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/ParameterAttributes.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;
840 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
841 /// this predicate to simplify operations downstream. Mask is known to be zero
842 /// for bits that V cannot have.
843 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
844 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
845 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
846 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
847 return (KnownZero & Mask) == Mask;
850 /// ShrinkDemandedConstant - Check to see if the specified operand of the
851 /// specified instruction is a constant integer. If so, check to see if there
852 /// are any bits set in the constant that are not demanded. If so, shrink the
853 /// constant and return true.
854 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
856 assert(I && "No instruction?");
857 assert(OpNo < I->getNumOperands() && "Operand index too large");
859 // If the operand is not a constant integer, nothing to do.
860 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
861 if (!OpC) return false;
863 // If there are no bits set that aren't demanded, nothing to do.
864 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
865 if ((~Demanded & OpC->getValue()) == 0)
868 // This instruction is producing bits that are not demanded. Shrink the RHS.
869 Demanded &= OpC->getValue();
870 I->setOperand(OpNo, ConstantInt::get(Demanded));
874 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
875 // set of known zero and one bits, compute the maximum and minimum values that
876 // could have the specified known zero and known one bits, returning them in
878 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
879 const APInt& KnownZero,
880 const APInt& KnownOne,
881 APInt& Min, APInt& Max) {
882 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
883 assert(KnownZero.getBitWidth() == BitWidth &&
884 KnownOne.getBitWidth() == BitWidth &&
885 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
886 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
887 APInt UnknownBits = ~(KnownZero|KnownOne);
889 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
890 // bit if it is unknown.
892 Max = KnownOne|UnknownBits;
894 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
896 Max.clear(BitWidth-1);
900 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
901 // a set of known zero and one bits, compute the maximum and minimum values that
902 // could have the specified known zero and known one bits, returning them in
904 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
905 const APInt &KnownZero,
906 const APInt &KnownOne,
907 APInt &Min, APInt &Max) {
908 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
909 assert(KnownZero.getBitWidth() == BitWidth &&
910 KnownOne.getBitWidth() == BitWidth &&
911 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
912 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
913 APInt UnknownBits = ~(KnownZero|KnownOne);
915 // The minimum value is when the unknown bits are all zeros.
917 // The maximum value is when the unknown bits are all ones.
918 Max = KnownOne|UnknownBits;
921 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
922 /// value based on the demanded bits. When this function is called, it is known
923 /// that only the bits set in DemandedMask of the result of V are ever used
924 /// downstream. Consequently, depending on the mask and V, it may be possible
925 /// to replace V with a constant or one of its operands. In such cases, this
926 /// function does the replacement and returns true. In all other cases, it
927 /// returns false after analyzing the expression and setting KnownOne and known
928 /// to be one in the expression. KnownZero contains all the bits that are known
929 /// to be zero in the expression. These are provided to potentially allow the
930 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
931 /// the expression. KnownOne and KnownZero always follow the invariant that
932 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
933 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
934 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
935 /// and KnownOne must all be the same.
936 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
937 APInt& KnownZero, APInt& KnownOne,
939 assert(V != 0 && "Null pointer of Value???");
940 assert(Depth <= 6 && "Limit Search Depth");
941 uint32_t BitWidth = DemandedMask.getBitWidth();
942 const IntegerType *VTy = cast<IntegerType>(V->getType());
943 assert(VTy->getBitWidth() == BitWidth &&
944 KnownZero.getBitWidth() == BitWidth &&
945 KnownOne.getBitWidth() == BitWidth &&
946 "Value *V, DemandedMask, KnownZero and KnownOne \
947 must have same BitWidth");
948 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
949 // We know all of the bits for a constant!
950 KnownOne = CI->getValue() & DemandedMask;
951 KnownZero = ~KnownOne & DemandedMask;
957 if (!V->hasOneUse()) { // Other users may use these bits.
958 if (Depth != 0) { // Not at the root.
959 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
960 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
963 // If this is the root being simplified, allow it to have multiple uses,
964 // just set the DemandedMask to all bits.
965 DemandedMask = APInt::getAllOnesValue(BitWidth);
966 } else if (DemandedMask == 0) { // Not demanding any bits from V.
967 if (V != UndefValue::get(VTy))
968 return UpdateValueUsesWith(V, UndefValue::get(VTy));
970 } else if (Depth == 6) { // Limit search depth.
974 Instruction *I = dyn_cast<Instruction>(V);
975 if (!I) return false; // Only analyze instructions.
977 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
978 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
979 switch (I->getOpcode()) {
981 case Instruction::And:
982 // If either the LHS or the RHS are Zero, the result is zero.
983 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
984 RHSKnownZero, RHSKnownOne, Depth+1))
986 assert((RHSKnownZero & RHSKnownOne) == 0 &&
987 "Bits known to be one AND zero?");
989 // If something is known zero on the RHS, the bits aren't demanded on the
991 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
992 LHSKnownZero, LHSKnownOne, Depth+1))
994 assert((LHSKnownZero & LHSKnownOne) == 0 &&
995 "Bits known to be one AND zero?");
997 // If all of the demanded bits are known 1 on one side, return the other.
998 // These bits cannot contribute to the result of the 'and'.
999 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1000 (DemandedMask & ~LHSKnownZero))
1001 return UpdateValueUsesWith(I, I->getOperand(0));
1002 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1003 (DemandedMask & ~RHSKnownZero))
1004 return UpdateValueUsesWith(I, I->getOperand(1));
1006 // If all of the demanded bits in the inputs are known zeros, return zero.
1007 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1008 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1010 // If the RHS is a constant, see if we can simplify it.
1011 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1012 return UpdateValueUsesWith(I, I);
1014 // Output known-1 bits are only known if set in both the LHS & RHS.
1015 RHSKnownOne &= LHSKnownOne;
1016 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1017 RHSKnownZero |= LHSKnownZero;
1019 case Instruction::Or:
1020 // If either the LHS or the RHS are One, the result is One.
1021 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1022 RHSKnownZero, RHSKnownOne, Depth+1))
1024 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1025 "Bits known to be one AND zero?");
1026 // If something is known one on the RHS, the bits aren't demanded on the
1028 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1029 LHSKnownZero, LHSKnownOne, Depth+1))
1031 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1032 "Bits known to be one AND zero?");
1034 // If all of the demanded bits are known zero on one side, return the other.
1035 // These bits cannot contribute to the result of the 'or'.
1036 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1037 (DemandedMask & ~LHSKnownOne))
1038 return UpdateValueUsesWith(I, I->getOperand(0));
1039 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1040 (DemandedMask & ~RHSKnownOne))
1041 return UpdateValueUsesWith(I, I->getOperand(1));
1043 // If all of the potentially set bits on one side are known to be set on
1044 // the other side, just use the 'other' side.
1045 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1046 (DemandedMask & (~RHSKnownZero)))
1047 return UpdateValueUsesWith(I, I->getOperand(0));
1048 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1049 (DemandedMask & (~LHSKnownZero)))
1050 return UpdateValueUsesWith(I, I->getOperand(1));
1052 // If the RHS is a constant, see if we can simplify it.
1053 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1054 return UpdateValueUsesWith(I, I);
1056 // Output known-0 bits are only known if clear in both the LHS & RHS.
1057 RHSKnownZero &= LHSKnownZero;
1058 // Output known-1 are known to be set if set in either the LHS | RHS.
1059 RHSKnownOne |= LHSKnownOne;
1061 case Instruction::Xor: {
1062 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1063 RHSKnownZero, RHSKnownOne, Depth+1))
1065 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1066 "Bits known to be one AND zero?");
1067 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1068 LHSKnownZero, LHSKnownOne, Depth+1))
1070 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1071 "Bits known to be one AND zero?");
1073 // If all of the demanded bits are known zero on one side, return the other.
1074 // These bits cannot contribute to the result of the 'xor'.
1075 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1076 return UpdateValueUsesWith(I, I->getOperand(0));
1077 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1078 return UpdateValueUsesWith(I, I->getOperand(1));
1080 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1081 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1082 (RHSKnownOne & LHSKnownOne);
1083 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1084 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1085 (RHSKnownOne & LHSKnownZero);
1087 // If all of the demanded bits are known to be zero on one side or the
1088 // other, turn this into an *inclusive* or.
1089 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1090 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1092 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1094 InsertNewInstBefore(Or, *I);
1095 return UpdateValueUsesWith(I, Or);
1098 // If all of the demanded bits on one side are known, and all of the set
1099 // bits on that side are also known to be set on the other side, turn this
1100 // into an AND, as we know the bits will be cleared.
1101 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1102 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1104 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1105 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1107 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1108 InsertNewInstBefore(And, *I);
1109 return UpdateValueUsesWith(I, And);
1113 // If the RHS is a constant, see if we can simplify it.
1114 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1115 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1116 return UpdateValueUsesWith(I, I);
1118 RHSKnownZero = KnownZeroOut;
1119 RHSKnownOne = KnownOneOut;
1122 case Instruction::Select:
1123 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1124 RHSKnownZero, RHSKnownOne, Depth+1))
1126 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1127 LHSKnownZero, LHSKnownOne, Depth+1))
1129 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1130 "Bits known to be one AND zero?");
1131 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1132 "Bits known to be one AND zero?");
1134 // If the operands are constants, see if we can simplify them.
1135 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1136 return UpdateValueUsesWith(I, I);
1137 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1138 return UpdateValueUsesWith(I, I);
1140 // Only known if known in both the LHS and RHS.
1141 RHSKnownOne &= LHSKnownOne;
1142 RHSKnownZero &= LHSKnownZero;
1144 case Instruction::Trunc: {
1146 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1147 DemandedMask.zext(truncBf);
1148 RHSKnownZero.zext(truncBf);
1149 RHSKnownOne.zext(truncBf);
1150 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1151 RHSKnownZero, RHSKnownOne, Depth+1))
1153 DemandedMask.trunc(BitWidth);
1154 RHSKnownZero.trunc(BitWidth);
1155 RHSKnownOne.trunc(BitWidth);
1156 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1157 "Bits known to be one AND zero?");
1160 case Instruction::BitCast:
1161 if (!I->getOperand(0)->getType()->isInteger())
1164 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1165 RHSKnownZero, RHSKnownOne, Depth+1))
1167 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1168 "Bits known to be one AND zero?");
1170 case Instruction::ZExt: {
1171 // Compute the bits in the result that are not present in the input.
1172 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1173 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1175 DemandedMask.trunc(SrcBitWidth);
1176 RHSKnownZero.trunc(SrcBitWidth);
1177 RHSKnownOne.trunc(SrcBitWidth);
1178 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1179 RHSKnownZero, RHSKnownOne, Depth+1))
1181 DemandedMask.zext(BitWidth);
1182 RHSKnownZero.zext(BitWidth);
1183 RHSKnownOne.zext(BitWidth);
1184 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1185 "Bits known to be one AND zero?");
1186 // The top bits are known to be zero.
1187 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1190 case Instruction::SExt: {
1191 // Compute the bits in the result that are not present in the input.
1192 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1193 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1195 APInt InputDemandedBits = DemandedMask &
1196 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1198 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1199 // If any of the sign extended bits are demanded, we know that the sign
1201 if ((NewBits & DemandedMask) != 0)
1202 InputDemandedBits.set(SrcBitWidth-1);
1204 InputDemandedBits.trunc(SrcBitWidth);
1205 RHSKnownZero.trunc(SrcBitWidth);
1206 RHSKnownOne.trunc(SrcBitWidth);
1207 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1208 RHSKnownZero, RHSKnownOne, Depth+1))
1210 InputDemandedBits.zext(BitWidth);
1211 RHSKnownZero.zext(BitWidth);
1212 RHSKnownOne.zext(BitWidth);
1213 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1214 "Bits known to be one AND zero?");
1216 // If the sign bit of the input is known set or clear, then we know the
1217 // top bits of the result.
1219 // If the input sign bit is known zero, or if the NewBits are not demanded
1220 // convert this into a zero extension.
1221 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1223 // Convert to ZExt cast
1224 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1225 return UpdateValueUsesWith(I, NewCast);
1226 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1227 RHSKnownOne |= NewBits;
1231 case Instruction::Add: {
1232 // Figure out what the input bits are. If the top bits of the and result
1233 // are not demanded, then the add doesn't demand them from its input
1235 uint32_t NLZ = DemandedMask.countLeadingZeros();
1237 // If there is a constant on the RHS, there are a variety of xformations
1239 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1240 // If null, this should be simplified elsewhere. Some of the xforms here
1241 // won't work if the RHS is zero.
1245 // If the top bit of the output is demanded, demand everything from the
1246 // input. Otherwise, we demand all the input bits except NLZ top bits.
1247 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1249 // Find information about known zero/one bits in the input.
1250 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1251 LHSKnownZero, LHSKnownOne, Depth+1))
1254 // If the RHS of the add has bits set that can't affect the input, reduce
1256 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1257 return UpdateValueUsesWith(I, I);
1259 // Avoid excess work.
1260 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1263 // Turn it into OR if input bits are zero.
1264 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1266 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1268 InsertNewInstBefore(Or, *I);
1269 return UpdateValueUsesWith(I, Or);
1272 // We can say something about the output known-zero and known-one bits,
1273 // depending on potential carries from the input constant and the
1274 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1275 // bits set and the RHS constant is 0x01001, then we know we have a known
1276 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1278 // To compute this, we first compute the potential carry bits. These are
1279 // the bits which may be modified. I'm not aware of a better way to do
1281 const APInt& RHSVal = RHS->getValue();
1282 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1284 // Now that we know which bits have carries, compute the known-1/0 sets.
1286 // Bits are known one if they are known zero in one operand and one in the
1287 // other, and there is no input carry.
1288 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1289 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1291 // Bits are known zero if they are known zero in both operands and there
1292 // is no input carry.
1293 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1295 // If the high-bits of this ADD are not demanded, then it does not demand
1296 // the high bits of its LHS or RHS.
1297 if (DemandedMask[BitWidth-1] == 0) {
1298 // Right fill the mask of bits for this ADD to demand the most
1299 // significant bit and all those below it.
1300 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1301 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1302 LHSKnownZero, LHSKnownOne, Depth+1))
1304 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1305 LHSKnownZero, LHSKnownOne, Depth+1))
1311 case Instruction::Sub:
1312 // If the high-bits of this SUB are not demanded, then it does not demand
1313 // the high bits of its LHS or RHS.
1314 if (DemandedMask[BitWidth-1] == 0) {
1315 // Right fill the mask of bits for this SUB to demand the most
1316 // significant bit and all those below it.
1317 uint32_t NLZ = DemandedMask.countLeadingZeros();
1318 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1319 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1320 LHSKnownZero, LHSKnownOne, Depth+1))
1322 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1323 LHSKnownZero, LHSKnownOne, Depth+1))
1327 case Instruction::Shl:
1328 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1329 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1330 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1331 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1332 RHSKnownZero, RHSKnownOne, Depth+1))
1334 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1335 "Bits known to be one AND zero?");
1336 RHSKnownZero <<= ShiftAmt;
1337 RHSKnownOne <<= ShiftAmt;
1338 // low bits known zero.
1340 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1343 case Instruction::LShr:
1344 // For a logical shift right
1345 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1346 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1348 // Unsigned shift right.
1349 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1350 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1351 RHSKnownZero, RHSKnownOne, Depth+1))
1353 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1354 "Bits known to be one AND zero?");
1355 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1356 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1358 // Compute the new bits that are at the top now.
1359 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1360 RHSKnownZero |= HighBits; // high bits known zero.
1364 case Instruction::AShr:
1365 // If this is an arithmetic shift right and only the low-bit is set, we can
1366 // always convert this into a logical shr, even if the shift amount is
1367 // variable. The low bit of the shift cannot be an input sign bit unless
1368 // the shift amount is >= the size of the datatype, which is undefined.
1369 if (DemandedMask == 1) {
1370 // Perform the logical shift right.
1371 Value *NewVal = BinaryOperator::createLShr(
1372 I->getOperand(0), I->getOperand(1), I->getName());
1373 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1374 return UpdateValueUsesWith(I, NewVal);
1377 // If the sign bit is the only bit demanded by this ashr, then there is no
1378 // need to do it, the shift doesn't change the high bit.
1379 if (DemandedMask.isSignBit())
1380 return UpdateValueUsesWith(I, I->getOperand(0));
1382 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1383 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1385 // Signed shift right.
1386 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1387 // If any of the "high bits" are demanded, we should set the sign bit as
1389 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1390 DemandedMaskIn.set(BitWidth-1);
1391 if (SimplifyDemandedBits(I->getOperand(0),
1393 RHSKnownZero, RHSKnownOne, Depth+1))
1395 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1396 "Bits known to be one AND zero?");
1397 // Compute the new bits that are at the top now.
1398 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1399 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1400 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1402 // Handle the sign bits.
1403 APInt SignBit(APInt::getSignBit(BitWidth));
1404 // Adjust to where it is now in the mask.
1405 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1407 // If the input sign bit is known to be zero, or if none of the top bits
1408 // are demanded, turn this into an unsigned shift right.
1409 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1410 (HighBits & ~DemandedMask) == HighBits) {
1411 // Perform the logical shift right.
1412 Value *NewVal = BinaryOperator::createLShr(
1413 I->getOperand(0), SA, I->getName());
1414 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1415 return UpdateValueUsesWith(I, NewVal);
1416 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1417 RHSKnownOne |= HighBits;
1423 // If the client is only demanding bits that we know, return the known
1425 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1426 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1431 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1432 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1433 /// actually used by the caller. This method analyzes which elements of the
1434 /// operand are undef and returns that information in UndefElts.
1436 /// If the information about demanded elements can be used to simplify the
1437 /// operation, the operation is simplified, then the resultant value is
1438 /// returned. This returns null if no change was made.
1439 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1440 uint64_t &UndefElts,
1442 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1443 assert(VWidth <= 64 && "Vector too wide to analyze!");
1444 uint64_t EltMask = ~0ULL >> (64-VWidth);
1445 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1446 "Invalid DemandedElts!");
1448 if (isa<UndefValue>(V)) {
1449 // If the entire vector is undefined, just return this info.
1450 UndefElts = EltMask;
1452 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1453 UndefElts = EltMask;
1454 return UndefValue::get(V->getType());
1458 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1459 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1460 Constant *Undef = UndefValue::get(EltTy);
1462 std::vector<Constant*> Elts;
1463 for (unsigned i = 0; i != VWidth; ++i)
1464 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1465 Elts.push_back(Undef);
1466 UndefElts |= (1ULL << i);
1467 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1468 Elts.push_back(Undef);
1469 UndefElts |= (1ULL << i);
1470 } else { // Otherwise, defined.
1471 Elts.push_back(CP->getOperand(i));
1474 // If we changed the constant, return it.
1475 Constant *NewCP = ConstantVector::get(Elts);
1476 return NewCP != CP ? NewCP : 0;
1477 } else if (isa<ConstantAggregateZero>(V)) {
1478 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1480 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1481 Constant *Zero = Constant::getNullValue(EltTy);
1482 Constant *Undef = UndefValue::get(EltTy);
1483 std::vector<Constant*> Elts;
1484 for (unsigned i = 0; i != VWidth; ++i)
1485 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1486 UndefElts = DemandedElts ^ EltMask;
1487 return ConstantVector::get(Elts);
1490 if (!V->hasOneUse()) { // Other users may use these bits.
1491 if (Depth != 0) { // Not at the root.
1492 // TODO: Just compute the UndefElts information recursively.
1496 } else if (Depth == 10) { // Limit search depth.
1500 Instruction *I = dyn_cast<Instruction>(V);
1501 if (!I) return false; // Only analyze instructions.
1503 bool MadeChange = false;
1504 uint64_t UndefElts2;
1506 switch (I->getOpcode()) {
1509 case Instruction::InsertElement: {
1510 // If this is a variable index, we don't know which element it overwrites.
1511 // demand exactly the same input as we produce.
1512 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1514 // Note that we can't propagate undef elt info, because we don't know
1515 // which elt is getting updated.
1516 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1517 UndefElts2, Depth+1);
1518 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1522 // If this is inserting an element that isn't demanded, remove this
1524 unsigned IdxNo = Idx->getZExtValue();
1525 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1526 return AddSoonDeadInstToWorklist(*I, 0);
1528 // Otherwise, the element inserted overwrites whatever was there, so the
1529 // input demanded set is simpler than the output set.
1530 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1531 DemandedElts & ~(1ULL << IdxNo),
1532 UndefElts, Depth+1);
1533 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1535 // The inserted element is defined.
1536 UndefElts |= 1ULL << IdxNo;
1539 case Instruction::BitCast: {
1540 // Vector->vector casts only.
1541 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1543 unsigned InVWidth = VTy->getNumElements();
1544 uint64_t InputDemandedElts = 0;
1547 if (VWidth == InVWidth) {
1548 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1549 // elements as are demanded of us.
1551 InputDemandedElts = DemandedElts;
1552 } else if (VWidth > InVWidth) {
1556 // If there are more elements in the result than there are in the source,
1557 // then an input element is live if any of the corresponding output
1558 // elements are live.
1559 Ratio = VWidth/InVWidth;
1560 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1561 if (DemandedElts & (1ULL << OutIdx))
1562 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1568 // If there are more elements in the source than there are in the result,
1569 // then an input element is live if the corresponding output element is
1571 Ratio = InVWidth/VWidth;
1572 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1573 if (DemandedElts & (1ULL << InIdx/Ratio))
1574 InputDemandedElts |= 1ULL << InIdx;
1577 // div/rem demand all inputs, because they don't want divide by zero.
1578 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1579 UndefElts2, Depth+1);
1581 I->setOperand(0, TmpV);
1585 UndefElts = UndefElts2;
1586 if (VWidth > InVWidth) {
1587 assert(0 && "Unimp");
1588 // If there are more elements in the result than there are in the source,
1589 // then an output element is undef if the corresponding input element is
1591 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1592 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1593 UndefElts |= 1ULL << OutIdx;
1594 } else if (VWidth < InVWidth) {
1595 assert(0 && "Unimp");
1596 // If there are more elements in the source than there are in the result,
1597 // then a result element is undef if all of the corresponding input
1598 // elements are undef.
1599 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1600 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1601 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1602 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1606 case Instruction::And:
1607 case Instruction::Or:
1608 case Instruction::Xor:
1609 case Instruction::Add:
1610 case Instruction::Sub:
1611 case Instruction::Mul:
1612 // div/rem demand all inputs, because they don't want divide by zero.
1613 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1614 UndefElts, Depth+1);
1615 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1616 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1617 UndefElts2, Depth+1);
1618 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1620 // Output elements are undefined if both are undefined. Consider things
1621 // like undef&0. The result is known zero, not undef.
1622 UndefElts &= UndefElts2;
1625 case Instruction::Call: {
1626 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1628 switch (II->getIntrinsicID()) {
1631 // Binary vector operations that work column-wise. A dest element is a
1632 // function of the corresponding input elements from the two inputs.
1633 case Intrinsic::x86_sse_sub_ss:
1634 case Intrinsic::x86_sse_mul_ss:
1635 case Intrinsic::x86_sse_min_ss:
1636 case Intrinsic::x86_sse_max_ss:
1637 case Intrinsic::x86_sse2_sub_sd:
1638 case Intrinsic::x86_sse2_mul_sd:
1639 case Intrinsic::x86_sse2_min_sd:
1640 case Intrinsic::x86_sse2_max_sd:
1641 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1642 UndefElts, Depth+1);
1643 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1644 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1645 UndefElts2, Depth+1);
1646 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1648 // If only the low elt is demanded and this is a scalarizable intrinsic,
1649 // scalarize it now.
1650 if (DemandedElts == 1) {
1651 switch (II->getIntrinsicID()) {
1653 case Intrinsic::x86_sse_sub_ss:
1654 case Intrinsic::x86_sse_mul_ss:
1655 case Intrinsic::x86_sse2_sub_sd:
1656 case Intrinsic::x86_sse2_mul_sd:
1657 // TODO: Lower MIN/MAX/ABS/etc
1658 Value *LHS = II->getOperand(1);
1659 Value *RHS = II->getOperand(2);
1660 // Extract the element as scalars.
1661 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1662 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1664 switch (II->getIntrinsicID()) {
1665 default: assert(0 && "Case stmts out of sync!");
1666 case Intrinsic::x86_sse_sub_ss:
1667 case Intrinsic::x86_sse2_sub_sd:
1668 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1669 II->getName()), *II);
1671 case Intrinsic::x86_sse_mul_ss:
1672 case Intrinsic::x86_sse2_mul_sd:
1673 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1674 II->getName()), *II);
1679 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1681 InsertNewInstBefore(New, *II);
1682 AddSoonDeadInstToWorklist(*II, 0);
1687 // Output elements are undefined if both are undefined. Consider things
1688 // like undef&0. The result is known zero, not undef.
1689 UndefElts &= UndefElts2;
1695 return MadeChange ? I : 0;
1698 /// @returns true if the specified compare predicate is
1699 /// true when both operands are equal...
1700 /// @brief Determine if the icmp Predicate is true when both operands are equal
1701 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1702 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1703 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1704 pred == ICmpInst::ICMP_SLE;
1707 /// @returns true if the specified compare instruction is
1708 /// true when both operands are equal...
1709 /// @brief Determine if the ICmpInst returns true when both operands are equal
1710 static bool isTrueWhenEqual(ICmpInst &ICI) {
1711 return isTrueWhenEqual(ICI.getPredicate());
1714 /// AssociativeOpt - Perform an optimization on an associative operator. This
1715 /// function is designed to check a chain of associative operators for a
1716 /// potential to apply a certain optimization. Since the optimization may be
1717 /// applicable if the expression was reassociated, this checks the chain, then
1718 /// reassociates the expression as necessary to expose the optimization
1719 /// opportunity. This makes use of a special Functor, which must define
1720 /// 'shouldApply' and 'apply' methods.
1722 template<typename Functor>
1723 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1724 unsigned Opcode = Root.getOpcode();
1725 Value *LHS = Root.getOperand(0);
1727 // Quick check, see if the immediate LHS matches...
1728 if (F.shouldApply(LHS))
1729 return F.apply(Root);
1731 // Otherwise, if the LHS is not of the same opcode as the root, return.
1732 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1733 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1734 // Should we apply this transform to the RHS?
1735 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1737 // If not to the RHS, check to see if we should apply to the LHS...
1738 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1739 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1743 // If the functor wants to apply the optimization to the RHS of LHSI,
1744 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1746 BasicBlock *BB = Root.getParent();
1748 // Now all of the instructions are in the current basic block, go ahead
1749 // and perform the reassociation.
1750 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1752 // First move the selected RHS to the LHS of the root...
1753 Root.setOperand(0, LHSI->getOperand(1));
1755 // Make what used to be the LHS of the root be the user of the root...
1756 Value *ExtraOperand = TmpLHSI->getOperand(1);
1757 if (&Root == TmpLHSI) {
1758 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1761 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1762 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1763 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1764 BasicBlock::iterator ARI = &Root; ++ARI;
1765 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1768 // Now propagate the ExtraOperand down the chain of instructions until we
1770 while (TmpLHSI != LHSI) {
1771 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1772 // Move the instruction to immediately before the chain we are
1773 // constructing to avoid breaking dominance properties.
1774 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1775 BB->getInstList().insert(ARI, NextLHSI);
1778 Value *NextOp = NextLHSI->getOperand(1);
1779 NextLHSI->setOperand(1, ExtraOperand);
1781 ExtraOperand = NextOp;
1784 // Now that the instructions are reassociated, have the functor perform
1785 // the transformation...
1786 return F.apply(Root);
1789 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1795 // AddRHS - Implements: X + X --> X << 1
1798 AddRHS(Value *rhs) : RHS(rhs) {}
1799 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1800 Instruction *apply(BinaryOperator &Add) const {
1801 return BinaryOperator::createShl(Add.getOperand(0),
1802 ConstantInt::get(Add.getType(), 1));
1806 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1808 struct AddMaskingAnd {
1810 AddMaskingAnd(Constant *c) : C2(c) {}
1811 bool shouldApply(Value *LHS) const {
1813 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1814 ConstantExpr::getAnd(C1, C2)->isNullValue();
1816 Instruction *apply(BinaryOperator &Add) const {
1817 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1821 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1823 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1824 if (Constant *SOC = dyn_cast<Constant>(SO))
1825 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1827 return IC->InsertNewInstBefore(CastInst::create(
1828 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1831 // Figure out if the constant is the left or the right argument.
1832 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1833 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1835 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1837 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1838 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1841 Value *Op0 = SO, *Op1 = ConstOperand;
1843 std::swap(Op0, Op1);
1845 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1846 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1847 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1848 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1849 SO->getName()+".cmp");
1851 assert(0 && "Unknown binary instruction type!");
1854 return IC->InsertNewInstBefore(New, I);
1857 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1858 // constant as the other operand, try to fold the binary operator into the
1859 // select arguments. This also works for Cast instructions, which obviously do
1860 // not have a second operand.
1861 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1863 // Don't modify shared select instructions
1864 if (!SI->hasOneUse()) return 0;
1865 Value *TV = SI->getOperand(1);
1866 Value *FV = SI->getOperand(2);
1868 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1869 // Bool selects with constant operands can be folded to logical ops.
1870 if (SI->getType() == Type::Int1Ty) return 0;
1872 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1873 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1875 return new SelectInst(SI->getCondition(), SelectTrueVal,
1882 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1883 /// node as operand #0, see if we can fold the instruction into the PHI (which
1884 /// is only possible if all operands to the PHI are constants).
1885 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1886 PHINode *PN = cast<PHINode>(I.getOperand(0));
1887 unsigned NumPHIValues = PN->getNumIncomingValues();
1888 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1890 // Check to see if all of the operands of the PHI are constants. If there is
1891 // one non-constant value, remember the BB it is. If there is more than one
1892 // or if *it* is a PHI, bail out.
1893 BasicBlock *NonConstBB = 0;
1894 for (unsigned i = 0; i != NumPHIValues; ++i)
1895 if (!isa<Constant>(PN->getIncomingValue(i))) {
1896 if (NonConstBB) return 0; // More than one non-const value.
1897 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1898 NonConstBB = PN->getIncomingBlock(i);
1900 // If the incoming non-constant value is in I's block, we have an infinite
1902 if (NonConstBB == I.getParent())
1906 // If there is exactly one non-constant value, we can insert a copy of the
1907 // operation in that block. However, if this is a critical edge, we would be
1908 // inserting the computation one some other paths (e.g. inside a loop). Only
1909 // do this if the pred block is unconditionally branching into the phi block.
1911 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1912 if (!BI || !BI->isUnconditional()) return 0;
1915 // Okay, we can do the transformation: create the new PHI node.
1916 PHINode *NewPN = new PHINode(I.getType(), "");
1917 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1918 InsertNewInstBefore(NewPN, *PN);
1919 NewPN->takeName(PN);
1921 // Next, add all of the operands to the PHI.
1922 if (I.getNumOperands() == 2) {
1923 Constant *C = cast<Constant>(I.getOperand(1));
1924 for (unsigned i = 0; i != NumPHIValues; ++i) {
1926 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1927 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1928 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1930 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1932 assert(PN->getIncomingBlock(i) == NonConstBB);
1933 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1934 InV = BinaryOperator::create(BO->getOpcode(),
1935 PN->getIncomingValue(i), C, "phitmp",
1936 NonConstBB->getTerminator());
1937 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1938 InV = CmpInst::create(CI->getOpcode(),
1940 PN->getIncomingValue(i), C, "phitmp",
1941 NonConstBB->getTerminator());
1943 assert(0 && "Unknown binop!");
1945 AddToWorkList(cast<Instruction>(InV));
1947 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1950 CastInst *CI = cast<CastInst>(&I);
1951 const Type *RetTy = CI->getType();
1952 for (unsigned i = 0; i != NumPHIValues; ++i) {
1954 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1955 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1957 assert(PN->getIncomingBlock(i) == NonConstBB);
1958 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1959 I.getType(), "phitmp",
1960 NonConstBB->getTerminator());
1961 AddToWorkList(cast<Instruction>(InV));
1963 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1966 return ReplaceInstUsesWith(I, NewPN);
1970 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1971 /// value is never equal to -0.0.
1973 /// Note that this function will need to be revisited when we support nondefault
1976 static bool CannotBeNegativeZero(const Value *V) {
1977 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1978 return !CFP->getValueAPF().isNegZero();
1980 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1981 if (const Instruction *I = dyn_cast<Instruction>(V)) {
1982 if (I->getOpcode() == Instruction::Add &&
1983 isa<ConstantFP>(I->getOperand(1)) &&
1984 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1987 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1988 if (II->getIntrinsicID() == Intrinsic::sqrt)
1989 return CannotBeNegativeZero(II->getOperand(1));
1991 if (const CallInst *CI = dyn_cast<CallInst>(I))
1992 if (const Function *F = CI->getCalledFunction()) {
1993 if (F->isDeclaration()) {
1994 switch (F->getNameLen()) {
1995 case 3: // abs(x) != -0.0
1996 if (!strcmp(F->getNameStart(), "abs")) return true;
1998 case 4: // abs[lf](x) != -0.0
1999 if (!strcmp(F->getNameStart(), "absf")) return true;
2000 if (!strcmp(F->getNameStart(), "absl")) return true;
2011 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2012 bool Changed = SimplifyCommutative(I);
2013 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2015 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2016 // X + undef -> undef
2017 if (isa<UndefValue>(RHS))
2018 return ReplaceInstUsesWith(I, RHS);
2021 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2022 if (RHSC->isNullValue())
2023 return ReplaceInstUsesWith(I, LHS);
2024 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2025 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2026 (I.getType())->getValueAPF()))
2027 return ReplaceInstUsesWith(I, LHS);
2030 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2031 // X + (signbit) --> X ^ signbit
2032 const APInt& Val = CI->getValue();
2033 uint32_t BitWidth = Val.getBitWidth();
2034 if (Val == APInt::getSignBit(BitWidth))
2035 return BinaryOperator::createXor(LHS, RHS);
2037 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2038 // (X & 254)+1 -> (X&254)|1
2039 if (!isa<VectorType>(I.getType())) {
2040 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2041 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2042 KnownZero, KnownOne))
2047 if (isa<PHINode>(LHS))
2048 if (Instruction *NV = FoldOpIntoPhi(I))
2051 ConstantInt *XorRHS = 0;
2053 if (isa<ConstantInt>(RHSC) &&
2054 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2055 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2056 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2058 uint32_t Size = TySizeBits / 2;
2059 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2060 APInt CFF80Val(-C0080Val);
2062 if (TySizeBits > Size) {
2063 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2064 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2065 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2066 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2067 // This is a sign extend if the top bits are known zero.
2068 if (!MaskedValueIsZero(XorLHS,
2069 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2070 Size = 0; // Not a sign ext, but can't be any others either.
2075 C0080Val = APIntOps::lshr(C0080Val, Size);
2076 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2077 } while (Size >= 1);
2079 // FIXME: This shouldn't be necessary. When the backends can handle types
2080 // with funny bit widths then this whole cascade of if statements should
2081 // be removed. It is just here to get the size of the "middle" type back
2082 // up to something that the back ends can handle.
2083 const Type *MiddleType = 0;
2086 case 32: MiddleType = Type::Int32Ty; break;
2087 case 16: MiddleType = Type::Int16Ty; break;
2088 case 8: MiddleType = Type::Int8Ty; break;
2091 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2092 InsertNewInstBefore(NewTrunc, I);
2093 return new SExtInst(NewTrunc, I.getType(), I.getName());
2099 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2100 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2102 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2103 if (RHSI->getOpcode() == Instruction::Sub)
2104 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2105 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2107 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2108 if (LHSI->getOpcode() == Instruction::Sub)
2109 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2110 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2115 // -A + -B --> -(A + B)
2116 if (Value *LHSV = dyn_castNegVal(LHS)) {
2117 if (LHS->getType()->isIntOrIntVector()) {
2118 if (Value *RHSV = dyn_castNegVal(RHS)) {
2119 Instruction *NewAdd = BinaryOperator::createAdd(LHSV, RHSV, "sum");
2120 InsertNewInstBefore(NewAdd, I);
2121 return BinaryOperator::createNeg(NewAdd);
2125 return BinaryOperator::createSub(RHS, LHSV);
2129 if (!isa<Constant>(RHS))
2130 if (Value *V = dyn_castNegVal(RHS))
2131 return BinaryOperator::createSub(LHS, V);
2135 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2136 if (X == RHS) // X*C + X --> X * (C+1)
2137 return BinaryOperator::createMul(RHS, AddOne(C2));
2139 // X*C1 + X*C2 --> X * (C1+C2)
2141 if (X == dyn_castFoldableMul(RHS, C1))
2142 return BinaryOperator::createMul(X, Add(C1, C2));
2145 // X + X*C --> X * (C+1)
2146 if (dyn_castFoldableMul(RHS, C2) == LHS)
2147 return BinaryOperator::createMul(LHS, AddOne(C2));
2149 // X + ~X --> -1 since ~X = -X-1
2150 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2151 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2154 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2155 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2156 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2159 // W*X + Y*Z --> W * (X+Z) iff W == Y
2160 if (I.getType()->isIntOrIntVector()) {
2161 Value *W, *X, *Y, *Z;
2162 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2163 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2167 } else if (Y == X) {
2169 } else if (X == Z) {
2176 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, Z,
2177 LHS->getName()), I);
2178 return BinaryOperator::createMul(W, NewAdd);
2183 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2185 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2186 return BinaryOperator::createSub(SubOne(CRHS), X);
2188 // (X & FF00) + xx00 -> (X+xx00) & FF00
2189 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2190 Constant *Anded = And(CRHS, C2);
2191 if (Anded == CRHS) {
2192 // See if all bits from the first bit set in the Add RHS up are included
2193 // in the mask. First, get the rightmost bit.
2194 const APInt& AddRHSV = CRHS->getValue();
2196 // Form a mask of all bits from the lowest bit added through the top.
2197 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2199 // See if the and mask includes all of these bits.
2200 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2202 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2203 // Okay, the xform is safe. Insert the new add pronto.
2204 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2205 LHS->getName()), I);
2206 return BinaryOperator::createAnd(NewAdd, C2);
2211 // Try to fold constant add into select arguments.
2212 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2213 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2217 // add (cast *A to intptrtype) B ->
2218 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2220 CastInst *CI = dyn_cast<CastInst>(LHS);
2223 CI = dyn_cast<CastInst>(RHS);
2226 if (CI && CI->getType()->isSized() &&
2227 (CI->getType()->getPrimitiveSizeInBits() ==
2228 TD->getIntPtrType()->getPrimitiveSizeInBits())
2229 && isa<PointerType>(CI->getOperand(0)->getType())) {
2231 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2232 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2233 PointerType::get(Type::Int8Ty, AS), I);
2234 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2235 return new PtrToIntInst(I2, CI->getType());
2239 // add (select X 0 (sub n A)) A --> select X A n
2241 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2244 SI = dyn_cast<SelectInst>(RHS);
2247 if (SI && SI->hasOneUse()) {
2248 Value *TV = SI->getTrueValue();
2249 Value *FV = SI->getFalseValue();
2252 // Can we fold the add into the argument of the select?
2253 // We check both true and false select arguments for a matching subtract.
2254 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2255 A == Other) // Fold the add into the true select value.
2256 return new SelectInst(SI->getCondition(), N, A);
2257 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2258 A == Other) // Fold the add into the false select value.
2259 return new SelectInst(SI->getCondition(), A, N);
2263 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2264 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2265 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2266 return ReplaceInstUsesWith(I, LHS);
2268 return Changed ? &I : 0;
2271 // isSignBit - Return true if the value represented by the constant only has the
2272 // highest order bit set.
2273 static bool isSignBit(ConstantInt *CI) {
2274 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2275 return CI->getValue() == APInt::getSignBit(NumBits);
2278 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2279 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2281 if (Op0 == Op1) // sub X, X -> 0
2282 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2284 // If this is a 'B = x-(-A)', change to B = x+A...
2285 if (Value *V = dyn_castNegVal(Op1))
2286 return BinaryOperator::createAdd(Op0, V);
2288 if (isa<UndefValue>(Op0))
2289 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2290 if (isa<UndefValue>(Op1))
2291 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2293 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2294 // Replace (-1 - A) with (~A)...
2295 if (C->isAllOnesValue())
2296 return BinaryOperator::createNot(Op1);
2298 // C - ~X == X + (1+C)
2300 if (match(Op1, m_Not(m_Value(X))))
2301 return BinaryOperator::createAdd(X, AddOne(C));
2303 // -(X >>u 31) -> (X >>s 31)
2304 // -(X >>s 31) -> (X >>u 31)
2306 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2307 if (SI->getOpcode() == Instruction::LShr) {
2308 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2309 // Check to see if we are shifting out everything but the sign bit.
2310 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2311 SI->getType()->getPrimitiveSizeInBits()-1) {
2312 // Ok, the transformation is safe. Insert AShr.
2313 return BinaryOperator::create(Instruction::AShr,
2314 SI->getOperand(0), CU, SI->getName());
2318 else if (SI->getOpcode() == Instruction::AShr) {
2319 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2320 // Check to see if we are shifting out everything but the sign bit.
2321 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2322 SI->getType()->getPrimitiveSizeInBits()-1) {
2323 // Ok, the transformation is safe. Insert LShr.
2324 return BinaryOperator::createLShr(
2325 SI->getOperand(0), CU, SI->getName());
2332 // Try to fold constant sub into select arguments.
2333 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2334 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2337 if (isa<PHINode>(Op0))
2338 if (Instruction *NV = FoldOpIntoPhi(I))
2342 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2343 if (Op1I->getOpcode() == Instruction::Add &&
2344 !Op0->getType()->isFPOrFPVector()) {
2345 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2346 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2347 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2348 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2349 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2350 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2351 // C1-(X+C2) --> (C1-C2)-X
2352 return BinaryOperator::createSub(Subtract(CI1, CI2),
2353 Op1I->getOperand(0));
2357 if (Op1I->hasOneUse()) {
2358 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2359 // is not used by anyone else...
2361 if (Op1I->getOpcode() == Instruction::Sub &&
2362 !Op1I->getType()->isFPOrFPVector()) {
2363 // Swap the two operands of the subexpr...
2364 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2365 Op1I->setOperand(0, IIOp1);
2366 Op1I->setOperand(1, IIOp0);
2368 // Create the new top level add instruction...
2369 return BinaryOperator::createAdd(Op0, Op1);
2372 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2374 if (Op1I->getOpcode() == Instruction::And &&
2375 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2376 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2379 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2380 return BinaryOperator::createAnd(Op0, NewNot);
2383 // 0 - (X sdiv C) -> (X sdiv -C)
2384 if (Op1I->getOpcode() == Instruction::SDiv)
2385 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2387 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2388 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2389 ConstantExpr::getNeg(DivRHS));
2391 // X - X*C --> X * (1-C)
2392 ConstantInt *C2 = 0;
2393 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2394 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2395 return BinaryOperator::createMul(Op0, CP1);
2398 // X - ((X / Y) * Y) --> X % Y
2399 if (Op1I->getOpcode() == Instruction::Mul)
2400 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2401 if (Op0 == I->getOperand(0) &&
2402 Op1I->getOperand(1) == I->getOperand(1)) {
2403 if (I->getOpcode() == Instruction::SDiv)
2404 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2405 if (I->getOpcode() == Instruction::UDiv)
2406 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2411 if (!Op0->getType()->isFPOrFPVector())
2412 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2413 if (Op0I->getOpcode() == Instruction::Add) {
2414 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2415 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2416 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2417 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2418 } else if (Op0I->getOpcode() == Instruction::Sub) {
2419 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2420 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2425 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2426 if (X == Op1) // X*C - X --> X * (C-1)
2427 return BinaryOperator::createMul(Op1, SubOne(C1));
2429 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2430 if (X == dyn_castFoldableMul(Op1, C2))
2431 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2436 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2437 /// comparison only checks the sign bit. If it only checks the sign bit, set
2438 /// TrueIfSigned if the result of the comparison is true when the input value is
2440 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2441 bool &TrueIfSigned) {
2443 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2444 TrueIfSigned = true;
2445 return RHS->isZero();
2446 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2447 TrueIfSigned = true;
2448 return RHS->isAllOnesValue();
2449 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2450 TrueIfSigned = false;
2451 return RHS->isAllOnesValue();
2452 case ICmpInst::ICMP_UGT:
2453 // True if LHS u> RHS and RHS == high-bit-mask - 1
2454 TrueIfSigned = true;
2455 return RHS->getValue() ==
2456 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2457 case ICmpInst::ICMP_UGE:
2458 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2459 TrueIfSigned = true;
2460 return RHS->getValue() ==
2461 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2467 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2468 bool Changed = SimplifyCommutative(I);
2469 Value *Op0 = I.getOperand(0);
2471 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2472 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2474 // Simplify mul instructions with a constant RHS...
2475 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2476 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2478 // ((X << C1)*C2) == (X * (C2 << C1))
2479 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2480 if (SI->getOpcode() == Instruction::Shl)
2481 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2482 return BinaryOperator::createMul(SI->getOperand(0),
2483 ConstantExpr::getShl(CI, ShOp));
2486 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2487 if (CI->equalsInt(1)) // X * 1 == X
2488 return ReplaceInstUsesWith(I, Op0);
2489 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2490 return BinaryOperator::createNeg(Op0, I.getName());
2492 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2493 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2494 return BinaryOperator::createShl(Op0,
2495 ConstantInt::get(Op0->getType(), Val.logBase2()));
2497 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2498 if (Op1F->isNullValue())
2499 return ReplaceInstUsesWith(I, Op1);
2501 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2502 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2503 // We need a better interface for long double here.
2504 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2505 if (Op1F->isExactlyValue(1.0))
2506 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2509 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2510 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2511 isa<ConstantInt>(Op0I->getOperand(1))) {
2512 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2513 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2515 InsertNewInstBefore(Add, I);
2516 Value *C1C2 = ConstantExpr::getMul(Op1,
2517 cast<Constant>(Op0I->getOperand(1)));
2518 return BinaryOperator::createAdd(Add, C1C2);
2522 // Try to fold constant mul into select arguments.
2523 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2524 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2527 if (isa<PHINode>(Op0))
2528 if (Instruction *NV = FoldOpIntoPhi(I))
2532 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2533 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2534 return BinaryOperator::createMul(Op0v, Op1v);
2536 // If one of the operands of the multiply is a cast from a boolean value, then
2537 // we know the bool is either zero or one, so this is a 'masking' multiply.
2538 // See if we can simplify things based on how the boolean was originally
2540 CastInst *BoolCast = 0;
2541 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2542 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2545 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2546 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2549 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2550 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2551 const Type *SCOpTy = SCIOp0->getType();
2554 // If the icmp is true iff the sign bit of X is set, then convert this
2555 // multiply into a shift/and combination.
2556 if (isa<ConstantInt>(SCIOp1) &&
2557 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2559 // Shift the X value right to turn it into "all signbits".
2560 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2561 SCOpTy->getPrimitiveSizeInBits()-1);
2563 InsertNewInstBefore(
2564 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2565 BoolCast->getOperand(0)->getName()+
2568 // If the multiply type is not the same as the source type, sign extend
2569 // or truncate to the multiply type.
2570 if (I.getType() != V->getType()) {
2571 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2572 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2573 Instruction::CastOps opcode =
2574 (SrcBits == DstBits ? Instruction::BitCast :
2575 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2576 V = InsertCastBefore(opcode, V, I.getType(), I);
2579 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2580 return BinaryOperator::createAnd(V, OtherOp);
2585 return Changed ? &I : 0;
2588 /// This function implements the transforms on div instructions that work
2589 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2590 /// used by the visitors to those instructions.
2591 /// @brief Transforms common to all three div instructions
2592 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2593 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2595 // undef / X -> 0 for integer.
2596 // undef / X -> undef for FP (the undef could be a snan).
2597 if (isa<UndefValue>(Op0)) {
2598 if (Op0->getType()->isFPOrFPVector())
2599 return ReplaceInstUsesWith(I, Op0);
2600 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2603 // X / undef -> undef
2604 if (isa<UndefValue>(Op1))
2605 return ReplaceInstUsesWith(I, Op1);
2607 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2608 // This does not apply for fdiv.
2609 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2610 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2611 // the same basic block, then we replace the select with Y, and the
2612 // condition of the select with false (if the cond value is in the same BB).
2613 // If the select has uses other than the div, this allows them to be
2614 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2615 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2616 if (ST->isNullValue()) {
2617 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2618 if (CondI && CondI->getParent() == I.getParent())
2619 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2620 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2621 I.setOperand(1, SI->getOperand(2));
2623 UpdateValueUsesWith(SI, SI->getOperand(2));
2627 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2628 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2629 if (ST->isNullValue()) {
2630 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2631 if (CondI && CondI->getParent() == I.getParent())
2632 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2633 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2634 I.setOperand(1, SI->getOperand(1));
2636 UpdateValueUsesWith(SI, SI->getOperand(1));
2644 /// This function implements the transforms common to both integer division
2645 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2646 /// division instructions.
2647 /// @brief Common integer divide transforms
2648 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2649 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2651 if (Instruction *Common = commonDivTransforms(I))
2654 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2656 if (RHS->equalsInt(1))
2657 return ReplaceInstUsesWith(I, Op0);
2659 // (X / C1) / C2 -> X / (C1*C2)
2660 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2661 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2662 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2663 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2664 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2666 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2667 Multiply(RHS, LHSRHS));
2670 if (!RHS->isZero()) { // avoid X udiv 0
2671 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2672 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2674 if (isa<PHINode>(Op0))
2675 if (Instruction *NV = FoldOpIntoPhi(I))
2680 // 0 / X == 0, we don't need to preserve faults!
2681 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2682 if (LHS->equalsInt(0))
2683 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2688 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2689 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2691 // Handle the integer div common cases
2692 if (Instruction *Common = commonIDivTransforms(I))
2695 // X udiv C^2 -> X >> C
2696 // Check to see if this is an unsigned division with an exact power of 2,
2697 // if so, convert to a right shift.
2698 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2699 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2700 return BinaryOperator::createLShr(Op0,
2701 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2704 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2705 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2706 if (RHSI->getOpcode() == Instruction::Shl &&
2707 isa<ConstantInt>(RHSI->getOperand(0))) {
2708 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2709 if (C1.isPowerOf2()) {
2710 Value *N = RHSI->getOperand(1);
2711 const Type *NTy = N->getType();
2712 if (uint32_t C2 = C1.logBase2()) {
2713 Constant *C2V = ConstantInt::get(NTy, C2);
2714 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2716 return BinaryOperator::createLShr(Op0, N);
2721 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2722 // where C1&C2 are powers of two.
2723 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2724 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2725 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2726 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2727 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2728 // Compute the shift amounts
2729 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2730 // Construct the "on true" case of the select
2731 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2732 Instruction *TSI = BinaryOperator::createLShr(
2733 Op0, TC, SI->getName()+".t");
2734 TSI = InsertNewInstBefore(TSI, I);
2736 // Construct the "on false" case of the select
2737 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2738 Instruction *FSI = BinaryOperator::createLShr(
2739 Op0, FC, SI->getName()+".f");
2740 FSI = InsertNewInstBefore(FSI, I);
2742 // construct the select instruction and return it.
2743 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2749 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2750 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2752 // Handle the integer div common cases
2753 if (Instruction *Common = commonIDivTransforms(I))
2756 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2758 if (RHS->isAllOnesValue())
2759 return BinaryOperator::createNeg(Op0);
2762 if (Value *LHSNeg = dyn_castNegVal(Op0))
2763 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2766 // If the sign bits of both operands are zero (i.e. we can prove they are
2767 // unsigned inputs), turn this into a udiv.
2768 if (I.getType()->isInteger()) {
2769 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2770 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2771 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2772 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2779 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2780 return commonDivTransforms(I);
2783 /// GetFactor - If we can prove that the specified value is at least a multiple
2784 /// of some factor, return that factor.
2785 static Constant *GetFactor(Value *V) {
2786 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2789 // Unless we can be tricky, we know this is a multiple of 1.
2790 Constant *Result = ConstantInt::get(V->getType(), 1);
2792 Instruction *I = dyn_cast<Instruction>(V);
2793 if (!I) return Result;
2795 if (I->getOpcode() == Instruction::Mul) {
2796 // Handle multiplies by a constant, etc.
2797 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2798 GetFactor(I->getOperand(1)));
2799 } else if (I->getOpcode() == Instruction::Shl) {
2800 // (X<<C) -> X * (1 << C)
2801 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2802 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2803 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2805 } else if (I->getOpcode() == Instruction::And) {
2806 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2807 // X & 0xFFF0 is known to be a multiple of 16.
2808 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2809 if (Zeros != V->getType()->getPrimitiveSizeInBits())// don't shift by "32"
2810 return ConstantExpr::getShl(Result,
2811 ConstantInt::get(Result->getType(), Zeros));
2813 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2814 // Only handle int->int casts.
2815 if (!CI->isIntegerCast())
2817 Value *Op = CI->getOperand(0);
2818 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2823 /// This function implements the transforms on rem instructions that work
2824 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2825 /// is used by the visitors to those instructions.
2826 /// @brief Transforms common to all three rem instructions
2827 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2828 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2830 // 0 % X == 0 for integer, we don't need to preserve faults!
2831 if (Constant *LHS = dyn_cast<Constant>(Op0))
2832 if (LHS->isNullValue())
2833 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2835 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2836 if (I.getType()->isFPOrFPVector())
2837 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2838 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2840 if (isa<UndefValue>(Op1))
2841 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2843 // Handle cases involving: rem X, (select Cond, Y, Z)
2844 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2845 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2846 // the same basic block, then we replace the select with Y, and the
2847 // condition of the select with false (if the cond value is in the same
2848 // BB). If the select has uses other than the div, this allows them to be
2850 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2851 if (ST->isNullValue()) {
2852 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2853 if (CondI && CondI->getParent() == I.getParent())
2854 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2855 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2856 I.setOperand(1, SI->getOperand(2));
2858 UpdateValueUsesWith(SI, SI->getOperand(2));
2861 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2862 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2863 if (ST->isNullValue()) {
2864 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2865 if (CondI && CondI->getParent() == I.getParent())
2866 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2867 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2868 I.setOperand(1, SI->getOperand(1));
2870 UpdateValueUsesWith(SI, SI->getOperand(1));
2878 /// This function implements the transforms common to both integer remainder
2879 /// instructions (urem and srem). It is called by the visitors to those integer
2880 /// remainder instructions.
2881 /// @brief Common integer remainder transforms
2882 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2883 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2885 if (Instruction *common = commonRemTransforms(I))
2888 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2889 // X % 0 == undef, we don't need to preserve faults!
2890 if (RHS->equalsInt(0))
2891 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2893 if (RHS->equalsInt(1)) // X % 1 == 0
2894 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2896 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2897 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2898 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2900 } else if (isa<PHINode>(Op0I)) {
2901 if (Instruction *NV = FoldOpIntoPhi(I))
2904 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2905 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2906 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2913 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2914 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2916 if (Instruction *common = commonIRemTransforms(I))
2919 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2920 // X urem C^2 -> X and C
2921 // Check to see if this is an unsigned remainder with an exact power of 2,
2922 // if so, convert to a bitwise and.
2923 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2924 if (C->getValue().isPowerOf2())
2925 return BinaryOperator::createAnd(Op0, SubOne(C));
2928 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2929 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2930 if (RHSI->getOpcode() == Instruction::Shl &&
2931 isa<ConstantInt>(RHSI->getOperand(0))) {
2932 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2933 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2934 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2936 return BinaryOperator::createAnd(Op0, Add);
2941 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2942 // where C1&C2 are powers of two.
2943 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2944 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2945 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2946 // STO == 0 and SFO == 0 handled above.
2947 if ((STO->getValue().isPowerOf2()) &&
2948 (SFO->getValue().isPowerOf2())) {
2949 Value *TrueAnd = InsertNewInstBefore(
2950 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2951 Value *FalseAnd = InsertNewInstBefore(
2952 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2953 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2961 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2962 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2964 // Handle the integer rem common cases
2965 if (Instruction *common = commonIRemTransforms(I))
2968 if (Value *RHSNeg = dyn_castNegVal(Op1))
2969 if (!isa<ConstantInt>(RHSNeg) ||
2970 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2972 AddUsesToWorkList(I);
2973 I.setOperand(1, RHSNeg);
2977 // If the sign bits of both operands are zero (i.e. we can prove they are
2978 // unsigned inputs), turn this into a urem.
2979 if (I.getType()->isInteger()) {
2980 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2981 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2982 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2983 return BinaryOperator::createURem(Op0, Op1, I.getName());
2990 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2991 return commonRemTransforms(I);
2994 // isMaxValueMinusOne - return true if this is Max-1
2995 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2996 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2998 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2999 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
3002 // isMinValuePlusOne - return true if this is Min+1
3003 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
3005 return C->getValue() == 1; // unsigned
3007 // Calculate 1111111111000000000000
3008 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3009 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
3012 // isOneBitSet - Return true if there is exactly one bit set in the specified
3014 static bool isOneBitSet(const ConstantInt *CI) {
3015 return CI->getValue().isPowerOf2();
3018 // isHighOnes - Return true if the constant is of the form 1+0+.
3019 // This is the same as lowones(~X).
3020 static bool isHighOnes(const ConstantInt *CI) {
3021 return (~CI->getValue() + 1).isPowerOf2();
3024 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3025 /// are carefully arranged to allow folding of expressions such as:
3027 /// (A < B) | (A > B) --> (A != B)
3029 /// Note that this is only valid if the first and second predicates have the
3030 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3032 /// Three bits are used to represent the condition, as follows:
3037 /// <=> Value Definition
3038 /// 000 0 Always false
3045 /// 111 7 Always true
3047 static unsigned getICmpCode(const ICmpInst *ICI) {
3048 switch (ICI->getPredicate()) {
3050 case ICmpInst::ICMP_UGT: return 1; // 001
3051 case ICmpInst::ICMP_SGT: return 1; // 001
3052 case ICmpInst::ICMP_EQ: return 2; // 010
3053 case ICmpInst::ICMP_UGE: return 3; // 011
3054 case ICmpInst::ICMP_SGE: return 3; // 011
3055 case ICmpInst::ICMP_ULT: return 4; // 100
3056 case ICmpInst::ICMP_SLT: return 4; // 100
3057 case ICmpInst::ICMP_NE: return 5; // 101
3058 case ICmpInst::ICMP_ULE: return 6; // 110
3059 case ICmpInst::ICMP_SLE: return 6; // 110
3062 assert(0 && "Invalid ICmp predicate!");
3067 /// getICmpValue - This is the complement of getICmpCode, which turns an
3068 /// opcode and two operands into either a constant true or false, or a brand
3069 /// new ICmp instruction. The sign is passed in to determine which kind
3070 /// of predicate to use in new icmp instructions.
3071 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3073 default: assert(0 && "Illegal ICmp code!");
3074 case 0: return ConstantInt::getFalse();
3077 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3079 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3080 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3083 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3085 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3088 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3090 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3091 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3094 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3096 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3097 case 7: return ConstantInt::getTrue();
3101 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3102 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3103 (ICmpInst::isSignedPredicate(p1) &&
3104 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3105 (ICmpInst::isSignedPredicate(p2) &&
3106 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3110 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3111 struct FoldICmpLogical {
3114 ICmpInst::Predicate pred;
3115 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3116 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3117 pred(ICI->getPredicate()) {}
3118 bool shouldApply(Value *V) const {
3119 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3120 if (PredicatesFoldable(pred, ICI->getPredicate()))
3121 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3122 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3125 Instruction *apply(Instruction &Log) const {
3126 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3127 if (ICI->getOperand(0) != LHS) {
3128 assert(ICI->getOperand(1) == LHS);
3129 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3132 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3133 unsigned LHSCode = getICmpCode(ICI);
3134 unsigned RHSCode = getICmpCode(RHSICI);
3136 switch (Log.getOpcode()) {
3137 case Instruction::And: Code = LHSCode & RHSCode; break;
3138 case Instruction::Or: Code = LHSCode | RHSCode; break;
3139 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3140 default: assert(0 && "Illegal logical opcode!"); return 0;
3143 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3144 ICmpInst::isSignedPredicate(ICI->getPredicate());
3146 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3147 if (Instruction *I = dyn_cast<Instruction>(RV))
3149 // Otherwise, it's a constant boolean value...
3150 return IC.ReplaceInstUsesWith(Log, RV);
3153 } // end anonymous namespace
3155 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3156 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3157 // guaranteed to be a binary operator.
3158 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3160 ConstantInt *AndRHS,
3161 BinaryOperator &TheAnd) {
3162 Value *X = Op->getOperand(0);
3163 Constant *Together = 0;
3165 Together = And(AndRHS, OpRHS);
3167 switch (Op->getOpcode()) {
3168 case Instruction::Xor:
3169 if (Op->hasOneUse()) {
3170 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3171 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3172 InsertNewInstBefore(And, TheAnd);
3174 return BinaryOperator::createXor(And, Together);
3177 case Instruction::Or:
3178 if (Together == AndRHS) // (X | C) & C --> C
3179 return ReplaceInstUsesWith(TheAnd, AndRHS);
3181 if (Op->hasOneUse() && Together != OpRHS) {
3182 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3183 Instruction *Or = BinaryOperator::createOr(X, Together);
3184 InsertNewInstBefore(Or, TheAnd);
3186 return BinaryOperator::createAnd(Or, AndRHS);
3189 case Instruction::Add:
3190 if (Op->hasOneUse()) {
3191 // Adding a one to a single bit bit-field should be turned into an XOR
3192 // of the bit. First thing to check is to see if this AND is with a
3193 // single bit constant.
3194 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3196 // If there is only one bit set...
3197 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3198 // Ok, at this point, we know that we are masking the result of the
3199 // ADD down to exactly one bit. If the constant we are adding has
3200 // no bits set below this bit, then we can eliminate the ADD.
3201 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3203 // Check to see if any bits below the one bit set in AndRHSV are set.
3204 if ((AddRHS & (AndRHSV-1)) == 0) {
3205 // If not, the only thing that can effect the output of the AND is
3206 // the bit specified by AndRHSV. If that bit is set, the effect of
3207 // the XOR is to toggle the bit. If it is clear, then the ADD has
3209 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3210 TheAnd.setOperand(0, X);
3213 // Pull the XOR out of the AND.
3214 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3215 InsertNewInstBefore(NewAnd, TheAnd);
3216 NewAnd->takeName(Op);
3217 return BinaryOperator::createXor(NewAnd, AndRHS);
3224 case Instruction::Shl: {
3225 // We know that the AND will not produce any of the bits shifted in, so if
3226 // the anded constant includes them, clear them now!
3228 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3229 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3230 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3231 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3233 if (CI->getValue() == ShlMask) {
3234 // Masking out bits that the shift already masks
3235 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3236 } else if (CI != AndRHS) { // Reducing bits set in and.
3237 TheAnd.setOperand(1, CI);
3242 case Instruction::LShr:
3244 // We know that the AND will not produce any of the bits shifted in, so if
3245 // the anded constant includes them, clear them now! This only applies to
3246 // unsigned shifts, because a signed shr may bring in set bits!
3248 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3249 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3250 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3251 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3253 if (CI->getValue() == ShrMask) {
3254 // Masking out bits that the shift already masks.
3255 return ReplaceInstUsesWith(TheAnd, Op);
3256 } else if (CI != AndRHS) {
3257 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3262 case Instruction::AShr:
3264 // See if this is shifting in some sign extension, then masking it out
3266 if (Op->hasOneUse()) {
3267 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3268 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3269 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3270 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3271 if (C == AndRHS) { // Masking out bits shifted in.
3272 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3273 // Make the argument unsigned.
3274 Value *ShVal = Op->getOperand(0);
3275 ShVal = InsertNewInstBefore(
3276 BinaryOperator::createLShr(ShVal, OpRHS,
3277 Op->getName()), TheAnd);
3278 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3287 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3288 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3289 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3290 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3291 /// insert new instructions.
3292 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3293 bool isSigned, bool Inside,
3295 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3296 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3297 "Lo is not <= Hi in range emission code!");
3300 if (Lo == Hi) // Trivially false.
3301 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3303 // V >= Min && V < Hi --> V < Hi
3304 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3305 ICmpInst::Predicate pred = (isSigned ?
3306 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3307 return new ICmpInst(pred, V, Hi);
3310 // Emit V-Lo <u Hi-Lo
3311 Constant *NegLo = ConstantExpr::getNeg(Lo);
3312 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3313 InsertNewInstBefore(Add, IB);
3314 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3315 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3318 if (Lo == Hi) // Trivially true.
3319 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3321 // V < Min || V >= Hi -> V > Hi-1
3322 Hi = SubOne(cast<ConstantInt>(Hi));
3323 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3324 ICmpInst::Predicate pred = (isSigned ?
3325 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3326 return new ICmpInst(pred, V, Hi);
3329 // Emit V-Lo >u Hi-1-Lo
3330 // Note that Hi has already had one subtracted from it, above.
3331 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3332 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3333 InsertNewInstBefore(Add, IB);
3334 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3335 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3338 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3339 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3340 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3341 // not, since all 1s are not contiguous.
3342 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3343 const APInt& V = Val->getValue();
3344 uint32_t BitWidth = Val->getType()->getBitWidth();
3345 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3347 // look for the first zero bit after the run of ones
3348 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3349 // look for the first non-zero bit
3350 ME = V.getActiveBits();
3354 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3355 /// where isSub determines whether the operator is a sub. If we can fold one of
3356 /// the following xforms:
3358 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3359 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3360 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3362 /// return (A +/- B).
3364 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3365 ConstantInt *Mask, bool isSub,
3367 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3368 if (!LHSI || LHSI->getNumOperands() != 2 ||
3369 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3371 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3373 switch (LHSI->getOpcode()) {
3375 case Instruction::And:
3376 if (And(N, Mask) == Mask) {
3377 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3378 if ((Mask->getValue().countLeadingZeros() +
3379 Mask->getValue().countPopulation()) ==
3380 Mask->getValue().getBitWidth())
3383 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3384 // part, we don't need any explicit masks to take them out of A. If that
3385 // is all N is, ignore it.
3386 uint32_t MB = 0, ME = 0;
3387 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3388 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3389 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3390 if (MaskedValueIsZero(RHS, Mask))
3395 case Instruction::Or:
3396 case Instruction::Xor:
3397 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3398 if ((Mask->getValue().countLeadingZeros() +
3399 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3400 && And(N, Mask)->isZero())
3407 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3409 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3410 return InsertNewInstBefore(New, I);
3413 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3414 bool Changed = SimplifyCommutative(I);
3415 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3417 if (isa<UndefValue>(Op1)) // X & undef -> 0
3418 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3422 return ReplaceInstUsesWith(I, Op1);
3424 // See if we can simplify any instructions used by the instruction whose sole
3425 // purpose is to compute bits we don't care about.
3426 if (!isa<VectorType>(I.getType())) {
3427 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3428 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3429 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3430 KnownZero, KnownOne))
3433 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3434 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3435 return ReplaceInstUsesWith(I, I.getOperand(0));
3436 } else if (isa<ConstantAggregateZero>(Op1)) {
3437 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3441 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3442 const APInt& AndRHSMask = AndRHS->getValue();
3443 APInt NotAndRHS(~AndRHSMask);
3445 // Optimize a variety of ((val OP C1) & C2) combinations...
3446 if (isa<BinaryOperator>(Op0)) {
3447 Instruction *Op0I = cast<Instruction>(Op0);
3448 Value *Op0LHS = Op0I->getOperand(0);
3449 Value *Op0RHS = Op0I->getOperand(1);
3450 switch (Op0I->getOpcode()) {
3451 case Instruction::Xor:
3452 case Instruction::Or:
3453 // If the mask is only needed on one incoming arm, push it up.
3454 if (Op0I->hasOneUse()) {
3455 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3456 // Not masking anything out for the LHS, move to RHS.
3457 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3458 Op0RHS->getName()+".masked");
3459 InsertNewInstBefore(NewRHS, I);
3460 return BinaryOperator::create(
3461 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3463 if (!isa<Constant>(Op0RHS) &&
3464 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3465 // Not masking anything out for the RHS, move to LHS.
3466 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3467 Op0LHS->getName()+".masked");
3468 InsertNewInstBefore(NewLHS, I);
3469 return BinaryOperator::create(
3470 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3475 case Instruction::Add:
3476 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3477 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3478 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3479 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3480 return BinaryOperator::createAnd(V, AndRHS);
3481 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3482 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3485 case Instruction::Sub:
3486 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3487 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3488 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3489 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3490 return BinaryOperator::createAnd(V, AndRHS);
3494 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3495 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3497 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3498 // If this is an integer truncation or change from signed-to-unsigned, and
3499 // if the source is an and/or with immediate, transform it. This
3500 // frequently occurs for bitfield accesses.
3501 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3502 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3503 CastOp->getNumOperands() == 2)
3504 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3505 if (CastOp->getOpcode() == Instruction::And) {
3506 // Change: and (cast (and X, C1) to T), C2
3507 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3508 // This will fold the two constants together, which may allow
3509 // other simplifications.
3510 Instruction *NewCast = CastInst::createTruncOrBitCast(
3511 CastOp->getOperand(0), I.getType(),
3512 CastOp->getName()+".shrunk");
3513 NewCast = InsertNewInstBefore(NewCast, I);
3514 // trunc_or_bitcast(C1)&C2
3515 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3516 C3 = ConstantExpr::getAnd(C3, AndRHS);
3517 return BinaryOperator::createAnd(NewCast, C3);
3518 } else if (CastOp->getOpcode() == Instruction::Or) {
3519 // Change: and (cast (or X, C1) to T), C2
3520 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3521 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3522 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3523 return ReplaceInstUsesWith(I, AndRHS);
3529 // Try to fold constant and into select arguments.
3530 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3531 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3533 if (isa<PHINode>(Op0))
3534 if (Instruction *NV = FoldOpIntoPhi(I))
3538 Value *Op0NotVal = dyn_castNotVal(Op0);
3539 Value *Op1NotVal = dyn_castNotVal(Op1);
3541 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3542 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3544 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3545 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3546 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3547 I.getName()+".demorgan");
3548 InsertNewInstBefore(Or, I);
3549 return BinaryOperator::createNot(Or);
3553 Value *A = 0, *B = 0, *C = 0, *D = 0;
3554 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3555 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3556 return ReplaceInstUsesWith(I, Op1);
3558 // (A|B) & ~(A&B) -> A^B
3559 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3560 if ((A == C && B == D) || (A == D && B == C))
3561 return BinaryOperator::createXor(A, B);
3565 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3566 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3567 return ReplaceInstUsesWith(I, Op0);
3569 // ~(A&B) & (A|B) -> A^B
3570 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3571 if ((A == C && B == D) || (A == D && B == C))
3572 return BinaryOperator::createXor(A, B);
3576 if (Op0->hasOneUse() &&
3577 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3578 if (A == Op1) { // (A^B)&A -> A&(A^B)
3579 I.swapOperands(); // Simplify below
3580 std::swap(Op0, Op1);
3581 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3582 cast<BinaryOperator>(Op0)->swapOperands();
3583 I.swapOperands(); // Simplify below
3584 std::swap(Op0, Op1);
3587 if (Op1->hasOneUse() &&
3588 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3589 if (B == Op0) { // B&(A^B) -> B&(B^A)
3590 cast<BinaryOperator>(Op1)->swapOperands();
3593 if (A == Op0) { // A&(A^B) -> A & ~B
3594 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3595 InsertNewInstBefore(NotB, I);
3596 return BinaryOperator::createAnd(A, NotB);
3601 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3602 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3603 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3606 Value *LHSVal, *RHSVal;
3607 ConstantInt *LHSCst, *RHSCst;
3608 ICmpInst::Predicate LHSCC, RHSCC;
3609 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3610 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3611 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3612 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3613 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3614 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3615 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3616 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3618 // Don't try to fold ICMP_SLT + ICMP_ULT.
3619 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3620 ICmpInst::isSignedPredicate(LHSCC) ==
3621 ICmpInst::isSignedPredicate(RHSCC))) {
3622 // Ensure that the larger constant is on the RHS.
3623 ICmpInst::Predicate GT;
3624 if (ICmpInst::isSignedPredicate(LHSCC) ||
3625 (ICmpInst::isEquality(LHSCC) &&
3626 ICmpInst::isSignedPredicate(RHSCC)))
3627 GT = ICmpInst::ICMP_SGT;
3629 GT = ICmpInst::ICMP_UGT;
3631 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3632 ICmpInst *LHS = cast<ICmpInst>(Op0);
3633 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3634 std::swap(LHS, RHS);
3635 std::swap(LHSCst, RHSCst);
3636 std::swap(LHSCC, RHSCC);
3639 // At this point, we know we have have two icmp instructions
3640 // comparing a value against two constants and and'ing the result
3641 // together. Because of the above check, we know that we only have
3642 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3643 // (from the FoldICmpLogical check above), that the two constants
3644 // are not equal and that the larger constant is on the RHS
3645 assert(LHSCst != RHSCst && "Compares not folded above?");
3648 default: assert(0 && "Unknown integer condition code!");
3649 case ICmpInst::ICMP_EQ:
3651 default: assert(0 && "Unknown integer condition code!");
3652 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3653 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3654 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3655 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3656 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3657 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3658 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3659 return ReplaceInstUsesWith(I, LHS);
3661 case ICmpInst::ICMP_NE:
3663 default: assert(0 && "Unknown integer condition code!");
3664 case ICmpInst::ICMP_ULT:
3665 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3666 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3667 break; // (X != 13 & X u< 15) -> no change
3668 case ICmpInst::ICMP_SLT:
3669 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3670 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3671 break; // (X != 13 & X s< 15) -> no change
3672 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3673 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3674 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3675 return ReplaceInstUsesWith(I, RHS);
3676 case ICmpInst::ICMP_NE:
3677 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3678 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3679 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3680 LHSVal->getName()+".off");
3681 InsertNewInstBefore(Add, I);
3682 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3683 ConstantInt::get(Add->getType(), 1));
3685 break; // (X != 13 & X != 15) -> no change
3688 case ICmpInst::ICMP_ULT:
3690 default: assert(0 && "Unknown integer condition code!");
3691 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3692 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3693 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3694 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3696 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3697 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3698 return ReplaceInstUsesWith(I, LHS);
3699 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3703 case ICmpInst::ICMP_SLT:
3705 default: assert(0 && "Unknown integer condition code!");
3706 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3707 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3708 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3709 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3711 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3712 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3713 return ReplaceInstUsesWith(I, LHS);
3714 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3718 case ICmpInst::ICMP_UGT:
3720 default: assert(0 && "Unknown integer condition code!");
3721 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3722 return ReplaceInstUsesWith(I, LHS);
3723 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3724 return ReplaceInstUsesWith(I, RHS);
3725 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3727 case ICmpInst::ICMP_NE:
3728 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3729 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3730 break; // (X u> 13 & X != 15) -> no change
3731 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3732 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3734 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3738 case ICmpInst::ICMP_SGT:
3740 default: assert(0 && "Unknown integer condition code!");
3741 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3742 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3743 return ReplaceInstUsesWith(I, RHS);
3744 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3746 case ICmpInst::ICMP_NE:
3747 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3748 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3749 break; // (X s> 13 & X != 15) -> no change
3750 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3751 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3753 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3761 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3762 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3763 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3764 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3765 const Type *SrcTy = Op0C->getOperand(0)->getType();
3766 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3767 // Only do this if the casts both really cause code to be generated.
3768 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3770 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3772 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3773 Op1C->getOperand(0),
3775 InsertNewInstBefore(NewOp, I);
3776 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3780 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3781 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3782 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3783 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3784 SI0->getOperand(1) == SI1->getOperand(1) &&
3785 (SI0->hasOneUse() || SI1->hasOneUse())) {
3786 Instruction *NewOp =
3787 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3789 SI0->getName()), I);
3790 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3791 SI1->getOperand(1));
3795 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3796 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3797 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3798 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3799 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3800 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3801 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3802 // If either of the constants are nans, then the whole thing returns
3804 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3805 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3806 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3807 RHS->getOperand(0));
3812 return Changed ? &I : 0;
3815 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3816 /// in the result. If it does, and if the specified byte hasn't been filled in
3817 /// yet, fill it in and return false.
3818 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3819 Instruction *I = dyn_cast<Instruction>(V);
3820 if (I == 0) return true;
3822 // If this is an or instruction, it is an inner node of the bswap.
3823 if (I->getOpcode() == Instruction::Or)
3824 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3825 CollectBSwapParts(I->getOperand(1), ByteValues);
3827 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3828 // If this is a shift by a constant int, and it is "24", then its operand
3829 // defines a byte. We only handle unsigned types here.
3830 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3831 // Not shifting the entire input by N-1 bytes?
3832 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3833 8*(ByteValues.size()-1))
3837 if (I->getOpcode() == Instruction::Shl) {
3838 // X << 24 defines the top byte with the lowest of the input bytes.
3839 DestNo = ByteValues.size()-1;
3841 // X >>u 24 defines the low byte with the highest of the input bytes.
3845 // If the destination byte value is already defined, the values are or'd
3846 // together, which isn't a bswap (unless it's an or of the same bits).
3847 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3849 ByteValues[DestNo] = I->getOperand(0);
3853 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3855 Value *Shift = 0, *ShiftLHS = 0;
3856 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3857 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3858 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3860 Instruction *SI = cast<Instruction>(Shift);
3862 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3863 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3864 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3867 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3869 if (AndAmt->getValue().getActiveBits() > 64)
3871 uint64_t AndAmtVal = AndAmt->getZExtValue();
3872 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3873 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3875 // Unknown mask for bswap.
3876 if (DestByte == ByteValues.size()) return true;
3878 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3880 if (SI->getOpcode() == Instruction::Shl)
3881 SrcByte = DestByte - ShiftBytes;
3883 SrcByte = DestByte + ShiftBytes;
3885 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3886 if (SrcByte != ByteValues.size()-DestByte-1)
3889 // If the destination byte value is already defined, the values are or'd
3890 // together, which isn't a bswap (unless it's an or of the same bits).
3891 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3893 ByteValues[DestByte] = SI->getOperand(0);
3897 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3898 /// If so, insert the new bswap intrinsic and return it.
3899 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3900 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3901 if (!ITy || ITy->getBitWidth() % 16)
3902 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3904 /// ByteValues - For each byte of the result, we keep track of which value
3905 /// defines each byte.
3906 SmallVector<Value*, 8> ByteValues;
3907 ByteValues.resize(ITy->getBitWidth()/8);
3909 // Try to find all the pieces corresponding to the bswap.
3910 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3911 CollectBSwapParts(I.getOperand(1), ByteValues))
3914 // Check to see if all of the bytes come from the same value.
3915 Value *V = ByteValues[0];
3916 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3918 // Check to make sure that all of the bytes come from the same value.
3919 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3920 if (ByteValues[i] != V)
3922 const Type *Tys[] = { ITy };
3923 Module *M = I.getParent()->getParent()->getParent();
3924 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3925 return new CallInst(F, V);
3929 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3930 bool Changed = SimplifyCommutative(I);
3931 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3933 if (isa<UndefValue>(Op1)) // X | undef -> -1
3934 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3938 return ReplaceInstUsesWith(I, Op0);
3940 // See if we can simplify any instructions used by the instruction whose sole
3941 // purpose is to compute bits we don't care about.
3942 if (!isa<VectorType>(I.getType())) {
3943 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3944 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3945 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3946 KnownZero, KnownOne))
3948 } else if (isa<ConstantAggregateZero>(Op1)) {
3949 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3950 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3951 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3952 return ReplaceInstUsesWith(I, I.getOperand(1));
3958 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3959 ConstantInt *C1 = 0; Value *X = 0;
3960 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3961 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3962 Instruction *Or = BinaryOperator::createOr(X, RHS);
3963 InsertNewInstBefore(Or, I);
3965 return BinaryOperator::createAnd(Or,
3966 ConstantInt::get(RHS->getValue() | C1->getValue()));
3969 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3970 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3971 Instruction *Or = BinaryOperator::createOr(X, RHS);
3972 InsertNewInstBefore(Or, I);
3974 return BinaryOperator::createXor(Or,
3975 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3978 // Try to fold constant and into select arguments.
3979 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3980 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3982 if (isa<PHINode>(Op0))
3983 if (Instruction *NV = FoldOpIntoPhi(I))
3987 Value *A = 0, *B = 0;
3988 ConstantInt *C1 = 0, *C2 = 0;
3990 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3991 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3992 return ReplaceInstUsesWith(I, Op1);
3993 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3994 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3995 return ReplaceInstUsesWith(I, Op0);
3997 // (A | B) | C and A | (B | C) -> bswap if possible.
3998 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3999 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4000 match(Op1, m_Or(m_Value(), m_Value())) ||
4001 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4002 match(Op1, m_Shift(m_Value(), m_Value())))) {
4003 if (Instruction *BSwap = MatchBSwap(I))
4007 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4008 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4009 MaskedValueIsZero(Op1, C1->getValue())) {
4010 Instruction *NOr = BinaryOperator::createOr(A, Op1);
4011 InsertNewInstBefore(NOr, I);
4013 return BinaryOperator::createXor(NOr, C1);
4016 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4017 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4018 MaskedValueIsZero(Op0, C1->getValue())) {
4019 Instruction *NOr = BinaryOperator::createOr(A, Op0);
4020 InsertNewInstBefore(NOr, I);
4022 return BinaryOperator::createXor(NOr, C1);
4026 Value *C = 0, *D = 0;
4027 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4028 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4029 Value *V1 = 0, *V2 = 0, *V3 = 0;
4030 C1 = dyn_cast<ConstantInt>(C);
4031 C2 = dyn_cast<ConstantInt>(D);
4032 if (C1 && C2) { // (A & C1)|(B & C2)
4033 // If we have: ((V + N) & C1) | (V & C2)
4034 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4035 // replace with V+N.
4036 if (C1->getValue() == ~C2->getValue()) {
4037 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4038 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4039 // Add commutes, try both ways.
4040 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4041 return ReplaceInstUsesWith(I, A);
4042 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4043 return ReplaceInstUsesWith(I, A);
4045 // Or commutes, try both ways.
4046 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4047 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4048 // Add commutes, try both ways.
4049 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4050 return ReplaceInstUsesWith(I, B);
4051 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4052 return ReplaceInstUsesWith(I, B);
4055 V1 = 0; V2 = 0; V3 = 0;
4058 // Check to see if we have any common things being and'ed. If so, find the
4059 // terms for V1 & (V2|V3).
4060 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4061 if (A == B) // (A & C)|(A & D) == A & (C|D)
4062 V1 = A, V2 = C, V3 = D;
4063 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4064 V1 = A, V2 = B, V3 = C;
4065 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4066 V1 = C, V2 = A, V3 = D;
4067 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4068 V1 = C, V2 = A, V3 = B;
4072 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
4073 return BinaryOperator::createAnd(V1, Or);
4078 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4079 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4080 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4081 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4082 SI0->getOperand(1) == SI1->getOperand(1) &&
4083 (SI0->hasOneUse() || SI1->hasOneUse())) {
4084 Instruction *NewOp =
4085 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4087 SI0->getName()), I);
4088 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4089 SI1->getOperand(1));
4093 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4094 if (A == Op1) // ~A | A == -1
4095 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4099 // Note, A is still live here!
4100 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4102 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4104 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4105 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4106 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4107 I.getName()+".demorgan"), I);
4108 return BinaryOperator::createNot(And);
4112 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4113 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4114 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4117 Value *LHSVal, *RHSVal;
4118 ConstantInt *LHSCst, *RHSCst;
4119 ICmpInst::Predicate LHSCC, RHSCC;
4120 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4121 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4122 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4123 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4124 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4125 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4126 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4127 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4128 // We can't fold (ugt x, C) | (sgt x, C2).
4129 PredicatesFoldable(LHSCC, RHSCC)) {
4130 // Ensure that the larger constant is on the RHS.
4131 ICmpInst *LHS = cast<ICmpInst>(Op0);
4133 if (ICmpInst::isSignedPredicate(LHSCC))
4134 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4136 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4139 std::swap(LHS, RHS);
4140 std::swap(LHSCst, RHSCst);
4141 std::swap(LHSCC, RHSCC);
4144 // At this point, we know we have have two icmp instructions
4145 // comparing a value against two constants and or'ing the result
4146 // together. Because of the above check, we know that we only have
4147 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4148 // FoldICmpLogical check above), that the two constants are not
4150 assert(LHSCst != RHSCst && "Compares not folded above?");
4153 default: assert(0 && "Unknown integer condition code!");
4154 case ICmpInst::ICMP_EQ:
4156 default: assert(0 && "Unknown integer condition code!");
4157 case ICmpInst::ICMP_EQ:
4158 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4159 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4160 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4161 LHSVal->getName()+".off");
4162 InsertNewInstBefore(Add, I);
4163 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4164 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4166 break; // (X == 13 | X == 15) -> no change
4167 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4168 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4170 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4171 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4172 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4173 return ReplaceInstUsesWith(I, RHS);
4176 case ICmpInst::ICMP_NE:
4178 default: assert(0 && "Unknown integer condition code!");
4179 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4180 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4181 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4182 return ReplaceInstUsesWith(I, LHS);
4183 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4184 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4185 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4186 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4189 case ICmpInst::ICMP_ULT:
4191 default: assert(0 && "Unknown integer condition code!");
4192 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4194 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4195 // If RHSCst is [us]MAXINT, it is always false. Not handling
4196 // this can cause overflow.
4197 if (RHSCst->isMaxValue(false))
4198 return ReplaceInstUsesWith(I, LHS);
4199 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4201 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4203 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4204 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4205 return ReplaceInstUsesWith(I, RHS);
4206 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4210 case ICmpInst::ICMP_SLT:
4212 default: assert(0 && "Unknown integer condition code!");
4213 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4215 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4216 // If RHSCst is [us]MAXINT, it is always false. Not handling
4217 // this can cause overflow.
4218 if (RHSCst->isMaxValue(true))
4219 return ReplaceInstUsesWith(I, LHS);
4220 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4222 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4224 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4225 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4226 return ReplaceInstUsesWith(I, RHS);
4227 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4231 case ICmpInst::ICMP_UGT:
4233 default: assert(0 && "Unknown integer condition code!");
4234 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4235 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4236 return ReplaceInstUsesWith(I, LHS);
4237 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4239 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4240 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4241 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4242 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4246 case ICmpInst::ICMP_SGT:
4248 default: assert(0 && "Unknown integer condition code!");
4249 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4250 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4251 return ReplaceInstUsesWith(I, LHS);
4252 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4254 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4255 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4256 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4257 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4265 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4266 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4267 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4268 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4269 const Type *SrcTy = Op0C->getOperand(0)->getType();
4270 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4271 // Only do this if the casts both really cause code to be generated.
4272 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4274 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4276 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4277 Op1C->getOperand(0),
4279 InsertNewInstBefore(NewOp, I);
4280 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4286 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4287 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4288 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4289 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4290 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4291 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4292 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4293 // If either of the constants are nans, then the whole thing returns
4295 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4296 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4298 // Otherwise, no need to compare the two constants, compare the
4300 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4301 RHS->getOperand(0));
4306 return Changed ? &I : 0;
4309 // XorSelf - Implements: X ^ X --> 0
4312 XorSelf(Value *rhs) : RHS(rhs) {}
4313 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4314 Instruction *apply(BinaryOperator &Xor) const {
4320 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4321 bool Changed = SimplifyCommutative(I);
4322 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4324 if (isa<UndefValue>(Op1))
4325 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4327 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4328 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4329 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4330 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4333 // See if we can simplify any instructions used by the instruction whose sole
4334 // purpose is to compute bits we don't care about.
4335 if (!isa<VectorType>(I.getType())) {
4336 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4337 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4338 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4339 KnownZero, KnownOne))
4341 } else if (isa<ConstantAggregateZero>(Op1)) {
4342 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4345 // Is this a ~ operation?
4346 if (Value *NotOp = dyn_castNotVal(&I)) {
4347 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4348 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4349 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4350 if (Op0I->getOpcode() == Instruction::And ||
4351 Op0I->getOpcode() == Instruction::Or) {
4352 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4353 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4355 BinaryOperator::createNot(Op0I->getOperand(1),
4356 Op0I->getOperand(1)->getName()+".not");
4357 InsertNewInstBefore(NotY, I);
4358 if (Op0I->getOpcode() == Instruction::And)
4359 return BinaryOperator::createOr(Op0NotVal, NotY);
4361 return BinaryOperator::createAnd(Op0NotVal, NotY);
4368 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4369 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4370 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4371 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4372 return new ICmpInst(ICI->getInversePredicate(),
4373 ICI->getOperand(0), ICI->getOperand(1));
4375 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4376 return new FCmpInst(FCI->getInversePredicate(),
4377 FCI->getOperand(0), FCI->getOperand(1));
4380 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4381 // ~(c-X) == X-c-1 == X+(-c-1)
4382 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4383 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4384 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4385 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4386 ConstantInt::get(I.getType(), 1));
4387 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4390 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4391 if (Op0I->getOpcode() == Instruction::Add) {
4392 // ~(X-c) --> (-c-1)-X
4393 if (RHS->isAllOnesValue()) {
4394 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4395 return BinaryOperator::createSub(
4396 ConstantExpr::getSub(NegOp0CI,
4397 ConstantInt::get(I.getType(), 1)),
4398 Op0I->getOperand(0));
4399 } else if (RHS->getValue().isSignBit()) {
4400 // (X + C) ^ signbit -> (X + C + signbit)
4401 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4402 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4405 } else if (Op0I->getOpcode() == Instruction::Or) {
4406 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4407 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4408 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4409 // Anything in both C1 and C2 is known to be zero, remove it from
4411 Constant *CommonBits = And(Op0CI, RHS);
4412 NewRHS = ConstantExpr::getAnd(NewRHS,
4413 ConstantExpr::getNot(CommonBits));
4414 AddToWorkList(Op0I);
4415 I.setOperand(0, Op0I->getOperand(0));
4416 I.setOperand(1, NewRHS);
4423 // Try to fold constant and into select arguments.
4424 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4425 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4427 if (isa<PHINode>(Op0))
4428 if (Instruction *NV = FoldOpIntoPhi(I))
4432 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4434 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4436 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4438 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4441 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4444 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4445 if (A == Op0) { // B^(B|A) == (A|B)^B
4446 Op1I->swapOperands();
4448 std::swap(Op0, Op1);
4449 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4450 I.swapOperands(); // Simplified below.
4451 std::swap(Op0, Op1);
4453 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4454 if (Op0 == A) // A^(A^B) == B
4455 return ReplaceInstUsesWith(I, B);
4456 else if (Op0 == B) // A^(B^A) == B
4457 return ReplaceInstUsesWith(I, A);
4458 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4459 if (A == Op0) { // A^(A&B) -> A^(B&A)
4460 Op1I->swapOperands();
4463 if (B == Op0) { // A^(B&A) -> (B&A)^A
4464 I.swapOperands(); // Simplified below.
4465 std::swap(Op0, Op1);
4470 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4473 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4474 if (A == Op1) // (B|A)^B == (A|B)^B
4476 if (B == Op1) { // (A|B)^B == A & ~B
4478 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4479 return BinaryOperator::createAnd(A, NotB);
4481 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4482 if (Op1 == A) // (A^B)^A == B
4483 return ReplaceInstUsesWith(I, B);
4484 else if (Op1 == B) // (B^A)^A == B
4485 return ReplaceInstUsesWith(I, A);
4486 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4487 if (A == Op1) // (A&B)^A -> (B&A)^A
4489 if (B == Op1 && // (B&A)^A == ~B & A
4490 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4492 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4493 return BinaryOperator::createAnd(N, Op1);
4498 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4499 if (Op0I && Op1I && Op0I->isShift() &&
4500 Op0I->getOpcode() == Op1I->getOpcode() &&
4501 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4502 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4503 Instruction *NewOp =
4504 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4505 Op1I->getOperand(0),
4506 Op0I->getName()), I);
4507 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4508 Op1I->getOperand(1));
4512 Value *A, *B, *C, *D;
4513 // (A & B)^(A | B) -> A ^ B
4514 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4515 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4516 if ((A == C && B == D) || (A == D && B == C))
4517 return BinaryOperator::createXor(A, B);
4519 // (A | B)^(A & B) -> A ^ B
4520 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4521 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4522 if ((A == C && B == D) || (A == D && B == C))
4523 return BinaryOperator::createXor(A, B);
4527 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4528 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4529 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4530 // (X & Y)^(X & Y) -> (Y^Z) & X
4531 Value *X = 0, *Y = 0, *Z = 0;
4533 X = A, Y = B, Z = D;
4535 X = A, Y = B, Z = C;
4537 X = B, Y = A, Z = D;
4539 X = B, Y = A, Z = C;
4542 Instruction *NewOp =
4543 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4544 return BinaryOperator::createAnd(NewOp, X);
4549 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4550 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4551 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4554 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4555 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4556 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4557 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4558 const Type *SrcTy = Op0C->getOperand(0)->getType();
4559 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4560 // Only do this if the casts both really cause code to be generated.
4561 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4563 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4565 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4566 Op1C->getOperand(0),
4568 InsertNewInstBefore(NewOp, I);
4569 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4573 return Changed ? &I : 0;
4576 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4577 /// overflowed for this type.
4578 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4579 ConstantInt *In2, bool IsSigned = false) {
4580 Result = cast<ConstantInt>(Add(In1, In2));
4583 if (In2->getValue().isNegative())
4584 return Result->getValue().sgt(In1->getValue());
4586 return Result->getValue().slt(In1->getValue());
4588 return Result->getValue().ult(In1->getValue());
4591 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4592 /// code necessary to compute the offset from the base pointer (without adding
4593 /// in the base pointer). Return the result as a signed integer of intptr size.
4594 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4595 TargetData &TD = IC.getTargetData();
4596 gep_type_iterator GTI = gep_type_begin(GEP);
4597 const Type *IntPtrTy = TD.getIntPtrType();
4598 Value *Result = Constant::getNullValue(IntPtrTy);
4600 // Build a mask for high order bits.
4601 unsigned IntPtrWidth = TD.getPointerSize()*8;
4602 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4604 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4605 Value *Op = GEP->getOperand(i);
4606 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4607 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4608 if (OpC->isZero()) continue;
4610 // Handle a struct index, which adds its field offset to the pointer.
4611 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4612 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4614 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4615 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4617 Result = IC.InsertNewInstBefore(
4618 BinaryOperator::createAdd(Result,
4619 ConstantInt::get(IntPtrTy, Size),
4620 GEP->getName()+".offs"), I);
4624 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4625 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4626 Scale = ConstantExpr::getMul(OC, Scale);
4627 if (Constant *RC = dyn_cast<Constant>(Result))
4628 Result = ConstantExpr::getAdd(RC, Scale);
4630 // Emit an add instruction.
4631 Result = IC.InsertNewInstBefore(
4632 BinaryOperator::createAdd(Result, Scale,
4633 GEP->getName()+".offs"), I);
4637 // Convert to correct type.
4638 if (Op->getType() != IntPtrTy) {
4639 if (Constant *OpC = dyn_cast<Constant>(Op))
4640 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4642 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4643 Op->getName()+".c"), I);
4646 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4647 if (Constant *OpC = dyn_cast<Constant>(Op))
4648 Op = ConstantExpr::getMul(OpC, Scale);
4649 else // We'll let instcombine(mul) convert this to a shl if possible.
4650 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4651 GEP->getName()+".idx"), I);
4654 // Emit an add instruction.
4655 if (isa<Constant>(Op) && isa<Constant>(Result))
4656 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4657 cast<Constant>(Result));
4659 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4660 GEP->getName()+".offs"), I);
4665 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4666 /// else. At this point we know that the GEP is on the LHS of the comparison.
4667 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4668 ICmpInst::Predicate Cond,
4670 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4672 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4673 if (isa<PointerType>(CI->getOperand(0)->getType()))
4674 RHS = CI->getOperand(0);
4676 Value *PtrBase = GEPLHS->getOperand(0);
4677 if (PtrBase == RHS) {
4678 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4679 // This transformation is valid because we know pointers can't overflow.
4680 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4681 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4682 Constant::getNullValue(Offset->getType()));
4683 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4684 // If the base pointers are different, but the indices are the same, just
4685 // compare the base pointer.
4686 if (PtrBase != GEPRHS->getOperand(0)) {
4687 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4688 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4689 GEPRHS->getOperand(0)->getType();
4691 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4692 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4693 IndicesTheSame = false;
4697 // If all indices are the same, just compare the base pointers.
4699 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4700 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4702 // Otherwise, the base pointers are different and the indices are
4703 // different, bail out.
4707 // If one of the GEPs has all zero indices, recurse.
4708 bool AllZeros = true;
4709 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4710 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4711 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4716 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4717 ICmpInst::getSwappedPredicate(Cond), I);
4719 // If the other GEP has all zero indices, recurse.
4721 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4722 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4723 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4728 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4730 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4731 // If the GEPs only differ by one index, compare it.
4732 unsigned NumDifferences = 0; // Keep track of # differences.
4733 unsigned DiffOperand = 0; // The operand that differs.
4734 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4735 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4736 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4737 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4738 // Irreconcilable differences.
4742 if (NumDifferences++) break;
4747 if (NumDifferences == 0) // SAME GEP?
4748 return ReplaceInstUsesWith(I, // No comparison is needed here.
4749 ConstantInt::get(Type::Int1Ty,
4750 isTrueWhenEqual(Cond)));
4752 else if (NumDifferences == 1) {
4753 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4754 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4755 // Make sure we do a signed comparison here.
4756 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4760 // Only lower this if the icmp is the only user of the GEP or if we expect
4761 // the result to fold to a constant!
4762 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4763 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4764 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4765 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4766 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4767 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4773 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4774 bool Changed = SimplifyCompare(I);
4775 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4777 // Fold trivial predicates.
4778 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4779 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4780 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4781 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4783 // Simplify 'fcmp pred X, X'
4785 switch (I.getPredicate()) {
4786 default: assert(0 && "Unknown predicate!");
4787 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4788 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4789 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4790 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4791 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4792 case FCmpInst::FCMP_OLT: // True if ordered and less than
4793 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4794 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4796 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4797 case FCmpInst::FCMP_ULT: // True if unordered or less than
4798 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4799 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4800 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4801 I.setPredicate(FCmpInst::FCMP_UNO);
4802 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4805 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4806 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4807 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4808 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4809 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4810 I.setPredicate(FCmpInst::FCMP_ORD);
4811 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4816 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4817 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4819 // Handle fcmp with constant RHS
4820 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4821 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4822 switch (LHSI->getOpcode()) {
4823 case Instruction::PHI:
4824 if (Instruction *NV = FoldOpIntoPhi(I))
4827 case Instruction::Select:
4828 // If either operand of the select is a constant, we can fold the
4829 // comparison into the select arms, which will cause one to be
4830 // constant folded and the select turned into a bitwise or.
4831 Value *Op1 = 0, *Op2 = 0;
4832 if (LHSI->hasOneUse()) {
4833 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4834 // Fold the known value into the constant operand.
4835 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4836 // Insert a new FCmp of the other select operand.
4837 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4838 LHSI->getOperand(2), RHSC,
4840 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4841 // Fold the known value into the constant operand.
4842 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4843 // Insert a new FCmp of the other select operand.
4844 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4845 LHSI->getOperand(1), RHSC,
4851 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4856 return Changed ? &I : 0;
4859 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4860 bool Changed = SimplifyCompare(I);
4861 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4862 const Type *Ty = Op0->getType();
4866 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4867 isTrueWhenEqual(I)));
4869 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4870 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4872 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4873 // addresses never equal each other! We already know that Op0 != Op1.
4874 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4875 isa<ConstantPointerNull>(Op0)) &&
4876 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4877 isa<ConstantPointerNull>(Op1)))
4878 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4879 !isTrueWhenEqual(I)));
4881 // icmp's with boolean values can always be turned into bitwise operations
4882 if (Ty == Type::Int1Ty) {
4883 switch (I.getPredicate()) {
4884 default: assert(0 && "Invalid icmp instruction!");
4885 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4886 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4887 InsertNewInstBefore(Xor, I);
4888 return BinaryOperator::createNot(Xor);
4890 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4891 return BinaryOperator::createXor(Op0, Op1);
4893 case ICmpInst::ICMP_UGT:
4894 case ICmpInst::ICMP_SGT:
4895 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4897 case ICmpInst::ICMP_ULT:
4898 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4899 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4900 InsertNewInstBefore(Not, I);
4901 return BinaryOperator::createAnd(Not, Op1);
4903 case ICmpInst::ICMP_UGE:
4904 case ICmpInst::ICMP_SGE:
4905 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4907 case ICmpInst::ICMP_ULE:
4908 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4909 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4910 InsertNewInstBefore(Not, I);
4911 return BinaryOperator::createOr(Not, Op1);
4916 // See if we are doing a comparison between a constant and an instruction that
4917 // can be folded into the comparison.
4918 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4921 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
4922 if (I.isEquality() && CI->isNullValue() &&
4923 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
4924 // (icmp cond A B) if cond is equality
4925 return new ICmpInst(I.getPredicate(), A, B);
4928 switch (I.getPredicate()) {
4930 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4931 if (CI->isMinValue(false))
4932 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4933 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4934 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4935 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4936 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4937 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4938 if (CI->isMinValue(true))
4939 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4940 ConstantInt::getAllOnesValue(Op0->getType()));
4944 case ICmpInst::ICMP_SLT:
4945 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4946 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4947 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4948 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4949 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4950 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4953 case ICmpInst::ICMP_UGT:
4954 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4955 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4956 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4957 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4958 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4959 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4961 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4962 if (CI->isMaxValue(true))
4963 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4964 ConstantInt::getNullValue(Op0->getType()));
4967 case ICmpInst::ICMP_SGT:
4968 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4969 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4970 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4971 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4972 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4973 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4976 case ICmpInst::ICMP_ULE:
4977 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4978 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4979 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4980 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4981 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4982 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4985 case ICmpInst::ICMP_SLE:
4986 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4987 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4988 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4989 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4990 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4991 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4994 case ICmpInst::ICMP_UGE:
4995 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4996 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4997 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4998 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4999 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5000 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5003 case ICmpInst::ICMP_SGE:
5004 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5005 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5006 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5007 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5008 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5009 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5013 // If we still have a icmp le or icmp ge instruction, turn it into the
5014 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5015 // already been handled above, this requires little checking.
5017 switch (I.getPredicate()) {
5019 case ICmpInst::ICMP_ULE:
5020 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5021 case ICmpInst::ICMP_SLE:
5022 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5023 case ICmpInst::ICMP_UGE:
5024 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5025 case ICmpInst::ICMP_SGE:
5026 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5029 // See if we can fold the comparison based on bits known to be zero or one
5030 // in the input. If this comparison is a normal comparison, it demands all
5031 // bits, if it is a sign bit comparison, it only demands the sign bit.
5034 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5036 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5037 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5038 if (SimplifyDemandedBits(Op0,
5039 isSignBit ? APInt::getSignBit(BitWidth)
5040 : APInt::getAllOnesValue(BitWidth),
5041 KnownZero, KnownOne, 0))
5044 // Given the known and unknown bits, compute a range that the LHS could be
5046 if ((KnownOne | KnownZero) != 0) {
5047 // Compute the Min, Max and RHS values based on the known bits. For the
5048 // EQ and NE we use unsigned values.
5049 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5050 const APInt& RHSVal = CI->getValue();
5051 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5052 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5055 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5058 switch (I.getPredicate()) { // LE/GE have been folded already.
5059 default: assert(0 && "Unknown icmp opcode!");
5060 case ICmpInst::ICMP_EQ:
5061 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5062 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5064 case ICmpInst::ICMP_NE:
5065 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5066 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5068 case ICmpInst::ICMP_ULT:
5069 if (Max.ult(RHSVal))
5070 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5071 if (Min.uge(RHSVal))
5072 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5074 case ICmpInst::ICMP_UGT:
5075 if (Min.ugt(RHSVal))
5076 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5077 if (Max.ule(RHSVal))
5078 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5080 case ICmpInst::ICMP_SLT:
5081 if (Max.slt(RHSVal))
5082 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5083 if (Min.sgt(RHSVal))
5084 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5086 case ICmpInst::ICMP_SGT:
5087 if (Min.sgt(RHSVal))
5088 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5089 if (Max.sle(RHSVal))
5090 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5095 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5096 // instruction, see if that instruction also has constants so that the
5097 // instruction can be folded into the icmp
5098 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5099 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5103 // Handle icmp with constant (but not simple integer constant) RHS
5104 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5105 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5106 switch (LHSI->getOpcode()) {
5107 case Instruction::GetElementPtr:
5108 if (RHSC->isNullValue()) {
5109 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5110 bool isAllZeros = true;
5111 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5112 if (!isa<Constant>(LHSI->getOperand(i)) ||
5113 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5118 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5119 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5123 case Instruction::PHI:
5124 if (Instruction *NV = FoldOpIntoPhi(I))
5127 case Instruction::Select: {
5128 // If either operand of the select is a constant, we can fold the
5129 // comparison into the select arms, which will cause one to be
5130 // constant folded and the select turned into a bitwise or.
5131 Value *Op1 = 0, *Op2 = 0;
5132 if (LHSI->hasOneUse()) {
5133 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5134 // Fold the known value into the constant operand.
5135 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5136 // Insert a new ICmp of the other select operand.
5137 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5138 LHSI->getOperand(2), RHSC,
5140 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5141 // Fold the known value into the constant operand.
5142 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5143 // Insert a new ICmp of the other select operand.
5144 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5145 LHSI->getOperand(1), RHSC,
5151 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5154 case Instruction::Malloc:
5155 // If we have (malloc != null), and if the malloc has a single use, we
5156 // can assume it is successful and remove the malloc.
5157 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5158 AddToWorkList(LHSI);
5159 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5160 !isTrueWhenEqual(I)));
5166 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5167 if (User *GEP = dyn_castGetElementPtr(Op0))
5168 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5170 if (User *GEP = dyn_castGetElementPtr(Op1))
5171 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5172 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5175 // Test to see if the operands of the icmp are casted versions of other
5176 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5178 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5179 if (isa<PointerType>(Op0->getType()) &&
5180 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5181 // We keep moving the cast from the left operand over to the right
5182 // operand, where it can often be eliminated completely.
5183 Op0 = CI->getOperand(0);
5185 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5186 // so eliminate it as well.
5187 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5188 Op1 = CI2->getOperand(0);
5190 // If Op1 is a constant, we can fold the cast into the constant.
5191 if (Op0->getType() != Op1->getType()) {
5192 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5193 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5195 // Otherwise, cast the RHS right before the icmp
5196 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5199 return new ICmpInst(I.getPredicate(), Op0, Op1);
5203 if (isa<CastInst>(Op0)) {
5204 // Handle the special case of: icmp (cast bool to X), <cst>
5205 // This comes up when you have code like
5208 // For generality, we handle any zero-extension of any operand comparison
5209 // with a constant or another cast from the same type.
5210 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5211 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5215 if (I.isEquality()) {
5216 Value *A, *B, *C, *D;
5217 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5218 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5219 Value *OtherVal = A == Op1 ? B : A;
5220 return new ICmpInst(I.getPredicate(), OtherVal,
5221 Constant::getNullValue(A->getType()));
5224 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5225 // A^c1 == C^c2 --> A == C^(c1^c2)
5226 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5227 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5228 if (Op1->hasOneUse()) {
5229 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5230 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5231 return new ICmpInst(I.getPredicate(), A,
5232 InsertNewInstBefore(Xor, I));
5235 // A^B == A^D -> B == D
5236 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5237 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5238 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5239 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5243 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5244 (A == Op0 || B == Op0)) {
5245 // A == (A^B) -> B == 0
5246 Value *OtherVal = A == Op0 ? B : A;
5247 return new ICmpInst(I.getPredicate(), OtherVal,
5248 Constant::getNullValue(A->getType()));
5250 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5251 // (A-B) == A -> B == 0
5252 return new ICmpInst(I.getPredicate(), B,
5253 Constant::getNullValue(B->getType()));
5255 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5256 // A == (A-B) -> B == 0
5257 return new ICmpInst(I.getPredicate(), B,
5258 Constant::getNullValue(B->getType()));
5261 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5262 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5263 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5264 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5265 Value *X = 0, *Y = 0, *Z = 0;
5268 X = B; Y = D; Z = A;
5269 } else if (A == D) {
5270 X = B; Y = C; Z = A;
5271 } else if (B == C) {
5272 X = A; Y = D; Z = B;
5273 } else if (B == D) {
5274 X = A; Y = C; Z = B;
5277 if (X) { // Build (X^Y) & Z
5278 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5279 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5280 I.setOperand(0, Op1);
5281 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5286 return Changed ? &I : 0;
5290 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5291 /// and CmpRHS are both known to be integer constants.
5292 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5293 ConstantInt *DivRHS) {
5294 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5295 const APInt &CmpRHSV = CmpRHS->getValue();
5297 // FIXME: If the operand types don't match the type of the divide
5298 // then don't attempt this transform. The code below doesn't have the
5299 // logic to deal with a signed divide and an unsigned compare (and
5300 // vice versa). This is because (x /s C1) <s C2 produces different
5301 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5302 // (x /u C1) <u C2. Simply casting the operands and result won't
5303 // work. :( The if statement below tests that condition and bails
5305 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5306 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5308 if (DivRHS->isZero())
5309 return 0; // The ProdOV computation fails on divide by zero.
5311 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5312 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5313 // C2 (CI). By solving for X we can turn this into a range check
5314 // instead of computing a divide.
5315 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5317 // Determine if the product overflows by seeing if the product is
5318 // not equal to the divide. Make sure we do the same kind of divide
5319 // as in the LHS instruction that we're folding.
5320 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5321 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5323 // Get the ICmp opcode
5324 ICmpInst::Predicate Pred = ICI.getPredicate();
5326 // Figure out the interval that is being checked. For example, a comparison
5327 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5328 // Compute this interval based on the constants involved and the signedness of
5329 // the compare/divide. This computes a half-open interval, keeping track of
5330 // whether either value in the interval overflows. After analysis each
5331 // overflow variable is set to 0 if it's corresponding bound variable is valid
5332 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5333 int LoOverflow = 0, HiOverflow = 0;
5334 ConstantInt *LoBound = 0, *HiBound = 0;
5337 if (!DivIsSigned) { // udiv
5338 // e.g. X/5 op 3 --> [15, 20)
5340 HiOverflow = LoOverflow = ProdOV;
5342 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5343 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5344 if (CmpRHSV == 0) { // (X / pos) op 0
5345 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5346 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5348 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5349 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5350 HiOverflow = LoOverflow = ProdOV;
5352 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5353 } else { // (X / pos) op neg
5354 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5355 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5356 LoOverflow = AddWithOverflow(LoBound, Prod,
5357 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5358 HiBound = AddOne(Prod);
5359 HiOverflow = ProdOV ? -1 : 0;
5361 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5362 if (CmpRHSV == 0) { // (X / neg) op 0
5363 // e.g. X/-5 op 0 --> [-4, 5)
5364 LoBound = AddOne(DivRHS);
5365 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5366 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5367 HiOverflow = 1; // [INTMIN+1, overflow)
5368 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5370 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5371 // e.g. X/-5 op 3 --> [-19, -14)
5372 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5374 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5375 HiBound = AddOne(Prod);
5376 } else { // (X / neg) op neg
5377 // e.g. X/-5 op -3 --> [15, 20)
5379 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5380 HiBound = Subtract(Prod, DivRHS);
5383 // Dividing by a negative swaps the condition. LT <-> GT
5384 Pred = ICmpInst::getSwappedPredicate(Pred);
5387 Value *X = DivI->getOperand(0);
5389 default: assert(0 && "Unhandled icmp opcode!");
5390 case ICmpInst::ICMP_EQ:
5391 if (LoOverflow && HiOverflow)
5392 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5393 else if (HiOverflow)
5394 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5395 ICmpInst::ICMP_UGE, X, LoBound);
5396 else if (LoOverflow)
5397 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5398 ICmpInst::ICMP_ULT, X, HiBound);
5400 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5401 case ICmpInst::ICMP_NE:
5402 if (LoOverflow && HiOverflow)
5403 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5404 else if (HiOverflow)
5405 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5406 ICmpInst::ICMP_ULT, X, LoBound);
5407 else if (LoOverflow)
5408 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5409 ICmpInst::ICMP_UGE, X, HiBound);
5411 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5412 case ICmpInst::ICMP_ULT:
5413 case ICmpInst::ICMP_SLT:
5414 if (LoOverflow == +1) // Low bound is greater than input range.
5415 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5416 if (LoOverflow == -1) // Low bound is less than input range.
5417 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5418 return new ICmpInst(Pred, X, LoBound);
5419 case ICmpInst::ICMP_UGT:
5420 case ICmpInst::ICMP_SGT:
5421 if (HiOverflow == +1) // High bound greater than input range.
5422 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5423 else if (HiOverflow == -1) // High bound less than input range.
5424 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5425 if (Pred == ICmpInst::ICMP_UGT)
5426 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5428 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5433 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5435 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5438 const APInt &RHSV = RHS->getValue();
5440 switch (LHSI->getOpcode()) {
5441 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5442 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5443 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5445 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5446 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5447 Value *CompareVal = LHSI->getOperand(0);
5449 // If the sign bit of the XorCST is not set, there is no change to
5450 // the operation, just stop using the Xor.
5451 if (!XorCST->getValue().isNegative()) {
5452 ICI.setOperand(0, CompareVal);
5453 AddToWorkList(LHSI);
5457 // Was the old condition true if the operand is positive?
5458 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5460 // If so, the new one isn't.
5461 isTrueIfPositive ^= true;
5463 if (isTrueIfPositive)
5464 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5466 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5470 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5471 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5472 LHSI->getOperand(0)->hasOneUse()) {
5473 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5475 // If the LHS is an AND of a truncating cast, we can widen the
5476 // and/compare to be the input width without changing the value
5477 // produced, eliminating a cast.
5478 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5479 // We can do this transformation if either the AND constant does not
5480 // have its sign bit set or if it is an equality comparison.
5481 // Extending a relational comparison when we're checking the sign
5482 // bit would not work.
5483 if (Cast->hasOneUse() &&
5484 (ICI.isEquality() ||
5485 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5487 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5488 APInt NewCST = AndCST->getValue();
5489 NewCST.zext(BitWidth);
5491 NewCI.zext(BitWidth);
5492 Instruction *NewAnd =
5493 BinaryOperator::createAnd(Cast->getOperand(0),
5494 ConstantInt::get(NewCST),LHSI->getName());
5495 InsertNewInstBefore(NewAnd, ICI);
5496 return new ICmpInst(ICI.getPredicate(), NewAnd,
5497 ConstantInt::get(NewCI));
5501 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5502 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5503 // happens a LOT in code produced by the C front-end, for bitfield
5505 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5506 if (Shift && !Shift->isShift())
5510 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5511 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5512 const Type *AndTy = AndCST->getType(); // Type of the and.
5514 // We can fold this as long as we can't shift unknown bits
5515 // into the mask. This can only happen with signed shift
5516 // rights, as they sign-extend.
5518 bool CanFold = Shift->isLogicalShift();
5520 // To test for the bad case of the signed shr, see if any
5521 // of the bits shifted in could be tested after the mask.
5522 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5523 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5525 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5526 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5527 AndCST->getValue()) == 0)
5533 if (Shift->getOpcode() == Instruction::Shl)
5534 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5536 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5538 // Check to see if we are shifting out any of the bits being
5540 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5541 // If we shifted bits out, the fold is not going to work out.
5542 // As a special case, check to see if this means that the
5543 // result is always true or false now.
5544 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5545 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5546 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5547 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5549 ICI.setOperand(1, NewCst);
5550 Constant *NewAndCST;
5551 if (Shift->getOpcode() == Instruction::Shl)
5552 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5554 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5555 LHSI->setOperand(1, NewAndCST);
5556 LHSI->setOperand(0, Shift->getOperand(0));
5557 AddToWorkList(Shift); // Shift is dead.
5558 AddUsesToWorkList(ICI);
5564 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5565 // preferable because it allows the C<<Y expression to be hoisted out
5566 // of a loop if Y is invariant and X is not.
5567 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5568 ICI.isEquality() && !Shift->isArithmeticShift() &&
5569 isa<Instruction>(Shift->getOperand(0))) {
5572 if (Shift->getOpcode() == Instruction::LShr) {
5573 NS = BinaryOperator::createShl(AndCST,
5574 Shift->getOperand(1), "tmp");
5576 // Insert a logical shift.
5577 NS = BinaryOperator::createLShr(AndCST,
5578 Shift->getOperand(1), "tmp");
5580 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5582 // Compute X & (C << Y).
5583 Instruction *NewAnd =
5584 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5585 InsertNewInstBefore(NewAnd, ICI);
5587 ICI.setOperand(0, NewAnd);
5593 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5594 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5597 uint32_t TypeBits = RHSV.getBitWidth();
5599 // Check that the shift amount is in range. If not, don't perform
5600 // undefined shifts. When the shift is visited it will be
5602 if (ShAmt->uge(TypeBits))
5605 if (ICI.isEquality()) {
5606 // If we are comparing against bits always shifted out, the
5607 // comparison cannot succeed.
5609 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5610 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5611 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5612 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5613 return ReplaceInstUsesWith(ICI, Cst);
5616 if (LHSI->hasOneUse()) {
5617 // Otherwise strength reduce the shift into an and.
5618 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5620 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5623 BinaryOperator::createAnd(LHSI->getOperand(0),
5624 Mask, LHSI->getName()+".mask");
5625 Value *And = InsertNewInstBefore(AndI, ICI);
5626 return new ICmpInst(ICI.getPredicate(), And,
5627 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5631 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5632 bool TrueIfSigned = false;
5633 if (LHSI->hasOneUse() &&
5634 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5635 // (X << 31) <s 0 --> (X&1) != 0
5636 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5637 (TypeBits-ShAmt->getZExtValue()-1));
5639 BinaryOperator::createAnd(LHSI->getOperand(0),
5640 Mask, LHSI->getName()+".mask");
5641 Value *And = InsertNewInstBefore(AndI, ICI);
5643 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5644 And, Constant::getNullValue(And->getType()));
5649 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5650 case Instruction::AShr: {
5651 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5654 if (ICI.isEquality()) {
5655 // Check that the shift amount is in range. If not, don't perform
5656 // undefined shifts. When the shift is visited it will be
5658 uint32_t TypeBits = RHSV.getBitWidth();
5659 if (ShAmt->uge(TypeBits))
5661 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5663 // If we are comparing against bits always shifted out, the
5664 // comparison cannot succeed.
5665 APInt Comp = RHSV << ShAmtVal;
5666 if (LHSI->getOpcode() == Instruction::LShr)
5667 Comp = Comp.lshr(ShAmtVal);
5669 Comp = Comp.ashr(ShAmtVal);
5671 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5672 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5673 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5674 return ReplaceInstUsesWith(ICI, Cst);
5677 if (LHSI->hasOneUse() || RHSV == 0) {
5678 // Otherwise strength reduce the shift into an and.
5679 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5680 Constant *Mask = ConstantInt::get(Val);
5683 BinaryOperator::createAnd(LHSI->getOperand(0),
5684 Mask, LHSI->getName()+".mask");
5685 Value *And = InsertNewInstBefore(AndI, ICI);
5686 return new ICmpInst(ICI.getPredicate(), And,
5687 ConstantExpr::getShl(RHS, ShAmt));
5693 case Instruction::SDiv:
5694 case Instruction::UDiv:
5695 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5696 // Fold this div into the comparison, producing a range check.
5697 // Determine, based on the divide type, what the range is being
5698 // checked. If there is an overflow on the low or high side, remember
5699 // it, otherwise compute the range [low, hi) bounding the new value.
5700 // See: InsertRangeTest above for the kinds of replacements possible.
5701 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5702 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5707 case Instruction::Add:
5708 // Fold: icmp pred (add, X, C1), C2
5710 if (!ICI.isEquality()) {
5711 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5713 const APInt &LHSV = LHSC->getValue();
5715 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
5718 if (ICI.isSignedPredicate()) {
5719 if (CR.getLower().isSignBit()) {
5720 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
5721 ConstantInt::get(CR.getUpper()));
5722 } else if (CR.getUpper().isSignBit()) {
5723 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
5724 ConstantInt::get(CR.getLower()));
5727 if (CR.getLower().isMinValue()) {
5728 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
5729 ConstantInt::get(CR.getUpper()));
5730 } else if (CR.getUpper().isMinValue()) {
5731 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
5732 ConstantInt::get(CR.getLower()));
5739 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5740 if (ICI.isEquality()) {
5741 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5743 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5744 // the second operand is a constant, simplify a bit.
5745 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5746 switch (BO->getOpcode()) {
5747 case Instruction::SRem:
5748 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5749 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5750 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5751 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5752 Instruction *NewRem =
5753 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5755 InsertNewInstBefore(NewRem, ICI);
5756 return new ICmpInst(ICI.getPredicate(), NewRem,
5757 Constant::getNullValue(BO->getType()));
5761 case Instruction::Add:
5762 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5763 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5764 if (BO->hasOneUse())
5765 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5766 Subtract(RHS, BOp1C));
5767 } else if (RHSV == 0) {
5768 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5769 // efficiently invertible, or if the add has just this one use.
5770 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5772 if (Value *NegVal = dyn_castNegVal(BOp1))
5773 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5774 else if (Value *NegVal = dyn_castNegVal(BOp0))
5775 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5776 else if (BO->hasOneUse()) {
5777 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5778 InsertNewInstBefore(Neg, ICI);
5780 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5784 case Instruction::Xor:
5785 // For the xor case, we can xor two constants together, eliminating
5786 // the explicit xor.
5787 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5788 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5789 ConstantExpr::getXor(RHS, BOC));
5792 case Instruction::Sub:
5793 // Replace (([sub|xor] A, B) != 0) with (A != B)
5795 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5799 case Instruction::Or:
5800 // If bits are being or'd in that are not present in the constant we
5801 // are comparing against, then the comparison could never succeed!
5802 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5803 Constant *NotCI = ConstantExpr::getNot(RHS);
5804 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5805 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5810 case Instruction::And:
5811 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5812 // If bits are being compared against that are and'd out, then the
5813 // comparison can never succeed!
5814 if ((RHSV & ~BOC->getValue()) != 0)
5815 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5818 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5819 if (RHS == BOC && RHSV.isPowerOf2())
5820 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5821 ICmpInst::ICMP_NE, LHSI,
5822 Constant::getNullValue(RHS->getType()));
5824 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5825 if (isSignBit(BOC)) {
5826 Value *X = BO->getOperand(0);
5827 Constant *Zero = Constant::getNullValue(X->getType());
5828 ICmpInst::Predicate pred = isICMP_NE ?
5829 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5830 return new ICmpInst(pred, X, Zero);
5833 // ((X & ~7) == 0) --> X < 8
5834 if (RHSV == 0 && isHighOnes(BOC)) {
5835 Value *X = BO->getOperand(0);
5836 Constant *NegX = ConstantExpr::getNeg(BOC);
5837 ICmpInst::Predicate pred = isICMP_NE ?
5838 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5839 return new ICmpInst(pred, X, NegX);
5844 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5845 // Handle icmp {eq|ne} <intrinsic>, intcst.
5846 if (II->getIntrinsicID() == Intrinsic::bswap) {
5848 ICI.setOperand(0, II->getOperand(1));
5849 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5853 } else { // Not a ICMP_EQ/ICMP_NE
5854 // If the LHS is a cast from an integral value of the same size,
5855 // then since we know the RHS is a constant, try to simlify.
5856 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5857 Value *CastOp = Cast->getOperand(0);
5858 const Type *SrcTy = CastOp->getType();
5859 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5860 if (SrcTy->isInteger() &&
5861 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5862 // If this is an unsigned comparison, try to make the comparison use
5863 // smaller constant values.
5864 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5865 // X u< 128 => X s> -1
5866 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5867 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5868 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5869 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5870 // X u> 127 => X s< 0
5871 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5872 Constant::getNullValue(SrcTy));
5880 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5881 /// We only handle extending casts so far.
5883 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5884 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5885 Value *LHSCIOp = LHSCI->getOperand(0);
5886 const Type *SrcTy = LHSCIOp->getType();
5887 const Type *DestTy = LHSCI->getType();
5890 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5891 // integer type is the same size as the pointer type.
5892 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5893 getTargetData().getPointerSizeInBits() ==
5894 cast<IntegerType>(DestTy)->getBitWidth()) {
5896 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5897 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5898 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5899 RHSOp = RHSC->getOperand(0);
5900 // If the pointer types don't match, insert a bitcast.
5901 if (LHSCIOp->getType() != RHSOp->getType())
5902 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
5906 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5909 // The code below only handles extension cast instructions, so far.
5911 if (LHSCI->getOpcode() != Instruction::ZExt &&
5912 LHSCI->getOpcode() != Instruction::SExt)
5915 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5916 bool isSignedCmp = ICI.isSignedPredicate();
5918 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5919 // Not an extension from the same type?
5920 RHSCIOp = CI->getOperand(0);
5921 if (RHSCIOp->getType() != LHSCIOp->getType())
5924 // If the signedness of the two casts doesn't agree (i.e. one is a sext
5925 // and the other is a zext), then we can't handle this.
5926 if (CI->getOpcode() != LHSCI->getOpcode())
5929 // Deal with equality cases early.
5930 if (ICI.isEquality())
5931 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5933 // A signed comparison of sign extended values simplifies into a
5934 // signed comparison.
5935 if (isSignedCmp && isSignedExt)
5936 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5938 // The other three cases all fold into an unsigned comparison.
5939 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
5942 // If we aren't dealing with a constant on the RHS, exit early
5943 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5947 // Compute the constant that would happen if we truncated to SrcTy then
5948 // reextended to DestTy.
5949 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5950 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5952 // If the re-extended constant didn't change...
5954 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5955 // For example, we might have:
5956 // %A = sext short %X to uint
5957 // %B = icmp ugt uint %A, 1330
5958 // It is incorrect to transform this into
5959 // %B = icmp ugt short %X, 1330
5960 // because %A may have negative value.
5962 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5963 // OR operation is EQ/NE.
5964 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5965 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5970 // The re-extended constant changed so the constant cannot be represented
5971 // in the shorter type. Consequently, we cannot emit a simple comparison.
5973 // First, handle some easy cases. We know the result cannot be equal at this
5974 // point so handle the ICI.isEquality() cases
5975 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5976 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5977 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5978 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5980 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5981 // should have been folded away previously and not enter in here.
5984 // We're performing a signed comparison.
5985 if (cast<ConstantInt>(CI)->getValue().isNegative())
5986 Result = ConstantInt::getFalse(); // X < (small) --> false
5988 Result = ConstantInt::getTrue(); // X < (large) --> true
5990 // We're performing an unsigned comparison.
5992 // We're performing an unsigned comp with a sign extended value.
5993 // This is true if the input is >= 0. [aka >s -1]
5994 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5995 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5996 NegOne, ICI.getName()), ICI);
5998 // Unsigned extend & unsigned compare -> always true.
5999 Result = ConstantInt::getTrue();
6003 // Finally, return the value computed.
6004 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6005 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6006 return ReplaceInstUsesWith(ICI, Result);
6008 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6009 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6010 "ICmp should be folded!");
6011 if (Constant *CI = dyn_cast<Constant>(Result))
6012 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6014 return BinaryOperator::createNot(Result);
6018 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6019 return commonShiftTransforms(I);
6022 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6023 return commonShiftTransforms(I);
6026 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6027 if (Instruction *R = commonShiftTransforms(I))
6030 Value *Op0 = I.getOperand(0);
6032 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6033 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6034 if (CSI->isAllOnesValue())
6035 return ReplaceInstUsesWith(I, CSI);
6037 // See if we can turn a signed shr into an unsigned shr.
6038 if (MaskedValueIsZero(Op0,
6039 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6040 return BinaryOperator::createLShr(Op0, I.getOperand(1));
6045 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6046 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6047 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6049 // shl X, 0 == X and shr X, 0 == X
6050 // shl 0, X == 0 and shr 0, X == 0
6051 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6052 Op0 == Constant::getNullValue(Op0->getType()))
6053 return ReplaceInstUsesWith(I, Op0);
6055 if (isa<UndefValue>(Op0)) {
6056 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6057 return ReplaceInstUsesWith(I, Op0);
6058 else // undef << X -> 0, undef >>u X -> 0
6059 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6061 if (isa<UndefValue>(Op1)) {
6062 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6063 return ReplaceInstUsesWith(I, Op0);
6064 else // X << undef, X >>u undef -> 0
6065 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6068 // Try to fold constant and into select arguments.
6069 if (isa<Constant>(Op0))
6070 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6071 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6074 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6075 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6080 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6081 BinaryOperator &I) {
6082 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6084 // See if we can simplify any instructions used by the instruction whose sole
6085 // purpose is to compute bits we don't care about.
6086 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6087 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6088 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6089 KnownZero, KnownOne))
6092 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6093 // of a signed value.
6095 if (Op1->uge(TypeBits)) {
6096 if (I.getOpcode() != Instruction::AShr)
6097 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6099 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6104 // ((X*C1) << C2) == (X * (C1 << C2))
6105 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6106 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6107 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6108 return BinaryOperator::createMul(BO->getOperand(0),
6109 ConstantExpr::getShl(BOOp, Op1));
6111 // Try to fold constant and into select arguments.
6112 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6113 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6115 if (isa<PHINode>(Op0))
6116 if (Instruction *NV = FoldOpIntoPhi(I))
6119 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6120 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6121 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6122 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6123 // place. Don't try to do this transformation in this case. Also, we
6124 // require that the input operand is a shift-by-constant so that we have
6125 // confidence that the shifts will get folded together. We could do this
6126 // xform in more cases, but it is unlikely to be profitable.
6127 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6128 isa<ConstantInt>(TrOp->getOperand(1))) {
6129 // Okay, we'll do this xform. Make the shift of shift.
6130 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6131 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6133 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6135 // For logical shifts, the truncation has the effect of making the high
6136 // part of the register be zeros. Emulate this by inserting an AND to
6137 // clear the top bits as needed. This 'and' will usually be zapped by
6138 // other xforms later if dead.
6139 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6140 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6141 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6143 // The mask we constructed says what the trunc would do if occurring
6144 // between the shifts. We want to know the effect *after* the second
6145 // shift. We know that it is a logical shift by a constant, so adjust the
6146 // mask as appropriate.
6147 if (I.getOpcode() == Instruction::Shl)
6148 MaskV <<= Op1->getZExtValue();
6150 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6151 MaskV = MaskV.lshr(Op1->getZExtValue());
6154 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6156 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6158 // Return the value truncated to the interesting size.
6159 return new TruncInst(And, I.getType());
6163 if (Op0->hasOneUse()) {
6164 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6165 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6168 switch (Op0BO->getOpcode()) {
6170 case Instruction::Add:
6171 case Instruction::And:
6172 case Instruction::Or:
6173 case Instruction::Xor: {
6174 // These operators commute.
6175 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6176 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6177 match(Op0BO->getOperand(1),
6178 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6179 Instruction *YS = BinaryOperator::createShl(
6180 Op0BO->getOperand(0), Op1,
6182 InsertNewInstBefore(YS, I); // (Y << C)
6184 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6185 Op0BO->getOperand(1)->getName());
6186 InsertNewInstBefore(X, I); // (X + (Y << C))
6187 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6188 return BinaryOperator::createAnd(X, ConstantInt::get(
6189 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6192 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6193 Value *Op0BOOp1 = Op0BO->getOperand(1);
6194 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6196 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6197 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6199 Instruction *YS = BinaryOperator::createShl(
6200 Op0BO->getOperand(0), Op1,
6202 InsertNewInstBefore(YS, I); // (Y << C)
6204 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6205 V1->getName()+".mask");
6206 InsertNewInstBefore(XM, I); // X & (CC << C)
6208 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6213 case Instruction::Sub: {
6214 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6215 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6216 match(Op0BO->getOperand(0),
6217 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6218 Instruction *YS = BinaryOperator::createShl(
6219 Op0BO->getOperand(1), Op1,
6221 InsertNewInstBefore(YS, I); // (Y << C)
6223 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6224 Op0BO->getOperand(0)->getName());
6225 InsertNewInstBefore(X, I); // (X + (Y << C))
6226 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6227 return BinaryOperator::createAnd(X, ConstantInt::get(
6228 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6231 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6232 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6233 match(Op0BO->getOperand(0),
6234 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6235 m_ConstantInt(CC))) && V2 == Op1 &&
6236 cast<BinaryOperator>(Op0BO->getOperand(0))
6237 ->getOperand(0)->hasOneUse()) {
6238 Instruction *YS = BinaryOperator::createShl(
6239 Op0BO->getOperand(1), Op1,
6241 InsertNewInstBefore(YS, I); // (Y << C)
6243 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6244 V1->getName()+".mask");
6245 InsertNewInstBefore(XM, I); // X & (CC << C)
6247 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6255 // If the operand is an bitwise operator with a constant RHS, and the
6256 // shift is the only use, we can pull it out of the shift.
6257 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6258 bool isValid = true; // Valid only for And, Or, Xor
6259 bool highBitSet = false; // Transform if high bit of constant set?
6261 switch (Op0BO->getOpcode()) {
6262 default: isValid = false; break; // Do not perform transform!
6263 case Instruction::Add:
6264 isValid = isLeftShift;
6266 case Instruction::Or:
6267 case Instruction::Xor:
6270 case Instruction::And:
6275 // If this is a signed shift right, and the high bit is modified
6276 // by the logical operation, do not perform the transformation.
6277 // The highBitSet boolean indicates the value of the high bit of
6278 // the constant which would cause it to be modified for this
6281 if (isValid && I.getOpcode() == Instruction::AShr)
6282 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6285 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6287 Instruction *NewShift =
6288 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6289 InsertNewInstBefore(NewShift, I);
6290 NewShift->takeName(Op0BO);
6292 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6299 // Find out if this is a shift of a shift by a constant.
6300 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6301 if (ShiftOp && !ShiftOp->isShift())
6304 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6305 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6306 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6307 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6308 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6309 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6310 Value *X = ShiftOp->getOperand(0);
6312 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6313 if (AmtSum > TypeBits)
6316 const IntegerType *Ty = cast<IntegerType>(I.getType());
6318 // Check for (X << c1) << c2 and (X >> c1) >> c2
6319 if (I.getOpcode() == ShiftOp->getOpcode()) {
6320 return BinaryOperator::create(I.getOpcode(), X,
6321 ConstantInt::get(Ty, AmtSum));
6322 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6323 I.getOpcode() == Instruction::AShr) {
6324 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6325 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6326 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6327 I.getOpcode() == Instruction::LShr) {
6328 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6329 Instruction *Shift =
6330 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6331 InsertNewInstBefore(Shift, I);
6333 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6334 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6337 // Okay, if we get here, one shift must be left, and the other shift must be
6338 // right. See if the amounts are equal.
6339 if (ShiftAmt1 == ShiftAmt2) {
6340 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6341 if (I.getOpcode() == Instruction::Shl) {
6342 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6343 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6345 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6346 if (I.getOpcode() == Instruction::LShr) {
6347 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6348 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6350 // We can simplify ((X << C) >>s C) into a trunc + sext.
6351 // NOTE: we could do this for any C, but that would make 'unusual' integer
6352 // types. For now, just stick to ones well-supported by the code
6354 const Type *SExtType = 0;
6355 switch (Ty->getBitWidth() - ShiftAmt1) {
6362 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6367 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6368 InsertNewInstBefore(NewTrunc, I);
6369 return new SExtInst(NewTrunc, Ty);
6371 // Otherwise, we can't handle it yet.
6372 } else if (ShiftAmt1 < ShiftAmt2) {
6373 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6375 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6376 if (I.getOpcode() == Instruction::Shl) {
6377 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6378 ShiftOp->getOpcode() == Instruction::AShr);
6379 Instruction *Shift =
6380 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6381 InsertNewInstBefore(Shift, I);
6383 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6384 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6387 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6388 if (I.getOpcode() == Instruction::LShr) {
6389 assert(ShiftOp->getOpcode() == Instruction::Shl);
6390 Instruction *Shift =
6391 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6392 InsertNewInstBefore(Shift, I);
6394 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6395 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6398 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6400 assert(ShiftAmt2 < ShiftAmt1);
6401 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6403 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6404 if (I.getOpcode() == Instruction::Shl) {
6405 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6406 ShiftOp->getOpcode() == Instruction::AShr);
6407 Instruction *Shift =
6408 BinaryOperator::create(ShiftOp->getOpcode(), X,
6409 ConstantInt::get(Ty, ShiftDiff));
6410 InsertNewInstBefore(Shift, I);
6412 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6413 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6416 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6417 if (I.getOpcode() == Instruction::LShr) {
6418 assert(ShiftOp->getOpcode() == Instruction::Shl);
6419 Instruction *Shift =
6420 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6421 InsertNewInstBefore(Shift, I);
6423 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6424 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6427 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6434 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6435 /// expression. If so, decompose it, returning some value X, such that Val is
6438 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6440 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6441 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6442 Offset = CI->getZExtValue();
6444 return ConstantInt::get(Type::Int32Ty, 0);
6445 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6446 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6447 if (I->getOpcode() == Instruction::Shl) {
6448 // This is a value scaled by '1 << the shift amt'.
6449 Scale = 1U << RHS->getZExtValue();
6451 return I->getOperand(0);
6452 } else if (I->getOpcode() == Instruction::Mul) {
6453 // This value is scaled by 'RHS'.
6454 Scale = RHS->getZExtValue();
6456 return I->getOperand(0);
6457 } else if (I->getOpcode() == Instruction::Add) {
6458 // We have X+C. Check to see if we really have (X*C2)+C1,
6459 // where C1 is divisible by C2.
6462 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6463 Offset += RHS->getZExtValue();
6470 // Otherwise, we can't look past this.
6477 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6478 /// try to eliminate the cast by moving the type information into the alloc.
6479 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6480 AllocationInst &AI) {
6481 const PointerType *PTy = cast<PointerType>(CI.getType());
6483 // Remove any uses of AI that are dead.
6484 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6486 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6487 Instruction *User = cast<Instruction>(*UI++);
6488 if (isInstructionTriviallyDead(User)) {
6489 while (UI != E && *UI == User)
6490 ++UI; // If this instruction uses AI more than once, don't break UI.
6493 DOUT << "IC: DCE: " << *User;
6494 EraseInstFromFunction(*User);
6498 // Get the type really allocated and the type casted to.
6499 const Type *AllocElTy = AI.getAllocatedType();
6500 const Type *CastElTy = PTy->getElementType();
6501 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6503 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6504 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6505 if (CastElTyAlign < AllocElTyAlign) return 0;
6507 // If the allocation has multiple uses, only promote it if we are strictly
6508 // increasing the alignment of the resultant allocation. If we keep it the
6509 // same, we open the door to infinite loops of various kinds.
6510 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6512 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6513 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6514 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6516 // See if we can satisfy the modulus by pulling a scale out of the array
6518 unsigned ArraySizeScale;
6520 Value *NumElements = // See if the array size is a decomposable linear expr.
6521 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6523 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6525 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6526 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6528 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6533 // If the allocation size is constant, form a constant mul expression
6534 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6535 if (isa<ConstantInt>(NumElements))
6536 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6537 // otherwise multiply the amount and the number of elements
6538 else if (Scale != 1) {
6539 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6540 Amt = InsertNewInstBefore(Tmp, AI);
6544 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6545 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6546 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6547 Amt = InsertNewInstBefore(Tmp, AI);
6550 AllocationInst *New;
6551 if (isa<MallocInst>(AI))
6552 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6554 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6555 InsertNewInstBefore(New, AI);
6558 // If the allocation has multiple uses, insert a cast and change all things
6559 // that used it to use the new cast. This will also hack on CI, but it will
6561 if (!AI.hasOneUse()) {
6562 AddUsesToWorkList(AI);
6563 // New is the allocation instruction, pointer typed. AI is the original
6564 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6565 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6566 InsertNewInstBefore(NewCast, AI);
6567 AI.replaceAllUsesWith(NewCast);
6569 return ReplaceInstUsesWith(CI, New);
6572 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6573 /// and return it as type Ty without inserting any new casts and without
6574 /// changing the computed value. This is used by code that tries to decide
6575 /// whether promoting or shrinking integer operations to wider or smaller types
6576 /// will allow us to eliminate a truncate or extend.
6578 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6579 /// extension operation if Ty is larger.
6580 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6581 unsigned CastOpc, int &NumCastsRemoved) {
6582 // We can always evaluate constants in another type.
6583 if (isa<ConstantInt>(V))
6586 Instruction *I = dyn_cast<Instruction>(V);
6587 if (!I) return false;
6589 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6591 // If this is an extension or truncate, we can often eliminate it.
6592 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6593 // If this is a cast from the destination type, we can trivially eliminate
6594 // it, and this will remove a cast overall.
6595 if (I->getOperand(0)->getType() == Ty) {
6596 // If the first operand is itself a cast, and is eliminable, do not count
6597 // this as an eliminable cast. We would prefer to eliminate those two
6599 if (!isa<CastInst>(I->getOperand(0)))
6605 // We can't extend or shrink something that has multiple uses: doing so would
6606 // require duplicating the instruction in general, which isn't profitable.
6607 if (!I->hasOneUse()) return false;
6609 switch (I->getOpcode()) {
6610 case Instruction::Add:
6611 case Instruction::Sub:
6612 case Instruction::And:
6613 case Instruction::Or:
6614 case Instruction::Xor:
6615 // These operators can all arbitrarily be extended or truncated.
6616 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6618 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6621 case Instruction::Mul:
6622 // A multiply can be truncated by truncating its operands.
6623 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6624 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6626 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6629 case Instruction::Shl:
6630 // If we are truncating the result of this SHL, and if it's a shift of a
6631 // constant amount, we can always perform a SHL in a smaller type.
6632 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6633 uint32_t BitWidth = Ty->getBitWidth();
6634 if (BitWidth < OrigTy->getBitWidth() &&
6635 CI->getLimitedValue(BitWidth) < BitWidth)
6636 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6640 case Instruction::LShr:
6641 // If this is a truncate of a logical shr, we can truncate it to a smaller
6642 // lshr iff we know that the bits we would otherwise be shifting in are
6644 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6645 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6646 uint32_t BitWidth = Ty->getBitWidth();
6647 if (BitWidth < OrigBitWidth &&
6648 MaskedValueIsZero(I->getOperand(0),
6649 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6650 CI->getLimitedValue(BitWidth) < BitWidth) {
6651 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6656 case Instruction::ZExt:
6657 case Instruction::SExt:
6658 case Instruction::Trunc:
6659 // If this is the same kind of case as our original (e.g. zext+zext), we
6660 // can safely replace it. Note that replacing it does not reduce the number
6661 // of casts in the input.
6662 if (I->getOpcode() == CastOpc)
6667 // TODO: Can handle more cases here.
6674 /// EvaluateInDifferentType - Given an expression that
6675 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6676 /// evaluate the expression.
6677 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6679 if (Constant *C = dyn_cast<Constant>(V))
6680 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6682 // Otherwise, it must be an instruction.
6683 Instruction *I = cast<Instruction>(V);
6684 Instruction *Res = 0;
6685 switch (I->getOpcode()) {
6686 case Instruction::Add:
6687 case Instruction::Sub:
6688 case Instruction::Mul:
6689 case Instruction::And:
6690 case Instruction::Or:
6691 case Instruction::Xor:
6692 case Instruction::AShr:
6693 case Instruction::LShr:
6694 case Instruction::Shl: {
6695 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6696 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6697 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6698 LHS, RHS, I->getName());
6701 case Instruction::Trunc:
6702 case Instruction::ZExt:
6703 case Instruction::SExt:
6704 // If the source type of the cast is the type we're trying for then we can
6705 // just return the source. There's no need to insert it because it is not
6707 if (I->getOperand(0)->getType() == Ty)
6708 return I->getOperand(0);
6710 // Otherwise, must be the same type of case, so just reinsert a new one.
6711 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6715 // TODO: Can handle more cases here.
6716 assert(0 && "Unreachable!");
6720 return InsertNewInstBefore(Res, *I);
6723 /// @brief Implement the transforms common to all CastInst visitors.
6724 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6725 Value *Src = CI.getOperand(0);
6727 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6728 // eliminate it now.
6729 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6730 if (Instruction::CastOps opc =
6731 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6732 // The first cast (CSrc) is eliminable so we need to fix up or replace
6733 // the second cast (CI). CSrc will then have a good chance of being dead.
6734 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6738 // If we are casting a select then fold the cast into the select
6739 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6740 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6743 // If we are casting a PHI then fold the cast into the PHI
6744 if (isa<PHINode>(Src))
6745 if (Instruction *NV = FoldOpIntoPhi(CI))
6751 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6752 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6753 Value *Src = CI.getOperand(0);
6755 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6756 // If casting the result of a getelementptr instruction with no offset, turn
6757 // this into a cast of the original pointer!
6758 if (GEP->hasAllZeroIndices()) {
6759 // Changing the cast operand is usually not a good idea but it is safe
6760 // here because the pointer operand is being replaced with another
6761 // pointer operand so the opcode doesn't need to change.
6763 CI.setOperand(0, GEP->getOperand(0));
6767 // If the GEP has a single use, and the base pointer is a bitcast, and the
6768 // GEP computes a constant offset, see if we can convert these three
6769 // instructions into fewer. This typically happens with unions and other
6770 // non-type-safe code.
6771 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6772 if (GEP->hasAllConstantIndices()) {
6773 // We are guaranteed to get a constant from EmitGEPOffset.
6774 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6775 int64_t Offset = OffsetV->getSExtValue();
6777 // Get the base pointer input of the bitcast, and the type it points to.
6778 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6779 const Type *GEPIdxTy =
6780 cast<PointerType>(OrigBase->getType())->getElementType();
6781 if (GEPIdxTy->isSized()) {
6782 SmallVector<Value*, 8> NewIndices;
6784 // Start with the index over the outer type. Note that the type size
6785 // might be zero (even if the offset isn't zero) if the indexed type
6786 // is something like [0 x {int, int}]
6787 const Type *IntPtrTy = TD->getIntPtrType();
6788 int64_t FirstIdx = 0;
6789 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6790 FirstIdx = Offset/TySize;
6793 // Handle silly modulus not returning values values [0..TySize).
6797 assert(Offset >= 0);
6799 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6802 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6804 // Index into the types. If we fail, set OrigBase to null.
6806 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6807 const StructLayout *SL = TD->getStructLayout(STy);
6808 if (Offset < (int64_t)SL->getSizeInBytes()) {
6809 unsigned Elt = SL->getElementContainingOffset(Offset);
6810 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6812 Offset -= SL->getElementOffset(Elt);
6813 GEPIdxTy = STy->getElementType(Elt);
6815 // Otherwise, we can't index into this, bail out.
6819 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6820 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6821 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6822 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6825 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6827 GEPIdxTy = STy->getElementType();
6829 // Otherwise, we can't index into this, bail out.
6835 // If we were able to index down into an element, create the GEP
6836 // and bitcast the result. This eliminates one bitcast, potentially
6838 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6840 NewIndices.end(), "");
6841 InsertNewInstBefore(NGEP, CI);
6842 NGEP->takeName(GEP);
6844 if (isa<BitCastInst>(CI))
6845 return new BitCastInst(NGEP, CI.getType());
6846 assert(isa<PtrToIntInst>(CI));
6847 return new PtrToIntInst(NGEP, CI.getType());
6854 return commonCastTransforms(CI);
6859 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6860 /// integer types. This function implements the common transforms for all those
6862 /// @brief Implement the transforms common to CastInst with integer operands
6863 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6864 if (Instruction *Result = commonCastTransforms(CI))
6867 Value *Src = CI.getOperand(0);
6868 const Type *SrcTy = Src->getType();
6869 const Type *DestTy = CI.getType();
6870 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6871 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6873 // See if we can simplify any instructions used by the LHS whose sole
6874 // purpose is to compute bits we don't care about.
6875 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6876 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6877 KnownZero, KnownOne))
6880 // If the source isn't an instruction or has more than one use then we
6881 // can't do anything more.
6882 Instruction *SrcI = dyn_cast<Instruction>(Src);
6883 if (!SrcI || !Src->hasOneUse())
6886 // Attempt to propagate the cast into the instruction for int->int casts.
6887 int NumCastsRemoved = 0;
6888 if (!isa<BitCastInst>(CI) &&
6889 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6890 CI.getOpcode(), NumCastsRemoved)) {
6891 // If this cast is a truncate, evaluting in a different type always
6892 // eliminates the cast, so it is always a win. If this is a zero-extension,
6893 // we need to do an AND to maintain the clear top-part of the computation,
6894 // so we require that the input have eliminated at least one cast. If this
6895 // is a sign extension, we insert two new casts (to do the extension) so we
6896 // require that two casts have been eliminated.
6898 switch (CI.getOpcode()) {
6900 // All the others use floating point so we shouldn't actually
6901 // get here because of the check above.
6902 assert(0 && "Unknown cast type");
6903 case Instruction::Trunc:
6906 case Instruction::ZExt:
6907 DoXForm = NumCastsRemoved >= 1;
6909 case Instruction::SExt:
6910 DoXForm = NumCastsRemoved >= 2;
6915 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6916 CI.getOpcode() == Instruction::SExt);
6917 assert(Res->getType() == DestTy);
6918 switch (CI.getOpcode()) {
6919 default: assert(0 && "Unknown cast type!");
6920 case Instruction::Trunc:
6921 case Instruction::BitCast:
6922 // Just replace this cast with the result.
6923 return ReplaceInstUsesWith(CI, Res);
6924 case Instruction::ZExt: {
6925 // We need to emit an AND to clear the high bits.
6926 assert(SrcBitSize < DestBitSize && "Not a zext?");
6927 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6929 return BinaryOperator::createAnd(Res, C);
6931 case Instruction::SExt:
6932 // We need to emit a cast to truncate, then a cast to sext.
6933 return CastInst::create(Instruction::SExt,
6934 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6940 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6941 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6943 switch (SrcI->getOpcode()) {
6944 case Instruction::Add:
6945 case Instruction::Mul:
6946 case Instruction::And:
6947 case Instruction::Or:
6948 case Instruction::Xor:
6949 // If we are discarding information, rewrite.
6950 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6951 // Don't insert two casts if they cannot be eliminated. We allow
6952 // two casts to be inserted if the sizes are the same. This could
6953 // only be converting signedness, which is a noop.
6954 if (DestBitSize == SrcBitSize ||
6955 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6956 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6957 Instruction::CastOps opcode = CI.getOpcode();
6958 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6959 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6960 return BinaryOperator::create(
6961 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6965 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6966 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6967 SrcI->getOpcode() == Instruction::Xor &&
6968 Op1 == ConstantInt::getTrue() &&
6969 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6970 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6971 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6974 case Instruction::SDiv:
6975 case Instruction::UDiv:
6976 case Instruction::SRem:
6977 case Instruction::URem:
6978 // If we are just changing the sign, rewrite.
6979 if (DestBitSize == SrcBitSize) {
6980 // Don't insert two casts if they cannot be eliminated. We allow
6981 // two casts to be inserted if the sizes are the same. This could
6982 // only be converting signedness, which is a noop.
6983 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6984 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6985 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6987 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6989 return BinaryOperator::create(
6990 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6995 case Instruction::Shl:
6996 // Allow changing the sign of the source operand. Do not allow
6997 // changing the size of the shift, UNLESS the shift amount is a
6998 // constant. We must not change variable sized shifts to a smaller
6999 // size, because it is undefined to shift more bits out than exist
7001 if (DestBitSize == SrcBitSize ||
7002 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7003 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7004 Instruction::BitCast : Instruction::Trunc);
7005 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7006 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7007 return BinaryOperator::createShl(Op0c, Op1c);
7010 case Instruction::AShr:
7011 // If this is a signed shr, and if all bits shifted in are about to be
7012 // truncated off, turn it into an unsigned shr to allow greater
7014 if (DestBitSize < SrcBitSize &&
7015 isa<ConstantInt>(Op1)) {
7016 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7017 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7018 // Insert the new logical shift right.
7019 return BinaryOperator::createLShr(Op0, Op1);
7027 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7028 if (Instruction *Result = commonIntCastTransforms(CI))
7031 Value *Src = CI.getOperand(0);
7032 const Type *Ty = CI.getType();
7033 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7034 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7036 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7037 switch (SrcI->getOpcode()) {
7039 case Instruction::LShr:
7040 // We can shrink lshr to something smaller if we know the bits shifted in
7041 // are already zeros.
7042 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7043 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7045 // Get a mask for the bits shifting in.
7046 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7047 Value* SrcIOp0 = SrcI->getOperand(0);
7048 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7049 if (ShAmt >= DestBitWidth) // All zeros.
7050 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7052 // Okay, we can shrink this. Truncate the input, then return a new
7054 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7055 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7057 return BinaryOperator::createLShr(V1, V2);
7059 } else { // This is a variable shr.
7061 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7062 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7063 // loop-invariant and CSE'd.
7064 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7065 Value *One = ConstantInt::get(SrcI->getType(), 1);
7067 Value *V = InsertNewInstBefore(
7068 BinaryOperator::createShl(One, SrcI->getOperand(1),
7070 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7071 SrcI->getOperand(0),
7073 Value *Zero = Constant::getNullValue(V->getType());
7074 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7084 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7085 // If one of the common conversion will work ..
7086 if (Instruction *Result = commonIntCastTransforms(CI))
7089 Value *Src = CI.getOperand(0);
7091 // If this is a cast of a cast
7092 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7093 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7094 // types and if the sizes are just right we can convert this into a logical
7095 // 'and' which will be much cheaper than the pair of casts.
7096 if (isa<TruncInst>(CSrc)) {
7097 // Get the sizes of the types involved
7098 Value *A = CSrc->getOperand(0);
7099 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7100 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7101 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7102 // If we're actually extending zero bits and the trunc is a no-op
7103 if (MidSize < DstSize && SrcSize == DstSize) {
7104 // Replace both of the casts with an And of the type mask.
7105 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7106 Constant *AndConst = ConstantInt::get(AndValue);
7108 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7109 // Unfortunately, if the type changed, we need to cast it back.
7110 if (And->getType() != CI.getType()) {
7111 And->setName(CSrc->getName()+".mask");
7112 InsertNewInstBefore(And, CI);
7113 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7120 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7121 // If we are just checking for a icmp eq of a single bit and zext'ing it
7122 // to an integer, then shift the bit to the appropriate place and then
7123 // cast to integer to avoid the comparison.
7124 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7125 const APInt &Op1CV = Op1C->getValue();
7127 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7128 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7129 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7130 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7131 Value *In = ICI->getOperand(0);
7132 Value *Sh = ConstantInt::get(In->getType(),
7133 In->getType()->getPrimitiveSizeInBits()-1);
7134 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7135 In->getName()+".lobit"),
7137 if (In->getType() != CI.getType())
7138 In = CastInst::createIntegerCast(In, CI.getType(),
7139 false/*ZExt*/, "tmp", &CI);
7141 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7142 Constant *One = ConstantInt::get(In->getType(), 1);
7143 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7144 In->getName()+".not"),
7148 return ReplaceInstUsesWith(CI, In);
7153 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7154 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7155 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7156 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7157 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7158 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7159 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7160 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7161 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7162 // This only works for EQ and NE
7163 ICI->isEquality()) {
7164 // If Op1C some other power of two, convert:
7165 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7166 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7167 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7168 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7170 APInt KnownZeroMask(~KnownZero);
7171 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7172 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7173 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7174 // (X&4) == 2 --> false
7175 // (X&4) != 2 --> true
7176 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7177 Res = ConstantExpr::getZExt(Res, CI.getType());
7178 return ReplaceInstUsesWith(CI, Res);
7181 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7182 Value *In = ICI->getOperand(0);
7184 // Perform a logical shr by shiftamt.
7185 // Insert the shift to put the result in the low bit.
7186 In = InsertNewInstBefore(
7187 BinaryOperator::createLShr(In,
7188 ConstantInt::get(In->getType(), ShiftAmt),
7189 In->getName()+".lobit"), CI);
7192 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7193 Constant *One = ConstantInt::get(In->getType(), 1);
7194 In = BinaryOperator::createXor(In, One, "tmp");
7195 InsertNewInstBefore(cast<Instruction>(In), CI);
7198 if (CI.getType() == In->getType())
7199 return ReplaceInstUsesWith(CI, In);
7201 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7209 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7210 if (Instruction *I = commonIntCastTransforms(CI))
7213 Value *Src = CI.getOperand(0);
7215 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7216 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7217 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7218 // If we are just checking for a icmp eq of a single bit and zext'ing it
7219 // to an integer, then shift the bit to the appropriate place and then
7220 // cast to integer to avoid the comparison.
7221 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7222 const APInt &Op1CV = Op1C->getValue();
7224 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7225 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7226 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7227 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7228 Value *In = ICI->getOperand(0);
7229 Value *Sh = ConstantInt::get(In->getType(),
7230 In->getType()->getPrimitiveSizeInBits()-1);
7231 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7232 In->getName()+".lobit"),
7234 if (In->getType() != CI.getType())
7235 In = CastInst::createIntegerCast(In, CI.getType(),
7236 true/*SExt*/, "tmp", &CI);
7238 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7239 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7240 In->getName()+".not"), CI);
7242 return ReplaceInstUsesWith(CI, In);
7250 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7251 /// in the specified FP type without changing its value.
7252 static Constant *FitsInFPType(ConstantFP *CFP, const Type *FPTy,
7253 const fltSemantics &Sem) {
7254 APFloat F = CFP->getValueAPF();
7255 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7256 return ConstantFP::get(FPTy, F);
7260 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7261 /// through it until we get the source value.
7262 static Value *LookThroughFPExtensions(Value *V) {
7263 if (Instruction *I = dyn_cast<Instruction>(V))
7264 if (I->getOpcode() == Instruction::FPExt)
7265 return LookThroughFPExtensions(I->getOperand(0));
7267 // If this value is a constant, return the constant in the smallest FP type
7268 // that can accurately represent it. This allows us to turn
7269 // (float)((double)X+2.0) into x+2.0f.
7270 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7271 if (CFP->getType() == Type::PPC_FP128Ty)
7272 return V; // No constant folding of this.
7273 // See if the value can be truncated to float and then reextended.
7274 if (Value *V = FitsInFPType(CFP, Type::FloatTy, APFloat::IEEEsingle))
7276 if (CFP->getType() == Type::DoubleTy)
7277 return V; // Won't shrink.
7278 if (Value *V = FitsInFPType(CFP, Type::DoubleTy, APFloat::IEEEdouble))
7280 // Don't try to shrink to various long double types.
7286 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7287 if (Instruction *I = commonCastTransforms(CI))
7290 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7291 // smaller than the destination type, we can eliminate the truncate by doing
7292 // the add as the smaller type. This applies to add/sub/mul/div as well as
7293 // many builtins (sqrt, etc).
7294 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7295 if (OpI && OpI->hasOneUse()) {
7296 switch (OpI->getOpcode()) {
7298 case Instruction::Add:
7299 case Instruction::Sub:
7300 case Instruction::Mul:
7301 case Instruction::FDiv:
7302 case Instruction::FRem:
7303 const Type *SrcTy = OpI->getType();
7304 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7305 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7306 if (LHSTrunc->getType() != SrcTy &&
7307 RHSTrunc->getType() != SrcTy) {
7308 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7309 // If the source types were both smaller than the destination type of
7310 // the cast, do this xform.
7311 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7312 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7313 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7315 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7317 return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7326 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7327 return commonCastTransforms(CI);
7330 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7331 return commonCastTransforms(CI);
7334 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7335 return commonCastTransforms(CI);
7338 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7339 return commonCastTransforms(CI);
7342 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7343 return commonCastTransforms(CI);
7346 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7347 return commonPointerCastTransforms(CI);
7350 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7351 if (Instruction *I = commonCastTransforms(CI))
7354 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7355 if (!DestPointee->isSized()) return 0;
7357 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7360 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7361 m_ConstantInt(Cst)))) {
7362 // If the source and destination operands have the same type, see if this
7363 // is a single-index GEP.
7364 if (X->getType() == CI.getType()) {
7365 // Get the size of the pointee type.
7366 uint64_t Size = TD->getABITypeSizeInBits(DestPointee);
7368 // Convert the constant to intptr type.
7369 APInt Offset = Cst->getValue();
7370 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7372 // If Offset is evenly divisible by Size, we can do this xform.
7373 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7374 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7375 return new GetElementPtrInst(X, ConstantInt::get(Offset));
7378 // TODO: Could handle other cases, e.g. where add is indexing into field of
7380 } else if (CI.getOperand(0)->hasOneUse() &&
7381 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7382 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7383 // "inttoptr+GEP" instead of "add+intptr".
7385 // Get the size of the pointee type.
7386 uint64_t Size = TD->getABITypeSize(DestPointee);
7388 // Convert the constant to intptr type.
7389 APInt Offset = Cst->getValue();
7390 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7392 // If Offset is evenly divisible by Size, we can do this xform.
7393 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7394 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7396 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7398 return new GetElementPtrInst(P, ConstantInt::get(Offset), "tmp");
7404 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7405 // If the operands are integer typed then apply the integer transforms,
7406 // otherwise just apply the common ones.
7407 Value *Src = CI.getOperand(0);
7408 const Type *SrcTy = Src->getType();
7409 const Type *DestTy = CI.getType();
7411 if (SrcTy->isInteger() && DestTy->isInteger()) {
7412 if (Instruction *Result = commonIntCastTransforms(CI))
7414 } else if (isa<PointerType>(SrcTy)) {
7415 if (Instruction *I = commonPointerCastTransforms(CI))
7418 if (Instruction *Result = commonCastTransforms(CI))
7423 // Get rid of casts from one type to the same type. These are useless and can
7424 // be replaced by the operand.
7425 if (DestTy == Src->getType())
7426 return ReplaceInstUsesWith(CI, Src);
7428 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7429 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7430 const Type *DstElTy = DstPTy->getElementType();
7431 const Type *SrcElTy = SrcPTy->getElementType();
7433 // If we are casting a malloc or alloca to a pointer to a type of the same
7434 // size, rewrite the allocation instruction to allocate the "right" type.
7435 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7436 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7439 // If the source and destination are pointers, and this cast is equivalent
7440 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7441 // This can enhance SROA and other transforms that want type-safe pointers.
7442 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7443 unsigned NumZeros = 0;
7444 while (SrcElTy != DstElTy &&
7445 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7446 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7447 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7451 // If we found a path from the src to dest, create the getelementptr now.
7452 if (SrcElTy == DstElTy) {
7453 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7454 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7455 ((Instruction*) NULL));
7459 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7460 if (SVI->hasOneUse()) {
7461 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7462 // a bitconvert to a vector with the same # elts.
7463 if (isa<VectorType>(DestTy) &&
7464 cast<VectorType>(DestTy)->getNumElements() ==
7465 SVI->getType()->getNumElements()) {
7467 // If either of the operands is a cast from CI.getType(), then
7468 // evaluating the shuffle in the casted destination's type will allow
7469 // us to eliminate at least one cast.
7470 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7471 Tmp->getOperand(0)->getType() == DestTy) ||
7472 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7473 Tmp->getOperand(0)->getType() == DestTy)) {
7474 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7475 SVI->getOperand(0), DestTy, &CI);
7476 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7477 SVI->getOperand(1), DestTy, &CI);
7478 // Return a new shuffle vector. Use the same element ID's, as we
7479 // know the vector types match #elts.
7480 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7488 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7490 /// %D = select %cond, %C, %A
7492 /// %C = select %cond, %B, 0
7495 /// Assuming that the specified instruction is an operand to the select, return
7496 /// a bitmask indicating which operands of this instruction are foldable if they
7497 /// equal the other incoming value of the select.
7499 static unsigned GetSelectFoldableOperands(Instruction *I) {
7500 switch (I->getOpcode()) {
7501 case Instruction::Add:
7502 case Instruction::Mul:
7503 case Instruction::And:
7504 case Instruction::Or:
7505 case Instruction::Xor:
7506 return 3; // Can fold through either operand.
7507 case Instruction::Sub: // Can only fold on the amount subtracted.
7508 case Instruction::Shl: // Can only fold on the shift amount.
7509 case Instruction::LShr:
7510 case Instruction::AShr:
7513 return 0; // Cannot fold
7517 /// GetSelectFoldableConstant - For the same transformation as the previous
7518 /// function, return the identity constant that goes into the select.
7519 static Constant *GetSelectFoldableConstant(Instruction *I) {
7520 switch (I->getOpcode()) {
7521 default: assert(0 && "This cannot happen!"); abort();
7522 case Instruction::Add:
7523 case Instruction::Sub:
7524 case Instruction::Or:
7525 case Instruction::Xor:
7526 case Instruction::Shl:
7527 case Instruction::LShr:
7528 case Instruction::AShr:
7529 return Constant::getNullValue(I->getType());
7530 case Instruction::And:
7531 return Constant::getAllOnesValue(I->getType());
7532 case Instruction::Mul:
7533 return ConstantInt::get(I->getType(), 1);
7537 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7538 /// have the same opcode and only one use each. Try to simplify this.
7539 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7541 if (TI->getNumOperands() == 1) {
7542 // If this is a non-volatile load or a cast from the same type,
7545 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7548 return 0; // unknown unary op.
7551 // Fold this by inserting a select from the input values.
7552 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7553 FI->getOperand(0), SI.getName()+".v");
7554 InsertNewInstBefore(NewSI, SI);
7555 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7559 // Only handle binary operators here.
7560 if (!isa<BinaryOperator>(TI))
7563 // Figure out if the operations have any operands in common.
7564 Value *MatchOp, *OtherOpT, *OtherOpF;
7566 if (TI->getOperand(0) == FI->getOperand(0)) {
7567 MatchOp = TI->getOperand(0);
7568 OtherOpT = TI->getOperand(1);
7569 OtherOpF = FI->getOperand(1);
7570 MatchIsOpZero = true;
7571 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7572 MatchOp = TI->getOperand(1);
7573 OtherOpT = TI->getOperand(0);
7574 OtherOpF = FI->getOperand(0);
7575 MatchIsOpZero = false;
7576 } else if (!TI->isCommutative()) {
7578 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7579 MatchOp = TI->getOperand(0);
7580 OtherOpT = TI->getOperand(1);
7581 OtherOpF = FI->getOperand(0);
7582 MatchIsOpZero = true;
7583 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7584 MatchOp = TI->getOperand(1);
7585 OtherOpT = TI->getOperand(0);
7586 OtherOpF = FI->getOperand(1);
7587 MatchIsOpZero = true;
7592 // If we reach here, they do have operations in common.
7593 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7594 OtherOpF, SI.getName()+".v");
7595 InsertNewInstBefore(NewSI, SI);
7597 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7599 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7601 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7603 assert(0 && "Shouldn't get here");
7607 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7608 Value *CondVal = SI.getCondition();
7609 Value *TrueVal = SI.getTrueValue();
7610 Value *FalseVal = SI.getFalseValue();
7612 // select true, X, Y -> X
7613 // select false, X, Y -> Y
7614 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7615 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7617 // select C, X, X -> X
7618 if (TrueVal == FalseVal)
7619 return ReplaceInstUsesWith(SI, TrueVal);
7621 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7622 return ReplaceInstUsesWith(SI, FalseVal);
7623 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7624 return ReplaceInstUsesWith(SI, TrueVal);
7625 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7626 if (isa<Constant>(TrueVal))
7627 return ReplaceInstUsesWith(SI, TrueVal);
7629 return ReplaceInstUsesWith(SI, FalseVal);
7632 if (SI.getType() == Type::Int1Ty) {
7633 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7634 if (C->getZExtValue()) {
7635 // Change: A = select B, true, C --> A = or B, C
7636 return BinaryOperator::createOr(CondVal, FalseVal);
7638 // Change: A = select B, false, C --> A = and !B, C
7640 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7641 "not."+CondVal->getName()), SI);
7642 return BinaryOperator::createAnd(NotCond, FalseVal);
7644 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7645 if (C->getZExtValue() == false) {
7646 // Change: A = select B, C, false --> A = and B, C
7647 return BinaryOperator::createAnd(CondVal, TrueVal);
7649 // Change: A = select B, C, true --> A = or !B, C
7651 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7652 "not."+CondVal->getName()), SI);
7653 return BinaryOperator::createOr(NotCond, TrueVal);
7657 // select a, b, a -> a&b
7658 // select a, a, b -> a|b
7659 if (CondVal == TrueVal)
7660 return BinaryOperator::createOr(CondVal, FalseVal);
7661 else if (CondVal == FalseVal)
7662 return BinaryOperator::createAnd(CondVal, TrueVal);
7665 // Selecting between two integer constants?
7666 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7667 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7668 // select C, 1, 0 -> zext C to int
7669 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7670 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7671 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7672 // select C, 0, 1 -> zext !C to int
7674 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7675 "not."+CondVal->getName()), SI);
7676 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7679 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7681 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7683 // (x <s 0) ? -1 : 0 -> ashr x, 31
7684 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7685 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7686 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7687 // The comparison constant and the result are not neccessarily the
7688 // same width. Make an all-ones value by inserting a AShr.
7689 Value *X = IC->getOperand(0);
7690 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7691 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7692 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7694 InsertNewInstBefore(SRA, SI);
7696 // Finally, convert to the type of the select RHS. We figure out
7697 // if this requires a SExt, Trunc or BitCast based on the sizes.
7698 Instruction::CastOps opc = Instruction::BitCast;
7699 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7700 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7701 if (SRASize < SISize)
7702 opc = Instruction::SExt;
7703 else if (SRASize > SISize)
7704 opc = Instruction::Trunc;
7705 return CastInst::create(opc, SRA, SI.getType());
7710 // If one of the constants is zero (we know they can't both be) and we
7711 // have an icmp instruction with zero, and we have an 'and' with the
7712 // non-constant value, eliminate this whole mess. This corresponds to
7713 // cases like this: ((X & 27) ? 27 : 0)
7714 if (TrueValC->isZero() || FalseValC->isZero())
7715 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7716 cast<Constant>(IC->getOperand(1))->isNullValue())
7717 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7718 if (ICA->getOpcode() == Instruction::And &&
7719 isa<ConstantInt>(ICA->getOperand(1)) &&
7720 (ICA->getOperand(1) == TrueValC ||
7721 ICA->getOperand(1) == FalseValC) &&
7722 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7723 // Okay, now we know that everything is set up, we just don't
7724 // know whether we have a icmp_ne or icmp_eq and whether the
7725 // true or false val is the zero.
7726 bool ShouldNotVal = !TrueValC->isZero();
7727 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7730 V = InsertNewInstBefore(BinaryOperator::create(
7731 Instruction::Xor, V, ICA->getOperand(1)), SI);
7732 return ReplaceInstUsesWith(SI, V);
7737 // See if we are selecting two values based on a comparison of the two values.
7738 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7739 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7740 // Transform (X == Y) ? X : Y -> Y
7741 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7742 // This is not safe in general for floating point:
7743 // consider X== -0, Y== +0.
7744 // It becomes safe if either operand is a nonzero constant.
7745 ConstantFP *CFPt, *CFPf;
7746 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7747 !CFPt->getValueAPF().isZero()) ||
7748 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7749 !CFPf->getValueAPF().isZero()))
7750 return ReplaceInstUsesWith(SI, FalseVal);
7752 // Transform (X != Y) ? X : Y -> X
7753 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7754 return ReplaceInstUsesWith(SI, TrueVal);
7755 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7757 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7758 // Transform (X == Y) ? Y : X -> X
7759 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7760 // This is not safe in general for floating point:
7761 // consider X== -0, Y== +0.
7762 // It becomes safe if either operand is a nonzero constant.
7763 ConstantFP *CFPt, *CFPf;
7764 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7765 !CFPt->getValueAPF().isZero()) ||
7766 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7767 !CFPf->getValueAPF().isZero()))
7768 return ReplaceInstUsesWith(SI, FalseVal);
7770 // Transform (X != Y) ? Y : X -> Y
7771 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7772 return ReplaceInstUsesWith(SI, TrueVal);
7773 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7777 // See if we are selecting two values based on a comparison of the two values.
7778 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7779 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7780 // Transform (X == Y) ? X : Y -> Y
7781 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7782 return ReplaceInstUsesWith(SI, FalseVal);
7783 // Transform (X != Y) ? X : Y -> X
7784 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7785 return ReplaceInstUsesWith(SI, TrueVal);
7786 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7788 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7789 // Transform (X == Y) ? Y : X -> X
7790 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7791 return ReplaceInstUsesWith(SI, FalseVal);
7792 // Transform (X != Y) ? Y : X -> Y
7793 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7794 return ReplaceInstUsesWith(SI, TrueVal);
7795 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7799 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7800 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7801 if (TI->hasOneUse() && FI->hasOneUse()) {
7802 Instruction *AddOp = 0, *SubOp = 0;
7804 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7805 if (TI->getOpcode() == FI->getOpcode())
7806 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7809 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7810 // even legal for FP.
7811 if (TI->getOpcode() == Instruction::Sub &&
7812 FI->getOpcode() == Instruction::Add) {
7813 AddOp = FI; SubOp = TI;
7814 } else if (FI->getOpcode() == Instruction::Sub &&
7815 TI->getOpcode() == Instruction::Add) {
7816 AddOp = TI; SubOp = FI;
7820 Value *OtherAddOp = 0;
7821 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7822 OtherAddOp = AddOp->getOperand(1);
7823 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7824 OtherAddOp = AddOp->getOperand(0);
7828 // So at this point we know we have (Y -> OtherAddOp):
7829 // select C, (add X, Y), (sub X, Z)
7830 Value *NegVal; // Compute -Z
7831 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7832 NegVal = ConstantExpr::getNeg(C);
7834 NegVal = InsertNewInstBefore(
7835 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7838 Value *NewTrueOp = OtherAddOp;
7839 Value *NewFalseOp = NegVal;
7841 std::swap(NewTrueOp, NewFalseOp);
7842 Instruction *NewSel =
7843 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7845 NewSel = InsertNewInstBefore(NewSel, SI);
7846 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7851 // See if we can fold the select into one of our operands.
7852 if (SI.getType()->isInteger()) {
7853 // See the comment above GetSelectFoldableOperands for a description of the
7854 // transformation we are doing here.
7855 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7856 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7857 !isa<Constant>(FalseVal))
7858 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7859 unsigned OpToFold = 0;
7860 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7862 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7867 Constant *C = GetSelectFoldableConstant(TVI);
7868 Instruction *NewSel =
7869 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7870 InsertNewInstBefore(NewSel, SI);
7871 NewSel->takeName(TVI);
7872 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7873 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7875 assert(0 && "Unknown instruction!!");
7880 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7881 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7882 !isa<Constant>(TrueVal))
7883 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7884 unsigned OpToFold = 0;
7885 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7887 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7892 Constant *C = GetSelectFoldableConstant(FVI);
7893 Instruction *NewSel =
7894 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7895 InsertNewInstBefore(NewSel, SI);
7896 NewSel->takeName(FVI);
7897 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7898 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7900 assert(0 && "Unknown instruction!!");
7905 if (BinaryOperator::isNot(CondVal)) {
7906 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7907 SI.setOperand(1, FalseVal);
7908 SI.setOperand(2, TrueVal);
7915 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7916 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7917 /// and it is more than the alignment of the ultimate object, see if we can
7918 /// increase the alignment of the ultimate object, making this check succeed.
7919 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7920 unsigned PrefAlign = 0) {
7921 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7922 unsigned Align = GV->getAlignment();
7923 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7924 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7926 // If there is a large requested alignment and we can, bump up the alignment
7928 if (PrefAlign > Align && GV->hasInitializer()) {
7929 GV->setAlignment(PrefAlign);
7933 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7934 unsigned Align = AI->getAlignment();
7935 if (Align == 0 && TD) {
7936 if (isa<AllocaInst>(AI))
7937 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7938 else if (isa<MallocInst>(AI)) {
7939 // Malloc returns maximally aligned memory.
7940 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7943 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7946 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7950 // If there is a requested alignment and if this is an alloca, round up. We
7951 // don't do this for malloc, because some systems can't respect the request.
7952 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7953 AI->setAlignment(PrefAlign);
7957 } else if (isa<BitCastInst>(V) ||
7958 (isa<ConstantExpr>(V) &&
7959 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7960 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7962 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7963 // If all indexes are zero, it is just the alignment of the base pointer.
7964 bool AllZeroOperands = true;
7965 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7966 if (!isa<Constant>(GEPI->getOperand(i)) ||
7967 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7968 AllZeroOperands = false;
7972 if (AllZeroOperands) {
7973 // Treat this like a bitcast.
7974 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7977 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7978 if (BaseAlignment == 0) return 0;
7980 // Otherwise, if the base alignment is >= the alignment we expect for the
7981 // base pointer type, then we know that the resultant pointer is aligned at
7982 // least as much as its type requires.
7985 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7986 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7987 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7988 if (Align <= BaseAlignment) {
7989 const Type *GEPTy = GEPI->getType();
7990 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7991 Align = std::min(Align, (unsigned)
7992 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
8000 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8001 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
8002 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
8003 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8004 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8006 if (CopyAlign < MinAlign) {
8007 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8011 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8013 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8014 if (MemOpLength == 0) return 0;
8016 // Source and destination pointer types are always "i8*" for intrinsic. See
8017 // if the size is something we can handle with a single primitive load/store.
8018 // A single load+store correctly handles overlapping memory in the memmove
8020 unsigned Size = MemOpLength->getZExtValue();
8021 if (Size == 0 || Size > 8 || (Size&(Size-1)))
8022 return 0; // If not 1/2/4/8 bytes, exit.
8024 // Use an integer load+store unless we can find something better.
8025 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8027 // Memcpy forces the use of i8* for the source and destination. That means
8028 // that if you're using memcpy to move one double around, you'll get a cast
8029 // from double* to i8*. We'd much rather use a double load+store rather than
8030 // an i64 load+store, here because this improves the odds that the source or
8031 // dest address will be promotable. See if we can find a better type than the
8032 // integer datatype.
8033 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8034 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8035 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8036 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8037 // down through these levels if so.
8038 while (!SrcETy->isFirstClassType()) {
8039 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8040 if (STy->getNumElements() == 1)
8041 SrcETy = STy->getElementType(0);
8044 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8045 if (ATy->getNumElements() == 1)
8046 SrcETy = ATy->getElementType();
8053 if (SrcETy->isFirstClassType())
8054 NewPtrTy = PointerType::getUnqual(SrcETy);
8059 // If the memcpy/memmove provides better alignment info than we can
8061 SrcAlign = std::max(SrcAlign, CopyAlign);
8062 DstAlign = std::max(DstAlign, CopyAlign);
8064 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8065 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8066 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8067 InsertNewInstBefore(L, *MI);
8068 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8070 // Set the size of the copy to 0, it will be deleted on the next iteration.
8071 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8075 /// visitCallInst - CallInst simplification. This mostly only handles folding
8076 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8077 /// the heavy lifting.
8079 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8080 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8081 if (!II) return visitCallSite(&CI);
8083 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8085 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8086 bool Changed = false;
8088 // memmove/cpy/set of zero bytes is a noop.
8089 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8090 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8092 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8093 if (CI->getZExtValue() == 1) {
8094 // Replace the instruction with just byte operations. We would
8095 // transform other cases to loads/stores, but we don't know if
8096 // alignment is sufficient.
8100 // If we have a memmove and the source operation is a constant global,
8101 // then the source and dest pointers can't alias, so we can change this
8102 // into a call to memcpy.
8103 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8104 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8105 if (GVSrc->isConstant()) {
8106 Module *M = CI.getParent()->getParent()->getParent();
8107 Intrinsic::ID MemCpyID;
8108 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8109 MemCpyID = Intrinsic::memcpy_i32;
8111 MemCpyID = Intrinsic::memcpy_i64;
8112 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8117 // If we can determine a pointer alignment that is bigger than currently
8118 // set, update the alignment.
8119 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8120 if (Instruction *I = SimplifyMemTransfer(MI))
8122 } else if (isa<MemSetInst>(MI)) {
8123 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
8124 if (MI->getAlignment()->getZExtValue() < Alignment) {
8125 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8130 if (Changed) return II;
8132 switch (II->getIntrinsicID()) {
8134 case Intrinsic::ppc_altivec_lvx:
8135 case Intrinsic::ppc_altivec_lvxl:
8136 case Intrinsic::x86_sse_loadu_ps:
8137 case Intrinsic::x86_sse2_loadu_pd:
8138 case Intrinsic::x86_sse2_loadu_dq:
8139 // Turn PPC lvx -> load if the pointer is known aligned.
8140 // Turn X86 loadups -> load if the pointer is known aligned.
8141 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8142 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8143 PointerType::getUnqual(II->getType()),
8145 return new LoadInst(Ptr);
8148 case Intrinsic::ppc_altivec_stvx:
8149 case Intrinsic::ppc_altivec_stvxl:
8150 // Turn stvx -> store if the pointer is known aligned.
8151 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
8152 const Type *OpPtrTy =
8153 PointerType::getUnqual(II->getOperand(1)->getType());
8154 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8155 return new StoreInst(II->getOperand(1), Ptr);
8158 case Intrinsic::x86_sse_storeu_ps:
8159 case Intrinsic::x86_sse2_storeu_pd:
8160 case Intrinsic::x86_sse2_storeu_dq:
8161 case Intrinsic::x86_sse2_storel_dq:
8162 // Turn X86 storeu -> store if the pointer is known aligned.
8163 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8164 const Type *OpPtrTy =
8165 PointerType::getUnqual(II->getOperand(2)->getType());
8166 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8167 return new StoreInst(II->getOperand(2), Ptr);
8171 case Intrinsic::x86_sse_cvttss2si: {
8172 // These intrinsics only demands the 0th element of its input vector. If
8173 // we can simplify the input based on that, do so now.
8175 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8177 II->setOperand(1, V);
8183 case Intrinsic::ppc_altivec_vperm:
8184 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8185 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8186 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8188 // Check that all of the elements are integer constants or undefs.
8189 bool AllEltsOk = true;
8190 for (unsigned i = 0; i != 16; ++i) {
8191 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8192 !isa<UndefValue>(Mask->getOperand(i))) {
8199 // Cast the input vectors to byte vectors.
8200 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8201 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8202 Value *Result = UndefValue::get(Op0->getType());
8204 // Only extract each element once.
8205 Value *ExtractedElts[32];
8206 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8208 for (unsigned i = 0; i != 16; ++i) {
8209 if (isa<UndefValue>(Mask->getOperand(i)))
8211 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8212 Idx &= 31; // Match the hardware behavior.
8214 if (ExtractedElts[Idx] == 0) {
8216 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8217 InsertNewInstBefore(Elt, CI);
8218 ExtractedElts[Idx] = Elt;
8221 // Insert this value into the result vector.
8222 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
8223 InsertNewInstBefore(cast<Instruction>(Result), CI);
8225 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8230 case Intrinsic::stackrestore: {
8231 // If the save is right next to the restore, remove the restore. This can
8232 // happen when variable allocas are DCE'd.
8233 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8234 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8235 BasicBlock::iterator BI = SS;
8237 return EraseInstFromFunction(CI);
8241 // Scan down this block to see if there is another stack restore in the
8242 // same block without an intervening call/alloca.
8243 BasicBlock::iterator BI = II;
8244 TerminatorInst *TI = II->getParent()->getTerminator();
8245 bool CannotRemove = false;
8246 for (++BI; &*BI != TI; ++BI) {
8247 if (isa<AllocaInst>(BI)) {
8248 CannotRemove = true;
8251 if (isa<CallInst>(BI)) {
8252 if (!isa<IntrinsicInst>(BI)) {
8253 CannotRemove = true;
8256 // If there is a stackrestore below this one, remove this one.
8257 return EraseInstFromFunction(CI);
8261 // If the stack restore is in a return/unwind block and if there are no
8262 // allocas or calls between the restore and the return, nuke the restore.
8263 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8264 return EraseInstFromFunction(CI);
8270 return visitCallSite(II);
8273 // InvokeInst simplification
8275 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8276 return visitCallSite(&II);
8279 // visitCallSite - Improvements for call and invoke instructions.
8281 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8282 bool Changed = false;
8284 // If the callee is a constexpr cast of a function, attempt to move the cast
8285 // to the arguments of the call/invoke.
8286 if (transformConstExprCastCall(CS)) return 0;
8288 Value *Callee = CS.getCalledValue();
8290 if (Function *CalleeF = dyn_cast<Function>(Callee))
8291 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8292 Instruction *OldCall = CS.getInstruction();
8293 // If the call and callee calling conventions don't match, this call must
8294 // be unreachable, as the call is undefined.
8295 new StoreInst(ConstantInt::getTrue(),
8296 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8298 if (!OldCall->use_empty())
8299 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8300 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8301 return EraseInstFromFunction(*OldCall);
8305 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8306 // This instruction is not reachable, just remove it. We insert a store to
8307 // undef so that we know that this code is not reachable, despite the fact
8308 // that we can't modify the CFG here.
8309 new StoreInst(ConstantInt::getTrue(),
8310 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8311 CS.getInstruction());
8313 if (!CS.getInstruction()->use_empty())
8314 CS.getInstruction()->
8315 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8317 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8318 // Don't break the CFG, insert a dummy cond branch.
8319 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8320 ConstantInt::getTrue(), II);
8322 return EraseInstFromFunction(*CS.getInstruction());
8325 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8326 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8327 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8328 return transformCallThroughTrampoline(CS);
8330 const PointerType *PTy = cast<PointerType>(Callee->getType());
8331 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8332 if (FTy->isVarArg()) {
8333 // See if we can optimize any arguments passed through the varargs area of
8335 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8336 E = CS.arg_end(); I != E; ++I)
8337 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8338 // If this cast does not effect the value passed through the varargs
8339 // area, we can eliminate the use of the cast.
8340 Value *Op = CI->getOperand(0);
8341 if (CI->isLosslessCast()) {
8348 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8349 // Inline asm calls cannot throw - mark them 'nounwind'.
8350 CS.setDoesNotThrow();
8354 return Changed ? CS.getInstruction() : 0;
8357 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8358 // attempt to move the cast to the arguments of the call/invoke.
8360 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8361 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8362 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8363 if (CE->getOpcode() != Instruction::BitCast ||
8364 !isa<Function>(CE->getOperand(0)))
8366 Function *Callee = cast<Function>(CE->getOperand(0));
8367 Instruction *Caller = CS.getInstruction();
8368 const ParamAttrsList* CallerPAL = CS.getParamAttrs();
8370 // Okay, this is a cast from a function to a different type. Unless doing so
8371 // would cause a type conversion of one of our arguments, change this call to
8372 // be a direct call with arguments casted to the appropriate types.
8374 const FunctionType *FT = Callee->getFunctionType();
8375 const Type *OldRetTy = Caller->getType();
8377 // Check to see if we are changing the return type...
8378 if (OldRetTy != FT->getReturnType()) {
8379 if (Callee->isDeclaration() && !Caller->use_empty() &&
8380 // Conversion is ok if changing from pointer to int of same size.
8381 !(isa<PointerType>(FT->getReturnType()) &&
8382 TD->getIntPtrType() == OldRetTy))
8383 return false; // Cannot transform this return value.
8385 if (!Caller->use_empty() &&
8386 // void -> non-void is handled specially
8387 FT->getReturnType() != Type::VoidTy &&
8388 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8389 return false; // Cannot transform this return value.
8391 if (CallerPAL && !Caller->use_empty()) {
8392 ParameterAttributes RAttrs = CallerPAL->getParamAttrs(0);
8393 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8394 return false; // Attribute not compatible with transformed value.
8397 // If the callsite is an invoke instruction, and the return value is used by
8398 // a PHI node in a successor, we cannot change the return type of the call
8399 // because there is no place to put the cast instruction (without breaking
8400 // the critical edge). Bail out in this case.
8401 if (!Caller->use_empty())
8402 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8403 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8405 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8406 if (PN->getParent() == II->getNormalDest() ||
8407 PN->getParent() == II->getUnwindDest())
8411 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8412 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8414 CallSite::arg_iterator AI = CS.arg_begin();
8415 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8416 const Type *ParamTy = FT->getParamType(i);
8417 const Type *ActTy = (*AI)->getType();
8419 if (!CastInst::isCastable(ActTy, ParamTy))
8420 return false; // Cannot transform this parameter value.
8423 ParameterAttributes PAttrs = CallerPAL->getParamAttrs(i + 1);
8424 if (PAttrs & ParamAttr::typeIncompatible(ParamTy))
8425 return false; // Attribute not compatible with transformed value.
8428 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8429 // Some conversions are safe even if we do not have a body.
8430 // Either we can cast directly, or we can upconvert the argument
8431 bool isConvertible = ActTy == ParamTy ||
8432 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8433 (ParamTy->isInteger() && ActTy->isInteger() &&
8434 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8435 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8436 && c->getValue().isStrictlyPositive());
8437 if (Callee->isDeclaration() && !isConvertible) return false;
8440 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8441 Callee->isDeclaration())
8442 return false; // Do not delete arguments unless we have a function body...
8444 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && CallerPAL)
8445 // In this case we have more arguments than the new function type, but we
8446 // won't be dropping them. Check that these extra arguments have attributes
8447 // that are compatible with being a vararg call argument.
8448 for (unsigned i = CallerPAL->size(); i; --i) {
8449 if (CallerPAL->getParamIndex(i - 1) <= FT->getNumParams())
8451 ParameterAttributes PAttrs = CallerPAL->getParamAttrsAtIndex(i - 1);
8452 if (PAttrs & ParamAttr::VarArgsIncompatible)
8456 // Okay, we decided that this is a safe thing to do: go ahead and start
8457 // inserting cast instructions as necessary...
8458 std::vector<Value*> Args;
8459 Args.reserve(NumActualArgs);
8460 ParamAttrsVector attrVec;
8461 attrVec.reserve(NumCommonArgs);
8463 // Get any return attributes.
8464 ParameterAttributes RAttrs = CallerPAL ? CallerPAL->getParamAttrs(0) :
8467 // If the return value is not being used, the type may not be compatible
8468 // with the existing attributes. Wipe out any problematic attributes.
8469 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
8471 // Add the new return attributes.
8473 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8475 AI = CS.arg_begin();
8476 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8477 const Type *ParamTy = FT->getParamType(i);
8478 if ((*AI)->getType() == ParamTy) {
8479 Args.push_back(*AI);
8481 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8482 false, ParamTy, false);
8483 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8484 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8487 // Add any parameter attributes.
8488 ParameterAttributes PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) :
8491 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8494 // If the function takes more arguments than the call was taking, add them
8496 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8497 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8499 // If we are removing arguments to the function, emit an obnoxious warning...
8500 if (FT->getNumParams() < NumActualArgs) {
8501 if (!FT->isVarArg()) {
8502 cerr << "WARNING: While resolving call to function '"
8503 << Callee->getName() << "' arguments were dropped!\n";
8505 // Add all of the arguments in their promoted form to the arg list...
8506 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8507 const Type *PTy = getPromotedType((*AI)->getType());
8508 if (PTy != (*AI)->getType()) {
8509 // Must promote to pass through va_arg area!
8510 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8512 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8513 InsertNewInstBefore(Cast, *Caller);
8514 Args.push_back(Cast);
8516 Args.push_back(*AI);
8519 // Add any parameter attributes.
8520 ParameterAttributes PAttrs = CallerPAL ?
8521 CallerPAL->getParamAttrs(i + 1) :
8524 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8529 if (FT->getReturnType() == Type::VoidTy)
8530 Caller->setName(""); // Void type should not have a name.
8532 const ParamAttrsList* NewCallerPAL = ParamAttrsList::get(attrVec);
8535 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8536 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8537 Args.begin(), Args.end(), Caller->getName(), Caller);
8538 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8539 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
8541 NC = new CallInst(Callee, Args.begin(), Args.end(),
8542 Caller->getName(), Caller);
8543 CallInst *CI = cast<CallInst>(Caller);
8544 if (CI->isTailCall())
8545 cast<CallInst>(NC)->setTailCall();
8546 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8547 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
8550 // Insert a cast of the return type as necessary.
8552 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8553 if (NV->getType() != Type::VoidTy) {
8554 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8556 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8558 // If this is an invoke instruction, we should insert it after the first
8559 // non-phi, instruction in the normal successor block.
8560 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8561 BasicBlock::iterator I = II->getNormalDest()->begin();
8562 while (isa<PHINode>(I)) ++I;
8563 InsertNewInstBefore(NC, *I);
8565 // Otherwise, it's a call, just insert cast right after the call instr
8566 InsertNewInstBefore(NC, *Caller);
8568 AddUsersToWorkList(*Caller);
8570 NV = UndefValue::get(Caller->getType());
8574 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8575 Caller->replaceAllUsesWith(NV);
8576 Caller->eraseFromParent();
8577 RemoveFromWorkList(Caller);
8581 // transformCallThroughTrampoline - Turn a call to a function created by the
8582 // init_trampoline intrinsic into a direct call to the underlying function.
8584 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8585 Value *Callee = CS.getCalledValue();
8586 const PointerType *PTy = cast<PointerType>(Callee->getType());
8587 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8588 const ParamAttrsList *Attrs = CS.getParamAttrs();
8590 // If the call already has the 'nest' attribute somewhere then give up -
8591 // otherwise 'nest' would occur twice after splicing in the chain.
8592 if (Attrs && Attrs->hasAttrSomewhere(ParamAttr::Nest))
8595 IntrinsicInst *Tramp =
8596 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8599 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8600 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8601 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8603 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8604 unsigned NestIdx = 1;
8605 const Type *NestTy = 0;
8606 ParameterAttributes NestAttr = ParamAttr::None;
8608 // Look for a parameter marked with the 'nest' attribute.
8609 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8610 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8611 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8612 // Record the parameter type and any other attributes.
8614 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8619 Instruction *Caller = CS.getInstruction();
8620 std::vector<Value*> NewArgs;
8621 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8623 ParamAttrsVector NewAttrs;
8624 NewAttrs.reserve(Attrs ? Attrs->size() + 1 : 1);
8626 // Insert the nest argument into the call argument list, which may
8627 // mean appending it. Likewise for attributes.
8629 // Add any function result attributes.
8630 ParameterAttributes Attr = Attrs ? Attrs->getParamAttrs(0) :
8633 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8637 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8639 if (Idx == NestIdx) {
8640 // Add the chain argument and attributes.
8641 Value *NestVal = Tramp->getOperand(3);
8642 if (NestVal->getType() != NestTy)
8643 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8644 NewArgs.push_back(NestVal);
8645 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8651 // Add the original argument and attributes.
8652 NewArgs.push_back(*I);
8653 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8656 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8662 // The trampoline may have been bitcast to a bogus type (FTy).
8663 // Handle this by synthesizing a new function type, equal to FTy
8664 // with the chain parameter inserted.
8666 std::vector<const Type*> NewTypes;
8667 NewTypes.reserve(FTy->getNumParams()+1);
8669 // Insert the chain's type into the list of parameter types, which may
8670 // mean appending it.
8673 FunctionType::param_iterator I = FTy->param_begin(),
8674 E = FTy->param_end();
8678 // Add the chain's type.
8679 NewTypes.push_back(NestTy);
8684 // Add the original type.
8685 NewTypes.push_back(*I);
8691 // Replace the trampoline call with a direct call. Let the generic
8692 // code sort out any function type mismatches.
8693 FunctionType *NewFTy =
8694 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8695 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8696 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8697 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8699 Instruction *NewCaller;
8700 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8701 NewCaller = new InvokeInst(NewCallee,
8702 II->getNormalDest(), II->getUnwindDest(),
8703 NewArgs.begin(), NewArgs.end(),
8704 Caller->getName(), Caller);
8705 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8706 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8708 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8709 Caller->getName(), Caller);
8710 if (cast<CallInst>(Caller)->isTailCall())
8711 cast<CallInst>(NewCaller)->setTailCall();
8712 cast<CallInst>(NewCaller)->
8713 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8714 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8716 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8717 Caller->replaceAllUsesWith(NewCaller);
8718 Caller->eraseFromParent();
8719 RemoveFromWorkList(Caller);
8724 // Replace the trampoline call with a direct call. Since there is no 'nest'
8725 // parameter, there is no need to adjust the argument list. Let the generic
8726 // code sort out any function type mismatches.
8727 Constant *NewCallee =
8728 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8729 CS.setCalledFunction(NewCallee);
8730 return CS.getInstruction();
8733 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8734 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8735 /// and a single binop.
8736 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8737 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8738 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8739 isa<CmpInst>(FirstInst));
8740 unsigned Opc = FirstInst->getOpcode();
8741 Value *LHSVal = FirstInst->getOperand(0);
8742 Value *RHSVal = FirstInst->getOperand(1);
8744 const Type *LHSType = LHSVal->getType();
8745 const Type *RHSType = RHSVal->getType();
8747 // Scan to see if all operands are the same opcode, all have one use, and all
8748 // kill their operands (i.e. the operands have one use).
8749 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8750 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8751 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8752 // Verify type of the LHS matches so we don't fold cmp's of different
8753 // types or GEP's with different index types.
8754 I->getOperand(0)->getType() != LHSType ||
8755 I->getOperand(1)->getType() != RHSType)
8758 // If they are CmpInst instructions, check their predicates
8759 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8760 if (cast<CmpInst>(I)->getPredicate() !=
8761 cast<CmpInst>(FirstInst)->getPredicate())
8764 // Keep track of which operand needs a phi node.
8765 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8766 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8769 // Otherwise, this is safe to transform, determine if it is profitable.
8771 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8772 // Indexes are often folded into load/store instructions, so we don't want to
8773 // hide them behind a phi.
8774 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8777 Value *InLHS = FirstInst->getOperand(0);
8778 Value *InRHS = FirstInst->getOperand(1);
8779 PHINode *NewLHS = 0, *NewRHS = 0;
8781 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8782 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8783 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8784 InsertNewInstBefore(NewLHS, PN);
8789 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8790 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8791 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8792 InsertNewInstBefore(NewRHS, PN);
8796 // Add all operands to the new PHIs.
8797 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8799 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8800 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8803 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8804 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8808 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8809 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8810 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8811 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8814 assert(isa<GetElementPtrInst>(FirstInst));
8815 return new GetElementPtrInst(LHSVal, RHSVal);
8819 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8820 /// of the block that defines it. This means that it must be obvious the value
8821 /// of the load is not changed from the point of the load to the end of the
8824 /// Finally, it is safe, but not profitable, to sink a load targetting a
8825 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8827 static bool isSafeToSinkLoad(LoadInst *L) {
8828 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8830 for (++BBI; BBI != E; ++BBI)
8831 if (BBI->mayWriteToMemory())
8834 // Check for non-address taken alloca. If not address-taken already, it isn't
8835 // profitable to do this xform.
8836 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8837 bool isAddressTaken = false;
8838 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8840 if (isa<LoadInst>(UI)) continue;
8841 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8842 // If storing TO the alloca, then the address isn't taken.
8843 if (SI->getOperand(1) == AI) continue;
8845 isAddressTaken = true;
8849 if (!isAddressTaken)
8857 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8858 // operator and they all are only used by the PHI, PHI together their
8859 // inputs, and do the operation once, to the result of the PHI.
8860 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8861 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8863 // Scan the instruction, looking for input operations that can be folded away.
8864 // If all input operands to the phi are the same instruction (e.g. a cast from
8865 // the same type or "+42") we can pull the operation through the PHI, reducing
8866 // code size and simplifying code.
8867 Constant *ConstantOp = 0;
8868 const Type *CastSrcTy = 0;
8869 bool isVolatile = false;
8870 if (isa<CastInst>(FirstInst)) {
8871 CastSrcTy = FirstInst->getOperand(0)->getType();
8872 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8873 // Can fold binop, compare or shift here if the RHS is a constant,
8874 // otherwise call FoldPHIArgBinOpIntoPHI.
8875 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8876 if (ConstantOp == 0)
8877 return FoldPHIArgBinOpIntoPHI(PN);
8878 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8879 isVolatile = LI->isVolatile();
8880 // We can't sink the load if the loaded value could be modified between the
8881 // load and the PHI.
8882 if (LI->getParent() != PN.getIncomingBlock(0) ||
8883 !isSafeToSinkLoad(LI))
8885 } else if (isa<GetElementPtrInst>(FirstInst)) {
8886 if (FirstInst->getNumOperands() == 2)
8887 return FoldPHIArgBinOpIntoPHI(PN);
8888 // Can't handle general GEPs yet.
8891 return 0; // Cannot fold this operation.
8894 // Check to see if all arguments are the same operation.
8895 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8896 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8897 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8898 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8901 if (I->getOperand(0)->getType() != CastSrcTy)
8902 return 0; // Cast operation must match.
8903 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8904 // We can't sink the load if the loaded value could be modified between
8905 // the load and the PHI.
8906 if (LI->isVolatile() != isVolatile ||
8907 LI->getParent() != PN.getIncomingBlock(i) ||
8908 !isSafeToSinkLoad(LI))
8910 } else if (I->getOperand(1) != ConstantOp) {
8915 // Okay, they are all the same operation. Create a new PHI node of the
8916 // correct type, and PHI together all of the LHS's of the instructions.
8917 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8918 PN.getName()+".in");
8919 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8921 Value *InVal = FirstInst->getOperand(0);
8922 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8924 // Add all operands to the new PHI.
8925 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8926 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8927 if (NewInVal != InVal)
8929 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8934 // The new PHI unions all of the same values together. This is really
8935 // common, so we handle it intelligently here for compile-time speed.
8939 InsertNewInstBefore(NewPN, PN);
8943 // Insert and return the new operation.
8944 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8945 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8946 else if (isa<LoadInst>(FirstInst))
8947 return new LoadInst(PhiVal, "", isVolatile);
8948 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8949 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8950 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8951 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8952 PhiVal, ConstantOp);
8954 assert(0 && "Unknown operation");
8958 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8960 static bool DeadPHICycle(PHINode *PN,
8961 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8962 if (PN->use_empty()) return true;
8963 if (!PN->hasOneUse()) return false;
8965 // Remember this node, and if we find the cycle, return.
8966 if (!PotentiallyDeadPHIs.insert(PN))
8969 // Don't scan crazily complex things.
8970 if (PotentiallyDeadPHIs.size() == 16)
8973 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8974 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8979 /// PHIsEqualValue - Return true if this phi node is always equal to
8980 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8981 /// z = some value; x = phi (y, z); y = phi (x, z)
8982 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8983 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8984 // See if we already saw this PHI node.
8985 if (!ValueEqualPHIs.insert(PN))
8988 // Don't scan crazily complex things.
8989 if (ValueEqualPHIs.size() == 16)
8992 // Scan the operands to see if they are either phi nodes or are equal to
8994 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8995 Value *Op = PN->getIncomingValue(i);
8996 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8997 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8999 } else if (Op != NonPhiInVal)
9007 // PHINode simplification
9009 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9010 // If LCSSA is around, don't mess with Phi nodes
9011 if (MustPreserveLCSSA) return 0;
9013 if (Value *V = PN.hasConstantValue())
9014 return ReplaceInstUsesWith(PN, V);
9016 // If all PHI operands are the same operation, pull them through the PHI,
9017 // reducing code size.
9018 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9019 PN.getIncomingValue(0)->hasOneUse())
9020 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9023 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9024 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9025 // PHI)... break the cycle.
9026 if (PN.hasOneUse()) {
9027 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9028 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9029 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9030 PotentiallyDeadPHIs.insert(&PN);
9031 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9032 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9035 // If this phi has a single use, and if that use just computes a value for
9036 // the next iteration of a loop, delete the phi. This occurs with unused
9037 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9038 // common case here is good because the only other things that catch this
9039 // are induction variable analysis (sometimes) and ADCE, which is only run
9041 if (PHIUser->hasOneUse() &&
9042 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9043 PHIUser->use_back() == &PN) {
9044 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9048 // We sometimes end up with phi cycles that non-obviously end up being the
9049 // same value, for example:
9050 // z = some value; x = phi (y, z); y = phi (x, z)
9051 // where the phi nodes don't necessarily need to be in the same block. Do a
9052 // quick check to see if the PHI node only contains a single non-phi value, if
9053 // so, scan to see if the phi cycle is actually equal to that value.
9055 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9056 // Scan for the first non-phi operand.
9057 while (InValNo != NumOperandVals &&
9058 isa<PHINode>(PN.getIncomingValue(InValNo)))
9061 if (InValNo != NumOperandVals) {
9062 Value *NonPhiInVal = PN.getOperand(InValNo);
9064 // Scan the rest of the operands to see if there are any conflicts, if so
9065 // there is no need to recursively scan other phis.
9066 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9067 Value *OpVal = PN.getIncomingValue(InValNo);
9068 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9072 // If we scanned over all operands, then we have one unique value plus
9073 // phi values. Scan PHI nodes to see if they all merge in each other or
9075 if (InValNo == NumOperandVals) {
9076 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9077 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9078 return ReplaceInstUsesWith(PN, NonPhiInVal);
9085 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9086 Instruction *InsertPoint,
9088 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9089 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9090 // We must cast correctly to the pointer type. Ensure that we
9091 // sign extend the integer value if it is smaller as this is
9092 // used for address computation.
9093 Instruction::CastOps opcode =
9094 (VTySize < PtrSize ? Instruction::SExt :
9095 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9096 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9100 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9101 Value *PtrOp = GEP.getOperand(0);
9102 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9103 // If so, eliminate the noop.
9104 if (GEP.getNumOperands() == 1)
9105 return ReplaceInstUsesWith(GEP, PtrOp);
9107 if (isa<UndefValue>(GEP.getOperand(0)))
9108 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9110 bool HasZeroPointerIndex = false;
9111 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9112 HasZeroPointerIndex = C->isNullValue();
9114 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9115 return ReplaceInstUsesWith(GEP, PtrOp);
9117 // Eliminate unneeded casts for indices.
9118 bool MadeChange = false;
9120 gep_type_iterator GTI = gep_type_begin(GEP);
9121 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
9122 if (isa<SequentialType>(*GTI)) {
9123 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
9124 if (CI->getOpcode() == Instruction::ZExt ||
9125 CI->getOpcode() == Instruction::SExt) {
9126 const Type *SrcTy = CI->getOperand(0)->getType();
9127 // We can eliminate a cast from i32 to i64 iff the target
9128 // is a 32-bit pointer target.
9129 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9131 GEP.setOperand(i, CI->getOperand(0));
9135 // If we are using a wider index than needed for this platform, shrink it
9136 // to what we need. If the incoming value needs a cast instruction,
9137 // insert it. This explicit cast can make subsequent optimizations more
9139 Value *Op = GEP.getOperand(i);
9140 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9141 if (Constant *C = dyn_cast<Constant>(Op)) {
9142 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
9145 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9147 GEP.setOperand(i, Op);
9153 if (MadeChange) return &GEP;
9155 // If this GEP instruction doesn't move the pointer, and if the input operand
9156 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9157 // real input to the dest type.
9158 if (GEP.hasAllZeroIndices()) {
9159 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9160 // If the bitcast is of an allocation, and the allocation will be
9161 // converted to match the type of the cast, don't touch this.
9162 if (isa<AllocationInst>(BCI->getOperand(0))) {
9163 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9164 if (Instruction *I = visitBitCast(*BCI)) {
9167 BCI->getParent()->getInstList().insert(BCI, I);
9168 ReplaceInstUsesWith(*BCI, I);
9173 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9177 // Combine Indices - If the source pointer to this getelementptr instruction
9178 // is a getelementptr instruction, combine the indices of the two
9179 // getelementptr instructions into a single instruction.
9181 SmallVector<Value*, 8> SrcGEPOperands;
9182 if (User *Src = dyn_castGetElementPtr(PtrOp))
9183 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9185 if (!SrcGEPOperands.empty()) {
9186 // Note that if our source is a gep chain itself that we wait for that
9187 // chain to be resolved before we perform this transformation. This
9188 // avoids us creating a TON of code in some cases.
9190 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9191 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9192 return 0; // Wait until our source is folded to completion.
9194 SmallVector<Value*, 8> Indices;
9196 // Find out whether the last index in the source GEP is a sequential idx.
9197 bool EndsWithSequential = false;
9198 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9199 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9200 EndsWithSequential = !isa<StructType>(*I);
9202 // Can we combine the two pointer arithmetics offsets?
9203 if (EndsWithSequential) {
9204 // Replace: gep (gep %P, long B), long A, ...
9205 // With: T = long A+B; gep %P, T, ...
9207 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9208 if (SO1 == Constant::getNullValue(SO1->getType())) {
9210 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9213 // If they aren't the same type, convert both to an integer of the
9214 // target's pointer size.
9215 if (SO1->getType() != GO1->getType()) {
9216 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9217 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9218 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9219 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9221 unsigned PS = TD->getPointerSizeInBits();
9222 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9223 // Convert GO1 to SO1's type.
9224 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9226 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9227 // Convert SO1 to GO1's type.
9228 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9230 const Type *PT = TD->getIntPtrType();
9231 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9232 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9236 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9237 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9239 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
9240 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9244 // Recycle the GEP we already have if possible.
9245 if (SrcGEPOperands.size() == 2) {
9246 GEP.setOperand(0, SrcGEPOperands[0]);
9247 GEP.setOperand(1, Sum);
9250 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9251 SrcGEPOperands.end()-1);
9252 Indices.push_back(Sum);
9253 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9255 } else if (isa<Constant>(*GEP.idx_begin()) &&
9256 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9257 SrcGEPOperands.size() != 1) {
9258 // Otherwise we can do the fold if the first index of the GEP is a zero
9259 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9260 SrcGEPOperands.end());
9261 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9264 if (!Indices.empty())
9265 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
9266 Indices.end(), GEP.getName());
9268 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9269 // GEP of global variable. If all of the indices for this GEP are
9270 // constants, we can promote this to a constexpr instead of an instruction.
9272 // Scan for nonconstants...
9273 SmallVector<Constant*, 8> Indices;
9274 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9275 for (; I != E && isa<Constant>(*I); ++I)
9276 Indices.push_back(cast<Constant>(*I));
9278 if (I == E) { // If they are all constants...
9279 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9280 &Indices[0],Indices.size());
9282 // Replace all uses of the GEP with the new constexpr...
9283 return ReplaceInstUsesWith(GEP, CE);
9285 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9286 if (!isa<PointerType>(X->getType())) {
9287 // Not interesting. Source pointer must be a cast from pointer.
9288 } else if (HasZeroPointerIndex) {
9289 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9290 // into : GEP [10 x i8]* X, i32 0, ...
9292 // This occurs when the program declares an array extern like "int X[];"
9294 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9295 const PointerType *XTy = cast<PointerType>(X->getType());
9296 if (const ArrayType *XATy =
9297 dyn_cast<ArrayType>(XTy->getElementType()))
9298 if (const ArrayType *CATy =
9299 dyn_cast<ArrayType>(CPTy->getElementType()))
9300 if (CATy->getElementType() == XATy->getElementType()) {
9301 // At this point, we know that the cast source type is a pointer
9302 // to an array of the same type as the destination pointer
9303 // array. Because the array type is never stepped over (there
9304 // is a leading zero) we can fold the cast into this GEP.
9305 GEP.setOperand(0, X);
9308 } else if (GEP.getNumOperands() == 2) {
9309 // Transform things like:
9310 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9311 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9312 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9313 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9314 if (isa<ArrayType>(SrcElTy) &&
9315 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9316 TD->getABITypeSize(ResElTy)) {
9318 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9319 Idx[1] = GEP.getOperand(1);
9320 Value *V = InsertNewInstBefore(
9321 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
9322 // V and GEP are both pointer types --> BitCast
9323 return new BitCastInst(V, GEP.getType());
9326 // Transform things like:
9327 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9328 // (where tmp = 8*tmp2) into:
9329 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9331 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9332 uint64_t ArrayEltSize =
9333 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9335 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9336 // allow either a mul, shift, or constant here.
9338 ConstantInt *Scale = 0;
9339 if (ArrayEltSize == 1) {
9340 NewIdx = GEP.getOperand(1);
9341 Scale = ConstantInt::get(NewIdx->getType(), 1);
9342 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9343 NewIdx = ConstantInt::get(CI->getType(), 1);
9345 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9346 if (Inst->getOpcode() == Instruction::Shl &&
9347 isa<ConstantInt>(Inst->getOperand(1))) {
9348 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9349 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9350 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9351 NewIdx = Inst->getOperand(0);
9352 } else if (Inst->getOpcode() == Instruction::Mul &&
9353 isa<ConstantInt>(Inst->getOperand(1))) {
9354 Scale = cast<ConstantInt>(Inst->getOperand(1));
9355 NewIdx = Inst->getOperand(0);
9359 // If the index will be to exactly the right offset with the scale taken
9360 // out, perform the transformation. Note, we don't know whether Scale is
9361 // signed or not. We'll use unsigned version of division/modulo
9362 // operation after making sure Scale doesn't have the sign bit set.
9363 if (Scale && Scale->getSExtValue() >= 0LL &&
9364 Scale->getZExtValue() % ArrayEltSize == 0) {
9365 Scale = ConstantInt::get(Scale->getType(),
9366 Scale->getZExtValue() / ArrayEltSize);
9367 if (Scale->getZExtValue() != 1) {
9368 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9370 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9371 NewIdx = InsertNewInstBefore(Sc, GEP);
9374 // Insert the new GEP instruction.
9376 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9378 Instruction *NewGEP =
9379 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
9380 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9381 // The NewGEP must be pointer typed, so must the old one -> BitCast
9382 return new BitCastInst(NewGEP, GEP.getType());
9391 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9392 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9393 if (AI.isArrayAllocation()) { // Check C != 1
9394 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9396 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9397 AllocationInst *New = 0;
9399 // Create and insert the replacement instruction...
9400 if (isa<MallocInst>(AI))
9401 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9403 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9404 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9407 InsertNewInstBefore(New, AI);
9409 // Scan to the end of the allocation instructions, to skip over a block of
9410 // allocas if possible...
9412 BasicBlock::iterator It = New;
9413 while (isa<AllocationInst>(*It)) ++It;
9415 // Now that I is pointing to the first non-allocation-inst in the block,
9416 // insert our getelementptr instruction...
9418 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9422 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9423 New->getName()+".sub", It);
9425 // Now make everything use the getelementptr instead of the original
9427 return ReplaceInstUsesWith(AI, V);
9428 } else if (isa<UndefValue>(AI.getArraySize())) {
9429 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9433 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9434 // Note that we only do this for alloca's, because malloc should allocate and
9435 // return a unique pointer, even for a zero byte allocation.
9436 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9437 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9438 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9443 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9444 Value *Op = FI.getOperand(0);
9446 // free undef -> unreachable.
9447 if (isa<UndefValue>(Op)) {
9448 // Insert a new store to null because we cannot modify the CFG here.
9449 new StoreInst(ConstantInt::getTrue(),
9450 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9451 return EraseInstFromFunction(FI);
9454 // If we have 'free null' delete the instruction. This can happen in stl code
9455 // when lots of inlining happens.
9456 if (isa<ConstantPointerNull>(Op))
9457 return EraseInstFromFunction(FI);
9459 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9460 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9461 FI.setOperand(0, CI->getOperand(0));
9465 // Change free (gep X, 0,0,0,0) into free(X)
9466 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9467 if (GEPI->hasAllZeroIndices()) {
9468 AddToWorkList(GEPI);
9469 FI.setOperand(0, GEPI->getOperand(0));
9474 // Change free(malloc) into nothing, if the malloc has a single use.
9475 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9476 if (MI->hasOneUse()) {
9477 EraseInstFromFunction(FI);
9478 return EraseInstFromFunction(*MI);
9485 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9486 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9487 const TargetData *TD) {
9488 User *CI = cast<User>(LI.getOperand(0));
9489 Value *CastOp = CI->getOperand(0);
9491 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9492 // Instead of loading constant c string, use corresponding integer value
9493 // directly if string length is small enough.
9494 const std::string &Str = CE->getOperand(0)->getStringValue();
9496 unsigned len = Str.length();
9497 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9498 unsigned numBits = Ty->getPrimitiveSizeInBits();
9499 // Replace LI with immediate integer store.
9500 if ((numBits >> 3) == len + 1) {
9501 APInt StrVal(numBits, 0);
9502 APInt SingleChar(numBits, 0);
9503 if (TD->isLittleEndian()) {
9504 for (signed i = len-1; i >= 0; i--) {
9505 SingleChar = (uint64_t) Str[i];
9506 StrVal = (StrVal << 8) | SingleChar;
9509 for (unsigned i = 0; i < len; i++) {
9510 SingleChar = (uint64_t) Str[i];
9511 StrVal = (StrVal << 8) | SingleChar;
9513 // Append NULL at the end.
9515 StrVal = (StrVal << 8) | SingleChar;
9517 Value *NL = ConstantInt::get(StrVal);
9518 return IC.ReplaceInstUsesWith(LI, NL);
9523 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9524 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9525 const Type *SrcPTy = SrcTy->getElementType();
9527 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9528 isa<VectorType>(DestPTy)) {
9529 // If the source is an array, the code below will not succeed. Check to
9530 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9532 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9533 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9534 if (ASrcTy->getNumElements() != 0) {
9536 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9537 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9538 SrcTy = cast<PointerType>(CastOp->getType());
9539 SrcPTy = SrcTy->getElementType();
9542 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9543 isa<VectorType>(SrcPTy)) &&
9544 // Do not allow turning this into a load of an integer, which is then
9545 // casted to a pointer, this pessimizes pointer analysis a lot.
9546 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9547 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9548 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9550 // Okay, we are casting from one integer or pointer type to another of
9551 // the same size. Instead of casting the pointer before the load, cast
9552 // the result of the loaded value.
9553 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9555 LI.isVolatile()),LI);
9556 // Now cast the result of the load.
9557 return new BitCastInst(NewLoad, LI.getType());
9564 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9565 /// from this value cannot trap. If it is not obviously safe to load from the
9566 /// specified pointer, we do a quick local scan of the basic block containing
9567 /// ScanFrom, to determine if the address is already accessed.
9568 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9569 // If it is an alloca it is always safe to load from.
9570 if (isa<AllocaInst>(V)) return true;
9572 // If it is a global variable it is mostly safe to load from.
9573 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9574 // Don't try to evaluate aliases. External weak GV can be null.
9575 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9577 // Otherwise, be a little bit agressive by scanning the local block where we
9578 // want to check to see if the pointer is already being loaded or stored
9579 // from/to. If so, the previous load or store would have already trapped,
9580 // so there is no harm doing an extra load (also, CSE will later eliminate
9581 // the load entirely).
9582 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9587 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9588 if (LI->getOperand(0) == V) return true;
9589 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9590 if (SI->getOperand(1) == V) return true;
9596 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9597 /// until we find the underlying object a pointer is referring to or something
9598 /// we don't understand. Note that the returned pointer may be offset from the
9599 /// input, because we ignore GEP indices.
9600 static Value *GetUnderlyingObject(Value *Ptr) {
9602 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9603 if (CE->getOpcode() == Instruction::BitCast ||
9604 CE->getOpcode() == Instruction::GetElementPtr)
9605 Ptr = CE->getOperand(0);
9608 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9609 Ptr = BCI->getOperand(0);
9610 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9611 Ptr = GEP->getOperand(0);
9618 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9619 Value *Op = LI.getOperand(0);
9621 // Attempt to improve the alignment.
9622 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9623 if (KnownAlign > LI.getAlignment())
9624 LI.setAlignment(KnownAlign);
9626 // load (cast X) --> cast (load X) iff safe
9627 if (isa<CastInst>(Op))
9628 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9631 // None of the following transforms are legal for volatile loads.
9632 if (LI.isVolatile()) return 0;
9634 if (&LI.getParent()->front() != &LI) {
9635 BasicBlock::iterator BBI = &LI; --BBI;
9636 // If the instruction immediately before this is a store to the same
9637 // address, do a simple form of store->load forwarding.
9638 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9639 if (SI->getOperand(1) == LI.getOperand(0))
9640 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9641 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9642 if (LIB->getOperand(0) == LI.getOperand(0))
9643 return ReplaceInstUsesWith(LI, LIB);
9646 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9647 const Value *GEPI0 = GEPI->getOperand(0);
9648 // TODO: Consider a target hook for valid address spaces for this xform.
9649 if (isa<ConstantPointerNull>(GEPI0) &&
9650 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
9651 // Insert a new store to null instruction before the load to indicate
9652 // that this code is not reachable. We do this instead of inserting
9653 // an unreachable instruction directly because we cannot modify the
9655 new StoreInst(UndefValue::get(LI.getType()),
9656 Constant::getNullValue(Op->getType()), &LI);
9657 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9661 if (Constant *C = dyn_cast<Constant>(Op)) {
9662 // load null/undef -> undef
9663 // TODO: Consider a target hook for valid address spaces for this xform.
9664 if (isa<UndefValue>(C) || (C->isNullValue() &&
9665 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
9666 // Insert a new store to null instruction before the load to indicate that
9667 // this code is not reachable. We do this instead of inserting an
9668 // unreachable instruction directly because we cannot modify the CFG.
9669 new StoreInst(UndefValue::get(LI.getType()),
9670 Constant::getNullValue(Op->getType()), &LI);
9671 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9674 // Instcombine load (constant global) into the value loaded.
9675 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9676 if (GV->isConstant() && !GV->isDeclaration())
9677 return ReplaceInstUsesWith(LI, GV->getInitializer());
9679 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9680 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
9681 if (CE->getOpcode() == Instruction::GetElementPtr) {
9682 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9683 if (GV->isConstant() && !GV->isDeclaration())
9685 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9686 return ReplaceInstUsesWith(LI, V);
9687 if (CE->getOperand(0)->isNullValue()) {
9688 // Insert a new store to null instruction before the load to indicate
9689 // that this code is not reachable. We do this instead of inserting
9690 // an unreachable instruction directly because we cannot modify the
9692 new StoreInst(UndefValue::get(LI.getType()),
9693 Constant::getNullValue(Op->getType()), &LI);
9694 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9697 } else if (CE->isCast()) {
9698 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9704 // If this load comes from anywhere in a constant global, and if the global
9705 // is all undef or zero, we know what it loads.
9706 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9707 if (GV->isConstant() && GV->hasInitializer()) {
9708 if (GV->getInitializer()->isNullValue())
9709 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9710 else if (isa<UndefValue>(GV->getInitializer()))
9711 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9715 if (Op->hasOneUse()) {
9716 // Change select and PHI nodes to select values instead of addresses: this
9717 // helps alias analysis out a lot, allows many others simplifications, and
9718 // exposes redundancy in the code.
9720 // Note that we cannot do the transformation unless we know that the
9721 // introduced loads cannot trap! Something like this is valid as long as
9722 // the condition is always false: load (select bool %C, int* null, int* %G),
9723 // but it would not be valid if we transformed it to load from null
9726 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9727 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9728 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9729 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9730 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9731 SI->getOperand(1)->getName()+".val"), LI);
9732 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9733 SI->getOperand(2)->getName()+".val"), LI);
9734 return new SelectInst(SI->getCondition(), V1, V2);
9737 // load (select (cond, null, P)) -> load P
9738 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9739 if (C->isNullValue()) {
9740 LI.setOperand(0, SI->getOperand(2));
9744 // load (select (cond, P, null)) -> load P
9745 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9746 if (C->isNullValue()) {
9747 LI.setOperand(0, SI->getOperand(1));
9755 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9757 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9758 User *CI = cast<User>(SI.getOperand(1));
9759 Value *CastOp = CI->getOperand(0);
9761 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9762 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9763 const Type *SrcPTy = SrcTy->getElementType();
9765 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9766 // If the source is an array, the code below will not succeed. Check to
9767 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9769 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9770 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9771 if (ASrcTy->getNumElements() != 0) {
9773 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9774 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9775 SrcTy = cast<PointerType>(CastOp->getType());
9776 SrcPTy = SrcTy->getElementType();
9779 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9780 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9781 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9783 // Okay, we are casting from one integer or pointer type to another of
9784 // the same size. Instead of casting the pointer before
9785 // the store, cast the value to be stored.
9787 Value *SIOp0 = SI.getOperand(0);
9788 Instruction::CastOps opcode = Instruction::BitCast;
9789 const Type* CastSrcTy = SIOp0->getType();
9790 const Type* CastDstTy = SrcPTy;
9791 if (isa<PointerType>(CastDstTy)) {
9792 if (CastSrcTy->isInteger())
9793 opcode = Instruction::IntToPtr;
9794 } else if (isa<IntegerType>(CastDstTy)) {
9795 if (isa<PointerType>(SIOp0->getType()))
9796 opcode = Instruction::PtrToInt;
9798 if (Constant *C = dyn_cast<Constant>(SIOp0))
9799 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9801 NewCast = IC.InsertNewInstBefore(
9802 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9804 return new StoreInst(NewCast, CastOp);
9811 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9812 Value *Val = SI.getOperand(0);
9813 Value *Ptr = SI.getOperand(1);
9815 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9816 EraseInstFromFunction(SI);
9821 // If the RHS is an alloca with a single use, zapify the store, making the
9823 if (Ptr->hasOneUse()) {
9824 if (isa<AllocaInst>(Ptr)) {
9825 EraseInstFromFunction(SI);
9830 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9831 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9832 GEP->getOperand(0)->hasOneUse()) {
9833 EraseInstFromFunction(SI);
9839 // Attempt to improve the alignment.
9840 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9841 if (KnownAlign > SI.getAlignment())
9842 SI.setAlignment(KnownAlign);
9844 // Do really simple DSE, to catch cases where there are several consequtive
9845 // stores to the same location, separated by a few arithmetic operations. This
9846 // situation often occurs with bitfield accesses.
9847 BasicBlock::iterator BBI = &SI;
9848 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9852 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9853 // Prev store isn't volatile, and stores to the same location?
9854 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9857 EraseInstFromFunction(*PrevSI);
9863 // If this is a load, we have to stop. However, if the loaded value is from
9864 // the pointer we're loading and is producing the pointer we're storing,
9865 // then *this* store is dead (X = load P; store X -> P).
9866 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9867 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9868 EraseInstFromFunction(SI);
9872 // Otherwise, this is a load from some other location. Stores before it
9877 // Don't skip over loads or things that can modify memory.
9878 if (BBI->mayWriteToMemory())
9883 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9885 // store X, null -> turns into 'unreachable' in SimplifyCFG
9886 if (isa<ConstantPointerNull>(Ptr)) {
9887 if (!isa<UndefValue>(Val)) {
9888 SI.setOperand(0, UndefValue::get(Val->getType()));
9889 if (Instruction *U = dyn_cast<Instruction>(Val))
9890 AddToWorkList(U); // Dropped a use.
9893 return 0; // Do not modify these!
9896 // store undef, Ptr -> noop
9897 if (isa<UndefValue>(Val)) {
9898 EraseInstFromFunction(SI);
9903 // If the pointer destination is a cast, see if we can fold the cast into the
9905 if (isa<CastInst>(Ptr))
9906 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9908 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9910 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9914 // If this store is the last instruction in the basic block, and if the block
9915 // ends with an unconditional branch, try to move it to the successor block.
9917 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9918 if (BI->isUnconditional())
9919 if (SimplifyStoreAtEndOfBlock(SI))
9920 return 0; // xform done!
9925 /// SimplifyStoreAtEndOfBlock - Turn things like:
9926 /// if () { *P = v1; } else { *P = v2 }
9927 /// into a phi node with a store in the successor.
9929 /// Simplify things like:
9930 /// *P = v1; if () { *P = v2; }
9931 /// into a phi node with a store in the successor.
9933 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9934 BasicBlock *StoreBB = SI.getParent();
9936 // Check to see if the successor block has exactly two incoming edges. If
9937 // so, see if the other predecessor contains a store to the same location.
9938 // if so, insert a PHI node (if needed) and move the stores down.
9939 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9941 // Determine whether Dest has exactly two predecessors and, if so, compute
9942 // the other predecessor.
9943 pred_iterator PI = pred_begin(DestBB);
9944 BasicBlock *OtherBB = 0;
9948 if (PI == pred_end(DestBB))
9951 if (*PI != StoreBB) {
9956 if (++PI != pred_end(DestBB))
9960 // Verify that the other block ends in a branch and is not otherwise empty.
9961 BasicBlock::iterator BBI = OtherBB->getTerminator();
9962 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9963 if (!OtherBr || BBI == OtherBB->begin())
9966 // If the other block ends in an unconditional branch, check for the 'if then
9967 // else' case. there is an instruction before the branch.
9968 StoreInst *OtherStore = 0;
9969 if (OtherBr->isUnconditional()) {
9970 // If this isn't a store, or isn't a store to the same location, bail out.
9972 OtherStore = dyn_cast<StoreInst>(BBI);
9973 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9976 // Otherwise, the other block ended with a conditional branch. If one of the
9977 // destinations is StoreBB, then we have the if/then case.
9978 if (OtherBr->getSuccessor(0) != StoreBB &&
9979 OtherBr->getSuccessor(1) != StoreBB)
9982 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9983 // if/then triangle. See if there is a store to the same ptr as SI that
9984 // lives in OtherBB.
9986 // Check to see if we find the matching store.
9987 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9988 if (OtherStore->getOperand(1) != SI.getOperand(1))
9992 // If we find something that may be using the stored value, or if we run
9993 // out of instructions, we can't do the xform.
9994 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9995 BBI == OtherBB->begin())
9999 // In order to eliminate the store in OtherBr, we have to
10000 // make sure nothing reads the stored value in StoreBB.
10001 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10002 // FIXME: This should really be AA driven.
10003 if (isa<LoadInst>(I) || I->mayWriteToMemory())
10008 // Insert a PHI node now if we need it.
10009 Value *MergedVal = OtherStore->getOperand(0);
10010 if (MergedVal != SI.getOperand(0)) {
10011 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
10012 PN->reserveOperandSpace(2);
10013 PN->addIncoming(SI.getOperand(0), SI.getParent());
10014 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10015 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10018 // Advance to a place where it is safe to insert the new store and
10020 BBI = DestBB->begin();
10021 while (isa<PHINode>(BBI)) ++BBI;
10022 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10023 OtherStore->isVolatile()), *BBI);
10025 // Nuke the old stores.
10026 EraseInstFromFunction(SI);
10027 EraseInstFromFunction(*OtherStore);
10033 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10034 // Change br (not X), label True, label False to: br X, label False, True
10036 BasicBlock *TrueDest;
10037 BasicBlock *FalseDest;
10038 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10039 !isa<Constant>(X)) {
10040 // Swap Destinations and condition...
10041 BI.setCondition(X);
10042 BI.setSuccessor(0, FalseDest);
10043 BI.setSuccessor(1, TrueDest);
10047 // Cannonicalize fcmp_one -> fcmp_oeq
10048 FCmpInst::Predicate FPred; Value *Y;
10049 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10050 TrueDest, FalseDest)))
10051 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10052 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10053 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10054 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10055 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10056 NewSCC->takeName(I);
10057 // Swap Destinations and condition...
10058 BI.setCondition(NewSCC);
10059 BI.setSuccessor(0, FalseDest);
10060 BI.setSuccessor(1, TrueDest);
10061 RemoveFromWorkList(I);
10062 I->eraseFromParent();
10063 AddToWorkList(NewSCC);
10067 // Cannonicalize icmp_ne -> icmp_eq
10068 ICmpInst::Predicate IPred;
10069 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10070 TrueDest, FalseDest)))
10071 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10072 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10073 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10074 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10075 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10076 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10077 NewSCC->takeName(I);
10078 // Swap Destinations and condition...
10079 BI.setCondition(NewSCC);
10080 BI.setSuccessor(0, FalseDest);
10081 BI.setSuccessor(1, TrueDest);
10082 RemoveFromWorkList(I);
10083 I->eraseFromParent();;
10084 AddToWorkList(NewSCC);
10091 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10092 Value *Cond = SI.getCondition();
10093 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10094 if (I->getOpcode() == Instruction::Add)
10095 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10096 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10097 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10098 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10100 SI.setOperand(0, I->getOperand(0));
10108 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10109 /// is to leave as a vector operation.
10110 static bool CheapToScalarize(Value *V, bool isConstant) {
10111 if (isa<ConstantAggregateZero>(V))
10113 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10114 if (isConstant) return true;
10115 // If all elts are the same, we can extract.
10116 Constant *Op0 = C->getOperand(0);
10117 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10118 if (C->getOperand(i) != Op0)
10122 Instruction *I = dyn_cast<Instruction>(V);
10123 if (!I) return false;
10125 // Insert element gets simplified to the inserted element or is deleted if
10126 // this is constant idx extract element and its a constant idx insertelt.
10127 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10128 isa<ConstantInt>(I->getOperand(2)))
10130 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10132 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10133 if (BO->hasOneUse() &&
10134 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10135 CheapToScalarize(BO->getOperand(1), isConstant)))
10137 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10138 if (CI->hasOneUse() &&
10139 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10140 CheapToScalarize(CI->getOperand(1), isConstant)))
10146 /// Read and decode a shufflevector mask.
10148 /// It turns undef elements into values that are larger than the number of
10149 /// elements in the input.
10150 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10151 unsigned NElts = SVI->getType()->getNumElements();
10152 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10153 return std::vector<unsigned>(NElts, 0);
10154 if (isa<UndefValue>(SVI->getOperand(2)))
10155 return std::vector<unsigned>(NElts, 2*NElts);
10157 std::vector<unsigned> Result;
10158 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10159 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
10160 if (isa<UndefValue>(CP->getOperand(i)))
10161 Result.push_back(NElts*2); // undef -> 8
10163 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
10167 /// FindScalarElement - Given a vector and an element number, see if the scalar
10168 /// value is already around as a register, for example if it were inserted then
10169 /// extracted from the vector.
10170 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10171 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10172 const VectorType *PTy = cast<VectorType>(V->getType());
10173 unsigned Width = PTy->getNumElements();
10174 if (EltNo >= Width) // Out of range access.
10175 return UndefValue::get(PTy->getElementType());
10177 if (isa<UndefValue>(V))
10178 return UndefValue::get(PTy->getElementType());
10179 else if (isa<ConstantAggregateZero>(V))
10180 return Constant::getNullValue(PTy->getElementType());
10181 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10182 return CP->getOperand(EltNo);
10183 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10184 // If this is an insert to a variable element, we don't know what it is.
10185 if (!isa<ConstantInt>(III->getOperand(2)))
10187 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10189 // If this is an insert to the element we are looking for, return the
10191 if (EltNo == IIElt)
10192 return III->getOperand(1);
10194 // Otherwise, the insertelement doesn't modify the value, recurse on its
10196 return FindScalarElement(III->getOperand(0), EltNo);
10197 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10198 unsigned InEl = getShuffleMask(SVI)[EltNo];
10200 return FindScalarElement(SVI->getOperand(0), InEl);
10201 else if (InEl < Width*2)
10202 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10204 return UndefValue::get(PTy->getElementType());
10207 // Otherwise, we don't know.
10211 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10213 // If vector val is undef, replace extract with scalar undef.
10214 if (isa<UndefValue>(EI.getOperand(0)))
10215 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10217 // If vector val is constant 0, replace extract with scalar 0.
10218 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10219 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10221 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10222 // If vector val is constant with uniform operands, replace EI
10223 // with that operand
10224 Constant *op0 = C->getOperand(0);
10225 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10226 if (C->getOperand(i) != op0) {
10231 return ReplaceInstUsesWith(EI, op0);
10234 // If extracting a specified index from the vector, see if we can recursively
10235 // find a previously computed scalar that was inserted into the vector.
10236 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10237 unsigned IndexVal = IdxC->getZExtValue();
10238 unsigned VectorWidth =
10239 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10241 // If this is extracting an invalid index, turn this into undef, to avoid
10242 // crashing the code below.
10243 if (IndexVal >= VectorWidth)
10244 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10246 // This instruction only demands the single element from the input vector.
10247 // If the input vector has a single use, simplify it based on this use
10249 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10250 uint64_t UndefElts;
10251 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10254 EI.setOperand(0, V);
10259 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10260 return ReplaceInstUsesWith(EI, Elt);
10262 // If the this extractelement is directly using a bitcast from a vector of
10263 // the same number of elements, see if we can find the source element from
10264 // it. In this case, we will end up needing to bitcast the scalars.
10265 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10266 if (const VectorType *VT =
10267 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10268 if (VT->getNumElements() == VectorWidth)
10269 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10270 return new BitCastInst(Elt, EI.getType());
10274 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10275 if (I->hasOneUse()) {
10276 // Push extractelement into predecessor operation if legal and
10277 // profitable to do so
10278 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10279 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10280 if (CheapToScalarize(BO, isConstantElt)) {
10281 ExtractElementInst *newEI0 =
10282 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10283 EI.getName()+".lhs");
10284 ExtractElementInst *newEI1 =
10285 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10286 EI.getName()+".rhs");
10287 InsertNewInstBefore(newEI0, EI);
10288 InsertNewInstBefore(newEI1, EI);
10289 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10291 } else if (isa<LoadInst>(I)) {
10293 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10294 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10295 PointerType::get(EI.getType(), AS),EI);
10296 GetElementPtrInst *GEP =
10297 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
10298 InsertNewInstBefore(GEP, EI);
10299 return new LoadInst(GEP);
10302 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10303 // Extracting the inserted element?
10304 if (IE->getOperand(2) == EI.getOperand(1))
10305 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10306 // If the inserted and extracted elements are constants, they must not
10307 // be the same value, extract from the pre-inserted value instead.
10308 if (isa<Constant>(IE->getOperand(2)) &&
10309 isa<Constant>(EI.getOperand(1))) {
10310 AddUsesToWorkList(EI);
10311 EI.setOperand(0, IE->getOperand(0));
10314 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10315 // If this is extracting an element from a shufflevector, figure out where
10316 // it came from and extract from the appropriate input element instead.
10317 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10318 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10320 if (SrcIdx < SVI->getType()->getNumElements())
10321 Src = SVI->getOperand(0);
10322 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10323 SrcIdx -= SVI->getType()->getNumElements();
10324 Src = SVI->getOperand(1);
10326 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10328 return new ExtractElementInst(Src, SrcIdx);
10335 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10336 /// elements from either LHS or RHS, return the shuffle mask and true.
10337 /// Otherwise, return false.
10338 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10339 std::vector<Constant*> &Mask) {
10340 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10341 "Invalid CollectSingleShuffleElements");
10342 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10344 if (isa<UndefValue>(V)) {
10345 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10347 } else if (V == LHS) {
10348 for (unsigned i = 0; i != NumElts; ++i)
10349 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10351 } else if (V == RHS) {
10352 for (unsigned i = 0; i != NumElts; ++i)
10353 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10355 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10356 // If this is an insert of an extract from some other vector, include it.
10357 Value *VecOp = IEI->getOperand(0);
10358 Value *ScalarOp = IEI->getOperand(1);
10359 Value *IdxOp = IEI->getOperand(2);
10361 if (!isa<ConstantInt>(IdxOp))
10363 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10365 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10366 // Okay, we can handle this if the vector we are insertinting into is
10367 // transitively ok.
10368 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10369 // If so, update the mask to reflect the inserted undef.
10370 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10373 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10374 if (isa<ConstantInt>(EI->getOperand(1)) &&
10375 EI->getOperand(0)->getType() == V->getType()) {
10376 unsigned ExtractedIdx =
10377 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10379 // This must be extracting from either LHS or RHS.
10380 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10381 // Okay, we can handle this if the vector we are insertinting into is
10382 // transitively ok.
10383 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10384 // If so, update the mask to reflect the inserted value.
10385 if (EI->getOperand(0) == LHS) {
10386 Mask[InsertedIdx & (NumElts-1)] =
10387 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10389 assert(EI->getOperand(0) == RHS);
10390 Mask[InsertedIdx & (NumElts-1)] =
10391 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10400 // TODO: Handle shufflevector here!
10405 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10406 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10407 /// that computes V and the LHS value of the shuffle.
10408 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10410 assert(isa<VectorType>(V->getType()) &&
10411 (RHS == 0 || V->getType() == RHS->getType()) &&
10412 "Invalid shuffle!");
10413 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10415 if (isa<UndefValue>(V)) {
10416 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10418 } else if (isa<ConstantAggregateZero>(V)) {
10419 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10421 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10422 // If this is an insert of an extract from some other vector, include it.
10423 Value *VecOp = IEI->getOperand(0);
10424 Value *ScalarOp = IEI->getOperand(1);
10425 Value *IdxOp = IEI->getOperand(2);
10427 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10428 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10429 EI->getOperand(0)->getType() == V->getType()) {
10430 unsigned ExtractedIdx =
10431 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10432 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10434 // Either the extracted from or inserted into vector must be RHSVec,
10435 // otherwise we'd end up with a shuffle of three inputs.
10436 if (EI->getOperand(0) == RHS || RHS == 0) {
10437 RHS = EI->getOperand(0);
10438 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10439 Mask[InsertedIdx & (NumElts-1)] =
10440 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10444 if (VecOp == RHS) {
10445 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10446 // Everything but the extracted element is replaced with the RHS.
10447 for (unsigned i = 0; i != NumElts; ++i) {
10448 if (i != InsertedIdx)
10449 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10454 // If this insertelement is a chain that comes from exactly these two
10455 // vectors, return the vector and the effective shuffle.
10456 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10457 return EI->getOperand(0);
10462 // TODO: Handle shufflevector here!
10464 // Otherwise, can't do anything fancy. Return an identity vector.
10465 for (unsigned i = 0; i != NumElts; ++i)
10466 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10470 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10471 Value *VecOp = IE.getOperand(0);
10472 Value *ScalarOp = IE.getOperand(1);
10473 Value *IdxOp = IE.getOperand(2);
10475 // Inserting an undef or into an undefined place, remove this.
10476 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10477 ReplaceInstUsesWith(IE, VecOp);
10479 // If the inserted element was extracted from some other vector, and if the
10480 // indexes are constant, try to turn this into a shufflevector operation.
10481 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10482 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10483 EI->getOperand(0)->getType() == IE.getType()) {
10484 unsigned NumVectorElts = IE.getType()->getNumElements();
10485 unsigned ExtractedIdx =
10486 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10487 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10489 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10490 return ReplaceInstUsesWith(IE, VecOp);
10492 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10493 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10495 // If we are extracting a value from a vector, then inserting it right
10496 // back into the same place, just use the input vector.
10497 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10498 return ReplaceInstUsesWith(IE, VecOp);
10500 // We could theoretically do this for ANY input. However, doing so could
10501 // turn chains of insertelement instructions into a chain of shufflevector
10502 // instructions, and right now we do not merge shufflevectors. As such,
10503 // only do this in a situation where it is clear that there is benefit.
10504 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10505 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10506 // the values of VecOp, except then one read from EIOp0.
10507 // Build a new shuffle mask.
10508 std::vector<Constant*> Mask;
10509 if (isa<UndefValue>(VecOp))
10510 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10512 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10513 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10516 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10517 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10518 ConstantVector::get(Mask));
10521 // If this insertelement isn't used by some other insertelement, turn it
10522 // (and any insertelements it points to), into one big shuffle.
10523 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10524 std::vector<Constant*> Mask;
10526 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10527 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10528 // We now have a shuffle of LHS, RHS, Mask.
10529 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10538 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10539 Value *LHS = SVI.getOperand(0);
10540 Value *RHS = SVI.getOperand(1);
10541 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10543 bool MadeChange = false;
10545 // Undefined shuffle mask -> undefined value.
10546 if (isa<UndefValue>(SVI.getOperand(2)))
10547 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10549 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10550 // the undef, change them to undefs.
10551 if (isa<UndefValue>(SVI.getOperand(1))) {
10552 // Scan to see if there are any references to the RHS. If so, replace them
10553 // with undef element refs and set MadeChange to true.
10554 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10555 if (Mask[i] >= e && Mask[i] != 2*e) {
10562 // Remap any references to RHS to use LHS.
10563 std::vector<Constant*> Elts;
10564 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10565 if (Mask[i] == 2*e)
10566 Elts.push_back(UndefValue::get(Type::Int32Ty));
10568 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10570 SVI.setOperand(2, ConstantVector::get(Elts));
10574 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10575 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10576 if (LHS == RHS || isa<UndefValue>(LHS)) {
10577 if (isa<UndefValue>(LHS) && LHS == RHS) {
10578 // shuffle(undef,undef,mask) -> undef.
10579 return ReplaceInstUsesWith(SVI, LHS);
10582 // Remap any references to RHS to use LHS.
10583 std::vector<Constant*> Elts;
10584 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10585 if (Mask[i] >= 2*e)
10586 Elts.push_back(UndefValue::get(Type::Int32Ty));
10588 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10589 (Mask[i] < e && isa<UndefValue>(LHS)))
10590 Mask[i] = 2*e; // Turn into undef.
10592 Mask[i] &= (e-1); // Force to LHS.
10593 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10596 SVI.setOperand(0, SVI.getOperand(1));
10597 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10598 SVI.setOperand(2, ConstantVector::get(Elts));
10599 LHS = SVI.getOperand(0);
10600 RHS = SVI.getOperand(1);
10604 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10605 bool isLHSID = true, isRHSID = true;
10607 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10608 if (Mask[i] >= e*2) continue; // Ignore undef values.
10609 // Is this an identity shuffle of the LHS value?
10610 isLHSID &= (Mask[i] == i);
10612 // Is this an identity shuffle of the RHS value?
10613 isRHSID &= (Mask[i]-e == i);
10616 // Eliminate identity shuffles.
10617 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10618 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10620 // If the LHS is a shufflevector itself, see if we can combine it with this
10621 // one without producing an unusual shuffle. Here we are really conservative:
10622 // we are absolutely afraid of producing a shuffle mask not in the input
10623 // program, because the code gen may not be smart enough to turn a merged
10624 // shuffle into two specific shuffles: it may produce worse code. As such,
10625 // we only merge two shuffles if the result is one of the two input shuffle
10626 // masks. In this case, merging the shuffles just removes one instruction,
10627 // which we know is safe. This is good for things like turning:
10628 // (splat(splat)) -> splat.
10629 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10630 if (isa<UndefValue>(RHS)) {
10631 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10633 std::vector<unsigned> NewMask;
10634 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10635 if (Mask[i] >= 2*e)
10636 NewMask.push_back(2*e);
10638 NewMask.push_back(LHSMask[Mask[i]]);
10640 // If the result mask is equal to the src shuffle or this shuffle mask, do
10641 // the replacement.
10642 if (NewMask == LHSMask || NewMask == Mask) {
10643 std::vector<Constant*> Elts;
10644 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10645 if (NewMask[i] >= e*2) {
10646 Elts.push_back(UndefValue::get(Type::Int32Ty));
10648 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10651 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10652 LHSSVI->getOperand(1),
10653 ConstantVector::get(Elts));
10658 return MadeChange ? &SVI : 0;
10664 /// TryToSinkInstruction - Try to move the specified instruction from its
10665 /// current block into the beginning of DestBlock, which can only happen if it's
10666 /// safe to move the instruction past all of the instructions between it and the
10667 /// end of its block.
10668 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10669 assert(I->hasOneUse() && "Invariants didn't hold!");
10671 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10672 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10674 // Do not sink alloca instructions out of the entry block.
10675 if (isa<AllocaInst>(I) && I->getParent() ==
10676 &DestBlock->getParent()->getEntryBlock())
10679 // We can only sink load instructions if there is nothing between the load and
10680 // the end of block that could change the value.
10681 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10682 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10684 if (Scan->mayWriteToMemory())
10688 BasicBlock::iterator InsertPos = DestBlock->begin();
10689 while (isa<PHINode>(InsertPos)) ++InsertPos;
10691 I->moveBefore(InsertPos);
10697 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10698 /// all reachable code to the worklist.
10700 /// This has a couple of tricks to make the code faster and more powerful. In
10701 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10702 /// them to the worklist (this significantly speeds up instcombine on code where
10703 /// many instructions are dead or constant). Additionally, if we find a branch
10704 /// whose condition is a known constant, we only visit the reachable successors.
10706 static void AddReachableCodeToWorklist(BasicBlock *BB,
10707 SmallPtrSet<BasicBlock*, 64> &Visited,
10709 const TargetData *TD) {
10710 std::vector<BasicBlock*> Worklist;
10711 Worklist.push_back(BB);
10713 while (!Worklist.empty()) {
10714 BB = Worklist.back();
10715 Worklist.pop_back();
10717 // We have now visited this block! If we've already been here, ignore it.
10718 if (!Visited.insert(BB)) continue;
10720 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10721 Instruction *Inst = BBI++;
10723 // DCE instruction if trivially dead.
10724 if (isInstructionTriviallyDead(Inst)) {
10726 DOUT << "IC: DCE: " << *Inst;
10727 Inst->eraseFromParent();
10731 // ConstantProp instruction if trivially constant.
10732 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10733 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10734 Inst->replaceAllUsesWith(C);
10736 Inst->eraseFromParent();
10740 IC.AddToWorkList(Inst);
10743 // Recursively visit successors. If this is a branch or switch on a
10744 // constant, only visit the reachable successor.
10745 TerminatorInst *TI = BB->getTerminator();
10746 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10747 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10748 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10749 Worklist.push_back(BI->getSuccessor(!CondVal));
10752 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10753 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10754 // See if this is an explicit destination.
10755 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10756 if (SI->getCaseValue(i) == Cond) {
10757 Worklist.push_back(SI->getSuccessor(i));
10761 // Otherwise it is the default destination.
10762 Worklist.push_back(SI->getSuccessor(0));
10767 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10768 Worklist.push_back(TI->getSuccessor(i));
10772 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10773 bool Changed = false;
10774 TD = &getAnalysis<TargetData>();
10776 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10777 << F.getNameStr() << "\n");
10780 // Do a depth-first traversal of the function, populate the worklist with
10781 // the reachable instructions. Ignore blocks that are not reachable. Keep
10782 // track of which blocks we visit.
10783 SmallPtrSet<BasicBlock*, 64> Visited;
10784 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10786 // Do a quick scan over the function. If we find any blocks that are
10787 // unreachable, remove any instructions inside of them. This prevents
10788 // the instcombine code from having to deal with some bad special cases.
10789 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10790 if (!Visited.count(BB)) {
10791 Instruction *Term = BB->getTerminator();
10792 while (Term != BB->begin()) { // Remove instrs bottom-up
10793 BasicBlock::iterator I = Term; --I;
10795 DOUT << "IC: DCE: " << *I;
10798 if (!I->use_empty())
10799 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10800 I->eraseFromParent();
10805 while (!Worklist.empty()) {
10806 Instruction *I = RemoveOneFromWorkList();
10807 if (I == 0) continue; // skip null values.
10809 // Check to see if we can DCE the instruction.
10810 if (isInstructionTriviallyDead(I)) {
10811 // Add operands to the worklist.
10812 if (I->getNumOperands() < 4)
10813 AddUsesToWorkList(*I);
10816 DOUT << "IC: DCE: " << *I;
10818 I->eraseFromParent();
10819 RemoveFromWorkList(I);
10823 // Instruction isn't dead, see if we can constant propagate it.
10824 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10825 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10827 // Add operands to the worklist.
10828 AddUsesToWorkList(*I);
10829 ReplaceInstUsesWith(*I, C);
10832 I->eraseFromParent();
10833 RemoveFromWorkList(I);
10837 // See if we can trivially sink this instruction to a successor basic block.
10838 if (I->hasOneUse()) {
10839 BasicBlock *BB = I->getParent();
10840 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10841 if (UserParent != BB) {
10842 bool UserIsSuccessor = false;
10843 // See if the user is one of our successors.
10844 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10845 if (*SI == UserParent) {
10846 UserIsSuccessor = true;
10850 // If the user is one of our immediate successors, and if that successor
10851 // only has us as a predecessors (we'd have to split the critical edge
10852 // otherwise), we can keep going.
10853 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10854 next(pred_begin(UserParent)) == pred_end(UserParent))
10855 // Okay, the CFG is simple enough, try to sink this instruction.
10856 Changed |= TryToSinkInstruction(I, UserParent);
10860 // Now that we have an instruction, try combining it to simplify it...
10864 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10865 if (Instruction *Result = visit(*I)) {
10867 // Should we replace the old instruction with a new one?
10869 DOUT << "IC: Old = " << *I
10870 << " New = " << *Result;
10872 // Everything uses the new instruction now.
10873 I->replaceAllUsesWith(Result);
10875 // Push the new instruction and any users onto the worklist.
10876 AddToWorkList(Result);
10877 AddUsersToWorkList(*Result);
10879 // Move the name to the new instruction first.
10880 Result->takeName(I);
10882 // Insert the new instruction into the basic block...
10883 BasicBlock *InstParent = I->getParent();
10884 BasicBlock::iterator InsertPos = I;
10886 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10887 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10890 InstParent->getInstList().insert(InsertPos, Result);
10892 // Make sure that we reprocess all operands now that we reduced their
10894 AddUsesToWorkList(*I);
10896 // Instructions can end up on the worklist more than once. Make sure
10897 // we do not process an instruction that has been deleted.
10898 RemoveFromWorkList(I);
10900 // Erase the old instruction.
10901 InstParent->getInstList().erase(I);
10904 DOUT << "IC: Mod = " << OrigI
10905 << " New = " << *I;
10908 // If the instruction was modified, it's possible that it is now dead.
10909 // if so, remove it.
10910 if (isInstructionTriviallyDead(I)) {
10911 // Make sure we process all operands now that we are reducing their
10913 AddUsesToWorkList(*I);
10915 // Instructions may end up in the worklist more than once. Erase all
10916 // occurrences of this instruction.
10917 RemoveFromWorkList(I);
10918 I->eraseFromParent();
10921 AddUsersToWorkList(*I);
10928 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10930 // Do an explicit clear, this shrinks the map if needed.
10931 WorklistMap.clear();
10936 bool InstCombiner::runOnFunction(Function &F) {
10937 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10939 bool EverMadeChange = false;
10941 // Iterate while there is work to do.
10942 unsigned Iteration = 0;
10943 while (DoOneIteration(F, Iteration++))
10944 EverMadeChange = true;
10945 return EverMadeChange;
10948 FunctionPass *llvm::createInstructionCombiningPass() {
10949 return new InstCombiner();