7071 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7072 // Insert the new logical shift right.
7073 return BinaryOperator::createLShr(Op0, Op1);
7081 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7082 if (Instruction *Result = commonIntCastTransforms(CI))
7085 Value *Src = CI.getOperand(0);
7086 const Type *Ty = CI.getType();
7087 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7088 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7090 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7091 switch (SrcI->getOpcode()) {
7093 case Instruction::LShr:
7094 // We can shrink lshr to something smaller if we know the bits shifted in
7095 // are already zeros.
7096 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7097 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7099 // Get a mask for the bits shifting in.
7100 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7101 Value* SrcIOp0 = SrcI->getOperand(0);
7102 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7103 if (ShAmt >= DestBitWidth) // All zeros.
7104 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7106 // Okay, we can shrink this. Truncate the input, then return a new
7108 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7109 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7111 return BinaryOperator::createLShr(V1, V2);
7113 } else { // This is a variable shr.
7115 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7116 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7117 // loop-invariant and CSE'd.
7118 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7119 Value *One = ConstantInt::get(SrcI->getType(), 1);
7121 Value *V = InsertNewInstBefore(
7122 BinaryOperator::createShl(One, SrcI->getOperand(1),
7124 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7125 SrcI->getOperand(0),
7127 Value *Zero = Constant::getNullValue(V->getType());
7128 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7138 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7139 // If one of the common conversion will work ..
7140 if (Instruction *Result = commonIntCastTransforms(CI))
7143 Value *Src = CI.getOperand(0);
7145 // If this is a cast of a cast
7146 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7147 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7148 // types and if the sizes are just right we can convert this into a logical
7149 // 'and' which will be much cheaper than the pair of casts.
7150 if (isa<TruncInst>(CSrc)) {
7151 // Get the sizes of the types involved
7152 Value *A = CSrc->getOperand(0);
7153 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7154 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7155 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7156 // If we're actually extending zero bits and the trunc is a no-op
7157 if (MidSize < DstSize && SrcSize == DstSize) {
7158 // Replace both of the casts with an And of the type mask.
7159 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7160 Constant *AndConst = ConstantInt::get(AndValue);
7162 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7163 // Unfortunately, if the type changed, we need to cast it back.
7164 if (And->getType() != CI.getType()) {
7165 And->setName(CSrc->getName()+".mask");
7166 InsertNewInstBefore(And, CI);
7167 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7174 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7175 // If we are just checking for a icmp eq of a single bit and zext'ing it
7176 // to an integer, then shift the bit to the appropriate place and then
7177 // cast to integer to avoid the comparison.
7178 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7179 const APInt &Op1CV = Op1C->getValue();
7181 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7182 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7183 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7184 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7185 Value *In = ICI->getOperand(0);
7186 Value *Sh = ConstantInt::get(In->getType(),
7187 In->getType()->getPrimitiveSizeInBits()-1);
7188 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7189 In->getName()+".lobit"),
7191 if (In->getType() != CI.getType())
7192 In = CastInst::createIntegerCast(In, CI.getType(),
7193 false/*ZExt*/, "tmp", &CI);
7195 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7196 Constant *One = ConstantInt::get(In->getType(), 1);
7197 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7198 In->getName()+".not"),
7202 return ReplaceInstUsesWith(CI, In);
7207 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7208 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7209 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7210 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7211 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7212 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7213 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7214 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7215 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7216 // This only works for EQ and NE
7217 ICI->isEquality()) {
7218 // If Op1C some other power of two, convert:
7219 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7220 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7221 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7222 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7224 APInt KnownZeroMask(~KnownZero);
7225 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7226 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7227 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7228 // (X&4) == 2 --> false
7229 // (X&4) != 2 --> true
7230 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7231 Res = ConstantExpr::getZExt(Res, CI.getType());
7232 return ReplaceInstUsesWith(CI, Res);
7235 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7236 Value *In = ICI->getOperand(0);
7238 // Perform a logical shr by shiftamt.
7239 // Insert the shift to put the result in the low bit.
7240 In = InsertNewInstBefore(
7241 BinaryOperator::createLShr(In,
7242 ConstantInt::get(In->getType(), ShiftAmt),
7243 In->getName()+".lobit"), CI);
7246 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7247 Constant *One = ConstantInt::get(In->getType(), 1);
7248 In = BinaryOperator::createXor(In, One, "tmp");
7249 InsertNewInstBefore(cast<Instruction>(In), CI);
7252 if (CI.getType() == In->getType())
7253 return ReplaceInstUsesWith(CI, In);
7255 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7263 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7264 if (Instruction *I = commonIntCastTransforms(CI))
7267 Value *Src = CI.getOperand(0);
7269 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7270 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7271 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7272 // If we are just checking for a icmp eq of a single bit and zext'ing it
7273 // to an integer, then shift the bit to the appropriate place and then
7274 // cast to integer to avoid the comparison.
7275 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7276 const APInt &Op1CV = Op1C->getValue();
7278 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7279 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7280 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7281 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7282 Value *In = ICI->getOperand(0);
7283 Value *Sh = ConstantInt::get(In->getType(),
7284 In->getType()->getPrimitiveSizeInBits()-1);
7285 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7286 In->getName()+".lobit"),
7288 if (In->getType() != CI.getType())
7289 In = CastInst::createIntegerCast(In, CI.getType(),
7290 true/*SExt*/, "tmp", &CI);
7292 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7293 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7294 In->getName()+".not"), CI);
7296 return ReplaceInstUsesWith(CI, In);
7304 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7305 /// in the specified FP type without changing its value.
7306 static Constant *FitsInFPType(ConstantFP *CFP, const Type *FPTy,
7307 const fltSemantics &Sem) {
7308 APFloat F = CFP->getValueAPF();
7309 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7310 return ConstantFP::get(FPTy, F);
7314 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7315 /// through it until we get the source value.
7316 static Value *LookThroughFPExtensions(Value *V) {
7317 if (Instruction *I = dyn_cast<Instruction>(V))
7318 if (I->getOpcode() == Instruction::FPExt)
7319 return LookThroughFPExtensions(I->getOperand(0));
7321 // If this value is a constant, return the constant in the smallest FP type
7322 // that can accurately represent it. This allows us to turn
7323 // (float)((double)X+2.0) into x+2.0f.
7324 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7325 if (CFP->getType() == Type::PPC_FP128Ty)
7326 return V; // No constant folding of this.
7327 // See if the value can be truncated to float and then reextended.
7328 if (Value *V = FitsInFPType(CFP, Type::FloatTy, APFloat::IEEEsingle))
7330 if (CFP->getType() == Type::DoubleTy)
7331 return V; // Won't shrink.
7332 if (Value *V = FitsInFPType(CFP, Type::DoubleTy, APFloat::IEEEdouble))
7334 // Don't try to shrink to various long double types.
7340 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7341 if (Instruction *I = commonCastTransforms(CI))
7344 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7345 // smaller than the destination type, we can eliminate the truncate by doing
7346 // the add as the smaller type. This applies to add/sub/mul/div as well as
7347 // many builtins (sqrt, etc).
7348 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7349 if (OpI && OpI->hasOneUse()) {
7350 switch (OpI->getOpcode()) {
7352 case Instruction::Add:
7353 case Instruction::Sub:
7354 case Instruction::Mul:
7355 case Instruction::FDiv:
7356 case Instruction::FRem:
7357 const Type *SrcTy = OpI->getType();
7358 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7359 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7360 if (LHSTrunc->getType() != SrcTy &&
7361 RHSTrunc->getType() != SrcTy) {
7362 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7363 // If the source types were both smaller than the destination type of
7364 // the cast, do this xform.
7365 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7366 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7367 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7369 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7371 return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7380 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7381 return commonCastTransforms(CI);
7384 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7385 return commonCastTransforms(CI);
7388 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7389 return commonCastTransforms(CI);
7392 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7393 return commonCastTransforms(CI);
7396 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7397 return commonCastTransforms(CI);
7400 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7401 return commonPointerCastTransforms(CI);
7404 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7405 if (Instruction *I = commonCastTransforms(CI))
7408 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7409 if (!DestPointee->isSized()) return 0;
7411 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7414 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7415 m_ConstantInt(Cst)))) {
7416 // If the source and destination operands have the same type, see if this
7417 // is a single-index GEP.
7418 if (X->getType() == CI.getType()) {
7419 // Get the size of the pointee type.
7420 uint64_t Size = TD->getABITypeSizeInBits(DestPointee);
7422 // Convert the constant to intptr type.
7423 APInt Offset = Cst->getValue();
7424 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7426 // If Offset is evenly divisible by Size, we can do this xform.
7427 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7428 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7429 return new GetElementPtrInst(X, ConstantInt::get(Offset));
7432 // TODO: Could handle other cases, e.g. where add is indexing into field of
7434 } else if (CI.getOperand(0)->hasOneUse() &&
7435 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7436 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7437 // "inttoptr+GEP" instead of "add+intptr".
7439 // Get the size of the pointee type.
7440 uint64_t Size = TD->getABITypeSize(DestPointee);
7442 // Convert the constant to intptr type.
7443 APInt Offset = Cst->getValue();
7444 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7446 // If Offset is evenly divisible by Size, we can do this xform.
7447 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7448 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7450 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7452 return new GetElementPtrInst(P, ConstantInt::get(Offset), "tmp");
7458 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7459 // If the operands are integer typed then apply the integer transforms,
7460 // otherwise just apply the common ones.
7461 Value *Src = CI.getOperand(0);
7462 const Type *SrcTy = Src->getType();
7463 const Type *DestTy = CI.getType();
7465 if (SrcTy->isInteger() && DestTy->isInteger()) {
7466 if (Instruction *Result = commonIntCastTransforms(CI))
7468 } else if (isa<PointerType>(SrcTy)) {
7469 if (Instruction *I = commonPointerCastTransforms(CI))
7472 if (Instruction *Result = commonCastTransforms(CI))
7477 // Get rid of casts from one type to the same type. These are useless and can
7478 // be replaced by the operand.
7479 if (DestTy == Src->getType())
7480 return ReplaceInstUsesWith(CI, Src);
7482 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7483 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7484 const Type *DstElTy = DstPTy->getElementType();
7485 const Type *SrcElTy = SrcPTy->getElementType();
7487 // If we are casting a malloc or alloca to a pointer to a type of the same
7488 // size, rewrite the allocation instruction to allocate the "right" type.
7489 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7490 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7493 // If the source and destination are pointers, and this cast is equivalent
7494 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7495 // This can enhance SROA and other transforms that want type-safe pointers.
7496 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7497 unsigned NumZeros = 0;
7498 while (SrcElTy != DstElTy &&
7499 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7500 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7501 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7505 // If we found a path from the src to dest, create the getelementptr now.
7506 if (SrcElTy == DstElTy) {
7507 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7508 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7509 ((Instruction*) NULL));
7513 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7514 if (SVI->hasOneUse()) {
7515 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7516 // a bitconvert to a vector with the same # elts.
7517 if (isa<VectorType>(DestTy) &&
7518 cast<VectorType>(DestTy)->getNumElements() ==
7519 SVI->getType()->getNumElements()) {
7521 // If either of the operands is a cast from CI.getType(), then
7522 // evaluating the shuffle in the casted destination's type will allow
7523 // us to eliminate at least one cast.
7524 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7525 Tmp->getOperand(0)->getType() == DestTy) ||
7526 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7527 Tmp->getOperand(0)->getType() == DestTy)) {
7528 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7529 SVI->getOperand(0), DestTy, &CI);
7530 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7531 SVI->getOperand(1), DestTy, &CI);
7532 // Return a new shuffle vector. Use the same element ID's, as we
7533 // know the vector types match #elts.
7534 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7542 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7544 /// %D = select %cond, %C, %A
7546 /// %C = select %cond, %B, 0
7549 /// Assuming that the specified instruction is an operand to the select, return
7550 /// a bitmask indicating which operands of this instruction are foldable if they
7551 /// equal the other incoming value of the select.
7553 static unsigned GetSelectFoldableOperands(Instruction *I) {
7554 switch (I->getOpcode()) {
7555 case Instruction::Add:
7556 case Instruction::Mul:
7557 case Instruction::And:
7558 case Instruction::Or:
7559 case Instruction::Xor:
7560 return 3; // Can fold through either operand.
7561 case Instruction::Sub: // Can only fold on the amount subtracted.
7562 case Instruction::Shl: // Can only fold on the shift amount.
7563 case Instruction::LShr:
7564 case Instruction::AShr:
7567 return 0; // Cannot fold
7571 /// GetSelectFoldableConstant - For the same transformation as the previous
7572 /// function, return the identity constant that goes into the select.
7573 static Constant *GetSelectFoldableConstant(Instruction *I) {
7574 switch (I->getOpcode()) {
7575 default: assert(0 && "This cannot happen!"); abort();
7576 case Instruction::Add:
7577 case Instruction::Sub:
7578 case Instruction::Or:
7579 case Instruction::Xor:
7580 case Instruction::Shl:
7581 case Instruction::LShr:
7582 case Instruction::AShr:
7583 return Constant::getNullValue(I->getType());
7584 case Instruction::And:
7585 return Constant::getAllOnesValue(I->getType());
7586 case Instruction::Mul:
7587 return ConstantInt::get(I->getType(), 1);
7591 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7592 /// have the same opcode and only one use each. Try to simplify this.
7593 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7595 if (TI->getNumOperands() == 1) {
7596 // If this is a non-volatile load or a cast from the same type,
7599 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7602 return 0; // unknown unary op.
7605 // Fold this by inserting a select from the input values.
7606 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7607 FI->getOperand(0), SI.getName()+".v");
7608 InsertNewInstBefore(NewSI, SI);
7609 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7613 // Only handle binary operators here.
7614 if (!isa<BinaryOperator>(TI))
7617 // Figure out if the operations have any operands in common.
7618 Value *MatchOp, *OtherOpT, *OtherOpF;
7620 if (TI->getOperand(0) == FI->getOperand(0)) {
7621 MatchOp = TI->getOperand(0);
7622 OtherOpT = TI->getOperand(1);
7623 OtherOpF = FI->getOperand(1);
7624 MatchIsOpZero = true;
7625 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7626 MatchOp = TI->getOperand(1);
7627 OtherOpT = TI->getOperand(0);
7628 OtherOpF = FI->getOperand(0);
7629 MatchIsOpZero = false;
7630 } else if (!TI->isCommutative()) {
7632 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7633 MatchOp = TI->getOperand(0);
7634 OtherOpT = TI->getOperand(1);
7635 OtherOpF = FI->getOperand(0);
7636 MatchIsOpZero = true;
7637 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7638 MatchOp = TI->getOperand(1);
7639 OtherOpT = TI->getOperand(0);
7640 OtherOpF = FI->getOperand(1);
7641 MatchIsOpZero = true;
7646 // If we reach here, they do have operations in common.
7647 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7648 OtherOpF, SI.getName()+".v");
7649 InsertNewInstBefore(NewSI, SI);
7651 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7653 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7655 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7657 assert(0 && "Shouldn't get here");
7661 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7662 Value *CondVal = SI.getCondition();
7663 Value *TrueVal = SI.getTrueValue();
7664 Value *FalseVal = SI.getFalseValue();
7666 // select true, X, Y -> X
7667 // select false, X, Y -> Y
7668 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7669 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7671 // select C, X, X -> X
7672 if (TrueVal == FalseVal)
7673 return ReplaceInstUsesWith(SI, TrueVal);
7675 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7676 return ReplaceInstUsesWith(SI, FalseVal);
7677 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7678 return ReplaceInstUsesWith(SI, TrueVal);
7679 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7680 if (isa<Constant>(TrueVal))
7681 return ReplaceInstUsesWith(SI, TrueVal);
7683 return ReplaceInstUsesWith(SI, FalseVal);
7686 if (SI.getType() == Type::Int1Ty) {
7687 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7688 if (C->getZExtValue()) {
7689 // Change: A = select B, true, C --> A = or B, C
7690 return BinaryOperator::createOr(CondVal, FalseVal);
7692 // Change: A = select B, false, C --> A = and !B, C
7694 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7695 "not."+CondVal->getName()), SI);
7696 return BinaryOperator::createAnd(NotCond, FalseVal);
7698 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7699 if (C->getZExtValue() == false) {
7700 // Change: A = select B, C, false --> A = and B, C
7701 return BinaryOperator::createAnd(CondVal, TrueVal);
7703 // Change: A = select B, C, true --> A = or !B, C
7705 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7706 "not."+CondVal->getName()), SI);
7707 return BinaryOperator::createOr(NotCond, TrueVal);
7711 // select a, b, a -> a&b
7712 // select a, a, b -> a|b
7713 if (CondVal == TrueVal)
7714 return BinaryOperator::createOr(CondVal, FalseVal);
7715 else if (CondVal == FalseVal)
7716 return BinaryOperator::createAnd(CondVal, TrueVal);
7719 // Selecting between two integer constants?
7720 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7721 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7722 // select C, 1, 0 -> zext C to int
7723 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7724 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7725 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7726 // select C, 0, 1 -> zext !C to int
7728 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7729 "not."+CondVal->getName()), SI);
7730 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7733 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7735 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7737 // (x <s 0) ? -1 : 0 -> ashr x, 31
7738 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7739 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7740 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7741 // The comparison constant and the result are not neccessarily the
7742 // same width. Make an all-ones value by inserting a AShr.
7743 Value *X = IC->getOperand(0);
7744 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7745 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7746 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7748 InsertNewInstBefore(SRA, SI);
7750 // Finally, convert to the type of the select RHS. We figure out
7751 // if this requires a SExt, Trunc or BitCast based on the sizes.
7752 Instruction::CastOps opc = Instruction::BitCast;
7753 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7754 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7755 if (SRASize < SISize)
7756 opc = Instruction::SExt;
7757 else if (SRASize > SISize)
7758 opc = Instruction::Trunc;
7759 return CastInst::create(opc, SRA, SI.getType());
7764 // If one of the constants is zero (we know they can't both be) and we
7765 // have an icmp instruction with zero, and we have an 'and' with the
7766 // non-constant value, eliminate this whole mess. This corresponds to
7767 // cases like this: ((X & 27) ? 27 : 0)
7768 if (TrueValC->isZero() || FalseValC->isZero())
7769 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7770 cast<Constant>(IC->getOperand(1))->isNullValue())
7771 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7772 if (ICA->getOpcode() == Instruction::And &&
7773 isa<ConstantInt>(ICA->getOperand(1)) &&
7774 (ICA->getOperand(1) == TrueValC ||
7775 ICA->getOperand(1) == FalseValC) &&
7776 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7777 // Okay, now we know that everything is set up, we just don't
7778 // know whether we have a icmp_ne or icmp_eq and whether the
7779 // true or false val is the zero.
7780 bool ShouldNotVal = !TrueValC->isZero();
7781 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7784 V = InsertNewInstBefore(BinaryOperator::create(
7785 Instruction::Xor, V, ICA->getOperand(1)), SI);
7786 return ReplaceInstUsesWith(SI, V);
7791 // See if we are selecting two values based on a comparison of the two values.
7792 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7793 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7794 // Transform (X == Y) ? X : Y -> Y
7795 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7796 // This is not safe in general for floating point:
7797 // consider X== -0, Y== +0.
7798 // It becomes safe if either operand is a nonzero constant.
7799 ConstantFP *CFPt, *CFPf;
7800 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7801 !CFPt->getValueAPF().isZero()) ||
7802 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7803 !CFPf->getValueAPF().isZero()))
7804 return ReplaceInstUsesWith(SI, FalseVal);
7806 // Transform (X != Y) ? X : Y -> X
7807 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7808 return ReplaceInstUsesWith(SI, TrueVal);
7809 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7811 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7812 // Transform (X == Y) ? Y : X -> X
7813 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7814 // This is not safe in general for floating point:
7815 // consider X== -0, Y== +0.
7816 // It becomes safe if either operand is a nonzero constant.
7817 ConstantFP *CFPt, *CFPf;
7818 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7819 !CFPt->getValueAPF().isZero()) ||
7820 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7821 !CFPf->getValueAPF().isZero()))
7822 return ReplaceInstUsesWith(SI, FalseVal);
7824 // Transform (X != Y) ? Y : X -> Y
7825 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7826 return ReplaceInstUsesWith(SI, TrueVal);
7827 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7831 // See if we are selecting two values based on a comparison of the two values.
7832 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7833 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7834 // Transform (X == Y) ? X : Y -> Y
7835 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7836 return ReplaceInstUsesWith(SI, FalseVal);
7837 // Transform (X != Y) ? X : Y -> X
7838 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7839 return ReplaceInstUsesWith(SI, TrueVal);
7840 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7842 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7843 // Transform (X == Y) ? Y : X -> X
7844 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7845 return ReplaceInstUsesWith(SI, FalseVal);
7846 // Transform (X != Y) ? Y : X -> Y
7847 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7848 return ReplaceInstUsesWith(SI, TrueVal);
7849 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7853 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7854 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7855 if (TI->hasOneUse() && FI->hasOneUse()) {
7856 Instruction *AddOp = 0, *SubOp = 0;
7858 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7859 if (TI->getOpcode() == FI->getOpcode())
7860 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7863 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7864 // even legal for FP.
7865 if (TI->getOpcode() == Instruction::Sub &&
7866 FI->getOpcode() == Instruction::Add) {
7867 AddOp = FI; SubOp = TI;
7868 } else if (FI->getOpcode() == Instruction::Sub &&
7869 TI->getOpcode() == Instruction::Add) {
7870 AddOp = TI; SubOp = FI;
7874 Value *OtherAddOp = 0;
7875 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7876 OtherAddOp = AddOp->getOperand(1);
7877 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7878 OtherAddOp = AddOp->getOperand(0);
7882 // So at this point we know we have (Y -> OtherAddOp):
7883 // select C, (add X, Y), (sub X, Z)
7884 Value *NegVal; // Compute -Z
7885 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7886 NegVal = ConstantExpr::getNeg(C);
7888 NegVal = InsertNewInstBefore(
7889 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7892 Value *NewTrueOp = OtherAddOp;
7893 Value *NewFalseOp = NegVal;
7895 std::swap(NewTrueOp, NewFalseOp);
7896 Instruction *NewSel =
7897 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7899 NewSel = InsertNewInstBefore(NewSel, SI);
7900 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7905 // See if we can fold the select into one of our operands.
7906 if (SI.getType()->isInteger()) {
7907 // See the comment above GetSelectFoldableOperands for a description of the
7908 // transformation we are doing here.
7909 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7910 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7911 !isa<Constant>(FalseVal))
7912 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7913 unsigned OpToFold = 0;
7914 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7916 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7921 Constant *C = GetSelectFoldableConstant(TVI);
7922 Instruction *NewSel =
7923 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7924 InsertNewInstBefore(NewSel, SI);
7925 NewSel->takeName(TVI);
7926 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7927 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7929 assert(0 && "Unknown instruction!!");
7934 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7935 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7936 !isa<Constant>(TrueVal))
7937 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7938 unsigned OpToFold = 0;
7939 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7941 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7946 Constant *C = GetSelectFoldableConstant(FVI);
7947 Instruction *NewSel =
7948 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7949 InsertNewInstBefore(NewSel, SI);
7950 NewSel->takeName(FVI);
7951 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7952 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7954 assert(0 && "Unknown instruction!!");
7959 if (BinaryOperator::isNot(CondVal)) {
7960 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7961 SI.setOperand(1, FalseVal);
7962 SI.setOperand(2, TrueVal);
7969 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7970 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7971 /// and it is more than the alignment of the ultimate object, see if we can
7972 /// increase the alignment of the ultimate object, making this check succeed.
7973 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7974 unsigned PrefAlign = 0) {
7975 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7976 unsigned Align = GV->getAlignment();
7977 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7978 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7980 // If there is a large requested alignment and we can, bump up the alignment
7982 if (PrefAlign > Align && GV->hasInitializer()) {
7983 GV->setAlignment(PrefAlign);
7987 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7988 unsigned Align = AI->getAlignment();
7989 if (Align == 0 && TD) {
7990 if (isa<AllocaInst>(AI))
7991 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7992 else if (isa<MallocInst>(AI)) {
7993 // Malloc returns maximally aligned memory.
7994 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7997 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
8000 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
8004 // If there is a requested alignment and if this is an alloca, round up. We
8005 // don't do this for malloc, because some systems can't respect the request.
8006 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
8007 AI->setAlignment(PrefAlign);
8011 } else if (isa<BitCastInst>(V) ||
8012 (isa<ConstantExpr>(V) &&
8013 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
8014 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
8016 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
8017 // If all indexes are zero, it is just the alignment of the base pointer.
8018 bool AllZeroOperands = true;
8019 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
8020 if (!isa<Constant>(GEPI->getOperand(i)) ||
8021 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
8022 AllZeroOperands = false;
8026 if (AllZeroOperands) {
8027 // Treat this like a bitcast.
8028 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
8031 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
8032 if (BaseAlignment == 0) return 0;
8034 // Otherwise, if the base alignment is >= the alignment we expect for the
8035 // base pointer type, then we know that the resultant pointer is aligned at
8036 // least as much as its type requires.
8039 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
8040 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
8041 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
8042 if (Align <= BaseAlignment) {
8043 const Type *GEPTy = GEPI->getType();
8044 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
8045 Align = std::min(Align, (unsigned)
8046 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
8054 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8055 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
8056 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
8057 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8058 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8060 if (CopyAlign < MinAlign) {
8061 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8065 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8067 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8068 if (MemOpLength == 0) return 0;
8070 // Source and destination pointer types are always "i8*" for intrinsic. See
8071 // if the size is something we can handle with a single primitive load/store.
8072 // A single load+store correctly handles overlapping memory in the memmove
8074 unsigned Size = MemOpLength->getZExtValue();
8075 if (Size == 0 || Size > 8 || (Size&(Size-1)))
8076 return 0; // If not 1/2/4/8 bytes, exit.
8078 // Use an integer load+store unless we can find something better.
8079 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8081 // Memcpy forces the use of i8* for the source and destination. That means
8082 // that if you're using memcpy to move one double around, you'll get a cast
8083 // from double* to i8*. We'd much rather use a double load+store rather than
8084 // an i64 load+store, here because this improves the odds that the source or
8085 // dest address will be promotable. See if we can find a better type than the
8086 // integer datatype.
8087 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8088 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8089 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8090 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8091 // down through these levels if so.
8092 while (!SrcETy->isFirstClassType()) {
8093 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8094 if (STy->getNumElements() == 1)
8095 SrcETy = STy->getElementType(0);
8098 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8099 if (ATy->getNumElements() == 1)
8100 SrcETy = ATy->getElementType();
8107 if (SrcETy->isFirstClassType())
8108 NewPtrTy = PointerType::getUnqual(SrcETy);
8113 // If the memcpy/memmove provides better alignment info than we can
8115 SrcAlign = std::max(SrcAlign, CopyAlign);
8116 DstAlign = std::max(DstAlign, CopyAlign);
8118 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8119 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8120 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8121 InsertNewInstBefore(L, *MI);
8122 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8124 // Set the size of the copy to 0, it will be deleted on the next iteration.
8125 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8129 /// visitCallInst - CallInst simplification. This mostly only handles folding
8130 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8131 /// the heavy lifting.
8133 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8134 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8135 if (!II) return visitCallSite(&CI);
8137 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8139 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8140 bool Changed = false;
8142 // memmove/cpy/set of zero bytes is a noop.
8143 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8144 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8146 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8147 if (CI->getZExtValue() == 1) {
8148 // Replace the instruction with just byte operations. We would
8149 // transform other cases to loads/stores, but we don't know if
8150 // alignment is sufficient.
8154 // If we have a memmove and the source operation is a constant global,
8155 // then the source and dest pointers can't alias, so we can change this
8156 // into a call to memcpy.
8157 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8158 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8159 if (GVSrc->isConstant()) {
8160 Module *M = CI.getParent()->getParent()->getParent();
8161 Intrinsic::ID MemCpyID;
8162 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8163 MemCpyID = Intrinsic::memcpy_i32;
8165 MemCpyID = Intrinsic::memcpy_i64;
8166 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8171 // If we can determine a pointer alignment that is bigger than currently
8172 // set, update the alignment.
8173 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8174 if (Instruction *I = SimplifyMemTransfer(MI))
8176 } else if (isa<MemSetInst>(MI)) {
8177 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
8178 if (MI->getAlignment()->getZExtValue() < Alignment) {
8179 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8184 if (Changed) return II;
8186 switch (II->getIntrinsicID()) {
8188 case Intrinsic::ppc_altivec_lvx:
8189 case Intrinsic::ppc_altivec_lvxl:
8190 case Intrinsic::x86_sse_loadu_ps:
8191 case Intrinsic::x86_sse2_loadu_pd:
8192 case Intrinsic::x86_sse2_loadu_dq:
8193 // Turn PPC lvx -> load if the pointer is known aligned.
8194 // Turn X86 loadups -> load if the pointer is known aligned.
8195 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8196 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8197 PointerType::getUnqual(II->getType()),
8199 return new LoadInst(Ptr);
8202 case Intrinsic::ppc_altivec_stvx:
8203 case Intrinsic::ppc_altivec_stvxl:
8204 // Turn stvx -> store if the pointer is known aligned.
8205 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
8206 const Type *OpPtrTy =
8207 PointerType::getUnqual(II->getOperand(1)->getType());
8208 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8209 return new StoreInst(II->getOperand(1), Ptr);
8212 case Intrinsic::x86_sse_storeu_ps:
8213 case Intrinsic::x86_sse2_storeu_pd:
8214 case Intrinsic::x86_sse2_storeu_dq:
8215 case Intrinsic::x86_sse2_storel_dq:
8216 // Turn X86 storeu -> store if the pointer is known aligned.
8217 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8218 const Type *OpPtrTy =
8219 PointerType::getUnqual(II->getOperand(2)->getType());
8220 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8221 return new StoreInst(II->getOperand(2), Ptr);
8225 case Intrinsic::x86_sse_cvttss2si: {
8226 // These intrinsics only demands the 0th element of its input vector. If
8227 // we can simplify the input based on that, do so now.
8229 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8231 II->setOperand(1, V);
8237 case Intrinsic::ppc_altivec_vperm:
8238 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8239 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8240 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8242 // Check that all of the elements are integer constants or undefs.
8243 bool AllEltsOk = true;
8244 for (unsigned i = 0; i != 16; ++i) {
8245 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8246 !isa<UndefValue>(Mask->getOperand(i))) {
8253 // Cast the input vectors to byte vectors.
8254 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8255 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8256 Value *Result = UndefValue::get(Op0->getType());
8258 // Only extract each element once.
8259 Value *ExtractedElts[32];
8260 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8262 for (unsigned i = 0; i != 16; ++i) {
8263 if (isa<UndefValue>(Mask->getOperand(i)))
8265 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8266 Idx &= 31; // Match the hardware behavior.
8268 if (ExtractedElts[Idx] == 0) {
8270 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8271 InsertNewInstBefore(Elt, CI);
8272 ExtractedElts[Idx] = Elt;
8275 // Insert this value into the result vector.
8276 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
8277 InsertNewInstBefore(cast<Instruction>(Result), CI);
8279 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8284 case Intrinsic::stackrestore: {
8285 // If the save is right next to the restore, remove the restore. This can
8286 // happen when variable allocas are DCE'd.
8287 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8288 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8289 BasicBlock::iterator BI = SS;
8291 return EraseInstFromFunction(CI);
8295 // Scan down this block to see if there is another stack restore in the
8296 // same block without an intervening call/alloca.
8297 BasicBlock::iterator BI = II;
8298 TerminatorInst *TI = II->getParent()->getTerminator();
8299 bool CannotRemove = false;
8300 for (++BI; &*BI != TI; ++BI) {
8301 if (isa<AllocaInst>(BI)) {
8302 CannotRemove = true;
8305 if (isa<CallInst>(BI)) {
8306 if (!isa<IntrinsicInst>(BI)) {
8307 CannotRemove = true;
8310 // If there is a stackrestore below this one, remove this one.
8311 return EraseInstFromFunction(CI);
8315 // If the stack restore is in a return/unwind block and if there are no
8316 // allocas or calls between the restore and the return, nuke the restore.
8317 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8318 return EraseInstFromFunction(CI);
8324 return visitCallSite(II);
8327 // InvokeInst simplification
8329 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8330 return visitCallSite(&II);
8333 // visitCallSite - Improvements for call and invoke instructions.
8335 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8336 bool Changed = false;
8338 // If the callee is a constexpr cast of a function, attempt to move the cast
8339 // to the arguments of the call/invoke.
8340 if (transformConstExprCastCall(CS)) return 0;
8342 Value *Callee = CS.getCalledValue();
8344 if (Function *CalleeF = dyn_cast<Function>(Callee))
8345 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8346 Instruction *OldCall = CS.getInstruction();
8347 // If the call and callee calling conventions don't match, this call must
8348 // be unreachable, as the call is undefined.
8349 new StoreInst(ConstantInt::getTrue(),
8350 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8352 if (!OldCall->use_empty())
8353 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8354 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8355 return EraseInstFromFunction(*OldCall);
8359 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8360 // This instruction is not reachable, just remove it. We insert a store to
8361 // undef so that we know that this code is not reachable, despite the fact
8362 // that we can't modify the CFG here.
8363 new StoreInst(ConstantInt::getTrue(),
8364 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8365 CS.getInstruction());
8367 if (!CS.getInstruction()->use_empty())
8368 CS.getInstruction()->
8369 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8371 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8372 // Don't break the CFG, insert a dummy cond branch.
8373 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8374 ConstantInt::getTrue(), II);
8376 return EraseInstFromFunction(*CS.getInstruction());
8379 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8380 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8381 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8382 return transformCallThroughTrampoline(CS);
8384 const PointerType *PTy = cast<PointerType>(Callee->getType());
8385 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8386 if (FTy->isVarArg()) {
8387 // See if we can optimize any arguments passed through the varargs area of
8389 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8390 E = CS.arg_end(); I != E; ++I)
8391 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8392 // If this cast does not effect the value passed through the varargs
8393 // area, we can eliminate the use of the cast.
8394 Value *Op = CI->getOperand(0);
8395 if (CI->isLosslessCast()) {
8402 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8403 // Inline asm calls cannot throw - mark them 'nounwind'.
8404 CS.setDoesNotThrow();
8408 return Changed ? CS.getInstruction() : 0;
8411 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8412 // attempt to move the cast to the arguments of the call/invoke.
8414 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8415 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8416 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8417 if (CE->getOpcode() != Instruction::BitCast ||
8418 !isa<Function>(CE->getOperand(0)))
8420 Function *Callee = cast<Function>(CE->getOperand(0));
8421 Instruction *Caller = CS.getInstruction();
8422 const ParamAttrsList* CallerPAL = CS.getParamAttrs();
8424 // Okay, this is a cast from a function to a different type. Unless doing so
8425 // would cause a type conversion of one of our arguments, change this call to
8426 // be a direct call with arguments casted to the appropriate types.
8428 const FunctionType *FT = Callee->getFunctionType();
8429 const Type *OldRetTy = Caller->getType();
8431 if (isa<StructType>(FT->getReturnType()))
8432 return false; // TODO: Handle multiple return values.
8434 // Check to see if we are changing the return type...
8435 if (OldRetTy != FT->getReturnType()) {
8436 if (Callee->isDeclaration() && !Caller->use_empty() &&
8437 // Conversion is ok if changing from pointer to int of same size.
8438 !(isa<PointerType>(FT->getReturnType()) &&
8439 TD->getIntPtrType() == OldRetTy))
8440 return false; // Cannot transform this return value.
8442 if (!Caller->use_empty() &&
8443 // void -> non-void is handled specially
8444 FT->getReturnType() != Type::VoidTy &&
8445 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8446 return false; // Cannot transform this return value.
8448 if (CallerPAL && !Caller->use_empty()) {
8449 ParameterAttributes RAttrs = CallerPAL->getParamAttrs(0);
8450 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8451 return false; // Attribute not compatible with transformed value.
8454 // If the callsite is an invoke instruction, and the return value is used by
8455 // a PHI node in a successor, we cannot change the return type of the call
8456 // because there is no place to put the cast instruction (without breaking
8457 // the critical edge). Bail out in this case.
8458 if (!Caller->use_empty())
8459 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8460 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8462 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8463 if (PN->getParent() == II->getNormalDest() ||
8464 PN->getParent() == II->getUnwindDest())
8468 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8469 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8471 CallSite::arg_iterator AI = CS.arg_begin();
8472 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8473 const Type *ParamTy = FT->getParamType(i);
8474 const Type *ActTy = (*AI)->getType();
8476 if (!CastInst::isCastable(ActTy, ParamTy))
8477 return false; // Cannot transform this parameter value.
8480 ParameterAttributes PAttrs = CallerPAL->getParamAttrs(i + 1);
8481 if (PAttrs & ParamAttr::typeIncompatible(ParamTy))
8482 return false; // Attribute not compatible with transformed value.
8485 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8486 // Some conversions are safe even if we do not have a body.
8487 // Either we can cast directly, or we can upconvert the argument
8488 bool isConvertible = ActTy == ParamTy ||
8489 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8490 (ParamTy->isInteger() && ActTy->isInteger() &&
8491 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8492 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8493 && c->getValue().isStrictlyPositive());
8494 if (Callee->isDeclaration() && !isConvertible) return false;
8497 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8498 Callee->isDeclaration())
8499 return false; // Do not delete arguments unless we have a function body...
8501 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && CallerPAL)
8502 // In this case we have more arguments than the new function type, but we
8503 // won't be dropping them. Check that these extra arguments have attributes
8504 // that are compatible with being a vararg call argument.
8505 for (unsigned i = CallerPAL->size(); i; --i) {
8506 if (CallerPAL->getParamIndex(i - 1) <= FT->getNumParams())
8508 ParameterAttributes PAttrs = CallerPAL->getParamAttrsAtIndex(i - 1);
8509 if (PAttrs & ParamAttr::VarArgsIncompatible)
8513 // Okay, we decided that this is a safe thing to do: go ahead and start
8514 // inserting cast instructions as necessary...
8515 std::vector<Value*> Args;
8516 Args.reserve(NumActualArgs);
8517 ParamAttrsVector attrVec;
8518 attrVec.reserve(NumCommonArgs);
8520 // Get any return attributes.
8521 ParameterAttributes RAttrs = CallerPAL ? CallerPAL->getParamAttrs(0) :
8524 // If the return value is not being used, the type may not be compatible
8525 // with the existing attributes. Wipe out any problematic attributes.
8526 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
8528 // Add the new return attributes.
8530 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8532 AI = CS.arg_begin();
8533 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8534 const Type *ParamTy = FT->getParamType(i);
8535 if ((*AI)->getType() == ParamTy) {
8536 Args.push_back(*AI);
8538 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8539 false, ParamTy, false);
8540 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8541 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8544 // Add any parameter attributes.
8545 ParameterAttributes PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) :
8548 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8551 // If the function takes more arguments than the call was taking, add them
8553 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8554 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8556 // If we are removing arguments to the function, emit an obnoxious warning...
8557 if (FT->getNumParams() < NumActualArgs) {
8558 if (!FT->isVarArg()) {
8559 cerr << "WARNING: While resolving call to function '"
8560 << Callee->getName() << "' arguments were dropped!\n";
8562 // Add all of the arguments in their promoted form to the arg list...
8563 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8564 const Type *PTy = getPromotedType((*AI)->getType());
8565 if (PTy != (*AI)->getType()) {
8566 // Must promote to pass through va_arg area!
8567 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8569 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8570 InsertNewInstBefore(Cast, *Caller);
8571 Args.push_back(Cast);
8573 Args.push_back(*AI);
8576 // Add any parameter attributes.
8577 ParameterAttributes PAttrs = CallerPAL ?
8578 CallerPAL->getParamAttrs(i + 1) :
8581 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8586 if (FT->getReturnType() == Type::VoidTy)
8587 Caller->setName(""); // Void type should not have a name.
8589 const ParamAttrsList* NewCallerPAL = ParamAttrsList::get(attrVec);
8592 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8593 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8594 Args.begin(), Args.end(), Caller->getName(), Caller);
8595 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8596 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
8598 NC = new CallInst(Callee, Args.begin(), Args.end(),
8599 Caller->getName(), Caller);
8600 CallInst *CI = cast<CallInst>(Caller);
8601 if (CI->isTailCall())
8602 cast<CallInst>(NC)->setTailCall();
8603 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8604 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
8607 // Insert a cast of the return type as necessary.
8609 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8610 if (NV->getType() != Type::VoidTy) {
8611 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8613 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8615 // If this is an invoke instruction, we should insert it after the first
8616 // non-phi, instruction in the normal successor block.
8617 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8618 BasicBlock::iterator I = II->getNormalDest()->begin();
8619 while (isa<PHINode>(I)) ++I;
8620 InsertNewInstBefore(NC, *I);
8622 // Otherwise, it's a call, just insert cast right after the call instr
8623 InsertNewInstBefore(NC, *Caller);
8625 AddUsersToWorkList(*Caller);
8627 NV = UndefValue::get(Caller->getType());
8631 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8632 Caller->replaceAllUsesWith(NV);
8633 Caller->eraseFromParent();
8634 RemoveFromWorkList(Caller);
8638 // transformCallThroughTrampoline - Turn a call to a function created by the
8639 // init_trampoline intrinsic into a direct call to the underlying function.
8641 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8642 Value *Callee = CS.getCalledValue();
8643 const PointerType *PTy = cast<PointerType>(Callee->getType());
8644 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8645 const ParamAttrsList *Attrs = CS.getParamAttrs();
8647 // If the call already has the 'nest' attribute somewhere then give up -
8648 // otherwise 'nest' would occur twice after splicing in the chain.
8649 if (Attrs && Attrs->hasAttrSomewhere(ParamAttr::Nest))
8652 IntrinsicInst *Tramp =
8653 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8656 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8657 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8658 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8660 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8661 unsigned NestIdx = 1;
8662 const Type *NestTy = 0;
8663 ParameterAttributes NestAttr = ParamAttr::None;
8665 // Look for a parameter marked with the 'nest' attribute.
8666 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8667 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8668 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8669 // Record the parameter type and any other attributes.
8671 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8676 Instruction *Caller = CS.getInstruction();
8677 std::vector<Value*> NewArgs;
8678 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8680 ParamAttrsVector NewAttrs;
8681 NewAttrs.reserve(Attrs ? Attrs->size() + 1 : 1);
8683 // Insert the nest argument into the call argument list, which may
8684 // mean appending it. Likewise for attributes.
8686 // Add any function result attributes.
8687 ParameterAttributes Attr = Attrs ? Attrs->getParamAttrs(0) :
8690 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8694 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8696 if (Idx == NestIdx) {
8697 // Add the chain argument and attributes.
8698 Value *NestVal = Tramp->getOperand(3);
8699 if (NestVal->getType() != NestTy)
8700 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8701 NewArgs.push_back(NestVal);
8702 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8708 // Add the original argument and attributes.
8709 NewArgs.push_back(*I);
8710 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8713 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8719 // The trampoline may have been bitcast to a bogus type (FTy).
8720 // Handle this by synthesizing a new function type, equal to FTy
8721 // with the chain parameter inserted.
8723 std::vector<const Type*> NewTypes;
8724 NewTypes.reserve(FTy->getNumParams()+1);
8726 // Insert the chain's type into the list of parameter types, which may
8727 // mean appending it.
8730 FunctionType::param_iterator I = FTy->param_begin(),
8731 E = FTy->param_end();
8735 // Add the chain's type.
8736 NewTypes.push_back(NestTy);
8741 // Add the original type.
8742 NewTypes.push_back(*I);
8748 // Replace the trampoline call with a direct call. Let the generic
8749 // code sort out any function type mismatches.
8750 FunctionType *NewFTy =
8751 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8752 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8753 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8754 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8756 Instruction *NewCaller;
8757 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8758 NewCaller = new InvokeInst(NewCallee,
8759 II->getNormalDest(), II->getUnwindDest(),
8760 NewArgs.begin(), NewArgs.end(),
8761 Caller->getName(), Caller);
8762 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8763 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8765 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8766 Caller->getName(), Caller);
8767 if (cast<CallInst>(Caller)->isTailCall())
8768 cast<CallInst>(NewCaller)->setTailCall();
8769 cast<CallInst>(NewCaller)->
8770 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8771 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8773 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8774 Caller->replaceAllUsesWith(NewCaller);
8775 Caller->eraseFromParent();
8776 RemoveFromWorkList(Caller);
8781 // Replace the trampoline call with a direct call. Since there is no 'nest'
8782 // parameter, there is no need to adjust the argument list. Let the generic
8783 // code sort out any function type mismatches.
8784 Constant *NewCallee =
8785 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8786 CS.setCalledFunction(NewCallee);
8787 return CS.getInstruction();
8790 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8791 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8792 /// and a single binop.
8793 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8794 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8795 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8796 isa<CmpInst>(FirstInst));
8797 unsigned Opc = FirstInst->getOpcode();
8798 Value *LHSVal = FirstInst->getOperand(0);
8799 Value *RHSVal = FirstInst->getOperand(1);
8801 const Type *LHSType = LHSVal->getType();
8802 const Type *RHSType = RHSVal->getType();
8804 // Scan to see if all operands are the same opcode, all have one use, and all
8805 // kill their operands (i.e. the operands have one use).
8806 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8807 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8808 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8809 // Verify type of the LHS matches so we don't fold cmp's of different
8810 // types or GEP's with different index types.
8811 I->getOperand(0)->getType() != LHSType ||
8812 I->getOperand(1)->getType() != RHSType)
8815 // If they are CmpInst instructions, check their predicates
8816 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8817 if (cast<CmpInst>(I)->getPredicate() !=
8818 cast<CmpInst>(FirstInst)->getPredicate())
8821 // Keep track of which operand needs a phi node.
8822 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8823 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8826 // Otherwise, this is safe to transform, determine if it is profitable.
8828 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8829 // Indexes are often folded into load/store instructions, so we don't want to
8830 // hide them behind a phi.
8831 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8834 Value *InLHS = FirstInst->getOperand(0);
8835 Value *InRHS = FirstInst->getOperand(1);
8836 PHINode *NewLHS = 0, *NewRHS = 0;
8838 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8839 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8840 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8841 InsertNewInstBefore(NewLHS, PN);
8846 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8847 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8848 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8849 InsertNewInstBefore(NewRHS, PN);
8853 // Add all operands to the new PHIs.
8854 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8856 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8857 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8860 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8861 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8865 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8866 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8867 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8868 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8871 assert(isa<GetElementPtrInst>(FirstInst));
8872 return new GetElementPtrInst(LHSVal, RHSVal);
8876 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8877 /// of the block that defines it. This means that it must be obvious the value
8878 /// of the load is not changed from the point of the load to the end of the
8881 /// Finally, it is safe, but not profitable, to sink a load targetting a
8882 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8884 static bool isSafeToSinkLoad(LoadInst *L) {
8885 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8887 for (++BBI; BBI != E; ++BBI)
8888 if (BBI->mayWriteToMemory())
8891 // Check for non-address taken alloca. If not address-taken already, it isn't
8892 // profitable to do this xform.
8893 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8894 bool isAddressTaken = false;
8895 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8897 if (isa<LoadInst>(UI)) continue;
8898 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8899 // If storing TO the alloca, then the address isn't taken.
8900 if (SI->getOperand(1) == AI) continue;
8902 isAddressTaken = true;
8906 if (!isAddressTaken)
8914 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8915 // operator and they all are only used by the PHI, PHI together their
8916 // inputs, and do the operation once, to the result of the PHI.
8917 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8918 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8920 // Scan the instruction, looking for input operations that can be folded away.
8921 // If all input operands to the phi are the same instruction (e.g. a cast from
8922 // the same type or "+42") we can pull the operation through the PHI, reducing
8923 // code size and simplifying code.
8924 Constant *ConstantOp = 0;
8925 const Type *CastSrcTy = 0;
8926 bool isVolatile = false;
8927 if (isa<CastInst>(FirstInst)) {
8928 CastSrcTy = FirstInst->getOperand(0)->getType();
8929 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8930 // Can fold binop, compare or shift here if the RHS is a constant,
8931 // otherwise call FoldPHIArgBinOpIntoPHI.
8932 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8933 if (ConstantOp == 0)
8934 return FoldPHIArgBinOpIntoPHI(PN);
8935 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8936 isVolatile = LI->isVolatile();
8937 // We can't sink the load if the loaded value could be modified between the
8938 // load and the PHI.
8939 if (LI->getParent() != PN.getIncomingBlock(0) ||
8940 !isSafeToSinkLoad(LI))
8942 } else if (isa<GetElementPtrInst>(FirstInst)) {
8943 if (FirstInst->getNumOperands() == 2)
8944 return FoldPHIArgBinOpIntoPHI(PN);
8945 // Can't handle general GEPs yet.
8948 return 0; // Cannot fold this operation.
8951 // Check to see if all arguments are the same operation.
8952 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8953 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8954 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8955 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8958 if (I->getOperand(0)->getType() != CastSrcTy)
8959 return 0; // Cast operation must match.
8960 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8961 // We can't sink the load if the loaded value could be modified between
8962 // the load and the PHI.
8963 if (LI->isVolatile() != isVolatile ||
8964 LI->getParent() != PN.getIncomingBlock(i) ||
8965 !isSafeToSinkLoad(LI))
8967 } else if (I->getOperand(1) != ConstantOp) {
8972 // Okay, they are all the same operation. Create a new PHI node of the
8973 // correct type, and PHI together all of the LHS's of the instructions.
8974 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8975 PN.getName()+".in");
8976 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8978 Value *InVal = FirstInst->getOperand(0);
8979 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8981 // Add all operands to the new PHI.
8982 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8983 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8984 if (NewInVal != InVal)
8986 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8991 // The new PHI unions all of the same values together. This is really
8992 // common, so we handle it intelligently here for compile-time speed.
8996 InsertNewInstBefore(NewPN, PN);
9000 // Insert and return the new operation.
9001 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9002 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
9003 else if (isa<LoadInst>(FirstInst))
9004 return new LoadInst(PhiVal, "", isVolatile);
9005 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9006 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
9007 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9008 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
9009 PhiVal, ConstantOp);
9011 assert(0 && "Unknown operation");
9015 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9017 static bool DeadPHICycle(PHINode *PN,
9018 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9019 if (PN->use_empty()) return true;
9020 if (!PN->hasOneUse()) return false;
9022 // Remember this node, and if we find the cycle, return.
9023 if (!PotentiallyDeadPHIs.insert(PN))
9026 // Don't scan crazily complex things.
9027 if (PotentiallyDeadPHIs.size() == 16)
9030 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9031 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9036 /// PHIsEqualValue - Return true if this phi node is always equal to
9037 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9038 /// z = some value; x = phi (y, z); y = phi (x, z)
9039 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9040 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9041 // See if we already saw this PHI node.
9042 if (!ValueEqualPHIs.insert(PN))
9045 // Don't scan crazily complex things.
9046 if (ValueEqualPHIs.size() == 16)
9049 // Scan the operands to see if they are either phi nodes or are equal to
9051 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9052 Value *Op = PN->getIncomingValue(i);
9053 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9054 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9056 } else if (Op != NonPhiInVal)
9064 // PHINode simplification
9066 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9067 // If LCSSA is around, don't mess with Phi nodes
9068 if (MustPreserveLCSSA) return 0;
9070 if (Value *V = PN.hasConstantValue())
9071 return ReplaceInstUsesWith(PN, V);
9073 // If all PHI operands are the same operation, pull them through the PHI,
9074 // reducing code size.
9075 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9076 PN.getIncomingValue(0)->hasOneUse())
9077 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9080 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9081 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9082 // PHI)... break the cycle.
9083 if (PN.hasOneUse()) {
9084 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9085 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9086 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9087 PotentiallyDeadPHIs.insert(&PN);
9088 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9089 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9092 // If this phi has a single use, and if that use just computes a value for
9093 // the next iteration of a loop, delete the phi. This occurs with unused
9094 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9095 // common case here is good because the only other things that catch this
9096 // are induction variable analysis (sometimes) and ADCE, which is only run
9098 if (PHIUser->hasOneUse() &&
9099 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9100 PHIUser->use_back() == &PN) {
9101 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9105 // We sometimes end up with phi cycles that non-obviously end up being the
9106 // same value, for example:
9107 // z = some value; x = phi (y, z); y = phi (x, z)
9108 // where the phi nodes don't necessarily need to be in the same block. Do a
9109 // quick check to see if the PHI node only contains a single non-phi value, if
9110 // so, scan to see if the phi cycle is actually equal to that value.
9112 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9113 // Scan for the first non-phi operand.
9114 while (InValNo != NumOperandVals &&
9115 isa<PHINode>(PN.getIncomingValue(InValNo)))
9118 if (InValNo != NumOperandVals) {
9119 Value *NonPhiInVal = PN.getOperand(InValNo);
9121 // Scan the rest of the operands to see if there are any conflicts, if so
9122 // there is no need to recursively scan other phis.
9123 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9124 Value *OpVal = PN.getIncomingValue(InValNo);
9125 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9129 // If we scanned over all operands, then we have one unique value plus
9130 // phi values. Scan PHI nodes to see if they all merge in each other or
9132 if (InValNo == NumOperandVals) {
9133 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9134 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9135 return ReplaceInstUsesWith(PN, NonPhiInVal);
9142 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9143 Instruction *InsertPoint,
9145 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9146 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9147 // We must cast correctly to the pointer type. Ensure that we
9148 // sign extend the integer value if it is smaller as this is
9149 // used for address computation.
9150 Instruction::CastOps opcode =
9151 (VTySize < PtrSize ? Instruction::SExt :
9152 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9153 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9157 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9158 Value *PtrOp = GEP.getOperand(0);
9159 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9160 // If so, eliminate the noop.
9161 if (GEP.getNumOperands() == 1)
9162 return ReplaceInstUsesWith(GEP, PtrOp);
9164 if (isa<UndefValue>(GEP.getOperand(0)))
9165 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9167 bool HasZeroPointerIndex = false;
9168 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9169 HasZeroPointerIndex = C->isNullValue();
9171 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9172 return ReplaceInstUsesWith(GEP, PtrOp);
9174 // Eliminate unneeded casts for indices.
9175 bool MadeChange = false;
9177 gep_type_iterator GTI = gep_type_begin(GEP);
9178 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
9179 if (isa<SequentialType>(*GTI)) {
9180 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
9181 if (CI->getOpcode() == Instruction::ZExt ||
9182 CI->getOpcode() == Instruction::SExt) {
9183 const Type *SrcTy = CI->getOperand(0)->getType();
9184 // We can eliminate a cast from i32 to i64 iff the target
9185 // is a 32-bit pointer target.
9186 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9188 GEP.setOperand(i, CI->getOperand(0));
9192 // If we are using a wider index than needed for this platform, shrink it
9193 // to what we need. If the incoming value needs a cast instruction,
9194 // insert it. This explicit cast can make subsequent optimizations more
9196 Value *Op = GEP.getOperand(i);
9197 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9198 if (Constant *C = dyn_cast<Constant>(Op)) {
9199 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
9202 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9204 GEP.setOperand(i, Op);
9210 if (MadeChange) return &GEP;
9212 // If this GEP instruction doesn't move the pointer, and if the input operand
9213 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9214 // real input to the dest type.
9215 if (GEP.hasAllZeroIndices()) {
9216 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9217 // If the bitcast is of an allocation, and the allocation will be
9218 // converted to match the type of the cast, don't touch this.
9219 if (isa<AllocationInst>(BCI->getOperand(0))) {
9220 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9221 if (Instruction *I = visitBitCast(*BCI)) {
9224 BCI->getParent()->getInstList().insert(BCI, I);
9225 ReplaceInstUsesWith(*BCI, I);
9230 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9234 // Combine Indices - If the source pointer to this getelementptr instruction
9235 // is a getelementptr instruction, combine the indices of the two
9236 // getelementptr instructions into a single instruction.
9238 SmallVector<Value*, 8> SrcGEPOperands;
9239 if (User *Src = dyn_castGetElementPtr(PtrOp))
9240 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9242 if (!SrcGEPOperands.empty()) {
9243 // Note that if our source is a gep chain itself that we wait for that
9244 // chain to be resolved before we perform this transformation. This
9245 // avoids us creating a TON of code in some cases.
9247 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9248 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9249 return 0; // Wait until our source is folded to completion.
9251 SmallVector<Value*, 8> Indices;
9253 // Find out whether the last index in the source GEP is a sequential idx.
9254 bool EndsWithSequential = false;
9255 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9256 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9257 EndsWithSequential = !isa<StructType>(*I);
9259 // Can we combine the two pointer arithmetics offsets?
9260 if (EndsWithSequential) {
9261 // Replace: gep (gep %P, long B), long A, ...
9262 // With: T = long A+B; gep %P, T, ...
9264 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9265 if (SO1 == Constant::getNullValue(SO1->getType())) {
9267 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9270 // If they aren't the same type, convert both to an integer of the
9271 // target's pointer size.
9272 if (SO1->getType() != GO1->getType()) {
9273 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9274 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9275 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9276 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9278 unsigned PS = TD->getPointerSizeInBits();
9279 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9280 // Convert GO1 to SO1's type.
9281 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9283 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9284 // Convert SO1 to GO1's type.
9285 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9287 const Type *PT = TD->getIntPtrType();
9288 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9289 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9293 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9294 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9296 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
9297 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9301 // Recycle the GEP we already have if possible.
9302 if (SrcGEPOperands.size() == 2) {
9303 GEP.setOperand(0, SrcGEPOperands[0]);
9304 GEP.setOperand(1, Sum);
9307 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9308 SrcGEPOperands.end()-1);
9309 Indices.push_back(Sum);
9310 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9312 } else if (isa<Constant>(*GEP.idx_begin()) &&
9313 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9314 SrcGEPOperands.size() != 1) {
9315 // Otherwise we can do the fold if the first index of the GEP is a zero
9316 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9317 SrcGEPOperands.end());
9318 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9321 if (!Indices.empty())
9322 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
9323 Indices.end(), GEP.getName());
9325 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9326 // GEP of global variable. If all of the indices for this GEP are
9327 // constants, we can promote this to a constexpr instead of an instruction.
9329 // Scan for nonconstants...
9330 SmallVector<Constant*, 8> Indices;
9331 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9332 for (; I != E && isa<Constant>(*I); ++I)
9333 Indices.push_back(cast<Constant>(*I));
9335 if (I == E) { // If they are all constants...
9336 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9337 &Indices[0],Indices.size());
9339 // Replace all uses of the GEP with the new constexpr...
9340 return ReplaceInstUsesWith(GEP, CE);
9342 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9343 if (!isa<PointerType>(X->getType())) {
9344 // Not interesting. Source pointer must be a cast from pointer.
9345 } else if (HasZeroPointerIndex) {
9346 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9347 // into : GEP [10 x i8]* X, i32 0, ...
9349 // This occurs when the program declares an array extern like "int X[];"
9351 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9352 const PointerType *XTy = cast<PointerType>(X->getType());
9353 if (const ArrayType *XATy =
9354 dyn_cast<ArrayType>(XTy->getElementType()))
9355 if (const ArrayType *CATy =
9356 dyn_cast<ArrayType>(CPTy->getElementType()))
9357 if (CATy->getElementType() == XATy->getElementType()) {
9358 // At this point, we know that the cast source type is a pointer
9359 // to an array of the same type as the destination pointer
9360 // array. Because the array type is never stepped over (there
9361 // is a leading zero) we can fold the cast into this GEP.
9362 GEP.setOperand(0, X);
9365 } else if (GEP.getNumOperands() == 2) {
9366 // Transform things like:
9367 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9368 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9369 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9370 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9371 if (isa<ArrayType>(SrcElTy) &&
9372 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9373 TD->getABITypeSize(ResElTy)) {
9375 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9376 Idx[1] = GEP.getOperand(1);
9377 Value *V = InsertNewInstBefore(
9378 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
9379 // V and GEP are both pointer types --> BitCast
9380 return new BitCastInst(V, GEP.getType());
9383 // Transform things like:
9384 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9385 // (where tmp = 8*tmp2) into:
9386 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9388 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9389 uint64_t ArrayEltSize =
9390 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9392 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9393 // allow either a mul, shift, or constant here.
9395 ConstantInt *Scale = 0;
9396 if (ArrayEltSize == 1) {
9397 NewIdx = GEP.getOperand(1);
9398 Scale = ConstantInt::get(NewIdx->getType(), 1);
9399 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9400 NewIdx = ConstantInt::get(CI->getType(), 1);
9402 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9403 if (Inst->getOpcode() == Instruction::Shl &&
9404 isa<ConstantInt>(Inst->getOperand(1))) {
9405 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9406 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9407 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9408 NewIdx = Inst->getOperand(0);
9409 } else if (Inst->getOpcode() == Instruction::Mul &&
9410 isa<ConstantInt>(Inst->getOperand(1))) {
9411 Scale = cast<ConstantInt>(Inst->getOperand(1));
9412 NewIdx = Inst->getOperand(0);
9416 // If the index will be to exactly the right offset with the scale taken
9417 // out, perform the transformation. Note, we don't know whether Scale is
9418 // signed or not. We'll use unsigned version of division/modulo
9419 // operation after making sure Scale doesn't have the sign bit set.
9420 if (Scale && Scale->getSExtValue() >= 0LL &&
9421 Scale->getZExtValue() % ArrayEltSize == 0) {
9422 Scale = ConstantInt::get(Scale->getType(),
9423 Scale->getZExtValue() / ArrayEltSize);
9424 if (Scale->getZExtValue() != 1) {
9425 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9427 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9428 NewIdx = InsertNewInstBefore(Sc, GEP);
9431 // Insert the new GEP instruction.
9433 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9435 Instruction *NewGEP =
9436 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
9437 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9438 // The NewGEP must be pointer typed, so must the old one -> BitCast
9439 return new BitCastInst(NewGEP, GEP.getType());
9448 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9449 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9450 if (AI.isArrayAllocation()) { // Check C != 1
9451 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9453 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9454 AllocationInst *New = 0;
9456 // Create and insert the replacement instruction...
9457 if (isa<MallocInst>(AI))
9458 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9460 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9461 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9464 InsertNewInstBefore(New, AI);
9466 // Scan to the end of the allocation instructions, to skip over a block of
9467 // allocas if possible...
9469 BasicBlock::iterator It = New;
9470 while (isa<AllocationInst>(*It)) ++It;
9472 // Now that I is pointing to the first non-allocation-inst in the block,
9473 // insert our getelementptr instruction...
9475 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9479 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9480 New->getName()+".sub", It);
9482 // Now make everything use the getelementptr instead of the original
9484 return ReplaceInstUsesWith(AI, V);
9485 } else if (isa<UndefValue>(AI.getArraySize())) {
9486 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9490 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9491 // Note that we only do this for alloca's, because malloc should allocate and
9492 // return a unique pointer, even for a zero byte allocation.
9493 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9494 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9495 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9500 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9501 Value *Op = FI.getOperand(0);
9503 // free undef -> unreachable.
9504 if (isa<UndefValue>(Op)) {
9505 // Insert a new store to null because we cannot modify the CFG here.
9506 new StoreInst(ConstantInt::getTrue(),
9507 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9508 return EraseInstFromFunction(FI);
9511 // If we have 'free null' delete the instruction. This can happen in stl code
9512 // when lots of inlining happens.
9513 if (isa<ConstantPointerNull>(Op))
9514 return EraseInstFromFunction(FI);
9516 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9517 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9518 FI.setOperand(0, CI->getOperand(0));
9522 // Change free (gep X, 0,0,0,0) into free(X)
9523 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9524 if (GEPI->hasAllZeroIndices()) {
9525 AddToWorkList(GEPI);
9526 FI.setOperand(0, GEPI->getOperand(0));
9531 // Change free(malloc) into nothing, if the malloc has a single use.
9532 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9533 if (MI->hasOneUse()) {
9534 EraseInstFromFunction(FI);
9535 return EraseInstFromFunction(*MI);
9542 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9543 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9544 const TargetData *TD) {
9545 User *CI = cast<User>(LI.getOperand(0));
9546 Value *CastOp = CI->getOperand(0);
9548 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9549 // Instead of loading constant c string, use corresponding integer value
9550 // directly if string length is small enough.
9551 const std::string &Str = CE->getOperand(0)->getStringValue();
9553 unsigned len = Str.length();
9554 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9555 unsigned numBits = Ty->getPrimitiveSizeInBits();
9556 // Replace LI with immediate integer store.
9557 if ((numBits >> 3) == len + 1) {
9558 APInt StrVal(numBits, 0);
9559 APInt SingleChar(numBits, 0);
9560 if (TD->isLittleEndian()) {
9561 for (signed i = len-1; i >= 0; i--) {
9562 SingleChar = (uint64_t) Str[i];
9563 StrVal = (StrVal << 8) | SingleChar;
9566 for (unsigned i = 0; i < len; i++) {
9567 SingleChar = (uint64_t) Str[i];
9568 StrVal = (StrVal << 8) | SingleChar;
9570 // Append NULL at the end.
9572 StrVal = (StrVal << 8) | SingleChar;
9574 Value *NL = ConstantInt::get(StrVal);
9575 return IC.ReplaceInstUsesWith(LI, NL);
9580 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9581 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9582 const Type *SrcPTy = SrcTy->getElementType();
9584 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9585 isa<VectorType>(DestPTy)) {
9586 // If the source is an array, the code below will not succeed. Check to
9587 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9589 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9590 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9591 if (ASrcTy->getNumElements() != 0) {
9593 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9594 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9595 SrcTy = cast<PointerType>(CastOp->getType());
9596 SrcPTy = SrcTy->getElementType();
9599 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9600 isa<VectorType>(SrcPTy)) &&
9601 // Do not allow turning this into a load of an integer, which is then
9602 // casted to a pointer, this pessimizes pointer analysis a lot.
9603 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9604 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9605 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9607 // Okay, we are casting from one integer or pointer type to another of
9608 // the same size. Instead of casting the pointer before the load, cast
9609 // the result of the loaded value.
9610 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9612 LI.isVolatile()),LI);
9613 // Now cast the result of the load.
9614 return new BitCastInst(NewLoad, LI.getType());
9621 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9622 /// from this value cannot trap. If it is not obviously safe to load from the
9623 /// specified pointer, we do a quick local scan of the basic block containing
9624 /// ScanFrom, to determine if the address is already accessed.
9625 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9626 // If it is an alloca it is always safe to load from.
9627 if (isa<AllocaInst>(V)) return true;
9629 // If it is a global variable it is mostly safe to load from.
9630 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9631 // Don't try to evaluate aliases. External weak GV can be null.
9632 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9634 // Otherwise, be a little bit agressive by scanning the local block where we
9635 // want to check to see if the pointer is already being loaded or stored
9636 // from/to. If so, the previous load or store would have already trapped,
9637 // so there is no harm doing an extra load (also, CSE will later eliminate
9638 // the load entirely).
9639 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9644 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9645 if (LI->getOperand(0) == V) return true;
9646 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9647 if (SI->getOperand(1) == V) return true;
9653 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9654 /// until we find the underlying object a pointer is referring to or something
9655 /// we don't understand. Note that the returned pointer may be offset from the
9656 /// input, because we ignore GEP indices.
9657 static Value *GetUnderlyingObject(Value *Ptr) {
9659 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9660 if (CE->getOpcode() == Instruction::BitCast ||
9661 CE->getOpcode() == Instruction::GetElementPtr)
9662 Ptr = CE->getOperand(0);
9665 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9666 Ptr = BCI->getOperand(0);
9667 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9668 Ptr = GEP->getOperand(0);
9675 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9676 Value *Op = LI.getOperand(0);
9678 // Attempt to improve the alignment.
9679 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9680 if (KnownAlign > LI.getAlignment())
9681 LI.setAlignment(KnownAlign);
9683 // load (cast X) --> cast (load X) iff safe
9684 if (isa<CastInst>(Op))
9685 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9688 // None of the following transforms are legal for volatile loads.
9689 if (LI.isVolatile()) return 0;
9691 if (&LI.getParent()->front() != &LI) {
9692 BasicBlock::iterator BBI = &LI; --BBI;
9693 // If the instruction immediately before this is a store to the same
9694 // address, do a simple form of store->load forwarding.
9695 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9696 if (SI->getOperand(1) == LI.getOperand(0))
9697 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9698 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9699 if (LIB->getOperand(0) == LI.getOperand(0))
9700 return ReplaceInstUsesWith(LI, LIB);
9703 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9704 const Value *GEPI0 = GEPI->getOperand(0);
9705 // TODO: Consider a target hook for valid address spaces for this xform.
9706 if (isa<ConstantPointerNull>(GEPI0) &&
9707 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
9708 // Insert a new store to null instruction before the load to indicate
9709 // that this code is not reachable. We do this instead of inserting
9710 // an unreachable instruction directly because we cannot modify the
9712 new StoreInst(UndefValue::get(LI.getType()),
9713 Constant::getNullValue(Op->getType()), &LI);
9714 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9718 if (Constant *C = dyn_cast<Constant>(Op)) {
9719 // load null/undef -> undef
9720 // TODO: Consider a target hook for valid address spaces for this xform.
9721 if (isa<UndefValue>(C) || (C->isNullValue() &&
9722 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
9723 // Insert a new store to null instruction before the load to indicate that
9724 // this code is not reachable. We do this instead of inserting an
9725 // unreachable instruction directly because we cannot modify the CFG.
9726 new StoreInst(UndefValue::get(LI.getType()),
9727 Constant::getNullValue(Op->getType()), &LI);
9728 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9731 // Instcombine load (constant global) into the value loaded.
9732 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9733 if (GV->isConstant() && !GV->isDeclaration())
9734 return ReplaceInstUsesWith(LI, GV->getInitializer());
9736 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9737 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
9738 if (CE->getOpcode() == Instruction::GetElementPtr) {
9739 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9740 if (GV->isConstant() && !GV->isDeclaration())
9742 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9743 return ReplaceInstUsesWith(LI, V);
9744 if (CE->getOperand(0)->isNullValue()) {
9745 // Insert a new store to null instruction before the load to indicate
9746 // that this code is not reachable. We do this instead of inserting
9747 // an unreachable instruction directly because we cannot modify the
9749 new StoreInst(UndefValue::get(LI.getType()),
9750 Constant::getNullValue(Op->getType()), &LI);
9751 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9754 } else if (CE->isCast()) {
9755 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9761 // If this load comes from anywhere in a constant global, and if the global
9762 // is all undef or zero, we know what it loads.
9763 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9764 if (GV->isConstant() && GV->hasInitializer()) {
9765 if (GV->getInitializer()->isNullValue())
9766 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9767 else if (isa<UndefValue>(GV->getInitializer()))
9768 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9772 if (Op->hasOneUse()) {
9773 // Change select and PHI nodes to select values instead of addresses: this
9774 // helps alias analysis out a lot, allows many others simplifications, and
9775 // exposes redundancy in the code.
9777 // Note that we cannot do the transformation unless we know that the
9778 // introduced loads cannot trap! Something like this is valid as long as
9779 // the condition is always false: load (select bool %C, int* null, int* %G),
9780 // but it would not be valid if we transformed it to load from null
9783 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9784 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9785 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9786 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9787 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9788 SI->getOperand(1)->getName()+".val"), LI);
9789 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9790 SI->getOperand(2)->getName()+".val"), LI);
9791 return new SelectInst(SI->getCondition(), V1, V2);
9794 // load (select (cond, null, P)) -> load P
9795 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9796 if (C->isNullValue()) {
9797 LI.setOperand(0, SI->getOperand(2));
9801 // load (select (cond, P, null)) -> load P
9802 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9803 if (C->isNullValue()) {
9804 LI.setOperand(0, SI->getOperand(1));
9812 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9814 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9815 User *CI = cast<User>(SI.getOperand(1));
9816 Value *CastOp = CI->getOperand(0);
9818 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9819 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9820 const Type *SrcPTy = SrcTy->getElementType();
9822 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9823 // If the source is an array, the code below will not succeed. Check to
9824 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9826 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9827 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9828 if (ASrcTy->getNumElements() != 0) {
9830 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9831 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9832 SrcTy = cast<PointerType>(CastOp->getType());
9833 SrcPTy = SrcTy->getElementType();
9836 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9837 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9838 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9840 // Okay, we are casting from one integer or pointer type to another of
9841 // the same size. Instead of casting the pointer before
9842 // the store, cast the value to be stored.
9844 Value *SIOp0 = SI.getOperand(0);
9845 Instruction::CastOps opcode = Instruction::BitCast;
9846 const Type* CastSrcTy = SIOp0->getType();
9847 const Type* CastDstTy = SrcPTy;
9848 if (isa<PointerType>(CastDstTy)) {
9849 if (CastSrcTy->isInteger())
9850 opcode = Instruction::IntToPtr;
9851 } else if (isa<IntegerType>(CastDstTy)) {
9852 if (isa<PointerType>(SIOp0->getType()))
9853 opcode = Instruction::PtrToInt;
9855 if (Constant *C = dyn_cast<Constant>(SIOp0))
9856 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9858 NewCast = IC.InsertNewInstBefore(
9859 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9861 return new StoreInst(NewCast, CastOp);
9868 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9869 Value *Val = SI.getOperand(0);
9870 Value *Ptr = SI.getOperand(1);
9872 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9873 EraseInstFromFunction(SI);
9878 // If the RHS is an alloca with a single use, zapify the store, making the
9880 if (Ptr->hasOneUse()) {
9881 if (isa<AllocaInst>(Ptr)) {
9882 EraseInstFromFunction(SI);
9887 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9888 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9889 GEP->getOperand(0)->hasOneUse()) {
9890 EraseInstFromFunction(SI);
9896 // Attempt to improve the alignment.
9897 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9898 if (KnownAlign > SI.getAlignment())
9899 SI.setAlignment(KnownAlign);
9901 // Do really simple DSE, to catch cases where there are several consequtive
9902 // stores to the same location, separated by a few arithmetic operations. This
9903 // situation often occurs with bitfield accesses.
9904 BasicBlock::iterator BBI = &SI;
9905 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9909 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9910 // Prev store isn't volatile, and stores to the same location?
9911 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9914 EraseInstFromFunction(*PrevSI);
9920 // If this is a load, we have to stop. However, if the loaded value is from
9921 // the pointer we're loading and is producing the pointer we're storing,
9922 // then *this* store is dead (X = load P; store X -> P).
9923 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9924 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9925 EraseInstFromFunction(SI);
9929 // Otherwise, this is a load from some other location. Stores before it
9934 // Don't skip over loads or things that can modify memory.
9935 if (BBI->mayWriteToMemory())
9940 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9942 // store X, null -> turns into 'unreachable' in SimplifyCFG
9943 if (isa<ConstantPointerNull>(Ptr)) {
9944 if (!isa<UndefValue>(Val)) {
9945 SI.setOperand(0, UndefValue::get(Val->getType()));
9946 if (Instruction *U = dyn_cast<Instruction>(Val))
9947 AddToWorkList(U); // Dropped a use.
9950 return 0; // Do not modify these!
9953 // store undef, Ptr -> noop
9954 if (isa<UndefValue>(Val)) {
9955 EraseInstFromFunction(SI);
9960 // If the pointer destination is a cast, see if we can fold the cast into the
9962 if (isa<CastInst>(Ptr))
9963 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9965 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9967 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9971 // If this store is the last instruction in the basic block, and if the block
9972 // ends with an unconditional branch, try to move it to the successor block.
9974 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9975 if (BI->isUnconditional())
9976 if (SimplifyStoreAtEndOfBlock(SI))
9977 return 0; // xform done!
9982 /// SimplifyStoreAtEndOfBlock - Turn things like:
9983 /// if () { *P = v1; } else { *P = v2 }
9984 /// into a phi node with a store in the successor.
9986 /// Simplify things like:
9987 /// *P = v1; if () { *P = v2; }
9988 /// into a phi node with a store in the successor.
9990 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9991 BasicBlock *StoreBB = SI.getParent();
9993 // Check to see if the successor block has exactly two incoming edges. If
9994 // so, see if the other predecessor contains a store to the same location.
9995 // if so, insert a PHI node (if needed) and move the stores down.
9996 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9998 // Determine whether Dest has exactly two predecessors and, if so, compute
9999 // the other predecessor.
10000 pred_iterator PI = pred_begin(DestBB);
10001 BasicBlock *OtherBB = 0;
10002 if (*PI != StoreBB)
10005 if (PI == pred_end(DestBB))
10008 if (*PI != StoreBB) {
10013 if (++PI != pred_end(DestBB))
10017 // Verify that the other block ends in a branch and is not otherwise empty.
10018 BasicBlock::iterator BBI = OtherBB->getTerminator();
10019 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10020 if (!OtherBr || BBI == OtherBB->begin())
10023 // If the other block ends in an unconditional branch, check for the 'if then
10024 // else' case. there is an instruction before the branch.
10025 StoreInst *OtherStore = 0;
10026 if (OtherBr->isUnconditional()) {
10027 // If this isn't a store, or isn't a store to the same location, bail out.
10029 OtherStore = dyn_cast<StoreInst>(BBI);
10030 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10033 // Otherwise, the other block ended with a conditional branch. If one of the
10034 // destinations is StoreBB, then we have the if/then case.
10035 if (OtherBr->getSuccessor(0) != StoreBB &&
10036 OtherBr->getSuccessor(1) != StoreBB)
10039 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10040 // if/then triangle. See if there is a store to the same ptr as SI that
10041 // lives in OtherBB.
10043 // Check to see if we find the matching store.
10044 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10045 if (OtherStore->getOperand(1) != SI.getOperand(1))
10049 // If we find something that may be using the stored value, or if we run
10050 // out of instructions, we can't do the xform.
10051 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
10052 BBI == OtherBB->begin())
10056 // In order to eliminate the store in OtherBr, we have to
10057 // make sure nothing reads the stored value in StoreBB.
10058 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10059 // FIXME: This should really be AA driven.
10060 if (isa<LoadInst>(I) || I->mayWriteToMemory())
10065 // Insert a PHI node now if we need it.
10066 Value *MergedVal = OtherStore->getOperand(0);
10067 if (MergedVal != SI.getOperand(0)) {
10068 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
10069 PN->reserveOperandSpace(2);
10070 PN->addIncoming(SI.getOperand(0), SI.getParent());
10071 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10072 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10075 // Advance to a place where it is safe to insert the new store and
10077 BBI = DestBB->begin();
10078 while (isa<PHINode>(BBI)) ++BBI;
10079 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10080 OtherStore->isVolatile()), *BBI);
10082 // Nuke the old stores.
10083 EraseInstFromFunction(SI);
10084 EraseInstFromFunction(*OtherStore);
10090 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10091 // Change br (not X), label True, label False to: br X, label False, True
10093 BasicBlock *TrueDest;
10094 BasicBlock *FalseDest;
10095 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10096 !isa<Constant>(X)) {
10097 // Swap Destinations and condition...
10098 BI.setCondition(X);
10099 BI.setSuccessor(0, FalseDest);
10100 BI.setSuccessor(1, TrueDest);
10104 // Cannonicalize fcmp_one -> fcmp_oeq
10105 FCmpInst::Predicate FPred; Value *Y;
10106 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10107 TrueDest, FalseDest)))
10108 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10109 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10110 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10111 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10112 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10113 NewSCC->takeName(I);
10114 // Swap Destinations and condition...
10115 BI.setCondition(NewSCC);
10116 BI.setSuccessor(0, FalseDest);
10117 BI.setSuccessor(1, TrueDest);
10118 RemoveFromWorkList(I);
10119 I->eraseFromParent();
10120 AddToWorkList(NewSCC);
10124 // Cannonicalize icmp_ne -> icmp_eq
10125 ICmpInst::Predicate IPred;
10126 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10127 TrueDest, FalseDest)))
10128 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10129 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10130 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10131 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10132 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10133 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10134 NewSCC->takeName(I);
10135 // Swap Destinations and condition...
10136 BI.setCondition(NewSCC);
10137 BI.setSuccessor(0, FalseDest);
10138 BI.setSuccessor(1, TrueDest);
10139 RemoveFromWorkList(I);
10140 I->eraseFromParent();;
10141 AddToWorkList(NewSCC);
10148 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10149 Value *Cond = SI.getCondition();
10150 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10151 if (I->getOpcode() == Instruction::Add)
10152 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10153 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10154 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10155 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10157 SI.setOperand(0, I->getOperand(0));
10165 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10166 /// is to leave as a vector operation.
10167 static bool CheapToScalarize(Value *V, bool isConstant) {
10168 if (isa<ConstantAggregateZero>(V))
10170 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10171 if (isConstant) return true;
10172 // If all elts are the same, we can extract.
10173 Constant *Op0 = C->getOperand(0);
10174 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10175 if (C->getOperand(i) != Op0)
10179 Instruction *I = dyn_cast<Instruction>(V);
10180 if (!I) return false;
10182 // Insert element gets simplified to the inserted element or is deleted if
10183 // this is constant idx extract element and its a constant idx insertelt.
10184 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10185 isa<ConstantInt>(I->getOperand(2)))
10187 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10189 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10190 if (BO->hasOneUse() &&
10191 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10192 CheapToScalarize(BO->getOperand(1), isConstant)))
10194 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10195 if (CI->hasOneUse() &&
10196 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10197 CheapToScalarize(CI->getOperand(1), isConstant)))
10203 /// Read and decode a shufflevector mask.
10205 /// It turns undef elements into values that are larger than the number of
10206 /// elements in the input.
10207 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10208 unsigned NElts = SVI->getType()->getNumElements();
10209 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10210 return std::vector<unsigned>(NElts, 0);
10211 if (isa<UndefValue>(SVI->getOperand(2)))
10212 return std::vector<unsigned>(NElts, 2*NElts);
10214 std::vector<unsigned> Result;
10215 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10216 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
10217 if (isa<UndefValue>(CP->getOperand(i)))
10218 Result.push_back(NElts*2); // undef -> 8
10220 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
10224 /// FindScalarElement - Given a vector and an element number, see if the scalar
10225 /// value is already around as a register, for example if it were inserted then
10226 /// extracted from the vector.
10227 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10228 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10229 const VectorType *PTy = cast<VectorType>(V->getType());
10230 unsigned Width = PTy->getNumElements();
10231 if (EltNo >= Width) // Out of range access.
10232 return UndefValue::get(PTy->getElementType());
10234 if (isa<UndefValue>(V))
10235 return UndefValue::get(PTy->getElementType());
10236 else if (isa<ConstantAggregateZero>(V))
10237 return Constant::getNullValue(PTy->getElementType());
10238 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10239 return CP->getOperand(EltNo);
10240 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10241 // If this is an insert to a variable element, we don't know what it is.
10242 if (!isa<ConstantInt>(III->getOperand(2)))
10244 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10246 // If this is an insert to the element we are looking for, return the
10248 if (EltNo == IIElt)
10249 return III->getOperand(1);
10251 // Otherwise, the insertelement doesn't modify the value, recurse on its
10253 return FindScalarElement(III->getOperand(0), EltNo);
10254 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10255 unsigned InEl = getShuffleMask(SVI)[EltNo];
10257 return FindScalarElement(SVI->getOperand(0), InEl);
10258 else if (InEl < Width*2)
10259 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10261 return UndefValue::get(PTy->getElementType());
10264 // Otherwise, we don't know.
10268 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10270 // If vector val is undef, replace extract with scalar undef.
10271 if (isa<UndefValue>(EI.getOperand(0)))
10272 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10274 // If vector val is constant 0, replace extract with scalar 0.
10275 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10276 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10278 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10279 // If vector val is constant with uniform operands, replace EI
10280 // with that operand
10281 Constant *op0 = C->getOperand(0);
10282 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10283 if (C->getOperand(i) != op0) {
10288 return ReplaceInstUsesWith(EI, op0);
10291 // If extracting a specified index from the vector, see if we can recursively
10292 // find a previously computed scalar that was inserted into the vector.
10293 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10294 unsigned IndexVal = IdxC->getZExtValue();
10295 unsigned VectorWidth =
10296 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10298 // If this is extracting an invalid index, turn this into undef, to avoid
10299 // crashing the code below.
10300 if (IndexVal >= VectorWidth)
10301 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10303 // This instruction only demands the single element from the input vector.
10304 // If the input vector has a single use, simplify it based on this use
10306 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10307 uint64_t UndefElts;
10308 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10311 EI.setOperand(0, V);
10316 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10317 return ReplaceInstUsesWith(EI, Elt);
10319 // If the this extractelement is directly using a bitcast from a vector of
10320 // the same number of elements, see if we can find the source element from
10321 // it. In this case, we will end up needing to bitcast the scalars.
10322 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10323 if (const VectorType *VT =
10324 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10325 if (VT->getNumElements() == VectorWidth)
10326 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10327 return new BitCastInst(Elt, EI.getType());
10331 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10332 if (I->hasOneUse()) {
10333 // Push extractelement into predecessor operation if legal and
10334 // profitable to do so
10335 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10336 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10337 if (CheapToScalarize(BO, isConstantElt)) {
10338 ExtractElementInst *newEI0 =
10339 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10340 EI.getName()+".lhs");
10341 ExtractElementInst *newEI1 =
10342 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10343 EI.getName()+".rhs");
10344 InsertNewInstBefore(newEI0, EI);
10345 InsertNewInstBefore(newEI1, EI);
10346 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10348 } else if (isa<LoadInst>(I)) {
10350 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10351 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10352 PointerType::get(EI.getType(), AS),EI);
10353 GetElementPtrInst *GEP =
10354 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
10355 InsertNewInstBefore(GEP, EI);
10356 return new LoadInst(GEP);
10359 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10360 // Extracting the inserted element?
10361 if (IE->getOperand(2) == EI.getOperand(1))
10362 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10363 // If the inserted and extracted elements are constants, they must not
10364 // be the same value, extract from the pre-inserted value instead.
10365 if (isa<Constant>(IE->getOperand(2)) &&
10366 isa<Constant>(EI.getOperand(1))) {
10367 AddUsesToWorkList(EI);
10368 EI.setOperand(0, IE->getOperand(0));
10371 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10372 // If this is extracting an element from a shufflevector, figure out where
10373 // it came from and extract from the appropriate input element instead.
10374 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10375 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10377 if (SrcIdx < SVI->getType()->getNumElements())
10378 Src = SVI->getOperand(0);
10379 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10380 SrcIdx -= SVI->getType()->getNumElements();
10381 Src = SVI->getOperand(1);
10383 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10385 return new ExtractElementInst(Src, SrcIdx);
10392 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10393 /// elements from either LHS or RHS, return the shuffle mask and true.
10394 /// Otherwise, return false.
10395 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10396 std::vector<Constant*> &Mask) {
10397 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10398 "Invalid CollectSingleShuffleElements");
10399 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10401 if (isa<UndefValue>(V)) {
10402 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10404 } else if (V == LHS) {
10405 for (unsigned i = 0; i != NumElts; ++i)
10406 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10408 } else if (V == RHS) {
10409 for (unsigned i = 0; i != NumElts; ++i)
10410 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10412 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10413 // If this is an insert of an extract from some other vector, include it.
10414 Value *VecOp = IEI->getOperand(0);
10415 Value *ScalarOp = IEI->getOperand(1);
10416 Value *IdxOp = IEI->getOperand(2);
10418 if (!isa<ConstantInt>(IdxOp))
10420 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10422 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10423 // Okay, we can handle this if the vector we are insertinting into is
10424 // transitively ok.
10425 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10426 // If so, update the mask to reflect the inserted undef.
10427 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10430 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10431 if (isa<ConstantInt>(EI->getOperand(1)) &&
10432 EI->getOperand(0)->getType() == V->getType()) {
10433 unsigned ExtractedIdx =
10434 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10436 // This must be extracting from either LHS or RHS.
10437 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10438 // Okay, we can handle this if the vector we are insertinting into is
10439 // transitively ok.
10440 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10441 // If so, update the mask to reflect the inserted value.
10442 if (EI->getOperand(0) == LHS) {
10443 Mask[InsertedIdx & (NumElts-1)] =
10444 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10446 assert(EI->getOperand(0) == RHS);
10447 Mask[InsertedIdx & (NumElts-1)] =
10448 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10457 // TODO: Handle shufflevector here!
10462 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10463 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10464 /// that computes V and the LHS value of the shuffle.
10465 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10467 assert(isa<VectorType>(V->getType()) &&
10468 (RHS == 0 || V->getType() == RHS->getType()) &&
10469 "Invalid shuffle!");
10470 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10472 if (isa<UndefValue>(V)) {
10473 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10475 } else if (isa<ConstantAggregateZero>(V)) {
10476 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10478 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10479 // If this is an insert of an extract from some other vector, include it.
10480 Value *VecOp = IEI->getOperand(0);
10481 Value *ScalarOp = IEI->getOperand(1);
10482 Value *IdxOp = IEI->getOperand(2);
10484 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10485 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10486 EI->getOperand(0)->getType() == V->getType()) {
10487 unsigned ExtractedIdx =
10488 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10489 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10491 // Either the extracted from or inserted into vector must be RHSVec,
10492 // otherwise we'd end up with a shuffle of three inputs.
10493 if (EI->getOperand(0) == RHS || RHS == 0) {
10494 RHS = EI->getOperand(0);
10495 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10496 Mask[InsertedIdx & (NumElts-1)] =
10497 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10501 if (VecOp == RHS) {
10502 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10503 // Everything but the extracted element is replaced with the RHS.
10504 for (unsigned i = 0; i != NumElts; ++i) {
10505 if (i != InsertedIdx)
10506 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10511 // If this insertelement is a chain that comes from exactly these two
10512 // vectors, return the vector and the effective shuffle.
10513 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10514 return EI->getOperand(0);
10519 // TODO: Handle shufflevector here!
10521 // Otherwise, can't do anything fancy. Return an identity vector.
10522 for (unsigned i = 0; i != NumElts; ++i)
10523 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10527 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10528 Value *VecOp = IE.getOperand(0);
10529 Value *ScalarOp = IE.getOperand(1);
10530 Value *IdxOp = IE.getOperand(2);
10532 // Inserting an undef or into an undefined place, remove this.
10533 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10534 ReplaceInstUsesWith(IE, VecOp);
10536 // If the inserted element was extracted from some other vector, and if the
10537 // indexes are constant, try to turn this into a shufflevector operation.
10538 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10539 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10540 EI->getOperand(0)->getType() == IE.getType()) {
10541 unsigned NumVectorElts = IE.getType()->getNumElements();
10542 unsigned ExtractedIdx =
10543 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10544 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10546 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10547 return ReplaceInstUsesWith(IE, VecOp);
10549 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10550 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10552 // If we are extracting a value from a vector, then inserting it right
10553 // back into the same place, just use the input vector.
10554 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10555 return ReplaceInstUsesWith(IE, VecOp);
10557 // We could theoretically do this for ANY input. However, doing so could
10558 // turn chains of insertelement instructions into a chain of shufflevector
10559 // instructions, and right now we do not merge shufflevectors. As such,
10560 // only do this in a situation where it is clear that there is benefit.
10561 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10562 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10563 // the values of VecOp, except then one read from EIOp0.
10564 // Build a new shuffle mask.
10565 std::vector<Constant*> Mask;
10566 if (isa<UndefValue>(VecOp))
10567 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10569 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10570 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10573 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10574 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10575 ConstantVector::get(Mask));
10578 // If this insertelement isn't used by some other insertelement, turn it
10579 // (and any insertelements it points to), into one big shuffle.
10580 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10581 std::vector<Constant*> Mask;
10583 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10584 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10585 // We now have a shuffle of LHS, RHS, Mask.
10586 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10595 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10596 Value *LHS = SVI.getOperand(0);
10597 Value *RHS = SVI.getOperand(1);
10598 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10600 bool MadeChange = false;
10602 // Undefined shuffle mask -> undefined value.
10603 if (isa<UndefValue>(SVI.getOperand(2)))
10604 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10606 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10607 // the undef, change them to undefs.
10608 if (isa<UndefValue>(SVI.getOperand(1))) {
10609 // Scan to see if there are any references to the RHS. If so, replace them
10610 // with undef element refs and set MadeChange to true.
10611 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10612 if (Mask[i] >= e && Mask[i] != 2*e) {
10619 // Remap any references to RHS to use LHS.
10620 std::vector<Constant*> Elts;
10621 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10622 if (Mask[i] == 2*e)
10623 Elts.push_back(UndefValue::get(Type::Int32Ty));
10625 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10627 SVI.setOperand(2, ConstantVector::get(Elts));
10631 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10632 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10633 if (LHS == RHS || isa<UndefValue>(LHS)) {
10634 if (isa<UndefValue>(LHS) && LHS == RHS) {
10635 // shuffle(undef,undef,mask) -> undef.
10636 return ReplaceInstUsesWith(SVI, LHS);
10639 // Remap any references to RHS to use LHS.
10640 std::vector<Constant*> Elts;
10641 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10642 if (Mask[i] >= 2*e)
10643 Elts.push_back(UndefValue::get(Type::Int32Ty));
10645 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10646 (Mask[i] < e && isa<UndefValue>(LHS)))
10647 Mask[i] = 2*e; // Turn into undef.
10649 Mask[i] &= (e-1); // Force to LHS.
10650 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10653 SVI.setOperand(0, SVI.getOperand(1));
10654 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10655 SVI.setOperand(2, ConstantVector::get(Elts));
10656 LHS = SVI.getOperand(0);
10657 RHS = SVI.getOperand(1);
10661 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10662 bool isLHSID = true, isRHSID = true;
10664 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10665 if (Mask[i] >= e*2) continue; // Ignore undef values.
10666 // Is this an identity shuffle of the LHS value?
10667 isLHSID &= (Mask[i] == i);
10669 // Is this an identity shuffle of the RHS value?
10670 isRHSID &= (Mask[i]-e == i);
10673 // Eliminate identity shuffles.
10674 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10675 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10677 // If the LHS is a shufflevector itself, see if we can combine it with this
10678 // one without producing an unusual shuffle. Here we are really conservative:
10679 // we are absolutely afraid of producing a shuffle mask not in the input
10680 // program, because the code gen may not be smart enough to turn a merged
10681 // shuffle into two specific shuffles: it may produce worse code. As such,
10682 // we only merge two shuffles if the result is one of the two input shuffle
10683 // masks. In this case, merging the shuffles just removes one instruction,
10684 // which we know is safe. This is good for things like turning:
10685 // (splat(splat)) -> splat.
10686 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10687 if (isa<UndefValue>(RHS)) {
10688 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10690 std::vector<unsigned> NewMask;
10691 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10692 if (Mask[i] >= 2*e)
10693 NewMask.push_back(2*e);
10695 NewMask.push_back(LHSMask[Mask[i]]);
10697 // If the result mask is equal to the src shuffle or this shuffle mask, do
10698 // the replacement.
10699 if (NewMask == LHSMask || NewMask == Mask) {
10700 std::vector<Constant*> Elts;
10701 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10702 if (NewMask[i] >= e*2) {
10703 Elts.push_back(UndefValue::get(Type::Int32Ty));
10705 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10708 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10709 LHSSVI->getOperand(1),
10710 ConstantVector::get(Elts));
10715 return MadeChange ? &SVI : 0;
10721 /// TryToSinkInstruction - Try to move the specified instruction from its
10722 /// current block into the beginning of DestBlock, which can only happen if it's
10723 /// safe to move the instruction past all of the instructions between it and the
10724 /// end of its block.
10725 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10726 assert(I->hasOneUse() && "Invariants didn't hold!");
10728 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10729 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10731 // Do not sink alloca instructions out of the entry block.
10732 if (isa<AllocaInst>(I) && I->getParent() ==
10733 &DestBlock->getParent()->getEntryBlock())
10736 // We can only sink load instructions if there is nothing between the load and
10737 // the end of block that could change the value.
10738 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10739 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10741 if (Scan->mayWriteToMemory())
10745 BasicBlock::iterator InsertPos = DestBlock->begin();
10746 while (isa<PHINode>(InsertPos)) ++InsertPos;
10748 I->moveBefore(InsertPos);
10754 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10755 /// all reachable code to the worklist.
10757 /// This has a couple of tricks to make the code faster and more powerful. In
10758 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10759 /// them to the worklist (this significantly speeds up instcombine on code where
10760 /// many instructions are dead or constant). Additionally, if we find a branch
10761 /// whose condition is a known constant, we only visit the reachable successors.
10763 static void AddReachableCodeToWorklist(BasicBlock *BB,
10764 SmallPtrSet<BasicBlock*, 64> &Visited,
10766 const TargetData *TD) {
10767 std::vector<BasicBlock*> Worklist;
10768 Worklist.push_back(BB);
10770 while (!Worklist.empty()) {
10771 BB = Worklist.back();
10772 Worklist.pop_back();
10774 // We have now visited this block! If we've already been here, ignore it.
10775 if (!Visited.insert(BB)) continue;
10777 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10778 Instruction *Inst = BBI++;
10780 // DCE instruction if trivially dead.
10781 if (isInstructionTriviallyDead(Inst)) {
10783 DOUT << "IC: DCE: " << *Inst;
10784 Inst->eraseFromParent();
10788 // ConstantProp instruction if trivially constant.
10789 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10790 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10791 Inst->replaceAllUsesWith(C);
10793 Inst->eraseFromParent();
10797 IC.AddToWorkList(Inst);
10800 // Recursively visit successors. If this is a branch or switch on a
10801 // constant, only visit the reachable successor.
10802 if (BB->getUnwindDest())
10803 Worklist.push_back(BB->getUnwindDest());
10804 TerminatorInst *TI = BB->getTerminator();
10805 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10806 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10807 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10808 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
10809 if (ReachableBB != BB->getUnwindDest())
10810 Worklist.push_back(ReachableBB);
10813 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10814 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10815 // See if this is an explicit destination.
10816 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10817 if (SI->getCaseValue(i) == Cond) {
10818 BasicBlock *ReachableBB = SI->getSuccessor(i);
10819 if (ReachableBB != BB->getUnwindDest())
10820 Worklist.push_back(ReachableBB);
10824 // Otherwise it is the default destination.
10825 Worklist.push_back(SI->getSuccessor(0));
10830 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10831 Worklist.push_back(TI->getSuccessor(i));
10835 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10836 bool Changed = false;
10837 TD = &getAnalysis<TargetData>();
10839 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10840 << F.getNameStr() << "\n");
10843 // Do a depth-first traversal of the function, populate the worklist with
10844 // the reachable instructions. Ignore blocks that are not reachable. Keep
10845 // track of which blocks we visit.
10846 SmallPtrSet<BasicBlock*, 64> Visited;
10847 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10849 // Do a quick scan over the function. If we find any blocks that are
10850 // unreachable, remove any instructions inside of them. This prevents
10851 // the instcombine code from having to deal with some bad special cases.
10852 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10853 if (!Visited.count(BB)) {
10854 Instruction *Term = BB->getTerminator();
10855 while (Term != BB->begin()) { // Remove instrs bottom-up
10856 BasicBlock::iterator I = Term; --I;
10858 DOUT << "IC: DCE: " << *I;
10861 if (!I->use_empty())
10862 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10863 I->eraseFromParent();
10868 while (!Worklist.empty()) {
10869 Instruction *I = RemoveOneFromWorkList();
10870 if (I == 0) continue; // skip null values.
10872 // Check to see if we can DCE the instruction.
10873 if (isInstructionTriviallyDead(I)) {
10874 // Add operands to the worklist.
10875 if (I->getNumOperands() < 4)
10876 AddUsesToWorkList(*I);
10879 DOUT << "IC: DCE: " << *I;
10881 I->eraseFromParent();
10882 RemoveFromWorkList(I);
10886 // Instruction isn't dead, see if we can constant propagate it.
10887 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10888 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10890 // Add operands to the worklist.
10891 AddUsesToWorkList(*I);
10892 ReplaceInstUsesWith(*I, C);
10895 I->eraseFromParent();
10896 RemoveFromWorkList(I);
10900 // See if we can trivially sink this instruction to a successor basic block.
10901 if (I->hasOneUse()) {
10902 BasicBlock *BB = I->getParent();
10903 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10904 if (UserParent != BB) {
10905 bool UserIsSuccessor = false;
10906 // See if the user is one of our successors.
10907 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10908 if (*SI == UserParent) {
10909 UserIsSuccessor = true;
10913 // If the user is one of our immediate successors, and if that successor
10914 // only has us as a predecessors (we'd have to split the critical edge
10915 // otherwise), we can keep going.
10916 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10917 next(pred_begin(UserParent)) == pred_end(UserParent))
10918 // Okay, the CFG is simple enough, try to sink this instruction.
10919 Changed |= TryToSinkInstruction(I, UserParent);
10923 // Now that we have an instruction, try combining it to simplify it...
10927 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10928 if (Instruction *Result = visit(*I)) {
10930 // Should we replace the old instruction with a new one?
10932 DOUT << "IC: Old = " << *I
10933 << " New = " << *Result;
10935 // Everything uses the new instruction now.
10936 I->replaceAllUsesWith(Result);
10938 // Push the new instruction and any users onto the worklist.
10939 AddToWorkList(Result);
10940 AddUsersToWorkList(*Result);
10942 // Move the name to the new instruction first.
10943 Result->takeName(I);
10945 // Insert the new instruction into the basic block...
10946 BasicBlock *InstParent = I->getParent();
10947 BasicBlock::iterator InsertPos = I;
10949 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10950 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10953 InstParent->getInstList().insert(InsertPos, Result);
10955 // Make sure that we reprocess all operands now that we reduced their
10957 AddUsesToWorkList(*I);
10959 // Instructions can end up on the worklist more than once. Make sure
10960 // we do not process an instruction that has been deleted.
10961 RemoveFromWorkList(I);
10963 // Erase the old instruction.
10964 InstParent->getInstList().erase(I);
10967 DOUT << "IC: Mod = " << OrigI
10968 << " New = " << *I;
10971 // If the instruction was modified, it's possible that it is now dead.
10972 // if so, remove it.
10973 if (isInstructionTriviallyDead(I)) {
10974 // Make sure we process all operands now that we are reducing their
10976 AddUsesToWorkList(*I);
10978 // Instructions may end up in the worklist more than once. Erase all
10979 // occurrences of this instruction.
10980 RemoveFromWorkList(I);
10981 I->eraseFromParent();
10984 AddUsersToWorkList(*I);
10991 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10993 // Do an explicit clear, this shrinks the map if needed.
10994 WorklistMap.clear();
10999 bool InstCombiner::runOnFunction(Function &F) {
11000 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11002 bool EverMadeChange = false;
11004 // Iterate while there is work to do.
11005 unsigned Iteration = 0;
11006 while (DoOneIteration(F, Iteration++))
11007 EverMadeChange = true;
11008 return EverMadeChange;
11011 FunctionPass *llvm::createInstructionCombiningPass() {
11012 return new InstCombiner();