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());
609 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
610 /// known to be either zero or one and return them in the KnownZero/KnownOne
611 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
613 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
614 /// we cannot optimize based on the assumption that it is zero without changing
615 /// it to be an explicit zero. If we don't change it to zero, other code could
616 /// optimized based on the contradictory assumption that it is non-zero.
617 /// Because instcombine aggressively folds operations with undef args anyway,
618 /// this won't lose us code quality.
619 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
620 APInt& KnownOne, unsigned Depth = 0) {
621 assert(V && "No Value?");
622 assert(Depth <= 6 && "Limit Search Depth");
623 uint32_t BitWidth = Mask.getBitWidth();
624 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
625 KnownZero.getBitWidth() == BitWidth &&
626 KnownOne.getBitWidth() == BitWidth &&
627 "V, Mask, KnownOne and KnownZero should have same BitWidth");
628 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
629 // We know all of the bits for a constant!
630 KnownOne = CI->getValue() & Mask;
631 KnownZero = ~KnownOne & Mask;
635 if (Depth == 6 || Mask == 0)
636 return; // Limit search depth.
638 Instruction *I = dyn_cast<Instruction>(V);
641 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
642 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
644 switch (I->getOpcode()) {
645 case Instruction::And: {
646 // If either the LHS or the RHS are Zero, the result is zero.
647 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
648 APInt Mask2(Mask & ~KnownZero);
649 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
650 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
651 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
653 // Output known-1 bits are only known if set in both the LHS & RHS.
654 KnownOne &= KnownOne2;
655 // Output known-0 are known to be clear if zero in either the LHS | RHS.
656 KnownZero |= KnownZero2;
659 case Instruction::Or: {
660 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
661 APInt Mask2(Mask & ~KnownOne);
662 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
663 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
664 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
666 // Output known-0 bits are only known if clear in both the LHS & RHS.
667 KnownZero &= KnownZero2;
668 // Output known-1 are known to be set if set in either the LHS | RHS.
669 KnownOne |= KnownOne2;
672 case Instruction::Xor: {
673 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
674 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
675 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
676 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
678 // Output known-0 bits are known if clear or set in both the LHS & RHS.
679 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
680 // Output known-1 are known to be set if set in only one of the LHS, RHS.
681 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
682 KnownZero = KnownZeroOut;
685 case Instruction::Select:
686 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
687 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
688 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
689 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
691 // Only known if known in both the LHS and RHS.
692 KnownOne &= KnownOne2;
693 KnownZero &= KnownZero2;
695 case Instruction::FPTrunc:
696 case Instruction::FPExt:
697 case Instruction::FPToUI:
698 case Instruction::FPToSI:
699 case Instruction::SIToFP:
700 case Instruction::PtrToInt:
701 case Instruction::UIToFP:
702 case Instruction::IntToPtr:
703 return; // Can't work with floating point or pointers
704 case Instruction::Trunc: {
705 // All these have integer operands
706 uint32_t SrcBitWidth =
707 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
709 MaskIn.zext(SrcBitWidth);
710 KnownZero.zext(SrcBitWidth);
711 KnownOne.zext(SrcBitWidth);
712 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
713 KnownZero.trunc(BitWidth);
714 KnownOne.trunc(BitWidth);
717 case Instruction::BitCast: {
718 const Type *SrcTy = I->getOperand(0)->getType();
719 if (SrcTy->isInteger()) {
720 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
725 case Instruction::ZExt: {
726 // Compute the bits in the result that are not present in the input.
727 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
728 uint32_t SrcBitWidth = SrcTy->getBitWidth();
731 MaskIn.trunc(SrcBitWidth);
732 KnownZero.trunc(SrcBitWidth);
733 KnownOne.trunc(SrcBitWidth);
734 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
735 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
736 // The top bits are known to be zero.
737 KnownZero.zext(BitWidth);
738 KnownOne.zext(BitWidth);
739 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
742 case Instruction::SExt: {
743 // Compute the bits in the result that are not present in the input.
744 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
745 uint32_t SrcBitWidth = SrcTy->getBitWidth();
748 MaskIn.trunc(SrcBitWidth);
749 KnownZero.trunc(SrcBitWidth);
750 KnownOne.trunc(SrcBitWidth);
751 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
752 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
753 KnownZero.zext(BitWidth);
754 KnownOne.zext(BitWidth);
756 // If the sign bit of the input is known set or clear, then we know the
757 // top bits of the result.
758 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
759 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
760 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
761 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
764 case Instruction::Shl:
765 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
766 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
767 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
768 APInt Mask2(Mask.lshr(ShiftAmt));
769 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
770 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
771 KnownZero <<= ShiftAmt;
772 KnownOne <<= ShiftAmt;
773 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
777 case Instruction::LShr:
778 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
779 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
780 // Compute the new bits that are at the top now.
781 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
783 // Unsigned shift right.
784 APInt Mask2(Mask.shl(ShiftAmt));
785 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
786 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
787 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
788 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
789 // high bits known zero.
790 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
794 case Instruction::AShr:
795 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
796 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
797 // Compute the new bits that are at the top now.
798 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
800 // Signed shift right.
801 APInt Mask2(Mask.shl(ShiftAmt));
802 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
803 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
804 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
805 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
807 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
808 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
809 KnownZero |= HighBits;
810 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
811 KnownOne |= HighBits;
818 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
819 /// this predicate to simplify operations downstream. Mask is known to be zero
820 /// for bits that V cannot have.
821 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
822 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
823 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
824 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
825 return (KnownZero & Mask) == Mask;
828 /// ShrinkDemandedConstant - Check to see if the specified operand of the
829 /// specified instruction is a constant integer. If so, check to see if there
830 /// are any bits set in the constant that are not demanded. If so, shrink the
831 /// constant and return true.
832 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
834 assert(I && "No instruction?");
835 assert(OpNo < I->getNumOperands() && "Operand index too large");
837 // If the operand is not a constant integer, nothing to do.
838 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
839 if (!OpC) return false;
841 // If there are no bits set that aren't demanded, nothing to do.
842 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
843 if ((~Demanded & OpC->getValue()) == 0)
846 // This instruction is producing bits that are not demanded. Shrink the RHS.
847 Demanded &= OpC->getValue();
848 I->setOperand(OpNo, ConstantInt::get(Demanded));
852 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
853 // set of known zero and one bits, compute the maximum and minimum values that
854 // could have the specified known zero and known one bits, returning them in
856 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
857 const APInt& KnownZero,
858 const APInt& KnownOne,
859 APInt& Min, APInt& Max) {
860 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
861 assert(KnownZero.getBitWidth() == BitWidth &&
862 KnownOne.getBitWidth() == BitWidth &&
863 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
864 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
865 APInt UnknownBits = ~(KnownZero|KnownOne);
867 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
868 // bit if it is unknown.
870 Max = KnownOne|UnknownBits;
872 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
874 Max.clear(BitWidth-1);
878 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
879 // a set of known zero and one bits, compute the maximum and minimum values that
880 // could have the specified known zero and known one bits, returning them in
882 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
883 const APInt &KnownZero,
884 const APInt &KnownOne,
885 APInt &Min, APInt &Max) {
886 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
887 assert(KnownZero.getBitWidth() == BitWidth &&
888 KnownOne.getBitWidth() == BitWidth &&
889 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
890 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
891 APInt UnknownBits = ~(KnownZero|KnownOne);
893 // The minimum value is when the unknown bits are all zeros.
895 // The maximum value is when the unknown bits are all ones.
896 Max = KnownOne|UnknownBits;
899 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
900 /// value based on the demanded bits. When this function is called, it is known
901 /// that only the bits set in DemandedMask of the result of V are ever used
902 /// downstream. Consequently, depending on the mask and V, it may be possible
903 /// to replace V with a constant or one of its operands. In such cases, this
904 /// function does the replacement and returns true. In all other cases, it
905 /// returns false after analyzing the expression and setting KnownOne and known
906 /// to be one in the expression. KnownZero contains all the bits that are known
907 /// to be zero in the expression. These are provided to potentially allow the
908 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
909 /// the expression. KnownOne and KnownZero always follow the invariant that
910 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
911 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
912 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
913 /// and KnownOne must all be the same.
914 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
915 APInt& KnownZero, APInt& KnownOne,
917 assert(V != 0 && "Null pointer of Value???");
918 assert(Depth <= 6 && "Limit Search Depth");
919 uint32_t BitWidth = DemandedMask.getBitWidth();
920 const IntegerType *VTy = cast<IntegerType>(V->getType());
921 assert(VTy->getBitWidth() == BitWidth &&
922 KnownZero.getBitWidth() == BitWidth &&
923 KnownOne.getBitWidth() == BitWidth &&
924 "Value *V, DemandedMask, KnownZero and KnownOne \
925 must have same BitWidth");
926 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
927 // We know all of the bits for a constant!
928 KnownOne = CI->getValue() & DemandedMask;
929 KnownZero = ~KnownOne & DemandedMask;
935 if (!V->hasOneUse()) { // Other users may use these bits.
936 if (Depth != 0) { // Not at the root.
937 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
938 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
941 // If this is the root being simplified, allow it to have multiple uses,
942 // just set the DemandedMask to all bits.
943 DemandedMask = APInt::getAllOnesValue(BitWidth);
944 } else if (DemandedMask == 0) { // Not demanding any bits from V.
945 if (V != UndefValue::get(VTy))
946 return UpdateValueUsesWith(V, UndefValue::get(VTy));
948 } else if (Depth == 6) { // Limit search depth.
952 Instruction *I = dyn_cast<Instruction>(V);
953 if (!I) return false; // Only analyze instructions.
955 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
956 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
957 switch (I->getOpcode()) {
959 case Instruction::And:
960 // If either the LHS or the RHS are Zero, the result is zero.
961 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
962 RHSKnownZero, RHSKnownOne, Depth+1))
964 assert((RHSKnownZero & RHSKnownOne) == 0 &&
965 "Bits known to be one AND zero?");
967 // If something is known zero on the RHS, the bits aren't demanded on the
969 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
970 LHSKnownZero, LHSKnownOne, Depth+1))
972 assert((LHSKnownZero & LHSKnownOne) == 0 &&
973 "Bits known to be one AND zero?");
975 // If all of the demanded bits are known 1 on one side, return the other.
976 // These bits cannot contribute to the result of the 'and'.
977 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
978 (DemandedMask & ~LHSKnownZero))
979 return UpdateValueUsesWith(I, I->getOperand(0));
980 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
981 (DemandedMask & ~RHSKnownZero))
982 return UpdateValueUsesWith(I, I->getOperand(1));
984 // If all of the demanded bits in the inputs are known zeros, return zero.
985 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
986 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
988 // If the RHS is a constant, see if we can simplify it.
989 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
990 return UpdateValueUsesWith(I, I);
992 // Output known-1 bits are only known if set in both the LHS & RHS.
993 RHSKnownOne &= LHSKnownOne;
994 // Output known-0 are known to be clear if zero in either the LHS | RHS.
995 RHSKnownZero |= LHSKnownZero;
997 case Instruction::Or:
998 // If either the LHS or the RHS are One, the result is One.
999 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1000 RHSKnownZero, RHSKnownOne, Depth+1))
1002 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1003 "Bits known to be one AND zero?");
1004 // If something is known one on the RHS, the bits aren't demanded on the
1006 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1007 LHSKnownZero, LHSKnownOne, Depth+1))
1009 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1010 "Bits known to be one AND zero?");
1012 // If all of the demanded bits are known zero on one side, return the other.
1013 // These bits cannot contribute to the result of the 'or'.
1014 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1015 (DemandedMask & ~LHSKnownOne))
1016 return UpdateValueUsesWith(I, I->getOperand(0));
1017 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1018 (DemandedMask & ~RHSKnownOne))
1019 return UpdateValueUsesWith(I, I->getOperand(1));
1021 // If all of the potentially set bits on one side are known to be set on
1022 // the other side, just use the 'other' side.
1023 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1024 (DemandedMask & (~RHSKnownZero)))
1025 return UpdateValueUsesWith(I, I->getOperand(0));
1026 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1027 (DemandedMask & (~LHSKnownZero)))
1028 return UpdateValueUsesWith(I, I->getOperand(1));
1030 // If the RHS is a constant, see if we can simplify it.
1031 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1032 return UpdateValueUsesWith(I, I);
1034 // Output known-0 bits are only known if clear in both the LHS & RHS.
1035 RHSKnownZero &= LHSKnownZero;
1036 // Output known-1 are known to be set if set in either the LHS | RHS.
1037 RHSKnownOne |= LHSKnownOne;
1039 case Instruction::Xor: {
1040 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1041 RHSKnownZero, RHSKnownOne, Depth+1))
1043 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1044 "Bits known to be one AND zero?");
1045 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1046 LHSKnownZero, LHSKnownOne, Depth+1))
1048 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1049 "Bits known to be one AND zero?");
1051 // If all of the demanded bits are known zero on one side, return the other.
1052 // These bits cannot contribute to the result of the 'xor'.
1053 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1054 return UpdateValueUsesWith(I, I->getOperand(0));
1055 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1056 return UpdateValueUsesWith(I, I->getOperand(1));
1058 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1059 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1060 (RHSKnownOne & LHSKnownOne);
1061 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1062 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1063 (RHSKnownOne & LHSKnownZero);
1065 // If all of the demanded bits are known to be zero on one side or the
1066 // other, turn this into an *inclusive* or.
1067 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1068 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1070 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1072 InsertNewInstBefore(Or, *I);
1073 return UpdateValueUsesWith(I, Or);
1076 // If all of the demanded bits on one side are known, and all of the set
1077 // bits on that side are also known to be set on the other side, turn this
1078 // into an AND, as we know the bits will be cleared.
1079 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1080 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1082 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1083 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1085 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1086 InsertNewInstBefore(And, *I);
1087 return UpdateValueUsesWith(I, And);
1091 // If the RHS is a constant, see if we can simplify it.
1092 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1093 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1094 return UpdateValueUsesWith(I, I);
1096 RHSKnownZero = KnownZeroOut;
1097 RHSKnownOne = KnownOneOut;
1100 case Instruction::Select:
1101 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1102 RHSKnownZero, RHSKnownOne, Depth+1))
1104 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1105 LHSKnownZero, LHSKnownOne, Depth+1))
1107 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1108 "Bits known to be one AND zero?");
1109 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1110 "Bits known to be one AND zero?");
1112 // If the operands are constants, see if we can simplify them.
1113 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1114 return UpdateValueUsesWith(I, I);
1115 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1116 return UpdateValueUsesWith(I, I);
1118 // Only known if known in both the LHS and RHS.
1119 RHSKnownOne &= LHSKnownOne;
1120 RHSKnownZero &= LHSKnownZero;
1122 case Instruction::Trunc: {
1124 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1125 DemandedMask.zext(truncBf);
1126 RHSKnownZero.zext(truncBf);
1127 RHSKnownOne.zext(truncBf);
1128 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1129 RHSKnownZero, RHSKnownOne, Depth+1))
1131 DemandedMask.trunc(BitWidth);
1132 RHSKnownZero.trunc(BitWidth);
1133 RHSKnownOne.trunc(BitWidth);
1134 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1135 "Bits known to be one AND zero?");
1138 case Instruction::BitCast:
1139 if (!I->getOperand(0)->getType()->isInteger())
1142 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1143 RHSKnownZero, RHSKnownOne, Depth+1))
1145 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1146 "Bits known to be one AND zero?");
1148 case Instruction::ZExt: {
1149 // Compute the bits in the result that are not present in the input.
1150 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1151 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1153 DemandedMask.trunc(SrcBitWidth);
1154 RHSKnownZero.trunc(SrcBitWidth);
1155 RHSKnownOne.trunc(SrcBitWidth);
1156 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1157 RHSKnownZero, RHSKnownOne, Depth+1))
1159 DemandedMask.zext(BitWidth);
1160 RHSKnownZero.zext(BitWidth);
1161 RHSKnownOne.zext(BitWidth);
1162 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1163 "Bits known to be one AND zero?");
1164 // The top bits are known to be zero.
1165 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1168 case Instruction::SExt: {
1169 // Compute the bits in the result that are not present in the input.
1170 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1171 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1173 APInt InputDemandedBits = DemandedMask &
1174 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1176 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1177 // If any of the sign extended bits are demanded, we know that the sign
1179 if ((NewBits & DemandedMask) != 0)
1180 InputDemandedBits.set(SrcBitWidth-1);
1182 InputDemandedBits.trunc(SrcBitWidth);
1183 RHSKnownZero.trunc(SrcBitWidth);
1184 RHSKnownOne.trunc(SrcBitWidth);
1185 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1186 RHSKnownZero, RHSKnownOne, Depth+1))
1188 InputDemandedBits.zext(BitWidth);
1189 RHSKnownZero.zext(BitWidth);
1190 RHSKnownOne.zext(BitWidth);
1191 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1192 "Bits known to be one AND zero?");
1194 // If the sign bit of the input is known set or clear, then we know the
1195 // top bits of the result.
1197 // If the input sign bit is known zero, or if the NewBits are not demanded
1198 // convert this into a zero extension.
1199 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1201 // Convert to ZExt cast
1202 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1203 return UpdateValueUsesWith(I, NewCast);
1204 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1205 RHSKnownOne |= NewBits;
1209 case Instruction::Add: {
1210 // Figure out what the input bits are. If the top bits of the and result
1211 // are not demanded, then the add doesn't demand them from its input
1213 uint32_t NLZ = DemandedMask.countLeadingZeros();
1215 // If there is a constant on the RHS, there are a variety of xformations
1217 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1218 // If null, this should be simplified elsewhere. Some of the xforms here
1219 // won't work if the RHS is zero.
1223 // If the top bit of the output is demanded, demand everything from the
1224 // input. Otherwise, we demand all the input bits except NLZ top bits.
1225 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1227 // Find information about known zero/one bits in the input.
1228 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1229 LHSKnownZero, LHSKnownOne, Depth+1))
1232 // If the RHS of the add has bits set that can't affect the input, reduce
1234 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1235 return UpdateValueUsesWith(I, I);
1237 // Avoid excess work.
1238 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1241 // Turn it into OR if input bits are zero.
1242 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1244 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1246 InsertNewInstBefore(Or, *I);
1247 return UpdateValueUsesWith(I, Or);
1250 // We can say something about the output known-zero and known-one bits,
1251 // depending on potential carries from the input constant and the
1252 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1253 // bits set and the RHS constant is 0x01001, then we know we have a known
1254 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1256 // To compute this, we first compute the potential carry bits. These are
1257 // the bits which may be modified. I'm not aware of a better way to do
1259 const APInt& RHSVal = RHS->getValue();
1260 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1262 // Now that we know which bits have carries, compute the known-1/0 sets.
1264 // Bits are known one if they are known zero in one operand and one in the
1265 // other, and there is no input carry.
1266 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1267 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1269 // Bits are known zero if they are known zero in both operands and there
1270 // is no input carry.
1271 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1273 // If the high-bits of this ADD are not demanded, then it does not demand
1274 // the high bits of its LHS or RHS.
1275 if (DemandedMask[BitWidth-1] == 0) {
1276 // Right fill the mask of bits for this ADD to demand the most
1277 // significant bit and all those below it.
1278 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1279 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1280 LHSKnownZero, LHSKnownOne, Depth+1))
1282 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1283 LHSKnownZero, LHSKnownOne, Depth+1))
1289 case Instruction::Sub:
1290 // If the high-bits of this SUB are not demanded, then it does not demand
1291 // the high bits of its LHS or RHS.
1292 if (DemandedMask[BitWidth-1] == 0) {
1293 // Right fill the mask of bits for this SUB to demand the most
1294 // significant bit and all those below it.
1295 uint32_t NLZ = DemandedMask.countLeadingZeros();
1296 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1297 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1298 LHSKnownZero, LHSKnownOne, Depth+1))
1300 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1301 LHSKnownZero, LHSKnownOne, Depth+1))
1305 case Instruction::Shl:
1306 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1307 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1308 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1309 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1310 RHSKnownZero, RHSKnownOne, Depth+1))
1312 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1313 "Bits known to be one AND zero?");
1314 RHSKnownZero <<= ShiftAmt;
1315 RHSKnownOne <<= ShiftAmt;
1316 // low bits known zero.
1318 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1321 case Instruction::LShr:
1322 // For a logical shift right
1323 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1324 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1326 // Unsigned shift right.
1327 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1328 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1329 RHSKnownZero, RHSKnownOne, Depth+1))
1331 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1332 "Bits known to be one AND zero?");
1333 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1334 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1336 // Compute the new bits that are at the top now.
1337 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1338 RHSKnownZero |= HighBits; // high bits known zero.
1342 case Instruction::AShr:
1343 // If this is an arithmetic shift right and only the low-bit is set, we can
1344 // always convert this into a logical shr, even if the shift amount is
1345 // variable. The low bit of the shift cannot be an input sign bit unless
1346 // the shift amount is >= the size of the datatype, which is undefined.
1347 if (DemandedMask == 1) {
1348 // Perform the logical shift right.
1349 Value *NewVal = BinaryOperator::createLShr(
1350 I->getOperand(0), I->getOperand(1), I->getName());
1351 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1352 return UpdateValueUsesWith(I, NewVal);
1355 // If the sign bit is the only bit demanded by this ashr, then there is no
1356 // need to do it, the shift doesn't change the high bit.
1357 if (DemandedMask.isSignBit())
1358 return UpdateValueUsesWith(I, I->getOperand(0));
1360 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1361 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1363 // Signed shift right.
1364 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1365 // If any of the "high bits" are demanded, we should set the sign bit as
1367 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1368 DemandedMaskIn.set(BitWidth-1);
1369 if (SimplifyDemandedBits(I->getOperand(0),
1371 RHSKnownZero, RHSKnownOne, Depth+1))
1373 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1374 "Bits known to be one AND zero?");
1375 // Compute the new bits that are at the top now.
1376 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1377 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1378 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1380 // Handle the sign bits.
1381 APInt SignBit(APInt::getSignBit(BitWidth));
1382 // Adjust to where it is now in the mask.
1383 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1385 // If the input sign bit is known to be zero, or if none of the top bits
1386 // are demanded, turn this into an unsigned shift right.
1387 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1388 (HighBits & ~DemandedMask) == HighBits) {
1389 // Perform the logical shift right.
1390 Value *NewVal = BinaryOperator::createLShr(
1391 I->getOperand(0), SA, I->getName());
1392 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1393 return UpdateValueUsesWith(I, NewVal);
1394 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1395 RHSKnownOne |= HighBits;
1401 // If the client is only demanding bits that we know, return the known
1403 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1404 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1409 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1410 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1411 /// actually used by the caller. This method analyzes which elements of the
1412 /// operand are undef and returns that information in UndefElts.
1414 /// If the information about demanded elements can be used to simplify the
1415 /// operation, the operation is simplified, then the resultant value is
1416 /// returned. This returns null if no change was made.
1417 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1418 uint64_t &UndefElts,
1420 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1421 assert(VWidth <= 64 && "Vector too wide to analyze!");
1422 uint64_t EltMask = ~0ULL >> (64-VWidth);
1423 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1424 "Invalid DemandedElts!");
1426 if (isa<UndefValue>(V)) {
1427 // If the entire vector is undefined, just return this info.
1428 UndefElts = EltMask;
1430 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1431 UndefElts = EltMask;
1432 return UndefValue::get(V->getType());
1436 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1437 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1438 Constant *Undef = UndefValue::get(EltTy);
1440 std::vector<Constant*> Elts;
1441 for (unsigned i = 0; i != VWidth; ++i)
1442 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1443 Elts.push_back(Undef);
1444 UndefElts |= (1ULL << i);
1445 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1446 Elts.push_back(Undef);
1447 UndefElts |= (1ULL << i);
1448 } else { // Otherwise, defined.
1449 Elts.push_back(CP->getOperand(i));
1452 // If we changed the constant, return it.
1453 Constant *NewCP = ConstantVector::get(Elts);
1454 return NewCP != CP ? NewCP : 0;
1455 } else if (isa<ConstantAggregateZero>(V)) {
1456 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1458 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1459 Constant *Zero = Constant::getNullValue(EltTy);
1460 Constant *Undef = UndefValue::get(EltTy);
1461 std::vector<Constant*> Elts;
1462 for (unsigned i = 0; i != VWidth; ++i)
1463 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1464 UndefElts = DemandedElts ^ EltMask;
1465 return ConstantVector::get(Elts);
1468 if (!V->hasOneUse()) { // Other users may use these bits.
1469 if (Depth != 0) { // Not at the root.
1470 // TODO: Just compute the UndefElts information recursively.
1474 } else if (Depth == 10) { // Limit search depth.
1478 Instruction *I = dyn_cast<Instruction>(V);
1479 if (!I) return false; // Only analyze instructions.
1481 bool MadeChange = false;
1482 uint64_t UndefElts2;
1484 switch (I->getOpcode()) {
1487 case Instruction::InsertElement: {
1488 // If this is a variable index, we don't know which element it overwrites.
1489 // demand exactly the same input as we produce.
1490 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1492 // Note that we can't propagate undef elt info, because we don't know
1493 // which elt is getting updated.
1494 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1495 UndefElts2, Depth+1);
1496 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1500 // If this is inserting an element that isn't demanded, remove this
1502 unsigned IdxNo = Idx->getZExtValue();
1503 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1504 return AddSoonDeadInstToWorklist(*I, 0);
1506 // Otherwise, the element inserted overwrites whatever was there, so the
1507 // input demanded set is simpler than the output set.
1508 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1509 DemandedElts & ~(1ULL << IdxNo),
1510 UndefElts, Depth+1);
1511 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1513 // The inserted element is defined.
1514 UndefElts |= 1ULL << IdxNo;
1517 case Instruction::BitCast: {
1518 // Vector->vector casts only.
1519 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1521 unsigned InVWidth = VTy->getNumElements();
1522 uint64_t InputDemandedElts = 0;
1525 if (VWidth == InVWidth) {
1526 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1527 // elements as are demanded of us.
1529 InputDemandedElts = DemandedElts;
1530 } else if (VWidth > InVWidth) {
1534 // If there are more elements in the result than there are in the source,
1535 // then an input element is live if any of the corresponding output
1536 // elements are live.
1537 Ratio = VWidth/InVWidth;
1538 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1539 if (DemandedElts & (1ULL << OutIdx))
1540 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1546 // If there are more elements in the source than there are in the result,
1547 // then an input element is live if the corresponding output element is
1549 Ratio = InVWidth/VWidth;
1550 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1551 if (DemandedElts & (1ULL << InIdx/Ratio))
1552 InputDemandedElts |= 1ULL << InIdx;
1555 // div/rem demand all inputs, because they don't want divide by zero.
1556 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1557 UndefElts2, Depth+1);
1559 I->setOperand(0, TmpV);
1563 UndefElts = UndefElts2;
1564 if (VWidth > InVWidth) {
1565 assert(0 && "Unimp");
1566 // If there are more elements in the result than there are in the source,
1567 // then an output element is undef if the corresponding input element is
1569 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1570 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1571 UndefElts |= 1ULL << OutIdx;
1572 } else if (VWidth < InVWidth) {
1573 assert(0 && "Unimp");
1574 // If there are more elements in the source than there are in the result,
1575 // then a result element is undef if all of the corresponding input
1576 // elements are undef.
1577 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1578 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1579 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1580 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1584 case Instruction::And:
1585 case Instruction::Or:
1586 case Instruction::Xor:
1587 case Instruction::Add:
1588 case Instruction::Sub:
1589 case Instruction::Mul:
1590 // div/rem demand all inputs, because they don't want divide by zero.
1591 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1592 UndefElts, Depth+1);
1593 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1594 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1595 UndefElts2, Depth+1);
1596 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1598 // Output elements are undefined if both are undefined. Consider things
1599 // like undef&0. The result is known zero, not undef.
1600 UndefElts &= UndefElts2;
1603 case Instruction::Call: {
1604 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1606 switch (II->getIntrinsicID()) {
1609 // Binary vector operations that work column-wise. A dest element is a
1610 // function of the corresponding input elements from the two inputs.
1611 case Intrinsic::x86_sse_sub_ss:
1612 case Intrinsic::x86_sse_mul_ss:
1613 case Intrinsic::x86_sse_min_ss:
1614 case Intrinsic::x86_sse_max_ss:
1615 case Intrinsic::x86_sse2_sub_sd:
1616 case Intrinsic::x86_sse2_mul_sd:
1617 case Intrinsic::x86_sse2_min_sd:
1618 case Intrinsic::x86_sse2_max_sd:
1619 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1620 UndefElts, Depth+1);
1621 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1622 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1623 UndefElts2, Depth+1);
1624 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1626 // If only the low elt is demanded and this is a scalarizable intrinsic,
1627 // scalarize it now.
1628 if (DemandedElts == 1) {
1629 switch (II->getIntrinsicID()) {
1631 case Intrinsic::x86_sse_sub_ss:
1632 case Intrinsic::x86_sse_mul_ss:
1633 case Intrinsic::x86_sse2_sub_sd:
1634 case Intrinsic::x86_sse2_mul_sd:
1635 // TODO: Lower MIN/MAX/ABS/etc
1636 Value *LHS = II->getOperand(1);
1637 Value *RHS = II->getOperand(2);
1638 // Extract the element as scalars.
1639 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1640 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1642 switch (II->getIntrinsicID()) {
1643 default: assert(0 && "Case stmts out of sync!");
1644 case Intrinsic::x86_sse_sub_ss:
1645 case Intrinsic::x86_sse2_sub_sd:
1646 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1647 II->getName()), *II);
1649 case Intrinsic::x86_sse_mul_ss:
1650 case Intrinsic::x86_sse2_mul_sd:
1651 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1652 II->getName()), *II);
1657 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1659 InsertNewInstBefore(New, *II);
1660 AddSoonDeadInstToWorklist(*II, 0);
1665 // Output elements are undefined if both are undefined. Consider things
1666 // like undef&0. The result is known zero, not undef.
1667 UndefElts &= UndefElts2;
1673 return MadeChange ? I : 0;
1676 /// @returns true if the specified compare predicate is
1677 /// true when both operands are equal...
1678 /// @brief Determine if the icmp Predicate is true when both operands are equal
1679 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1680 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1681 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1682 pred == ICmpInst::ICMP_SLE;
1685 /// @returns true if the specified compare instruction is
1686 /// true when both operands are equal...
1687 /// @brief Determine if the ICmpInst returns true when both operands are equal
1688 static bool isTrueWhenEqual(ICmpInst &ICI) {
1689 return isTrueWhenEqual(ICI.getPredicate());
1692 /// AssociativeOpt - Perform an optimization on an associative operator. This
1693 /// function is designed to check a chain of associative operators for a
1694 /// potential to apply a certain optimization. Since the optimization may be
1695 /// applicable if the expression was reassociated, this checks the chain, then
1696 /// reassociates the expression as necessary to expose the optimization
1697 /// opportunity. This makes use of a special Functor, which must define
1698 /// 'shouldApply' and 'apply' methods.
1700 template<typename Functor>
1701 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1702 unsigned Opcode = Root.getOpcode();
1703 Value *LHS = Root.getOperand(0);
1705 // Quick check, see if the immediate LHS matches...
1706 if (F.shouldApply(LHS))
1707 return F.apply(Root);
1709 // Otherwise, if the LHS is not of the same opcode as the root, return.
1710 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1711 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1712 // Should we apply this transform to the RHS?
1713 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1715 // If not to the RHS, check to see if we should apply to the LHS...
1716 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1717 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1721 // If the functor wants to apply the optimization to the RHS of LHSI,
1722 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1724 BasicBlock *BB = Root.getParent();
1726 // Now all of the instructions are in the current basic block, go ahead
1727 // and perform the reassociation.
1728 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1730 // First move the selected RHS to the LHS of the root...
1731 Root.setOperand(0, LHSI->getOperand(1));
1733 // Make what used to be the LHS of the root be the user of the root...
1734 Value *ExtraOperand = TmpLHSI->getOperand(1);
1735 if (&Root == TmpLHSI) {
1736 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1739 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1740 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1741 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1742 BasicBlock::iterator ARI = &Root; ++ARI;
1743 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1746 // Now propagate the ExtraOperand down the chain of instructions until we
1748 while (TmpLHSI != LHSI) {
1749 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1750 // Move the instruction to immediately before the chain we are
1751 // constructing to avoid breaking dominance properties.
1752 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1753 BB->getInstList().insert(ARI, NextLHSI);
1756 Value *NextOp = NextLHSI->getOperand(1);
1757 NextLHSI->setOperand(1, ExtraOperand);
1759 ExtraOperand = NextOp;
1762 // Now that the instructions are reassociated, have the functor perform
1763 // the transformation...
1764 return F.apply(Root);
1767 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1773 // AddRHS - Implements: X + X --> X << 1
1776 AddRHS(Value *rhs) : RHS(rhs) {}
1777 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1778 Instruction *apply(BinaryOperator &Add) const {
1779 return BinaryOperator::createShl(Add.getOperand(0),
1780 ConstantInt::get(Add.getType(), 1));
1784 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1786 struct AddMaskingAnd {
1788 AddMaskingAnd(Constant *c) : C2(c) {}
1789 bool shouldApply(Value *LHS) const {
1791 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1792 ConstantExpr::getAnd(C1, C2)->isNullValue();
1794 Instruction *apply(BinaryOperator &Add) const {
1795 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1799 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1801 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1802 if (Constant *SOC = dyn_cast<Constant>(SO))
1803 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1805 return IC->InsertNewInstBefore(CastInst::create(
1806 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1809 // Figure out if the constant is the left or the right argument.
1810 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1811 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1813 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1815 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1816 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1819 Value *Op0 = SO, *Op1 = ConstOperand;
1821 std::swap(Op0, Op1);
1823 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1824 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1825 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1826 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1827 SO->getName()+".cmp");
1829 assert(0 && "Unknown binary instruction type!");
1832 return IC->InsertNewInstBefore(New, I);
1835 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1836 // constant as the other operand, try to fold the binary operator into the
1837 // select arguments. This also works for Cast instructions, which obviously do
1838 // not have a second operand.
1839 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1841 // Don't modify shared select instructions
1842 if (!SI->hasOneUse()) return 0;
1843 Value *TV = SI->getOperand(1);
1844 Value *FV = SI->getOperand(2);
1846 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1847 // Bool selects with constant operands can be folded to logical ops.
1848 if (SI->getType() == Type::Int1Ty) return 0;
1850 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1851 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1853 return new SelectInst(SI->getCondition(), SelectTrueVal,
1860 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1861 /// node as operand #0, see if we can fold the instruction into the PHI (which
1862 /// is only possible if all operands to the PHI are constants).
1863 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1864 PHINode *PN = cast<PHINode>(I.getOperand(0));
1865 unsigned NumPHIValues = PN->getNumIncomingValues();
1866 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1868 // Check to see if all of the operands of the PHI are constants. If there is
1869 // one non-constant value, remember the BB it is. If there is more than one
1870 // or if *it* is a PHI, bail out.
1871 BasicBlock *NonConstBB = 0;
1872 for (unsigned i = 0; i != NumPHIValues; ++i)
1873 if (!isa<Constant>(PN->getIncomingValue(i))) {
1874 if (NonConstBB) return 0; // More than one non-const value.
1875 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1876 NonConstBB = PN->getIncomingBlock(i);
1878 // If the incoming non-constant value is in I's block, we have an infinite
1880 if (NonConstBB == I.getParent())
1884 // If there is exactly one non-constant value, we can insert a copy of the
1885 // operation in that block. However, if this is a critical edge, we would be
1886 // inserting the computation one some other paths (e.g. inside a loop). Only
1887 // do this if the pred block is unconditionally branching into the phi block.
1889 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1890 if (!BI || !BI->isUnconditional()) return 0;
1893 // Okay, we can do the transformation: create the new PHI node.
1894 PHINode *NewPN = new PHINode(I.getType(), "");
1895 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1896 InsertNewInstBefore(NewPN, *PN);
1897 NewPN->takeName(PN);
1899 // Next, add all of the operands to the PHI.
1900 if (I.getNumOperands() == 2) {
1901 Constant *C = cast<Constant>(I.getOperand(1));
1902 for (unsigned i = 0; i != NumPHIValues; ++i) {
1904 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1905 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1906 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1908 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1910 assert(PN->getIncomingBlock(i) == NonConstBB);
1911 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1912 InV = BinaryOperator::create(BO->getOpcode(),
1913 PN->getIncomingValue(i), C, "phitmp",
1914 NonConstBB->getTerminator());
1915 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1916 InV = CmpInst::create(CI->getOpcode(),
1918 PN->getIncomingValue(i), C, "phitmp",
1919 NonConstBB->getTerminator());
1921 assert(0 && "Unknown binop!");
1923 AddToWorkList(cast<Instruction>(InV));
1925 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1928 CastInst *CI = cast<CastInst>(&I);
1929 const Type *RetTy = CI->getType();
1930 for (unsigned i = 0; i != NumPHIValues; ++i) {
1932 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1933 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1935 assert(PN->getIncomingBlock(i) == NonConstBB);
1936 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1937 I.getType(), "phitmp",
1938 NonConstBB->getTerminator());
1939 AddToWorkList(cast<Instruction>(InV));
1941 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1944 return ReplaceInstUsesWith(I, NewPN);
1948 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1949 /// value is never equal to -0.0.
1951 /// Note that this function will need to be revisited when we support nondefault
1954 static bool CannotBeNegativeZero(const Value *V) {
1955 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1956 return !CFP->getValueAPF().isNegZero();
1958 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1959 if (const Instruction *I = dyn_cast<Instruction>(V)) {
1960 if (I->getOpcode() == Instruction::Add &&
1961 isa<ConstantFP>(I->getOperand(1)) &&
1962 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1965 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1966 if (II->getIntrinsicID() == Intrinsic::sqrt)
1967 return CannotBeNegativeZero(II->getOperand(1));
1969 if (const CallInst *CI = dyn_cast<CallInst>(I))
1970 if (const Function *F = CI->getCalledFunction()) {
1971 if (F->isDeclaration()) {
1972 switch (F->getNameLen()) {
1973 case 3: // abs(x) != -0.0
1974 if (!strcmp(F->getNameStart(), "abs")) return true;
1976 case 4: // abs[lf](x) != -0.0
1977 if (!strcmp(F->getNameStart(), "absf")) return true;
1978 if (!strcmp(F->getNameStart(), "absl")) return true;
1989 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1990 bool Changed = SimplifyCommutative(I);
1991 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1993 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1994 // X + undef -> undef
1995 if (isa<UndefValue>(RHS))
1996 return ReplaceInstUsesWith(I, RHS);
1999 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2000 if (RHSC->isNullValue())
2001 return ReplaceInstUsesWith(I, LHS);
2002 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2003 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2004 (I.getType())->getValueAPF()))
2005 return ReplaceInstUsesWith(I, LHS);
2008 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2009 // X + (signbit) --> X ^ signbit
2010 const APInt& Val = CI->getValue();
2011 uint32_t BitWidth = Val.getBitWidth();
2012 if (Val == APInt::getSignBit(BitWidth))
2013 return BinaryOperator::createXor(LHS, RHS);
2015 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2016 // (X & 254)+1 -> (X&254)|1
2017 if (!isa<VectorType>(I.getType())) {
2018 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2019 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2020 KnownZero, KnownOne))
2025 if (isa<PHINode>(LHS))
2026 if (Instruction *NV = FoldOpIntoPhi(I))
2029 ConstantInt *XorRHS = 0;
2031 if (isa<ConstantInt>(RHSC) &&
2032 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2033 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2034 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2036 uint32_t Size = TySizeBits / 2;
2037 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2038 APInt CFF80Val(-C0080Val);
2040 if (TySizeBits > Size) {
2041 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2042 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2043 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2044 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2045 // This is a sign extend if the top bits are known zero.
2046 if (!MaskedValueIsZero(XorLHS,
2047 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2048 Size = 0; // Not a sign ext, but can't be any others either.
2053 C0080Val = APIntOps::lshr(C0080Val, Size);
2054 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2055 } while (Size >= 1);
2057 // FIXME: This shouldn't be necessary. When the backends can handle types
2058 // with funny bit widths then this whole cascade of if statements should
2059 // be removed. It is just here to get the size of the "middle" type back
2060 // up to something that the back ends can handle.
2061 const Type *MiddleType = 0;
2064 case 32: MiddleType = Type::Int32Ty; break;
2065 case 16: MiddleType = Type::Int16Ty; break;
2066 case 8: MiddleType = Type::Int8Ty; break;
2069 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2070 InsertNewInstBefore(NewTrunc, I);
2071 return new SExtInst(NewTrunc, I.getType(), I.getName());
2077 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2078 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2080 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2081 if (RHSI->getOpcode() == Instruction::Sub)
2082 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2083 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2085 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2086 if (LHSI->getOpcode() == Instruction::Sub)
2087 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2088 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2093 // -A + -B --> -(A + B)
2094 if (Value *LHSV = dyn_castNegVal(LHS)) {
2095 if (LHS->getType()->isIntOrIntVector()) {
2096 if (Value *RHSV = dyn_castNegVal(RHS)) {
2097 Instruction *NewAdd = BinaryOperator::createAdd(LHSV, RHSV, "sum");
2098 InsertNewInstBefore(NewAdd, I);
2099 return BinaryOperator::createNeg(NewAdd);
2103 return BinaryOperator::createSub(RHS, LHSV);
2107 if (!isa<Constant>(RHS))
2108 if (Value *V = dyn_castNegVal(RHS))
2109 return BinaryOperator::createSub(LHS, V);
2113 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2114 if (X == RHS) // X*C + X --> X * (C+1)
2115 return BinaryOperator::createMul(RHS, AddOne(C2));
2117 // X*C1 + X*C2 --> X * (C1+C2)
2119 if (X == dyn_castFoldableMul(RHS, C1))
2120 return BinaryOperator::createMul(X, Add(C1, C2));
2123 // X + X*C --> X * (C+1)
2124 if (dyn_castFoldableMul(RHS, C2) == LHS)
2125 return BinaryOperator::createMul(LHS, AddOne(C2));
2127 // X + ~X --> -1 since ~X = -X-1
2128 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2129 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2132 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2133 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2134 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2137 // W*X + Y*Z --> W * (X+Z) iff W == Y
2138 if (I.getType()->isIntOrIntVector()) {
2139 Value *W, *X, *Y, *Z;
2140 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2141 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2145 } else if (Y == X) {
2147 } else if (X == Z) {
2154 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, Z,
2155 LHS->getName()), I);
2156 return BinaryOperator::createMul(W, NewAdd);
2161 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2163 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2164 return BinaryOperator::createSub(SubOne(CRHS), X);
2166 // (X & FF00) + xx00 -> (X+xx00) & FF00
2167 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2168 Constant *Anded = And(CRHS, C2);
2169 if (Anded == CRHS) {
2170 // See if all bits from the first bit set in the Add RHS up are included
2171 // in the mask. First, get the rightmost bit.
2172 const APInt& AddRHSV = CRHS->getValue();
2174 // Form a mask of all bits from the lowest bit added through the top.
2175 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2177 // See if the and mask includes all of these bits.
2178 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2180 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2181 // Okay, the xform is safe. Insert the new add pronto.
2182 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2183 LHS->getName()), I);
2184 return BinaryOperator::createAnd(NewAdd, C2);
2189 // Try to fold constant add into select arguments.
2190 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2191 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2195 // add (cast *A to intptrtype) B ->
2196 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2198 CastInst *CI = dyn_cast<CastInst>(LHS);
2201 CI = dyn_cast<CastInst>(RHS);
2204 if (CI && CI->getType()->isSized() &&
2205 (CI->getType()->getPrimitiveSizeInBits() ==
2206 TD->getIntPtrType()->getPrimitiveSizeInBits())
2207 && isa<PointerType>(CI->getOperand(0)->getType())) {
2209 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2210 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2211 PointerType::get(Type::Int8Ty, AS), I);
2212 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2213 return new PtrToIntInst(I2, CI->getType());
2217 // add (select X 0 (sub n A)) A --> select X A n
2219 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2222 SI = dyn_cast<SelectInst>(RHS);
2225 if (SI && SI->hasOneUse()) {
2226 Value *TV = SI->getTrueValue();
2227 Value *FV = SI->getFalseValue();
2230 // Can we fold the add into the argument of the select?
2231 // We check both true and false select arguments for a matching subtract.
2232 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2233 A == Other) // Fold the add into the true select value.
2234 return new SelectInst(SI->getCondition(), N, A);
2235 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2236 A == Other) // Fold the add into the false select value.
2237 return new SelectInst(SI->getCondition(), A, N);
2241 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2242 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2243 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2244 return ReplaceInstUsesWith(I, LHS);
2246 return Changed ? &I : 0;
2249 // isSignBit - Return true if the value represented by the constant only has the
2250 // highest order bit set.
2251 static bool isSignBit(ConstantInt *CI) {
2252 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2253 return CI->getValue() == APInt::getSignBit(NumBits);
2256 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2257 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2259 if (Op0 == Op1) // sub X, X -> 0
2260 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2262 // If this is a 'B = x-(-A)', change to B = x+A...
2263 if (Value *V = dyn_castNegVal(Op1))
2264 return BinaryOperator::createAdd(Op0, V);
2266 if (isa<UndefValue>(Op0))
2267 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2268 if (isa<UndefValue>(Op1))
2269 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2271 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2272 // Replace (-1 - A) with (~A)...
2273 if (C->isAllOnesValue())
2274 return BinaryOperator::createNot(Op1);
2276 // C - ~X == X + (1+C)
2278 if (match(Op1, m_Not(m_Value(X))))
2279 return BinaryOperator::createAdd(X, AddOne(C));
2281 // -(X >>u 31) -> (X >>s 31)
2282 // -(X >>s 31) -> (X >>u 31)
2284 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2285 if (SI->getOpcode() == Instruction::LShr) {
2286 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2287 // Check to see if we are shifting out everything but the sign bit.
2288 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2289 SI->getType()->getPrimitiveSizeInBits()-1) {
2290 // Ok, the transformation is safe. Insert AShr.
2291 return BinaryOperator::create(Instruction::AShr,
2292 SI->getOperand(0), CU, SI->getName());
2296 else if (SI->getOpcode() == Instruction::AShr) {
2297 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2298 // Check to see if we are shifting out everything but the sign bit.
2299 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2300 SI->getType()->getPrimitiveSizeInBits()-1) {
2301 // Ok, the transformation is safe. Insert LShr.
2302 return BinaryOperator::createLShr(
2303 SI->getOperand(0), CU, SI->getName());
2309 // Try to fold constant sub into select arguments.
2310 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2311 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2314 if (isa<PHINode>(Op0))
2315 if (Instruction *NV = FoldOpIntoPhi(I))
2319 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2320 if (Op1I->getOpcode() == Instruction::Add &&
2321 !Op0->getType()->isFPOrFPVector()) {
2322 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2323 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2324 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2325 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2326 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2327 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2328 // C1-(X+C2) --> (C1-C2)-X
2329 return BinaryOperator::createSub(Subtract(CI1, CI2),
2330 Op1I->getOperand(0));
2334 if (Op1I->hasOneUse()) {
2335 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2336 // is not used by anyone else...
2338 if (Op1I->getOpcode() == Instruction::Sub &&
2339 !Op1I->getType()->isFPOrFPVector()) {
2340 // Swap the two operands of the subexpr...
2341 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2342 Op1I->setOperand(0, IIOp1);
2343 Op1I->setOperand(1, IIOp0);
2345 // Create the new top level add instruction...
2346 return BinaryOperator::createAdd(Op0, Op1);
2349 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2351 if (Op1I->getOpcode() == Instruction::And &&
2352 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2353 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2356 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2357 return BinaryOperator::createAnd(Op0, NewNot);
2360 // 0 - (X sdiv C) -> (X sdiv -C)
2361 if (Op1I->getOpcode() == Instruction::SDiv)
2362 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2364 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2365 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2366 ConstantExpr::getNeg(DivRHS));
2368 // X - X*C --> X * (1-C)
2369 ConstantInt *C2 = 0;
2370 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2371 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2372 return BinaryOperator::createMul(Op0, CP1);
2375 // X - ((X / Y) * Y) --> X % Y
2376 if (Op1I->getOpcode() == Instruction::Mul)
2377 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2378 if (Op0 == I->getOperand(0) &&
2379 Op1I->getOperand(1) == I->getOperand(1)) {
2380 if (I->getOpcode() == Instruction::SDiv)
2381 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2382 if (I->getOpcode() == Instruction::UDiv)
2383 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2388 if (!Op0->getType()->isFPOrFPVector())
2389 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2390 if (Op0I->getOpcode() == Instruction::Add) {
2391 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2392 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2393 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2394 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2395 } else if (Op0I->getOpcode() == Instruction::Sub) {
2396 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2397 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2401 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2402 if (X == Op1) // X*C - X --> X * (C-1)
2403 return BinaryOperator::createMul(Op1, SubOne(C1));
2405 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2406 if (X == dyn_castFoldableMul(Op1, C2))
2407 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2412 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2413 /// comparison only checks the sign bit. If it only checks the sign bit, set
2414 /// TrueIfSigned if the result of the comparison is true when the input value is
2416 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2417 bool &TrueIfSigned) {
2419 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2420 TrueIfSigned = true;
2421 return RHS->isZero();
2422 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2423 TrueIfSigned = true;
2424 return RHS->isAllOnesValue();
2425 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2426 TrueIfSigned = false;
2427 return RHS->isAllOnesValue();
2428 case ICmpInst::ICMP_UGT:
2429 // True if LHS u> RHS and RHS == high-bit-mask - 1
2430 TrueIfSigned = true;
2431 return RHS->getValue() ==
2432 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2433 case ICmpInst::ICMP_UGE:
2434 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2435 TrueIfSigned = true;
2436 return RHS->getValue() ==
2437 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2443 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2444 bool Changed = SimplifyCommutative(I);
2445 Value *Op0 = I.getOperand(0);
2447 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2448 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2450 // Simplify mul instructions with a constant RHS...
2451 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2452 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2454 // ((X << C1)*C2) == (X * (C2 << C1))
2455 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2456 if (SI->getOpcode() == Instruction::Shl)
2457 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2458 return BinaryOperator::createMul(SI->getOperand(0),
2459 ConstantExpr::getShl(CI, ShOp));
2462 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2463 if (CI->equalsInt(1)) // X * 1 == X
2464 return ReplaceInstUsesWith(I, Op0);
2465 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2466 return BinaryOperator::createNeg(Op0, I.getName());
2468 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2469 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2470 return BinaryOperator::createShl(Op0,
2471 ConstantInt::get(Op0->getType(), Val.logBase2()));
2473 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2474 if (Op1F->isNullValue())
2475 return ReplaceInstUsesWith(I, Op1);
2477 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2478 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2479 // We need a better interface for long double here.
2480 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2481 if (Op1F->isExactlyValue(1.0))
2482 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2485 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2486 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2487 isa<ConstantInt>(Op0I->getOperand(1))) {
2488 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2489 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2491 InsertNewInstBefore(Add, I);
2492 Value *C1C2 = ConstantExpr::getMul(Op1,
2493 cast<Constant>(Op0I->getOperand(1)));
2494 return BinaryOperator::createAdd(Add, C1C2);
2498 // Try to fold constant mul into select arguments.
2499 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2500 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2503 if (isa<PHINode>(Op0))
2504 if (Instruction *NV = FoldOpIntoPhi(I))
2508 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2509 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2510 return BinaryOperator::createMul(Op0v, Op1v);
2512 // If one of the operands of the multiply is a cast from a boolean value, then
2513 // we know the bool is either zero or one, so this is a 'masking' multiply.
2514 // See if we can simplify things based on how the boolean was originally
2516 CastInst *BoolCast = 0;
2517 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2518 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2521 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2522 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2525 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2526 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2527 const Type *SCOpTy = SCIOp0->getType();
2530 // If the icmp is true iff the sign bit of X is set, then convert this
2531 // multiply into a shift/and combination.
2532 if (isa<ConstantInt>(SCIOp1) &&
2533 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2535 // Shift the X value right to turn it into "all signbits".
2536 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2537 SCOpTy->getPrimitiveSizeInBits()-1);
2539 InsertNewInstBefore(
2540 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2541 BoolCast->getOperand(0)->getName()+
2544 // If the multiply type is not the same as the source type, sign extend
2545 // or truncate to the multiply type.
2546 if (I.getType() != V->getType()) {
2547 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2548 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2549 Instruction::CastOps opcode =
2550 (SrcBits == DstBits ? Instruction::BitCast :
2551 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2552 V = InsertCastBefore(opcode, V, I.getType(), I);
2555 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2556 return BinaryOperator::createAnd(V, OtherOp);
2561 return Changed ? &I : 0;
2564 /// This function implements the transforms on div instructions that work
2565 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2566 /// used by the visitors to those instructions.
2567 /// @brief Transforms common to all three div instructions
2568 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2569 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2572 if (isa<UndefValue>(Op0))
2573 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2575 // X / undef -> undef
2576 if (isa<UndefValue>(Op1))
2577 return ReplaceInstUsesWith(I, Op1);
2579 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2580 // This does not apply for fdiv.
2581 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2582 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2583 // the same basic block, then we replace the select with Y, and the
2584 // condition of the select with false (if the cond value is in the same BB).
2585 // If the select has uses other than the div, this allows them to be
2586 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2587 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2588 if (ST->isNullValue()) {
2589 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2590 if (CondI && CondI->getParent() == I.getParent())
2591 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2592 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2593 I.setOperand(1, SI->getOperand(2));
2595 UpdateValueUsesWith(SI, SI->getOperand(2));
2599 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2600 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2601 if (ST->isNullValue()) {
2602 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2603 if (CondI && CondI->getParent() == I.getParent())
2604 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2605 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2606 I.setOperand(1, SI->getOperand(1));
2608 UpdateValueUsesWith(SI, SI->getOperand(1));
2616 /// This function implements the transforms common to both integer division
2617 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2618 /// division instructions.
2619 /// @brief Common integer divide transforms
2620 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2621 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2623 if (Instruction *Common = commonDivTransforms(I))
2626 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2628 if (RHS->equalsInt(1))
2629 return ReplaceInstUsesWith(I, Op0);
2631 // (X / C1) / C2 -> X / (C1*C2)
2632 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2633 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2634 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2635 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2636 Multiply(RHS, LHSRHS));
2639 if (!RHS->isZero()) { // avoid X udiv 0
2640 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2641 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2643 if (isa<PHINode>(Op0))
2644 if (Instruction *NV = FoldOpIntoPhi(I))
2649 // 0 / X == 0, we don't need to preserve faults!
2650 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2651 if (LHS->equalsInt(0))
2652 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2657 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2658 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2660 // Handle the integer div common cases
2661 if (Instruction *Common = commonIDivTransforms(I))
2664 // X udiv C^2 -> X >> C
2665 // Check to see if this is an unsigned division with an exact power of 2,
2666 // if so, convert to a right shift.
2667 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2668 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2669 return BinaryOperator::createLShr(Op0,
2670 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2673 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2674 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2675 if (RHSI->getOpcode() == Instruction::Shl &&
2676 isa<ConstantInt>(RHSI->getOperand(0))) {
2677 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2678 if (C1.isPowerOf2()) {
2679 Value *N = RHSI->getOperand(1);
2680 const Type *NTy = N->getType();
2681 if (uint32_t C2 = C1.logBase2()) {
2682 Constant *C2V = ConstantInt::get(NTy, C2);
2683 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2685 return BinaryOperator::createLShr(Op0, N);
2690 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2691 // where C1&C2 are powers of two.
2692 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2693 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2694 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2695 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2696 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2697 // Compute the shift amounts
2698 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2699 // Construct the "on true" case of the select
2700 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2701 Instruction *TSI = BinaryOperator::createLShr(
2702 Op0, TC, SI->getName()+".t");
2703 TSI = InsertNewInstBefore(TSI, I);
2705 // Construct the "on false" case of the select
2706 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2707 Instruction *FSI = BinaryOperator::createLShr(
2708 Op0, FC, SI->getName()+".f");
2709 FSI = InsertNewInstBefore(FSI, I);
2711 // construct the select instruction and return it.
2712 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2718 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2719 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2721 // Handle the integer div common cases
2722 if (Instruction *Common = commonIDivTransforms(I))
2725 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2727 if (RHS->isAllOnesValue())
2728 return BinaryOperator::createNeg(Op0);
2731 if (Value *LHSNeg = dyn_castNegVal(Op0))
2732 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2735 // If the sign bits of both operands are zero (i.e. we can prove they are
2736 // unsigned inputs), turn this into a udiv.
2737 if (I.getType()->isInteger()) {
2738 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2739 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2740 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2741 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2748 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2749 return commonDivTransforms(I);
2752 /// GetFactor - If we can prove that the specified value is at least a multiple
2753 /// of some factor, return that factor.
2754 static Constant *GetFactor(Value *V) {
2755 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2758 // Unless we can be tricky, we know this is a multiple of 1.
2759 Constant *Result = ConstantInt::get(V->getType(), 1);
2761 Instruction *I = dyn_cast<Instruction>(V);
2762 if (!I) return Result;
2764 if (I->getOpcode() == Instruction::Mul) {
2765 // Handle multiplies by a constant, etc.
2766 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2767 GetFactor(I->getOperand(1)));
2768 } else if (I->getOpcode() == Instruction::Shl) {
2769 // (X<<C) -> X * (1 << C)
2770 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2771 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2772 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2774 } else if (I->getOpcode() == Instruction::And) {
2775 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2776 // X & 0xFFF0 is known to be a multiple of 16.
2777 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2778 if (Zeros != V->getType()->getPrimitiveSizeInBits())// don't shift by "32"
2779 return ConstantExpr::getShl(Result,
2780 ConstantInt::get(Result->getType(), Zeros));
2782 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2783 // Only handle int->int casts.
2784 if (!CI->isIntegerCast())
2786 Value *Op = CI->getOperand(0);
2787 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2792 /// This function implements the transforms on rem instructions that work
2793 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2794 /// is used by the visitors to those instructions.
2795 /// @brief Transforms common to all three rem instructions
2796 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2797 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2799 // 0 % X == 0, we don't need to preserve faults!
2800 if (Constant *LHS = dyn_cast<Constant>(Op0))
2801 if (LHS->isNullValue())
2802 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2804 if (isa<UndefValue>(Op0)) // undef % X -> 0
2805 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2806 if (isa<UndefValue>(Op1))
2807 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2809 // Handle cases involving: rem X, (select Cond, Y, Z)
2810 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2811 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2812 // the same basic block, then we replace the select with Y, and the
2813 // condition of the select with false (if the cond value is in the same
2814 // BB). If the select has uses other than the div, this allows them to be
2816 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2817 if (ST->isNullValue()) {
2818 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2819 if (CondI && CondI->getParent() == I.getParent())
2820 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2821 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2822 I.setOperand(1, SI->getOperand(2));
2824 UpdateValueUsesWith(SI, SI->getOperand(2));
2827 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2828 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2829 if (ST->isNullValue()) {
2830 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2831 if (CondI && CondI->getParent() == I.getParent())
2832 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2833 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2834 I.setOperand(1, SI->getOperand(1));
2836 UpdateValueUsesWith(SI, SI->getOperand(1));
2844 /// This function implements the transforms common to both integer remainder
2845 /// instructions (urem and srem). It is called by the visitors to those integer
2846 /// remainder instructions.
2847 /// @brief Common integer remainder transforms
2848 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2849 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2851 if (Instruction *common = commonRemTransforms(I))
2854 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2855 // X % 0 == undef, we don't need to preserve faults!
2856 if (RHS->equalsInt(0))
2857 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2859 if (RHS->equalsInt(1)) // X % 1 == 0
2860 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2862 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2863 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2864 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2866 } else if (isa<PHINode>(Op0I)) {
2867 if (Instruction *NV = FoldOpIntoPhi(I))
2870 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2871 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2872 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2879 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2880 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2882 if (Instruction *common = commonIRemTransforms(I))
2885 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2886 // X urem C^2 -> X and C
2887 // Check to see if this is an unsigned remainder with an exact power of 2,
2888 // if so, convert to a bitwise and.
2889 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2890 if (C->getValue().isPowerOf2())
2891 return BinaryOperator::createAnd(Op0, SubOne(C));
2894 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2895 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2896 if (RHSI->getOpcode() == Instruction::Shl &&
2897 isa<ConstantInt>(RHSI->getOperand(0))) {
2898 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2899 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2900 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2902 return BinaryOperator::createAnd(Op0, Add);
2907 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2908 // where C1&C2 are powers of two.
2909 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2910 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2911 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2912 // STO == 0 and SFO == 0 handled above.
2913 if ((STO->getValue().isPowerOf2()) &&
2914 (SFO->getValue().isPowerOf2())) {
2915 Value *TrueAnd = InsertNewInstBefore(
2916 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2917 Value *FalseAnd = InsertNewInstBefore(
2918 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2919 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2927 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2928 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2930 // Handle the integer rem common cases
2931 if (Instruction *common = commonIRemTransforms(I))
2934 if (Value *RHSNeg = dyn_castNegVal(Op1))
2935 if (!isa<ConstantInt>(RHSNeg) ||
2936 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2938 AddUsesToWorkList(I);
2939 I.setOperand(1, RHSNeg);
2943 // If the sign bits of both operands are zero (i.e. we can prove they are
2944 // unsigned inputs), turn this into a urem.
2945 if (I.getType()->isInteger()) {
2946 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2947 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2948 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2949 return BinaryOperator::createURem(Op0, Op1, I.getName());
2956 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2957 return commonRemTransforms(I);
2960 // isMaxValueMinusOne - return true if this is Max-1
2961 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2962 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2964 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2965 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2968 // isMinValuePlusOne - return true if this is Min+1
2969 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2971 return C->getValue() == 1; // unsigned
2973 // Calculate 1111111111000000000000
2974 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2975 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2978 // isOneBitSet - Return true if there is exactly one bit set in the specified
2980 static bool isOneBitSet(const ConstantInt *CI) {
2981 return CI->getValue().isPowerOf2();
2984 // isHighOnes - Return true if the constant is of the form 1+0+.
2985 // This is the same as lowones(~X).
2986 static bool isHighOnes(const ConstantInt *CI) {
2987 return (~CI->getValue() + 1).isPowerOf2();
2990 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2991 /// are carefully arranged to allow folding of expressions such as:
2993 /// (A < B) | (A > B) --> (A != B)
2995 /// Note that this is only valid if the first and second predicates have the
2996 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2998 /// Three bits are used to represent the condition, as follows:
3003 /// <=> Value Definition
3004 /// 000 0 Always false
3011 /// 111 7 Always true
3013 static unsigned getICmpCode(const ICmpInst *ICI) {
3014 switch (ICI->getPredicate()) {
3016 case ICmpInst::ICMP_UGT: return 1; // 001
3017 case ICmpInst::ICMP_SGT: return 1; // 001
3018 case ICmpInst::ICMP_EQ: return 2; // 010
3019 case ICmpInst::ICMP_UGE: return 3; // 011
3020 case ICmpInst::ICMP_SGE: return 3; // 011
3021 case ICmpInst::ICMP_ULT: return 4; // 100
3022 case ICmpInst::ICMP_SLT: return 4; // 100
3023 case ICmpInst::ICMP_NE: return 5; // 101
3024 case ICmpInst::ICMP_ULE: return 6; // 110
3025 case ICmpInst::ICMP_SLE: return 6; // 110
3028 assert(0 && "Invalid ICmp predicate!");
3033 /// getICmpValue - This is the complement of getICmpCode, which turns an
3034 /// opcode and two operands into either a constant true or false, or a brand
3035 /// new ICmp instruction. The sign is passed in to determine which kind
3036 /// of predicate to use in new icmp instructions.
3037 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3039 default: assert(0 && "Illegal ICmp code!");
3040 case 0: return ConstantInt::getFalse();
3043 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3045 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3046 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3049 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3051 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3054 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3056 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3057 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3060 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3062 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3063 case 7: return ConstantInt::getTrue();
3067 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3068 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3069 (ICmpInst::isSignedPredicate(p1) &&
3070 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3071 (ICmpInst::isSignedPredicate(p2) &&
3072 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3076 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3077 struct FoldICmpLogical {
3080 ICmpInst::Predicate pred;
3081 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3082 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3083 pred(ICI->getPredicate()) {}
3084 bool shouldApply(Value *V) const {
3085 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3086 if (PredicatesFoldable(pred, ICI->getPredicate()))
3087 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
3088 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
3091 Instruction *apply(Instruction &Log) const {
3092 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3093 if (ICI->getOperand(0) != LHS) {
3094 assert(ICI->getOperand(1) == LHS);
3095 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3098 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3099 unsigned LHSCode = getICmpCode(ICI);
3100 unsigned RHSCode = getICmpCode(RHSICI);
3102 switch (Log.getOpcode()) {
3103 case Instruction::And: Code = LHSCode & RHSCode; break;
3104 case Instruction::Or: Code = LHSCode | RHSCode; break;
3105 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3106 default: assert(0 && "Illegal logical opcode!"); return 0;
3109 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3110 ICmpInst::isSignedPredicate(ICI->getPredicate());
3112 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3113 if (Instruction *I = dyn_cast<Instruction>(RV))
3115 // Otherwise, it's a constant boolean value...
3116 return IC.ReplaceInstUsesWith(Log, RV);
3119 } // end anonymous namespace
3121 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3122 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3123 // guaranteed to be a binary operator.
3124 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3126 ConstantInt *AndRHS,
3127 BinaryOperator &TheAnd) {
3128 Value *X = Op->getOperand(0);
3129 Constant *Together = 0;
3131 Together = And(AndRHS, OpRHS);
3133 switch (Op->getOpcode()) {
3134 case Instruction::Xor:
3135 if (Op->hasOneUse()) {
3136 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3137 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3138 InsertNewInstBefore(And, TheAnd);
3140 return BinaryOperator::createXor(And, Together);
3143 case Instruction::Or:
3144 if (Together == AndRHS) // (X | C) & C --> C
3145 return ReplaceInstUsesWith(TheAnd, AndRHS);
3147 if (Op->hasOneUse() && Together != OpRHS) {
3148 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3149 Instruction *Or = BinaryOperator::createOr(X, Together);
3150 InsertNewInstBefore(Or, TheAnd);
3152 return BinaryOperator::createAnd(Or, AndRHS);
3155 case Instruction::Add:
3156 if (Op->hasOneUse()) {
3157 // Adding a one to a single bit bit-field should be turned into an XOR
3158 // of the bit. First thing to check is to see if this AND is with a
3159 // single bit constant.
3160 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3162 // If there is only one bit set...
3163 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3164 // Ok, at this point, we know that we are masking the result of the
3165 // ADD down to exactly one bit. If the constant we are adding has
3166 // no bits set below this bit, then we can eliminate the ADD.
3167 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3169 // Check to see if any bits below the one bit set in AndRHSV are set.
3170 if ((AddRHS & (AndRHSV-1)) == 0) {
3171 // If not, the only thing that can effect the output of the AND is
3172 // the bit specified by AndRHSV. If that bit is set, the effect of
3173 // the XOR is to toggle the bit. If it is clear, then the ADD has
3175 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3176 TheAnd.setOperand(0, X);
3179 // Pull the XOR out of the AND.
3180 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3181 InsertNewInstBefore(NewAnd, TheAnd);
3182 NewAnd->takeName(Op);
3183 return BinaryOperator::createXor(NewAnd, AndRHS);
3190 case Instruction::Shl: {
3191 // We know that the AND will not produce any of the bits shifted in, so if
3192 // the anded constant includes them, clear them now!
3194 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3195 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3196 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3197 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3199 if (CI->getValue() == ShlMask) {
3200 // Masking out bits that the shift already masks
3201 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3202 } else if (CI != AndRHS) { // Reducing bits set in and.
3203 TheAnd.setOperand(1, CI);
3208 case Instruction::LShr:
3210 // We know that the AND will not produce any of the bits shifted in, so if
3211 // the anded constant includes them, clear them now! This only applies to
3212 // unsigned shifts, because a signed shr may bring in set bits!
3214 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3215 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3216 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3217 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3219 if (CI->getValue() == ShrMask) {
3220 // Masking out bits that the shift already masks.
3221 return ReplaceInstUsesWith(TheAnd, Op);
3222 } else if (CI != AndRHS) {
3223 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3228 case Instruction::AShr:
3230 // See if this is shifting in some sign extension, then masking it out
3232 if (Op->hasOneUse()) {
3233 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3234 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3235 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3236 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3237 if (C == AndRHS) { // Masking out bits shifted in.
3238 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3239 // Make the argument unsigned.
3240 Value *ShVal = Op->getOperand(0);
3241 ShVal = InsertNewInstBefore(
3242 BinaryOperator::createLShr(ShVal, OpRHS,
3243 Op->getName()), TheAnd);
3244 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3253 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3254 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3255 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3256 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3257 /// insert new instructions.
3258 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3259 bool isSigned, bool Inside,
3261 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3262 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3263 "Lo is not <= Hi in range emission code!");
3266 if (Lo == Hi) // Trivially false.
3267 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3269 // V >= Min && V < Hi --> V < Hi
3270 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3271 ICmpInst::Predicate pred = (isSigned ?
3272 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3273 return new ICmpInst(pred, V, Hi);
3276 // Emit V-Lo <u Hi-Lo
3277 Constant *NegLo = ConstantExpr::getNeg(Lo);
3278 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3279 InsertNewInstBefore(Add, IB);
3280 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3281 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3284 if (Lo == Hi) // Trivially true.
3285 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3287 // V < Min || V >= Hi -> V > Hi-1
3288 Hi = SubOne(cast<ConstantInt>(Hi));
3289 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3290 ICmpInst::Predicate pred = (isSigned ?
3291 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3292 return new ICmpInst(pred, V, Hi);
3295 // Emit V-Lo >u Hi-1-Lo
3296 // Note that Hi has already had one subtracted from it, above.
3297 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3298 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3299 InsertNewInstBefore(Add, IB);
3300 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3301 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3304 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3305 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3306 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3307 // not, since all 1s are not contiguous.
3308 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3309 const APInt& V = Val->getValue();
3310 uint32_t BitWidth = Val->getType()->getBitWidth();
3311 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3313 // look for the first zero bit after the run of ones
3314 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3315 // look for the first non-zero bit
3316 ME = V.getActiveBits();
3320 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3321 /// where isSub determines whether the operator is a sub. If we can fold one of
3322 /// the following xforms:
3324 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3325 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3326 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3328 /// return (A +/- B).
3330 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3331 ConstantInt *Mask, bool isSub,
3333 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3334 if (!LHSI || LHSI->getNumOperands() != 2 ||
3335 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3337 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3339 switch (LHSI->getOpcode()) {
3341 case Instruction::And:
3342 if (And(N, Mask) == Mask) {
3343 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3344 if ((Mask->getValue().countLeadingZeros() +
3345 Mask->getValue().countPopulation()) ==
3346 Mask->getValue().getBitWidth())
3349 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3350 // part, we don't need any explicit masks to take them out of A. If that
3351 // is all N is, ignore it.
3352 uint32_t MB = 0, ME = 0;
3353 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3354 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3355 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3356 if (MaskedValueIsZero(RHS, Mask))
3361 case Instruction::Or:
3362 case Instruction::Xor:
3363 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3364 if ((Mask->getValue().countLeadingZeros() +
3365 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3366 && And(N, Mask)->isZero())
3373 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3375 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3376 return InsertNewInstBefore(New, I);
3379 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3380 bool Changed = SimplifyCommutative(I);
3381 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3383 if (isa<UndefValue>(Op1)) // X & undef -> 0
3384 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3388 return ReplaceInstUsesWith(I, Op1);
3390 // See if we can simplify any instructions used by the instruction whose sole
3391 // purpose is to compute bits we don't care about.
3392 if (!isa<VectorType>(I.getType())) {
3393 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3394 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3395 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3396 KnownZero, KnownOne))
3399 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3400 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3401 return ReplaceInstUsesWith(I, I.getOperand(0));
3402 } else if (isa<ConstantAggregateZero>(Op1)) {
3403 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3407 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3408 const APInt& AndRHSMask = AndRHS->getValue();
3409 APInt NotAndRHS(~AndRHSMask);
3411 // Optimize a variety of ((val OP C1) & C2) combinations...
3412 if (isa<BinaryOperator>(Op0)) {
3413 Instruction *Op0I = cast<Instruction>(Op0);
3414 Value *Op0LHS = Op0I->getOperand(0);
3415 Value *Op0RHS = Op0I->getOperand(1);
3416 switch (Op0I->getOpcode()) {
3417 case Instruction::Xor:
3418 case Instruction::Or:
3419 // If the mask is only needed on one incoming arm, push it up.
3420 if (Op0I->hasOneUse()) {
3421 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3422 // Not masking anything out for the LHS, move to RHS.
3423 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3424 Op0RHS->getName()+".masked");
3425 InsertNewInstBefore(NewRHS, I);
3426 return BinaryOperator::create(
3427 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3429 if (!isa<Constant>(Op0RHS) &&
3430 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3431 // Not masking anything out for the RHS, move to LHS.
3432 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3433 Op0LHS->getName()+".masked");
3434 InsertNewInstBefore(NewLHS, I);
3435 return BinaryOperator::create(
3436 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3441 case Instruction::Add:
3442 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3443 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3444 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3445 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3446 return BinaryOperator::createAnd(V, AndRHS);
3447 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3448 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3451 case Instruction::Sub:
3452 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3453 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3454 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3455 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3456 return BinaryOperator::createAnd(V, AndRHS);
3460 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3461 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3463 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3464 // If this is an integer truncation or change from signed-to-unsigned, and
3465 // if the source is an and/or with immediate, transform it. This
3466 // frequently occurs for bitfield accesses.
3467 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3468 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3469 CastOp->getNumOperands() == 2)
3470 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3471 if (CastOp->getOpcode() == Instruction::And) {
3472 // Change: and (cast (and X, C1) to T), C2
3473 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3474 // This will fold the two constants together, which may allow
3475 // other simplifications.
3476 Instruction *NewCast = CastInst::createTruncOrBitCast(
3477 CastOp->getOperand(0), I.getType(),
3478 CastOp->getName()+".shrunk");
3479 NewCast = InsertNewInstBefore(NewCast, I);
3480 // trunc_or_bitcast(C1)&C2
3481 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3482 C3 = ConstantExpr::getAnd(C3, AndRHS);
3483 return BinaryOperator::createAnd(NewCast, C3);
3484 } else if (CastOp->getOpcode() == Instruction::Or) {
3485 // Change: and (cast (or X, C1) to T), C2
3486 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3487 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3488 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3489 return ReplaceInstUsesWith(I, AndRHS);
3494 // Try to fold constant and into select arguments.
3495 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3496 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3498 if (isa<PHINode>(Op0))
3499 if (Instruction *NV = FoldOpIntoPhi(I))
3503 Value *Op0NotVal = dyn_castNotVal(Op0);
3504 Value *Op1NotVal = dyn_castNotVal(Op1);
3506 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3507 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3509 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3510 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3511 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3512 I.getName()+".demorgan");
3513 InsertNewInstBefore(Or, I);
3514 return BinaryOperator::createNot(Or);
3518 Value *A = 0, *B = 0, *C = 0, *D = 0;
3519 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3520 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3521 return ReplaceInstUsesWith(I, Op1);
3523 // (A|B) & ~(A&B) -> A^B
3524 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3525 if ((A == C && B == D) || (A == D && B == C))
3526 return BinaryOperator::createXor(A, B);
3530 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3531 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3532 return ReplaceInstUsesWith(I, Op0);
3534 // ~(A&B) & (A|B) -> A^B
3535 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3536 if ((A == C && B == D) || (A == D && B == C))
3537 return BinaryOperator::createXor(A, B);
3541 if (Op0->hasOneUse() &&
3542 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3543 if (A == Op1) { // (A^B)&A -> A&(A^B)
3544 I.swapOperands(); // Simplify below
3545 std::swap(Op0, Op1);
3546 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3547 cast<BinaryOperator>(Op0)->swapOperands();
3548 I.swapOperands(); // Simplify below
3549 std::swap(Op0, Op1);
3552 if (Op1->hasOneUse() &&
3553 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3554 if (B == Op0) { // B&(A^B) -> B&(B^A)
3555 cast<BinaryOperator>(Op1)->swapOperands();
3558 if (A == Op0) { // A&(A^B) -> A & ~B
3559 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3560 InsertNewInstBefore(NotB, I);
3561 return BinaryOperator::createAnd(A, NotB);
3566 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3567 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3568 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3571 Value *LHSVal, *RHSVal;
3572 ConstantInt *LHSCst, *RHSCst;
3573 ICmpInst::Predicate LHSCC, RHSCC;
3574 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3575 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3576 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3577 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3578 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3579 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3580 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3581 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3583 // Don't try to fold ICMP_SLT + ICMP_ULT.
3584 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3585 ICmpInst::isSignedPredicate(LHSCC) ==
3586 ICmpInst::isSignedPredicate(RHSCC))) {
3587 // Ensure that the larger constant is on the RHS.
3588 ICmpInst::Predicate GT;
3589 if (ICmpInst::isSignedPredicate(LHSCC) ||
3590 (ICmpInst::isEquality(LHSCC) &&
3591 ICmpInst::isSignedPredicate(RHSCC)))
3592 GT = ICmpInst::ICMP_SGT;
3594 GT = ICmpInst::ICMP_UGT;
3596 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3597 ICmpInst *LHS = cast<ICmpInst>(Op0);
3598 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3599 std::swap(LHS, RHS);
3600 std::swap(LHSCst, RHSCst);
3601 std::swap(LHSCC, RHSCC);
3604 // At this point, we know we have have two icmp instructions
3605 // comparing a value against two constants and and'ing the result
3606 // together. Because of the above check, we know that we only have
3607 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3608 // (from the FoldICmpLogical check above), that the two constants
3609 // are not equal and that the larger constant is on the RHS
3610 assert(LHSCst != RHSCst && "Compares not folded above?");
3613 default: assert(0 && "Unknown integer condition code!");
3614 case ICmpInst::ICMP_EQ:
3616 default: assert(0 && "Unknown integer condition code!");
3617 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3618 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3619 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3620 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3621 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3622 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3623 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3624 return ReplaceInstUsesWith(I, LHS);
3626 case ICmpInst::ICMP_NE:
3628 default: assert(0 && "Unknown integer condition code!");
3629 case ICmpInst::ICMP_ULT:
3630 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3631 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3632 break; // (X != 13 & X u< 15) -> no change
3633 case ICmpInst::ICMP_SLT:
3634 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3635 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3636 break; // (X != 13 & X s< 15) -> no change
3637 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3638 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3639 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3640 return ReplaceInstUsesWith(I, RHS);
3641 case ICmpInst::ICMP_NE:
3642 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3643 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3644 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3645 LHSVal->getName()+".off");
3646 InsertNewInstBefore(Add, I);
3647 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3648 ConstantInt::get(Add->getType(), 1));
3650 break; // (X != 13 & X != 15) -> no change
3653 case ICmpInst::ICMP_ULT:
3655 default: assert(0 && "Unknown integer condition code!");
3656 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3657 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3658 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3659 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3661 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3662 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3663 return ReplaceInstUsesWith(I, LHS);
3664 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3668 case ICmpInst::ICMP_SLT:
3670 default: assert(0 && "Unknown integer condition code!");
3671 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3672 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3673 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3674 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3676 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3677 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3678 return ReplaceInstUsesWith(I, LHS);
3679 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3683 case ICmpInst::ICMP_UGT:
3685 default: assert(0 && "Unknown integer condition code!");
3686 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3687 return ReplaceInstUsesWith(I, LHS);
3688 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3689 return ReplaceInstUsesWith(I, RHS);
3690 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3692 case ICmpInst::ICMP_NE:
3693 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3694 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3695 break; // (X u> 13 & X != 15) -> no change
3696 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3697 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3699 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3703 case ICmpInst::ICMP_SGT:
3705 default: assert(0 && "Unknown integer condition code!");
3706 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3707 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3708 return ReplaceInstUsesWith(I, RHS);
3709 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3711 case ICmpInst::ICMP_NE:
3712 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3713 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3714 break; // (X s> 13 & X != 15) -> no change
3715 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3716 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3718 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3726 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3727 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3728 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3729 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3730 const Type *SrcTy = Op0C->getOperand(0)->getType();
3731 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3732 // Only do this if the casts both really cause code to be generated.
3733 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3735 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3737 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3738 Op1C->getOperand(0),
3740 InsertNewInstBefore(NewOp, I);
3741 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3745 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3746 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3747 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3748 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3749 SI0->getOperand(1) == SI1->getOperand(1) &&
3750 (SI0->hasOneUse() || SI1->hasOneUse())) {
3751 Instruction *NewOp =
3752 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3754 SI0->getName()), I);
3755 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3756 SI1->getOperand(1));
3760 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3761 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3762 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3763 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3764 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3765 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3766 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3767 // If either of the constants are nans, then the whole thing returns
3769 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3770 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3771 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3772 RHS->getOperand(0));
3777 return Changed ? &I : 0;
3780 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3781 /// in the result. If it does, and if the specified byte hasn't been filled in
3782 /// yet, fill it in and return false.
3783 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3784 Instruction *I = dyn_cast<Instruction>(V);
3785 if (I == 0) return true;
3787 // If this is an or instruction, it is an inner node of the bswap.
3788 if (I->getOpcode() == Instruction::Or)
3789 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3790 CollectBSwapParts(I->getOperand(1), ByteValues);
3792 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3793 // If this is a shift by a constant int, and it is "24", then its operand
3794 // defines a byte. We only handle unsigned types here.
3795 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3796 // Not shifting the entire input by N-1 bytes?
3797 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3798 8*(ByteValues.size()-1))
3802 if (I->getOpcode() == Instruction::Shl) {
3803 // X << 24 defines the top byte with the lowest of the input bytes.
3804 DestNo = ByteValues.size()-1;
3806 // X >>u 24 defines the low byte with the highest of the input bytes.
3810 // If the destination byte value is already defined, the values are or'd
3811 // together, which isn't a bswap (unless it's an or of the same bits).
3812 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3814 ByteValues[DestNo] = I->getOperand(0);
3818 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3820 Value *Shift = 0, *ShiftLHS = 0;
3821 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3822 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3823 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3825 Instruction *SI = cast<Instruction>(Shift);
3827 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3828 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3829 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3832 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3834 if (AndAmt->getValue().getActiveBits() > 64)
3836 uint64_t AndAmtVal = AndAmt->getZExtValue();
3837 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3838 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3840 // Unknown mask for bswap.
3841 if (DestByte == ByteValues.size()) return true;
3843 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3845 if (SI->getOpcode() == Instruction::Shl)
3846 SrcByte = DestByte - ShiftBytes;
3848 SrcByte = DestByte + ShiftBytes;
3850 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3851 if (SrcByte != ByteValues.size()-DestByte-1)
3854 // If the destination byte value is already defined, the values are or'd
3855 // together, which isn't a bswap (unless it's an or of the same bits).
3856 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3858 ByteValues[DestByte] = SI->getOperand(0);
3862 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3863 /// If so, insert the new bswap intrinsic and return it.
3864 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3865 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3866 if (!ITy || ITy->getBitWidth() % 16)
3867 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3869 /// ByteValues - For each byte of the result, we keep track of which value
3870 /// defines each byte.
3871 SmallVector<Value*, 8> ByteValues;
3872 ByteValues.resize(ITy->getBitWidth()/8);
3874 // Try to find all the pieces corresponding to the bswap.
3875 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3876 CollectBSwapParts(I.getOperand(1), ByteValues))
3879 // Check to see if all of the bytes come from the same value.
3880 Value *V = ByteValues[0];
3881 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3883 // Check to make sure that all of the bytes come from the same value.
3884 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3885 if (ByteValues[i] != V)
3887 const Type *Tys[] = { ITy };
3888 Module *M = I.getParent()->getParent()->getParent();
3889 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3890 return new CallInst(F, V);
3894 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3895 bool Changed = SimplifyCommutative(I);
3896 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3898 if (isa<UndefValue>(Op1)) // X | undef -> -1
3899 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3903 return ReplaceInstUsesWith(I, Op0);
3905 // See if we can simplify any instructions used by the instruction whose sole
3906 // purpose is to compute bits we don't care about.
3907 if (!isa<VectorType>(I.getType())) {
3908 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3909 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3910 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3911 KnownZero, KnownOne))
3913 } else if (isa<ConstantAggregateZero>(Op1)) {
3914 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3915 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3916 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3917 return ReplaceInstUsesWith(I, I.getOperand(1));
3923 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3924 ConstantInt *C1 = 0; Value *X = 0;
3925 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3926 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3927 Instruction *Or = BinaryOperator::createOr(X, RHS);
3928 InsertNewInstBefore(Or, I);
3930 return BinaryOperator::createAnd(Or,
3931 ConstantInt::get(RHS->getValue() | C1->getValue()));
3934 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3935 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3936 Instruction *Or = BinaryOperator::createOr(X, RHS);
3937 InsertNewInstBefore(Or, I);
3939 return BinaryOperator::createXor(Or,
3940 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3943 // Try to fold constant and into select arguments.
3944 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3945 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3947 if (isa<PHINode>(Op0))
3948 if (Instruction *NV = FoldOpIntoPhi(I))
3952 Value *A = 0, *B = 0;
3953 ConstantInt *C1 = 0, *C2 = 0;
3955 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3956 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3957 return ReplaceInstUsesWith(I, Op1);
3958 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3959 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3960 return ReplaceInstUsesWith(I, Op0);
3962 // (A | B) | C and A | (B | C) -> bswap if possible.
3963 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3964 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3965 match(Op1, m_Or(m_Value(), m_Value())) ||
3966 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3967 match(Op1, m_Shift(m_Value(), m_Value())))) {
3968 if (Instruction *BSwap = MatchBSwap(I))
3972 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3973 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3974 MaskedValueIsZero(Op1, C1->getValue())) {
3975 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3976 InsertNewInstBefore(NOr, I);
3978 return BinaryOperator::createXor(NOr, C1);
3981 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3982 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3983 MaskedValueIsZero(Op0, C1->getValue())) {
3984 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3985 InsertNewInstBefore(NOr, I);
3987 return BinaryOperator::createXor(NOr, C1);
3991 Value *C = 0, *D = 0;
3992 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3993 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3994 Value *V1 = 0, *V2 = 0, *V3 = 0;
3995 C1 = dyn_cast<ConstantInt>(C);
3996 C2 = dyn_cast<ConstantInt>(D);
3997 if (C1 && C2) { // (A & C1)|(B & C2)
3998 // If we have: ((V + N) & C1) | (V & C2)
3999 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4000 // replace with V+N.
4001 if (C1->getValue() == ~C2->getValue()) {
4002 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4003 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4004 // Add commutes, try both ways.
4005 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4006 return ReplaceInstUsesWith(I, A);
4007 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4008 return ReplaceInstUsesWith(I, A);
4010 // Or commutes, try both ways.
4011 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4012 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4013 // Add commutes, try both ways.
4014 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4015 return ReplaceInstUsesWith(I, B);
4016 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4017 return ReplaceInstUsesWith(I, B);
4020 V1 = 0; V2 = 0; V3 = 0;
4023 // Check to see if we have any common things being and'ed. If so, find the
4024 // terms for V1 & (V2|V3).
4025 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4026 if (A == B) // (A & C)|(A & D) == A & (C|D)
4027 V1 = A, V2 = C, V3 = D;
4028 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4029 V1 = A, V2 = B, V3 = C;
4030 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4031 V1 = C, V2 = A, V3 = D;
4032 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4033 V1 = C, V2 = A, V3 = B;
4037 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
4038 return BinaryOperator::createAnd(V1, Or);
4043 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4044 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4045 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4046 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4047 SI0->getOperand(1) == SI1->getOperand(1) &&
4048 (SI0->hasOneUse() || SI1->hasOneUse())) {
4049 Instruction *NewOp =
4050 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4052 SI0->getName()), I);
4053 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4054 SI1->getOperand(1));
4058 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4059 if (A == Op1) // ~A | A == -1
4060 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4064 // Note, A is still live here!
4065 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4067 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4069 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4070 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4071 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4072 I.getName()+".demorgan"), I);
4073 return BinaryOperator::createNot(And);
4077 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4078 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4079 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4082 Value *LHSVal, *RHSVal;
4083 ConstantInt *LHSCst, *RHSCst;
4084 ICmpInst::Predicate LHSCC, RHSCC;
4085 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4086 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4087 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4088 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4089 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4090 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4091 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4092 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4093 // We can't fold (ugt x, C) | (sgt x, C2).
4094 PredicatesFoldable(LHSCC, RHSCC)) {
4095 // Ensure that the larger constant is on the RHS.
4096 ICmpInst *LHS = cast<ICmpInst>(Op0);
4098 if (ICmpInst::isSignedPredicate(LHSCC))
4099 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4101 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4104 std::swap(LHS, RHS);
4105 std::swap(LHSCst, RHSCst);
4106 std::swap(LHSCC, RHSCC);
4109 // At this point, we know we have have two icmp instructions
4110 // comparing a value against two constants and or'ing the result
4111 // together. Because of the above check, we know that we only have
4112 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4113 // FoldICmpLogical check above), that the two constants are not
4115 assert(LHSCst != RHSCst && "Compares not folded above?");
4118 default: assert(0 && "Unknown integer condition code!");
4119 case ICmpInst::ICMP_EQ:
4121 default: assert(0 && "Unknown integer condition code!");
4122 case ICmpInst::ICMP_EQ:
4123 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4124 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4125 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4126 LHSVal->getName()+".off");
4127 InsertNewInstBefore(Add, I);
4128 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4129 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4131 break; // (X == 13 | X == 15) -> no change
4132 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4133 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4135 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4136 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4137 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4138 return ReplaceInstUsesWith(I, RHS);
4141 case ICmpInst::ICMP_NE:
4143 default: assert(0 && "Unknown integer condition code!");
4144 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4145 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4146 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4147 return ReplaceInstUsesWith(I, LHS);
4148 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4149 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4150 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4151 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4154 case ICmpInst::ICMP_ULT:
4156 default: assert(0 && "Unknown integer condition code!");
4157 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4159 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4160 // If RHSCst is [us]MAXINT, it is always false. Not handling
4161 // this can cause overflow.
4162 if (RHSCst->isMaxValue(false))
4163 return ReplaceInstUsesWith(I, LHS);
4164 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4166 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4168 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4169 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4170 return ReplaceInstUsesWith(I, RHS);
4171 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4175 case ICmpInst::ICMP_SLT:
4177 default: assert(0 && "Unknown integer condition code!");
4178 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4180 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4181 // If RHSCst is [us]MAXINT, it is always false. Not handling
4182 // this can cause overflow.
4183 if (RHSCst->isMaxValue(true))
4184 return ReplaceInstUsesWith(I, LHS);
4185 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4187 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4189 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4190 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4191 return ReplaceInstUsesWith(I, RHS);
4192 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4196 case ICmpInst::ICMP_UGT:
4198 default: assert(0 && "Unknown integer condition code!");
4199 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4200 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4201 return ReplaceInstUsesWith(I, LHS);
4202 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4204 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4205 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4206 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4207 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4211 case ICmpInst::ICMP_SGT:
4213 default: assert(0 && "Unknown integer condition code!");
4214 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4215 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4216 return ReplaceInstUsesWith(I, LHS);
4217 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4219 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4220 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4221 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4222 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4230 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4231 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4232 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4233 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4234 const Type *SrcTy = Op0C->getOperand(0)->getType();
4235 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4236 // Only do this if the casts both really cause code to be generated.
4237 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4239 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4241 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4242 Op1C->getOperand(0),
4244 InsertNewInstBefore(NewOp, I);
4245 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4251 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4252 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4253 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4254 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4255 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4256 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4257 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4258 // If either of the constants are nans, then the whole thing returns
4260 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4261 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4263 // Otherwise, no need to compare the two constants, compare the
4265 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4266 RHS->getOperand(0));
4271 return Changed ? &I : 0;
4274 // XorSelf - Implements: X ^ X --> 0
4277 XorSelf(Value *rhs) : RHS(rhs) {}
4278 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4279 Instruction *apply(BinaryOperator &Xor) const {
4285 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4286 bool Changed = SimplifyCommutative(I);
4287 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4289 if (isa<UndefValue>(Op1))
4290 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4292 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4293 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4294 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4295 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4298 // See if we can simplify any instructions used by the instruction whose sole
4299 // purpose is to compute bits we don't care about.
4300 if (!isa<VectorType>(I.getType())) {
4301 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4302 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4303 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4304 KnownZero, KnownOne))
4306 } else if (isa<ConstantAggregateZero>(Op1)) {
4307 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4310 // Is this a ~ operation?
4311 if (Value *NotOp = dyn_castNotVal(&I)) {
4312 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4313 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4314 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4315 if (Op0I->getOpcode() == Instruction::And ||
4316 Op0I->getOpcode() == Instruction::Or) {
4317 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4318 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4320 BinaryOperator::createNot(Op0I->getOperand(1),
4321 Op0I->getOperand(1)->getName()+".not");
4322 InsertNewInstBefore(NotY, I);
4323 if (Op0I->getOpcode() == Instruction::And)
4324 return BinaryOperator::createOr(Op0NotVal, NotY);
4326 return BinaryOperator::createAnd(Op0NotVal, NotY);
4333 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4334 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4335 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4336 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4337 return new ICmpInst(ICI->getInversePredicate(),
4338 ICI->getOperand(0), ICI->getOperand(1));
4340 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4341 return new FCmpInst(FCI->getInversePredicate(),
4342 FCI->getOperand(0), FCI->getOperand(1));
4345 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4346 // ~(c-X) == X-c-1 == X+(-c-1)
4347 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4348 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4349 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4350 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4351 ConstantInt::get(I.getType(), 1));
4352 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4355 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4356 if (Op0I->getOpcode() == Instruction::Add) {
4357 // ~(X-c) --> (-c-1)-X
4358 if (RHS->isAllOnesValue()) {
4359 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4360 return BinaryOperator::createSub(
4361 ConstantExpr::getSub(NegOp0CI,
4362 ConstantInt::get(I.getType(), 1)),
4363 Op0I->getOperand(0));
4364 } else if (RHS->getValue().isSignBit()) {
4365 // (X + C) ^ signbit -> (X + C + signbit)
4366 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4367 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4370 } else if (Op0I->getOpcode() == Instruction::Or) {
4371 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4372 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4373 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4374 // Anything in both C1 and C2 is known to be zero, remove it from
4376 Constant *CommonBits = And(Op0CI, RHS);
4377 NewRHS = ConstantExpr::getAnd(NewRHS,
4378 ConstantExpr::getNot(CommonBits));
4379 AddToWorkList(Op0I);
4380 I.setOperand(0, Op0I->getOperand(0));
4381 I.setOperand(1, NewRHS);
4387 // Try to fold constant and into select arguments.
4388 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4389 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4391 if (isa<PHINode>(Op0))
4392 if (Instruction *NV = FoldOpIntoPhi(I))
4396 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4398 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4400 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4402 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4405 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4408 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4409 if (A == Op0) { // B^(B|A) == (A|B)^B
4410 Op1I->swapOperands();
4412 std::swap(Op0, Op1);
4413 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4414 I.swapOperands(); // Simplified below.
4415 std::swap(Op0, Op1);
4417 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4418 if (Op0 == A) // A^(A^B) == B
4419 return ReplaceInstUsesWith(I, B);
4420 else if (Op0 == B) // A^(B^A) == B
4421 return ReplaceInstUsesWith(I, A);
4422 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4423 if (A == Op0) { // A^(A&B) -> A^(B&A)
4424 Op1I->swapOperands();
4427 if (B == Op0) { // A^(B&A) -> (B&A)^A
4428 I.swapOperands(); // Simplified below.
4429 std::swap(Op0, Op1);
4434 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4437 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4438 if (A == Op1) // (B|A)^B == (A|B)^B
4440 if (B == Op1) { // (A|B)^B == A & ~B
4442 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4443 return BinaryOperator::createAnd(A, NotB);
4445 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4446 if (Op1 == A) // (A^B)^A == B
4447 return ReplaceInstUsesWith(I, B);
4448 else if (Op1 == B) // (B^A)^A == B
4449 return ReplaceInstUsesWith(I, A);
4450 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4451 if (A == Op1) // (A&B)^A -> (B&A)^A
4453 if (B == Op1 && // (B&A)^A == ~B & A
4454 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4456 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4457 return BinaryOperator::createAnd(N, Op1);
4462 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4463 if (Op0I && Op1I && Op0I->isShift() &&
4464 Op0I->getOpcode() == Op1I->getOpcode() &&
4465 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4466 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4467 Instruction *NewOp =
4468 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4469 Op1I->getOperand(0),
4470 Op0I->getName()), I);
4471 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4472 Op1I->getOperand(1));
4476 Value *A, *B, *C, *D;
4477 // (A & B)^(A | B) -> A ^ B
4478 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4479 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4480 if ((A == C && B == D) || (A == D && B == C))
4481 return BinaryOperator::createXor(A, B);
4483 // (A | B)^(A & B) -> A ^ B
4484 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4485 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4486 if ((A == C && B == D) || (A == D && B == C))
4487 return BinaryOperator::createXor(A, B);
4491 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4492 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4493 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4494 // (X & Y)^(X & Y) -> (Y^Z) & X
4495 Value *X = 0, *Y = 0, *Z = 0;
4497 X = A, Y = B, Z = D;
4499 X = A, Y = B, Z = C;
4501 X = B, Y = A, Z = D;
4503 X = B, Y = A, Z = C;
4506 Instruction *NewOp =
4507 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4508 return BinaryOperator::createAnd(NewOp, X);
4513 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4514 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4515 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4518 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4519 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4520 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4521 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4522 const Type *SrcTy = Op0C->getOperand(0)->getType();
4523 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4524 // Only do this if the casts both really cause code to be generated.
4525 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4527 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4529 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4530 Op1C->getOperand(0),
4532 InsertNewInstBefore(NewOp, I);
4533 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4537 return Changed ? &I : 0;
4540 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4541 /// overflowed for this type.
4542 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4543 ConstantInt *In2, bool IsSigned = false) {
4544 Result = cast<ConstantInt>(Add(In1, In2));
4547 if (In2->getValue().isNegative())
4548 return Result->getValue().sgt(In1->getValue());
4550 return Result->getValue().slt(In1->getValue());
4552 return Result->getValue().ult(In1->getValue());
4555 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4556 /// code necessary to compute the offset from the base pointer (without adding
4557 /// in the base pointer). Return the result as a signed integer of intptr size.
4558 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4559 TargetData &TD = IC.getTargetData();
4560 gep_type_iterator GTI = gep_type_begin(GEP);
4561 const Type *IntPtrTy = TD.getIntPtrType();
4562 Value *Result = Constant::getNullValue(IntPtrTy);
4564 // Build a mask for high order bits.
4565 unsigned IntPtrWidth = TD.getPointerSize()*8;
4566 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4568 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4569 Value *Op = GEP->getOperand(i);
4570 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4571 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4572 if (OpC->isZero()) continue;
4574 // Handle a struct index, which adds its field offset to the pointer.
4575 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4576 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4578 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4579 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4581 Result = IC.InsertNewInstBefore(
4582 BinaryOperator::createAdd(Result,
4583 ConstantInt::get(IntPtrTy, Size),
4584 GEP->getName()+".offs"), I);
4588 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4589 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4590 Scale = ConstantExpr::getMul(OC, Scale);
4591 if (Constant *RC = dyn_cast<Constant>(Result))
4592 Result = ConstantExpr::getAdd(RC, Scale);
4594 // Emit an add instruction.
4595 Result = IC.InsertNewInstBefore(
4596 BinaryOperator::createAdd(Result, Scale,
4597 GEP->getName()+".offs"), I);
4601 // Convert to correct type.
4602 if (Op->getType() != IntPtrTy) {
4603 if (Constant *OpC = dyn_cast<Constant>(Op))
4604 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4606 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4607 Op->getName()+".c"), I);
4610 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4611 if (Constant *OpC = dyn_cast<Constant>(Op))
4612 Op = ConstantExpr::getMul(OpC, Scale);
4613 else // We'll let instcombine(mul) convert this to a shl if possible.
4614 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4615 GEP->getName()+".idx"), I);
4618 // Emit an add instruction.
4619 if (isa<Constant>(Op) && isa<Constant>(Result))
4620 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4621 cast<Constant>(Result));
4623 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4624 GEP->getName()+".offs"), I);
4629 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4630 /// else. At this point we know that the GEP is on the LHS of the comparison.
4631 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4632 ICmpInst::Predicate Cond,
4634 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4636 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4637 if (isa<PointerType>(CI->getOperand(0)->getType()))
4638 RHS = CI->getOperand(0);
4640 Value *PtrBase = GEPLHS->getOperand(0);
4641 if (PtrBase == RHS) {
4642 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4643 // This transformation is valid because we know pointers can't overflow.
4644 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4645 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4646 Constant::getNullValue(Offset->getType()));
4647 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4648 // If the base pointers are different, but the indices are the same, just
4649 // compare the base pointer.
4650 if (PtrBase != GEPRHS->getOperand(0)) {
4651 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4652 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4653 GEPRHS->getOperand(0)->getType();
4655 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4656 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4657 IndicesTheSame = false;
4661 // If all indices are the same, just compare the base pointers.
4663 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4664 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4666 // Otherwise, the base pointers are different and the indices are
4667 // different, bail out.
4671 // If one of the GEPs has all zero indices, recurse.
4672 bool AllZeros = true;
4673 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4674 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4675 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4680 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4681 ICmpInst::getSwappedPredicate(Cond), I);
4683 // If the other GEP has all zero indices, recurse.
4685 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4686 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4687 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4692 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4694 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4695 // If the GEPs only differ by one index, compare it.
4696 unsigned NumDifferences = 0; // Keep track of # differences.
4697 unsigned DiffOperand = 0; // The operand that differs.
4698 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4699 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4700 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4701 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4702 // Irreconcilable differences.
4706 if (NumDifferences++) break;
4711 if (NumDifferences == 0) // SAME GEP?
4712 return ReplaceInstUsesWith(I, // No comparison is needed here.
4713 ConstantInt::get(Type::Int1Ty,
4714 isTrueWhenEqual(Cond)));
4716 else if (NumDifferences == 1) {
4717 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4718 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4719 // Make sure we do a signed comparison here.
4720 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4724 // Only lower this if the icmp is the only user of the GEP or if we expect
4725 // the result to fold to a constant!
4726 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4727 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4728 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4729 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4730 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4731 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4737 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4738 bool Changed = SimplifyCompare(I);
4739 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4741 // Fold trivial predicates.
4742 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4743 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4744 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4745 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4747 // Simplify 'fcmp pred X, X'
4749 switch (I.getPredicate()) {
4750 default: assert(0 && "Unknown predicate!");
4751 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4752 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4753 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4754 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4755 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4756 case FCmpInst::FCMP_OLT: // True if ordered and less than
4757 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4758 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4760 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4761 case FCmpInst::FCMP_ULT: // True if unordered or less than
4762 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4763 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4764 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4765 I.setPredicate(FCmpInst::FCMP_UNO);
4766 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4769 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4770 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4771 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4772 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4773 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4774 I.setPredicate(FCmpInst::FCMP_ORD);
4775 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4780 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4781 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4783 // Handle fcmp with constant RHS
4784 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4785 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4786 switch (LHSI->getOpcode()) {
4787 case Instruction::PHI:
4788 if (Instruction *NV = FoldOpIntoPhi(I))
4791 case Instruction::Select:
4792 // If either operand of the select is a constant, we can fold the
4793 // comparison into the select arms, which will cause one to be
4794 // constant folded and the select turned into a bitwise or.
4795 Value *Op1 = 0, *Op2 = 0;
4796 if (LHSI->hasOneUse()) {
4797 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4798 // Fold the known value into the constant operand.
4799 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4800 // Insert a new FCmp of the other select operand.
4801 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4802 LHSI->getOperand(2), RHSC,
4804 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4805 // Fold the known value into the constant operand.
4806 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4807 // Insert a new FCmp of the other select operand.
4808 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4809 LHSI->getOperand(1), RHSC,
4815 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4820 return Changed ? &I : 0;
4823 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4824 bool Changed = SimplifyCompare(I);
4825 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4826 const Type *Ty = Op0->getType();
4830 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4831 isTrueWhenEqual(I)));
4833 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4834 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4836 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4837 // addresses never equal each other! We already know that Op0 != Op1.
4838 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4839 isa<ConstantPointerNull>(Op0)) &&
4840 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4841 isa<ConstantPointerNull>(Op1)))
4842 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4843 !isTrueWhenEqual(I)));
4845 // icmp's with boolean values can always be turned into bitwise operations
4846 if (Ty == Type::Int1Ty) {
4847 switch (I.getPredicate()) {
4848 default: assert(0 && "Invalid icmp instruction!");
4849 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4850 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4851 InsertNewInstBefore(Xor, I);
4852 return BinaryOperator::createNot(Xor);
4854 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4855 return BinaryOperator::createXor(Op0, Op1);
4857 case ICmpInst::ICMP_UGT:
4858 case ICmpInst::ICMP_SGT:
4859 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4861 case ICmpInst::ICMP_ULT:
4862 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4863 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4864 InsertNewInstBefore(Not, I);
4865 return BinaryOperator::createAnd(Not, Op1);
4867 case ICmpInst::ICMP_UGE:
4868 case ICmpInst::ICMP_SGE:
4869 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4871 case ICmpInst::ICMP_ULE:
4872 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4873 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4874 InsertNewInstBefore(Not, I);
4875 return BinaryOperator::createOr(Not, Op1);
4880 // See if we are doing a comparison between a constant and an instruction that
4881 // can be folded into the comparison.
4882 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4885 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
4886 if (I.isEquality() && CI->isNullValue() &&
4887 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
4888 // (icmp cond A B) if cond is equality
4889 return new ICmpInst(I.getPredicate(), A, B);
4892 switch (I.getPredicate()) {
4894 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4895 if (CI->isMinValue(false))
4896 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4897 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4898 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4899 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4900 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4901 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4902 if (CI->isMinValue(true))
4903 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4904 ConstantInt::getAllOnesValue(Op0->getType()));
4908 case ICmpInst::ICMP_SLT:
4909 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4910 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4911 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4912 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4913 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4914 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4917 case ICmpInst::ICMP_UGT:
4918 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4919 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4920 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4921 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4922 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4923 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4925 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4926 if (CI->isMaxValue(true))
4927 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4928 ConstantInt::getNullValue(Op0->getType()));
4931 case ICmpInst::ICMP_SGT:
4932 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4933 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4934 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4935 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4936 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4937 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4940 case ICmpInst::ICMP_ULE:
4941 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4942 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4943 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4944 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4945 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4946 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4949 case ICmpInst::ICMP_SLE:
4950 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4951 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4952 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4953 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4954 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4955 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4958 case ICmpInst::ICMP_UGE:
4959 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4960 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4961 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4962 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4963 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4964 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4967 case ICmpInst::ICMP_SGE:
4968 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4969 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4970 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4971 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4972 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4973 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4977 // If we still have a icmp le or icmp ge instruction, turn it into the
4978 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4979 // already been handled above, this requires little checking.
4981 switch (I.getPredicate()) {
4983 case ICmpInst::ICMP_ULE:
4984 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4985 case ICmpInst::ICMP_SLE:
4986 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4987 case ICmpInst::ICMP_UGE:
4988 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4989 case ICmpInst::ICMP_SGE:
4990 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4993 // See if we can fold the comparison based on bits known to be zero or one
4994 // in the input. If this comparison is a normal comparison, it demands all
4995 // bits, if it is a sign bit comparison, it only demands the sign bit.
4998 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5000 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5001 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5002 if (SimplifyDemandedBits(Op0,
5003 isSignBit ? APInt::getSignBit(BitWidth)
5004 : APInt::getAllOnesValue(BitWidth),
5005 KnownZero, KnownOne, 0))
5008 // Given the known and unknown bits, compute a range that the LHS could be
5010 if ((KnownOne | KnownZero) != 0) {
5011 // Compute the Min, Max and RHS values based on the known bits. For the
5012 // EQ and NE we use unsigned values.
5013 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5014 const APInt& RHSVal = CI->getValue();
5015 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5016 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5019 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5022 switch (I.getPredicate()) { // LE/GE have been folded already.
5023 default: assert(0 && "Unknown icmp opcode!");
5024 case ICmpInst::ICMP_EQ:
5025 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5026 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5028 case ICmpInst::ICMP_NE:
5029 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5030 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5032 case ICmpInst::ICMP_ULT:
5033 if (Max.ult(RHSVal))
5034 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5035 if (Min.uge(RHSVal))
5036 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5038 case ICmpInst::ICMP_UGT:
5039 if (Min.ugt(RHSVal))
5040 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5041 if (Max.ule(RHSVal))
5042 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5044 case ICmpInst::ICMP_SLT:
5045 if (Max.slt(RHSVal))
5046 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5047 if (Min.sgt(RHSVal))
5048 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5050 case ICmpInst::ICMP_SGT:
5051 if (Min.sgt(RHSVal))
5052 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5053 if (Max.sle(RHSVal))
5054 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5059 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5060 // instruction, see if that instruction also has constants so that the
5061 // instruction can be folded into the icmp
5062 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5063 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5067 // Handle icmp with constant (but not simple integer constant) RHS
5068 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5069 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5070 switch (LHSI->getOpcode()) {
5071 case Instruction::GetElementPtr:
5072 if (RHSC->isNullValue()) {
5073 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5074 bool isAllZeros = true;
5075 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5076 if (!isa<Constant>(LHSI->getOperand(i)) ||
5077 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5082 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5083 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5087 case Instruction::PHI:
5088 if (Instruction *NV = FoldOpIntoPhi(I))
5091 case Instruction::Select: {
5092 // If either operand of the select is a constant, we can fold the
5093 // comparison into the select arms, which will cause one to be
5094 // constant folded and the select turned into a bitwise or.
5095 Value *Op1 = 0, *Op2 = 0;
5096 if (LHSI->hasOneUse()) {
5097 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5098 // Fold the known value into the constant operand.
5099 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5100 // Insert a new ICmp of the other select operand.
5101 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5102 LHSI->getOperand(2), RHSC,
5104 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5105 // Fold the known value into the constant operand.
5106 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5107 // Insert a new ICmp of the other select operand.
5108 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5109 LHSI->getOperand(1), RHSC,
5115 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5118 case Instruction::Malloc:
5119 // If we have (malloc != null), and if the malloc has a single use, we
5120 // can assume it is successful and remove the malloc.
5121 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5122 AddToWorkList(LHSI);
5123 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5124 !isTrueWhenEqual(I)));
5130 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5131 if (User *GEP = dyn_castGetElementPtr(Op0))
5132 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5134 if (User *GEP = dyn_castGetElementPtr(Op1))
5135 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5136 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5139 // Test to see if the operands of the icmp are casted versions of other
5140 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5142 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5143 if (isa<PointerType>(Op0->getType()) &&
5144 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5145 // We keep moving the cast from the left operand over to the right
5146 // operand, where it can often be eliminated completely.
5147 Op0 = CI->getOperand(0);
5149 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5150 // so eliminate it as well.
5151 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5152 Op1 = CI2->getOperand(0);
5154 // If Op1 is a constant, we can fold the cast into the constant.
5155 if (Op0->getType() != Op1->getType())
5156 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5157 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5159 // Otherwise, cast the RHS right before the icmp
5160 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5162 return new ICmpInst(I.getPredicate(), Op0, Op1);
5166 if (isa<CastInst>(Op0)) {
5167 // Handle the special case of: icmp (cast bool to X), <cst>
5168 // This comes up when you have code like
5171 // For generality, we handle any zero-extension of any operand comparison
5172 // with a constant or another cast from the same type.
5173 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5174 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5178 if (I.isEquality()) {
5179 Value *A, *B, *C, *D;
5180 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5181 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5182 Value *OtherVal = A == Op1 ? B : A;
5183 return new ICmpInst(I.getPredicate(), OtherVal,
5184 Constant::getNullValue(A->getType()));
5187 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5188 // A^c1 == C^c2 --> A == C^(c1^c2)
5189 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5190 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5191 if (Op1->hasOneUse()) {
5192 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5193 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5194 return new ICmpInst(I.getPredicate(), A,
5195 InsertNewInstBefore(Xor, I));
5198 // A^B == A^D -> B == D
5199 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5200 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5201 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5202 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5206 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5207 (A == Op0 || B == Op0)) {
5208 // A == (A^B) -> B == 0
5209 Value *OtherVal = A == Op0 ? B : A;
5210 return new ICmpInst(I.getPredicate(), OtherVal,
5211 Constant::getNullValue(A->getType()));
5213 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5214 // (A-B) == A -> B == 0
5215 return new ICmpInst(I.getPredicate(), B,
5216 Constant::getNullValue(B->getType()));
5218 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5219 // A == (A-B) -> B == 0
5220 return new ICmpInst(I.getPredicate(), B,
5221 Constant::getNullValue(B->getType()));
5224 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5225 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5226 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5227 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5228 Value *X = 0, *Y = 0, *Z = 0;
5231 X = B; Y = D; Z = A;
5232 } else if (A == D) {
5233 X = B; Y = C; Z = A;
5234 } else if (B == C) {
5235 X = A; Y = D; Z = B;
5236 } else if (B == D) {
5237 X = A; Y = C; Z = B;
5240 if (X) { // Build (X^Y) & Z
5241 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5242 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5243 I.setOperand(0, Op1);
5244 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5249 return Changed ? &I : 0;
5253 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5254 /// and CmpRHS are both known to be integer constants.
5255 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5256 ConstantInt *DivRHS) {
5257 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5258 const APInt &CmpRHSV = CmpRHS->getValue();
5260 // FIXME: If the operand types don't match the type of the divide
5261 // then don't attempt this transform. The code below doesn't have the
5262 // logic to deal with a signed divide and an unsigned compare (and
5263 // vice versa). This is because (x /s C1) <s C2 produces different
5264 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5265 // (x /u C1) <u C2. Simply casting the operands and result won't
5266 // work. :( The if statement below tests that condition and bails
5268 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5269 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5271 if (DivRHS->isZero())
5272 return 0; // The ProdOV computation fails on divide by zero.
5274 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5275 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5276 // C2 (CI). By solving for X we can turn this into a range check
5277 // instead of computing a divide.
5278 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5280 // Determine if the product overflows by seeing if the product is
5281 // not equal to the divide. Make sure we do the same kind of divide
5282 // as in the LHS instruction that we're folding.
5283 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5284 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5286 // Get the ICmp opcode
5287 ICmpInst::Predicate Pred = ICI.getPredicate();
5289 // Figure out the interval that is being checked. For example, a comparison
5290 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5291 // Compute this interval based on the constants involved and the signedness of
5292 // the compare/divide. This computes a half-open interval, keeping track of
5293 // whether either value in the interval overflows. After analysis each
5294 // overflow variable is set to 0 if it's corresponding bound variable is valid
5295 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5296 int LoOverflow = 0, HiOverflow = 0;
5297 ConstantInt *LoBound = 0, *HiBound = 0;
5300 if (!DivIsSigned) { // udiv
5301 // e.g. X/5 op 3 --> [15, 20)
5303 HiOverflow = LoOverflow = ProdOV;
5305 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5306 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5307 if (CmpRHSV == 0) { // (X / pos) op 0
5308 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5309 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5311 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5312 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5313 HiOverflow = LoOverflow = ProdOV;
5315 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5316 } else { // (X / pos) op neg
5317 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5318 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5319 LoOverflow = AddWithOverflow(LoBound, Prod,
5320 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5321 HiBound = AddOne(Prod);
5322 HiOverflow = ProdOV ? -1 : 0;
5324 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5325 if (CmpRHSV == 0) { // (X / neg) op 0
5326 // e.g. X/-5 op 0 --> [-4, 5)
5327 LoBound = AddOne(DivRHS);
5328 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5329 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5330 HiOverflow = 1; // [INTMIN+1, overflow)
5331 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5333 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5334 // e.g. X/-5 op 3 --> [-19, -14)
5335 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5337 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5338 HiBound = AddOne(Prod);
5339 } else { // (X / neg) op neg
5340 // e.g. X/-5 op -3 --> [15, 20)
5342 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5343 HiBound = Subtract(Prod, DivRHS);
5346 // Dividing by a negative swaps the condition. LT <-> GT
5347 Pred = ICmpInst::getSwappedPredicate(Pred);
5350 Value *X = DivI->getOperand(0);
5352 default: assert(0 && "Unhandled icmp opcode!");
5353 case ICmpInst::ICMP_EQ:
5354 if (LoOverflow && HiOverflow)
5355 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5356 else if (HiOverflow)
5357 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5358 ICmpInst::ICMP_UGE, X, LoBound);
5359 else if (LoOverflow)
5360 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5361 ICmpInst::ICMP_ULT, X, HiBound);
5363 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5364 case ICmpInst::ICMP_NE:
5365 if (LoOverflow && HiOverflow)
5366 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5367 else if (HiOverflow)
5368 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5369 ICmpInst::ICMP_ULT, X, LoBound);
5370 else if (LoOverflow)
5371 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5372 ICmpInst::ICMP_UGE, X, HiBound);
5374 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5375 case ICmpInst::ICMP_ULT:
5376 case ICmpInst::ICMP_SLT:
5377 if (LoOverflow == +1) // Low bound is greater than input range.
5378 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5379 if (LoOverflow == -1) // Low bound is less than input range.
5380 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5381 return new ICmpInst(Pred, X, LoBound);
5382 case ICmpInst::ICMP_UGT:
5383 case ICmpInst::ICMP_SGT:
5384 if (HiOverflow == +1) // High bound greater than input range.
5385 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5386 else if (HiOverflow == -1) // High bound less than input range.
5387 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5388 if (Pred == ICmpInst::ICMP_UGT)
5389 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5391 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5396 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5398 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5401 const APInt &RHSV = RHS->getValue();
5403 switch (LHSI->getOpcode()) {
5404 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5405 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5406 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5408 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5409 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5410 Value *CompareVal = LHSI->getOperand(0);
5412 // If the sign bit of the XorCST is not set, there is no change to
5413 // the operation, just stop using the Xor.
5414 if (!XorCST->getValue().isNegative()) {
5415 ICI.setOperand(0, CompareVal);
5416 AddToWorkList(LHSI);
5420 // Was the old condition true if the operand is positive?
5421 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5423 // If so, the new one isn't.
5424 isTrueIfPositive ^= true;
5426 if (isTrueIfPositive)
5427 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5429 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5433 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5434 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5435 LHSI->getOperand(0)->hasOneUse()) {
5436 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5438 // If the LHS is an AND of a truncating cast, we can widen the
5439 // and/compare to be the input width without changing the value
5440 // produced, eliminating a cast.
5441 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5442 // We can do this transformation if either the AND constant does not
5443 // have its sign bit set or if it is an equality comparison.
5444 // Extending a relational comparison when we're checking the sign
5445 // bit would not work.
5446 if (Cast->hasOneUse() &&
5447 (ICI.isEquality() || AndCST->getValue().isNonNegative() &&
5448 RHSV.isNonNegative())) {
5450 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5451 APInt NewCST = AndCST->getValue();
5452 NewCST.zext(BitWidth);
5454 NewCI.zext(BitWidth);
5455 Instruction *NewAnd =
5456 BinaryOperator::createAnd(Cast->getOperand(0),
5457 ConstantInt::get(NewCST),LHSI->getName());
5458 InsertNewInstBefore(NewAnd, ICI);
5459 return new ICmpInst(ICI.getPredicate(), NewAnd,
5460 ConstantInt::get(NewCI));
5464 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5465 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5466 // happens a LOT in code produced by the C front-end, for bitfield
5468 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5469 if (Shift && !Shift->isShift())
5473 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5474 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5475 const Type *AndTy = AndCST->getType(); // Type of the and.
5477 // We can fold this as long as we can't shift unknown bits
5478 // into the mask. This can only happen with signed shift
5479 // rights, as they sign-extend.
5481 bool CanFold = Shift->isLogicalShift();
5483 // To test for the bad case of the signed shr, see if any
5484 // of the bits shifted in could be tested after the mask.
5485 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5486 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5488 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5489 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5490 AndCST->getValue()) == 0)
5496 if (Shift->getOpcode() == Instruction::Shl)
5497 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5499 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5501 // Check to see if we are shifting out any of the bits being
5503 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5504 // If we shifted bits out, the fold is not going to work out.
5505 // As a special case, check to see if this means that the
5506 // result is always true or false now.
5507 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5508 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5509 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5510 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5512 ICI.setOperand(1, NewCst);
5513 Constant *NewAndCST;
5514 if (Shift->getOpcode() == Instruction::Shl)
5515 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5517 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5518 LHSI->setOperand(1, NewAndCST);
5519 LHSI->setOperand(0, Shift->getOperand(0));
5520 AddToWorkList(Shift); // Shift is dead.
5521 AddUsesToWorkList(ICI);
5527 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5528 // preferable because it allows the C<<Y expression to be hoisted out
5529 // of a loop if Y is invariant and X is not.
5530 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5531 ICI.isEquality() && !Shift->isArithmeticShift() &&
5532 isa<Instruction>(Shift->getOperand(0))) {
5535 if (Shift->getOpcode() == Instruction::LShr) {
5536 NS = BinaryOperator::createShl(AndCST,
5537 Shift->getOperand(1), "tmp");
5539 // Insert a logical shift.
5540 NS = BinaryOperator::createLShr(AndCST,
5541 Shift->getOperand(1), "tmp");
5543 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5545 // Compute X & (C << Y).
5546 Instruction *NewAnd =
5547 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5548 InsertNewInstBefore(NewAnd, ICI);
5550 ICI.setOperand(0, NewAnd);
5556 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5557 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5560 uint32_t TypeBits = RHSV.getBitWidth();
5562 // Check that the shift amount is in range. If not, don't perform
5563 // undefined shifts. When the shift is visited it will be
5565 if (ShAmt->uge(TypeBits))
5568 if (ICI.isEquality()) {
5569 // If we are comparing against bits always shifted out, the
5570 // comparison cannot succeed.
5572 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5573 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5574 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5575 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5576 return ReplaceInstUsesWith(ICI, Cst);
5579 if (LHSI->hasOneUse()) {
5580 // Otherwise strength reduce the shift into an and.
5581 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5583 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5586 BinaryOperator::createAnd(LHSI->getOperand(0),
5587 Mask, LHSI->getName()+".mask");
5588 Value *And = InsertNewInstBefore(AndI, ICI);
5589 return new ICmpInst(ICI.getPredicate(), And,
5590 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5594 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5595 bool TrueIfSigned = false;
5596 if (LHSI->hasOneUse() &&
5597 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5598 // (X << 31) <s 0 --> (X&1) != 0
5599 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5600 (TypeBits-ShAmt->getZExtValue()-1));
5602 BinaryOperator::createAnd(LHSI->getOperand(0),
5603 Mask, LHSI->getName()+".mask");
5604 Value *And = InsertNewInstBefore(AndI, ICI);
5606 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5607 And, Constant::getNullValue(And->getType()));
5612 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5613 case Instruction::AShr: {
5614 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5617 if (ICI.isEquality()) {
5618 // Check that the shift amount is in range. If not, don't perform
5619 // undefined shifts. When the shift is visited it will be
5621 uint32_t TypeBits = RHSV.getBitWidth();
5622 if (ShAmt->uge(TypeBits))
5624 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5626 // If we are comparing against bits always shifted out, the
5627 // comparison cannot succeed.
5628 APInt Comp = RHSV << ShAmtVal;
5629 if (LHSI->getOpcode() == Instruction::LShr)
5630 Comp = Comp.lshr(ShAmtVal);
5632 Comp = Comp.ashr(ShAmtVal);
5634 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5635 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5636 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5637 return ReplaceInstUsesWith(ICI, Cst);
5640 if (LHSI->hasOneUse() || RHSV == 0) {
5641 // Otherwise strength reduce the shift into an and.
5642 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5643 Constant *Mask = ConstantInt::get(Val);
5646 BinaryOperator::createAnd(LHSI->getOperand(0),
5647 Mask, LHSI->getName()+".mask");
5648 Value *And = InsertNewInstBefore(AndI, ICI);
5649 return new ICmpInst(ICI.getPredicate(), And,
5650 ConstantExpr::getShl(RHS, ShAmt));
5656 case Instruction::SDiv:
5657 case Instruction::UDiv:
5658 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5659 // Fold this div into the comparison, producing a range check.
5660 // Determine, based on the divide type, what the range is being
5661 // checked. If there is an overflow on the low or high side, remember
5662 // it, otherwise compute the range [low, hi) bounding the new value.
5663 // See: InsertRangeTest above for the kinds of replacements possible.
5664 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5665 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5670 case Instruction::Add:
5671 // Fold: icmp pred (add, X, C1), C2
5673 if (!ICI.isEquality()) {
5674 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5676 const APInt &LHSV = LHSC->getValue();
5678 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
5681 if (ICI.isSignedPredicate()) {
5682 if (CR.getLower().isSignBit()) {
5683 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
5684 ConstantInt::get(CR.getUpper()));
5685 } else if (CR.getUpper().isSignBit()) {
5686 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
5687 ConstantInt::get(CR.getLower()));
5690 if (CR.getLower().isMinValue()) {
5691 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
5692 ConstantInt::get(CR.getUpper()));
5693 } else if (CR.getUpper().isMinValue()) {
5694 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
5695 ConstantInt::get(CR.getLower()));
5702 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5703 if (ICI.isEquality()) {
5704 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5706 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5707 // the second operand is a constant, simplify a bit.
5708 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5709 switch (BO->getOpcode()) {
5710 case Instruction::SRem:
5711 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5712 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5713 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5714 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5715 Instruction *NewRem =
5716 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5718 InsertNewInstBefore(NewRem, ICI);
5719 return new ICmpInst(ICI.getPredicate(), NewRem,
5720 Constant::getNullValue(BO->getType()));
5724 case Instruction::Add:
5725 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5726 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5727 if (BO->hasOneUse())
5728 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5729 Subtract(RHS, BOp1C));
5730 } else if (RHSV == 0) {
5731 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5732 // efficiently invertible, or if the add has just this one use.
5733 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5735 if (Value *NegVal = dyn_castNegVal(BOp1))
5736 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5737 else if (Value *NegVal = dyn_castNegVal(BOp0))
5738 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5739 else if (BO->hasOneUse()) {
5740 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5741 InsertNewInstBefore(Neg, ICI);
5743 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5747 case Instruction::Xor:
5748 // For the xor case, we can xor two constants together, eliminating
5749 // the explicit xor.
5750 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5751 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5752 ConstantExpr::getXor(RHS, BOC));
5755 case Instruction::Sub:
5756 // Replace (([sub|xor] A, B) != 0) with (A != B)
5758 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5762 case Instruction::Or:
5763 // If bits are being or'd in that are not present in the constant we
5764 // are comparing against, then the comparison could never succeed!
5765 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5766 Constant *NotCI = ConstantExpr::getNot(RHS);
5767 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5768 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5773 case Instruction::And:
5774 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5775 // If bits are being compared against that are and'd out, then the
5776 // comparison can never succeed!
5777 if ((RHSV & ~BOC->getValue()) != 0)
5778 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5781 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5782 if (RHS == BOC && RHSV.isPowerOf2())
5783 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5784 ICmpInst::ICMP_NE, LHSI,
5785 Constant::getNullValue(RHS->getType()));
5787 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5788 if (isSignBit(BOC)) {
5789 Value *X = BO->getOperand(0);
5790 Constant *Zero = Constant::getNullValue(X->getType());
5791 ICmpInst::Predicate pred = isICMP_NE ?
5792 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5793 return new ICmpInst(pred, X, Zero);
5796 // ((X & ~7) == 0) --> X < 8
5797 if (RHSV == 0 && isHighOnes(BOC)) {
5798 Value *X = BO->getOperand(0);
5799 Constant *NegX = ConstantExpr::getNeg(BOC);
5800 ICmpInst::Predicate pred = isICMP_NE ?
5801 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5802 return new ICmpInst(pred, X, NegX);
5807 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5808 // Handle icmp {eq|ne} <intrinsic>, intcst.
5809 if (II->getIntrinsicID() == Intrinsic::bswap) {
5811 ICI.setOperand(0, II->getOperand(1));
5812 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5816 } else { // Not a ICMP_EQ/ICMP_NE
5817 // If the LHS is a cast from an integral value of the same size,
5818 // then since we know the RHS is a constant, try to simlify.
5819 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5820 Value *CastOp = Cast->getOperand(0);
5821 const Type *SrcTy = CastOp->getType();
5822 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5823 if (SrcTy->isInteger() &&
5824 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5825 // If this is an unsigned comparison, try to make the comparison use
5826 // smaller constant values.
5827 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5828 // X u< 128 => X s> -1
5829 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5830 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5831 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5832 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5833 // X u> 127 => X s< 0
5834 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5835 Constant::getNullValue(SrcTy));
5843 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5844 /// We only handle extending casts so far.
5846 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5847 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5848 Value *LHSCIOp = LHSCI->getOperand(0);
5849 const Type *SrcTy = LHSCIOp->getType();
5850 const Type *DestTy = LHSCI->getType();
5853 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5854 // integer type is the same size as the pointer type.
5855 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5856 getTargetData().getPointerSizeInBits() ==
5857 cast<IntegerType>(DestTy)->getBitWidth()) {
5859 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5860 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5861 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5862 RHSOp = RHSC->getOperand(0);
5863 // If the pointer types don't match, insert a bitcast.
5864 if (LHSCIOp->getType() != RHSOp->getType())
5865 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
5869 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5872 // The code below only handles extension cast instructions, so far.
5874 if (LHSCI->getOpcode() != Instruction::ZExt &&
5875 LHSCI->getOpcode() != Instruction::SExt)
5878 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5879 bool isSignedCmp = ICI.isSignedPredicate();
5881 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5882 // Not an extension from the same type?
5883 RHSCIOp = CI->getOperand(0);
5884 if (RHSCIOp->getType() != LHSCIOp->getType())
5887 // If the signedness of the two casts doesn't agree (i.e. one is a sext
5888 // and the other is a zext), then we can't handle this.
5889 if (CI->getOpcode() != LHSCI->getOpcode())
5892 // Deal with equality cases early.
5893 if (ICI.isEquality())
5894 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5896 // A signed comparison of sign extended values simplifies into a
5897 // signed comparison.
5898 if (isSignedCmp && isSignedExt)
5899 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5901 // The other three cases all fold into an unsigned comparison.
5902 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
5905 // If we aren't dealing with a constant on the RHS, exit early
5906 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5910 // Compute the constant that would happen if we truncated to SrcTy then
5911 // reextended to DestTy.
5912 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5913 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5915 // If the re-extended constant didn't change...
5917 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5918 // For example, we might have:
5919 // %A = sext short %X to uint
5920 // %B = icmp ugt uint %A, 1330
5921 // It is incorrect to transform this into
5922 // %B = icmp ugt short %X, 1330
5923 // because %A may have negative value.
5925 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5926 // OR operation is EQ/NE.
5927 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5928 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5933 // The re-extended constant changed so the constant cannot be represented
5934 // in the shorter type. Consequently, we cannot emit a simple comparison.
5936 // First, handle some easy cases. We know the result cannot be equal at this
5937 // point so handle the ICI.isEquality() cases
5938 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5939 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5940 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5941 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5943 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5944 // should have been folded away previously and not enter in here.
5947 // We're performing a signed comparison.
5948 if (cast<ConstantInt>(CI)->getValue().isNegative())
5949 Result = ConstantInt::getFalse(); // X < (small) --> false
5951 Result = ConstantInt::getTrue(); // X < (large) --> true
5953 // We're performing an unsigned comparison.
5955 // We're performing an unsigned comp with a sign extended value.
5956 // This is true if the input is >= 0. [aka >s -1]
5957 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5958 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5959 NegOne, ICI.getName()), ICI);
5961 // Unsigned extend & unsigned compare -> always true.
5962 Result = ConstantInt::getTrue();
5966 // Finally, return the value computed.
5967 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5968 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5969 return ReplaceInstUsesWith(ICI, Result);
5971 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5972 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5973 "ICmp should be folded!");
5974 if (Constant *CI = dyn_cast<Constant>(Result))
5975 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5977 return BinaryOperator::createNot(Result);
5981 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5982 return commonShiftTransforms(I);
5985 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5986 return commonShiftTransforms(I);
5989 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5990 if (Instruction *R = commonShiftTransforms(I))
5993 Value *Op0 = I.getOperand(0);
5995 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5996 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5997 if (CSI->isAllOnesValue())
5998 return ReplaceInstUsesWith(I, CSI);
6000 // See if we can turn a signed shr into an unsigned shr.
6001 if (MaskedValueIsZero(Op0,
6002 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6003 return BinaryOperator::createLShr(Op0, I.getOperand(1));
6008 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6009 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6010 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6012 // shl X, 0 == X and shr X, 0 == X
6013 // shl 0, X == 0 and shr 0, X == 0
6014 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6015 Op0 == Constant::getNullValue(Op0->getType()))
6016 return ReplaceInstUsesWith(I, Op0);
6018 if (isa<UndefValue>(Op0)) {
6019 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6020 return ReplaceInstUsesWith(I, Op0);
6021 else // undef << X -> 0, undef >>u X -> 0
6022 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6024 if (isa<UndefValue>(Op1)) {
6025 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6026 return ReplaceInstUsesWith(I, Op0);
6027 else // X << undef, X >>u undef -> 0
6028 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6031 // Try to fold constant and into select arguments.
6032 if (isa<Constant>(Op0))
6033 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6034 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6037 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6038 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6043 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6044 BinaryOperator &I) {
6045 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6047 // See if we can simplify any instructions used by the instruction whose sole
6048 // purpose is to compute bits we don't care about.
6049 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6050 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6051 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6052 KnownZero, KnownOne))
6055 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6056 // of a signed value.
6058 if (Op1->uge(TypeBits)) {
6059 if (I.getOpcode() != Instruction::AShr)
6060 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6062 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6067 // ((X*C1) << C2) == (X * (C1 << C2))
6068 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6069 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6070 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6071 return BinaryOperator::createMul(BO->getOperand(0),
6072 ConstantExpr::getShl(BOOp, Op1));
6074 // Try to fold constant and into select arguments.
6075 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6076 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6078 if (isa<PHINode>(Op0))
6079 if (Instruction *NV = FoldOpIntoPhi(I))
6082 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6083 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6084 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6085 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6086 // place. Don't try to do this transformation in this case. Also, we
6087 // require that the input operand is a shift-by-constant so that we have
6088 // confidence that the shifts will get folded together. We could do this
6089 // xform in more cases, but it is unlikely to be profitable.
6090 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6091 isa<ConstantInt>(TrOp->getOperand(1))) {
6092 // Okay, we'll do this xform. Make the shift of shift.
6093 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6094 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6096 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6098 // For logical shifts, the truncation has the effect of making the high
6099 // part of the register be zeros. Emulate this by inserting an AND to
6100 // clear the top bits as needed. This 'and' will usually be zapped by
6101 // other xforms later if dead.
6102 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6103 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6104 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6106 // The mask we constructed says what the trunc would do if occurring
6107 // between the shifts. We want to know the effect *after* the second
6108 // shift. We know that it is a logical shift by a constant, so adjust the
6109 // mask as appropriate.
6110 if (I.getOpcode() == Instruction::Shl)
6111 MaskV <<= Op1->getZExtValue();
6113 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6114 MaskV = MaskV.lshr(Op1->getZExtValue());
6117 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6119 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6121 // Return the value truncated to the interesting size.
6122 return new TruncInst(And, I.getType());
6126 if (Op0->hasOneUse()) {
6127 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6128 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6131 switch (Op0BO->getOpcode()) {
6133 case Instruction::Add:
6134 case Instruction::And:
6135 case Instruction::Or:
6136 case Instruction::Xor: {
6137 // These operators commute.
6138 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6139 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6140 match(Op0BO->getOperand(1),
6141 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6142 Instruction *YS = BinaryOperator::createShl(
6143 Op0BO->getOperand(0), Op1,
6145 InsertNewInstBefore(YS, I); // (Y << C)
6147 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6148 Op0BO->getOperand(1)->getName());
6149 InsertNewInstBefore(X, I); // (X + (Y << C))
6150 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6151 return BinaryOperator::createAnd(X, ConstantInt::get(
6152 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6155 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6156 Value *Op0BOOp1 = Op0BO->getOperand(1);
6157 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6159 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6160 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6162 Instruction *YS = BinaryOperator::createShl(
6163 Op0BO->getOperand(0), Op1,
6165 InsertNewInstBefore(YS, I); // (Y << C)
6167 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6168 V1->getName()+".mask");
6169 InsertNewInstBefore(XM, I); // X & (CC << C)
6171 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6176 case Instruction::Sub: {
6177 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6178 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6179 match(Op0BO->getOperand(0),
6180 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6181 Instruction *YS = BinaryOperator::createShl(
6182 Op0BO->getOperand(1), Op1,
6184 InsertNewInstBefore(YS, I); // (Y << C)
6186 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6187 Op0BO->getOperand(0)->getName());
6188 InsertNewInstBefore(X, I); // (X + (Y << C))
6189 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6190 return BinaryOperator::createAnd(X, ConstantInt::get(
6191 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6194 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6195 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6196 match(Op0BO->getOperand(0),
6197 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6198 m_ConstantInt(CC))) && V2 == Op1 &&
6199 cast<BinaryOperator>(Op0BO->getOperand(0))
6200 ->getOperand(0)->hasOneUse()) {
6201 Instruction *YS = BinaryOperator::createShl(
6202 Op0BO->getOperand(1), Op1,
6204 InsertNewInstBefore(YS, I); // (Y << C)
6206 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6207 V1->getName()+".mask");
6208 InsertNewInstBefore(XM, I); // X & (CC << C)
6210 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6218 // If the operand is an bitwise operator with a constant RHS, and the
6219 // shift is the only use, we can pull it out of the shift.
6220 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6221 bool isValid = true; // Valid only for And, Or, Xor
6222 bool highBitSet = false; // Transform if high bit of constant set?
6224 switch (Op0BO->getOpcode()) {
6225 default: isValid = false; break; // Do not perform transform!
6226 case Instruction::Add:
6227 isValid = isLeftShift;
6229 case Instruction::Or:
6230 case Instruction::Xor:
6233 case Instruction::And:
6238 // If this is a signed shift right, and the high bit is modified
6239 // by the logical operation, do not perform the transformation.
6240 // The highBitSet boolean indicates the value of the high bit of
6241 // the constant which would cause it to be modified for this
6244 if (isValid && I.getOpcode() == Instruction::AShr)
6245 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6248 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6250 Instruction *NewShift =
6251 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6252 InsertNewInstBefore(NewShift, I);
6253 NewShift->takeName(Op0BO);
6255 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6262 // Find out if this is a shift of a shift by a constant.
6263 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6264 if (ShiftOp && !ShiftOp->isShift())
6267 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6268 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6269 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6270 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6271 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6272 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6273 Value *X = ShiftOp->getOperand(0);
6275 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6276 if (AmtSum > TypeBits)
6279 const IntegerType *Ty = cast<IntegerType>(I.getType());
6281 // Check for (X << c1) << c2 and (X >> c1) >> c2
6282 if (I.getOpcode() == ShiftOp->getOpcode()) {
6283 return BinaryOperator::create(I.getOpcode(), X,
6284 ConstantInt::get(Ty, AmtSum));
6285 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6286 I.getOpcode() == Instruction::AShr) {
6287 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6288 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6289 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6290 I.getOpcode() == Instruction::LShr) {
6291 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6292 Instruction *Shift =
6293 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6294 InsertNewInstBefore(Shift, I);
6296 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6297 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6300 // Okay, if we get here, one shift must be left, and the other shift must be
6301 // right. See if the amounts are equal.
6302 if (ShiftAmt1 == ShiftAmt2) {
6303 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6304 if (I.getOpcode() == Instruction::Shl) {
6305 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6306 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6308 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6309 if (I.getOpcode() == Instruction::LShr) {
6310 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6311 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6313 // We can simplify ((X << C) >>s C) into a trunc + sext.
6314 // NOTE: we could do this for any C, but that would make 'unusual' integer
6315 // types. For now, just stick to ones well-supported by the code
6317 const Type *SExtType = 0;
6318 switch (Ty->getBitWidth() - ShiftAmt1) {
6325 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6330 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6331 InsertNewInstBefore(NewTrunc, I);
6332 return new SExtInst(NewTrunc, Ty);
6334 // Otherwise, we can't handle it yet.
6335 } else if (ShiftAmt1 < ShiftAmt2) {
6336 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6338 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6339 if (I.getOpcode() == Instruction::Shl) {
6340 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6341 ShiftOp->getOpcode() == Instruction::AShr);
6342 Instruction *Shift =
6343 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6344 InsertNewInstBefore(Shift, I);
6346 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6347 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6350 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6351 if (I.getOpcode() == Instruction::LShr) {
6352 assert(ShiftOp->getOpcode() == Instruction::Shl);
6353 Instruction *Shift =
6354 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6355 InsertNewInstBefore(Shift, I);
6357 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6358 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6361 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6363 assert(ShiftAmt2 < ShiftAmt1);
6364 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6366 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6367 if (I.getOpcode() == Instruction::Shl) {
6368 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6369 ShiftOp->getOpcode() == Instruction::AShr);
6370 Instruction *Shift =
6371 BinaryOperator::create(ShiftOp->getOpcode(), X,
6372 ConstantInt::get(Ty, ShiftDiff));
6373 InsertNewInstBefore(Shift, I);
6375 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6376 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6379 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6380 if (I.getOpcode() == Instruction::LShr) {
6381 assert(ShiftOp->getOpcode() == Instruction::Shl);
6382 Instruction *Shift =
6383 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6384 InsertNewInstBefore(Shift, I);
6386 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6387 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6390 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6397 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6398 /// expression. If so, decompose it, returning some value X, such that Val is
6401 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6403 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6404 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6405 Offset = CI->getZExtValue();
6407 return ConstantInt::get(Type::Int32Ty, 0);
6408 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6409 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6410 if (I->getOpcode() == Instruction::Shl) {
6411 // This is a value scaled by '1 << the shift amt'.
6412 Scale = 1U << RHS->getZExtValue();
6414 return I->getOperand(0);
6415 } else if (I->getOpcode() == Instruction::Mul) {
6416 // This value is scaled by 'RHS'.
6417 Scale = RHS->getZExtValue();
6419 return I->getOperand(0);
6420 } else if (I->getOpcode() == Instruction::Add) {
6421 // We have X+C. Check to see if we really have (X*C2)+C1,
6422 // where C1 is divisible by C2.
6425 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6426 Offset += RHS->getZExtValue();
6433 // Otherwise, we can't look past this.
6440 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6441 /// try to eliminate the cast by moving the type information into the alloc.
6442 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6443 AllocationInst &AI) {
6444 const PointerType *PTy = cast<PointerType>(CI.getType());
6446 // Remove any uses of AI that are dead.
6447 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6449 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6450 Instruction *User = cast<Instruction>(*UI++);
6451 if (isInstructionTriviallyDead(User)) {
6452 while (UI != E && *UI == User)
6453 ++UI; // If this instruction uses AI more than once, don't break UI.
6456 DOUT << "IC: DCE: " << *User;
6457 EraseInstFromFunction(*User);
6461 // Get the type really allocated and the type casted to.
6462 const Type *AllocElTy = AI.getAllocatedType();
6463 const Type *CastElTy = PTy->getElementType();
6464 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6466 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6467 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6468 if (CastElTyAlign < AllocElTyAlign) return 0;
6470 // If the allocation has multiple uses, only promote it if we are strictly
6471 // increasing the alignment of the resultant allocation. If we keep it the
6472 // same, we open the door to infinite loops of various kinds.
6473 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6475 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6476 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6477 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6479 // See if we can satisfy the modulus by pulling a scale out of the array
6481 unsigned ArraySizeScale;
6483 Value *NumElements = // See if the array size is a decomposable linear expr.
6484 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6486 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6488 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6489 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6491 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6496 // If the allocation size is constant, form a constant mul expression
6497 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6498 if (isa<ConstantInt>(NumElements))
6499 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6500 // otherwise multiply the amount and the number of elements
6501 else if (Scale != 1) {
6502 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6503 Amt = InsertNewInstBefore(Tmp, AI);
6507 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6508 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6509 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6510 Amt = InsertNewInstBefore(Tmp, AI);
6513 AllocationInst *New;
6514 if (isa<MallocInst>(AI))
6515 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6517 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6518 InsertNewInstBefore(New, AI);
6521 // If the allocation has multiple uses, insert a cast and change all things
6522 // that used it to use the new cast. This will also hack on CI, but it will
6524 if (!AI.hasOneUse()) {
6525 AddUsesToWorkList(AI);
6526 // New is the allocation instruction, pointer typed. AI is the original
6527 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6528 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6529 InsertNewInstBefore(NewCast, AI);
6530 AI.replaceAllUsesWith(NewCast);
6532 return ReplaceInstUsesWith(CI, New);
6535 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6536 /// and return it as type Ty without inserting any new casts and without
6537 /// changing the computed value. This is used by code that tries to decide
6538 /// whether promoting or shrinking integer operations to wider or smaller types
6539 /// will allow us to eliminate a truncate or extend.
6541 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6542 /// extension operation if Ty is larger.
6543 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6544 unsigned CastOpc, int &NumCastsRemoved) {
6545 // We can always evaluate constants in another type.
6546 if (isa<ConstantInt>(V))
6549 Instruction *I = dyn_cast<Instruction>(V);
6550 if (!I) return false;
6552 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6554 // If this is an extension or truncate, we can often eliminate it.
6555 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6556 // If this is a cast from the destination type, we can trivially eliminate
6557 // it, and this will remove a cast overall.
6558 if (I->getOperand(0)->getType() == Ty) {
6559 // If the first operand is itself a cast, and is eliminable, do not count
6560 // this as an eliminable cast. We would prefer to eliminate those two
6562 if (!isa<CastInst>(I->getOperand(0)))
6568 // We can't extend or shrink something that has multiple uses: doing so would
6569 // require duplicating the instruction in general, which isn't profitable.
6570 if (!I->hasOneUse()) return false;
6572 switch (I->getOpcode()) {
6573 case Instruction::Add:
6574 case Instruction::Sub:
6575 case Instruction::And:
6576 case Instruction::Or:
6577 case Instruction::Xor:
6578 // These operators can all arbitrarily be extended or truncated.
6579 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6581 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6584 case Instruction::Mul:
6585 // A multiply can be truncated by truncating its operands.
6586 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6587 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6589 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6592 case Instruction::Shl:
6593 // If we are truncating the result of this SHL, and if it's a shift of a
6594 // constant amount, we can always perform a SHL in a smaller type.
6595 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6596 uint32_t BitWidth = Ty->getBitWidth();
6597 if (BitWidth < OrigTy->getBitWidth() &&
6598 CI->getLimitedValue(BitWidth) < BitWidth)
6599 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6603 case Instruction::LShr:
6604 // If this is a truncate of a logical shr, we can truncate it to a smaller
6605 // lshr iff we know that the bits we would otherwise be shifting in are
6607 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6608 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6609 uint32_t BitWidth = Ty->getBitWidth();
6610 if (BitWidth < OrigBitWidth &&
6611 MaskedValueIsZero(I->getOperand(0),
6612 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6613 CI->getLimitedValue(BitWidth) < BitWidth) {
6614 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6619 case Instruction::ZExt:
6620 case Instruction::SExt:
6621 case Instruction::Trunc:
6622 // If this is the same kind of case as our original (e.g. zext+zext), we
6623 // can safely replace it. Note that replacing it does not reduce the number
6624 // of casts in the input.
6625 if (I->getOpcode() == CastOpc)
6630 // TODO: Can handle more cases here.
6637 /// EvaluateInDifferentType - Given an expression that
6638 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6639 /// evaluate the expression.
6640 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6642 if (Constant *C = dyn_cast<Constant>(V))
6643 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6645 // Otherwise, it must be an instruction.
6646 Instruction *I = cast<Instruction>(V);
6647 Instruction *Res = 0;
6648 switch (I->getOpcode()) {
6649 case Instruction::Add:
6650 case Instruction::Sub:
6651 case Instruction::Mul:
6652 case Instruction::And:
6653 case Instruction::Or:
6654 case Instruction::Xor:
6655 case Instruction::AShr:
6656 case Instruction::LShr:
6657 case Instruction::Shl: {
6658 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6659 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6660 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6661 LHS, RHS, I->getName());
6664 case Instruction::Trunc:
6665 case Instruction::ZExt:
6666 case Instruction::SExt:
6667 // If the source type of the cast is the type we're trying for then we can
6668 // just return the source. There's no need to insert it because it is not
6670 if (I->getOperand(0)->getType() == Ty)
6671 return I->getOperand(0);
6673 // Otherwise, must be the same type of case, so just reinsert a new one.
6674 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6678 // TODO: Can handle more cases here.
6679 assert(0 && "Unreachable!");
6683 return InsertNewInstBefore(Res, *I);
6686 /// @brief Implement the transforms common to all CastInst visitors.
6687 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6688 Value *Src = CI.getOperand(0);
6690 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6691 // eliminate it now.
6692 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6693 if (Instruction::CastOps opc =
6694 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6695 // The first cast (CSrc) is eliminable so we need to fix up or replace
6696 // the second cast (CI). CSrc will then have a good chance of being dead.
6697 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6701 // If we are casting a select then fold the cast into the select
6702 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6703 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6706 // If we are casting a PHI then fold the cast into the PHI
6707 if (isa<PHINode>(Src))
6708 if (Instruction *NV = FoldOpIntoPhi(CI))
6714 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6715 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6716 Value *Src = CI.getOperand(0);
6718 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6719 // If casting the result of a getelementptr instruction with no offset, turn
6720 // this into a cast of the original pointer!
6721 if (GEP->hasAllZeroIndices()) {
6722 // Changing the cast operand is usually not a good idea but it is safe
6723 // here because the pointer operand is being replaced with another
6724 // pointer operand so the opcode doesn't need to change.
6726 CI.setOperand(0, GEP->getOperand(0));
6730 // If the GEP has a single use, and the base pointer is a bitcast, and the
6731 // GEP computes a constant offset, see if we can convert these three
6732 // instructions into fewer. This typically happens with unions and other
6733 // non-type-safe code.
6734 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6735 if (GEP->hasAllConstantIndices()) {
6736 // We are guaranteed to get a constant from EmitGEPOffset.
6737 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6738 int64_t Offset = OffsetV->getSExtValue();
6740 // Get the base pointer input of the bitcast, and the type it points to.
6741 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6742 const Type *GEPIdxTy =
6743 cast<PointerType>(OrigBase->getType())->getElementType();
6744 if (GEPIdxTy->isSized()) {
6745 SmallVector<Value*, 8> NewIndices;
6747 // Start with the index over the outer type. Note that the type size
6748 // might be zero (even if the offset isn't zero) if the indexed type
6749 // is something like [0 x {int, int}]
6750 const Type *IntPtrTy = TD->getIntPtrType();
6751 int64_t FirstIdx = 0;
6752 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6753 FirstIdx = Offset/TySize;
6756 // Handle silly modulus not returning values values [0..TySize).
6760 assert(Offset >= 0);
6762 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6765 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6767 // Index into the types. If we fail, set OrigBase to null.
6769 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6770 const StructLayout *SL = TD->getStructLayout(STy);
6771 if (Offset < (int64_t)SL->getSizeInBytes()) {
6772 unsigned Elt = SL->getElementContainingOffset(Offset);
6773 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6775 Offset -= SL->getElementOffset(Elt);
6776 GEPIdxTy = STy->getElementType(Elt);
6778 // Otherwise, we can't index into this, bail out.
6782 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6783 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6784 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6785 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6788 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6790 GEPIdxTy = STy->getElementType();
6792 // Otherwise, we can't index into this, bail out.
6798 // If we were able to index down into an element, create the GEP
6799 // and bitcast the result. This eliminates one bitcast, potentially
6801 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6803 NewIndices.end(), "");
6804 InsertNewInstBefore(NGEP, CI);
6805 NGEP->takeName(GEP);
6807 if (isa<BitCastInst>(CI))
6808 return new BitCastInst(NGEP, CI.getType());
6809 assert(isa<PtrToIntInst>(CI));
6810 return new PtrToIntInst(NGEP, CI.getType());
6817 return commonCastTransforms(CI);
6822 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6823 /// integer types. This function implements the common transforms for all those
6825 /// @brief Implement the transforms common to CastInst with integer operands
6826 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6827 if (Instruction *Result = commonCastTransforms(CI))
6830 Value *Src = CI.getOperand(0);
6831 const Type *SrcTy = Src->getType();
6832 const Type *DestTy = CI.getType();
6833 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6834 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6836 // See if we can simplify any instructions used by the LHS whose sole
6837 // purpose is to compute bits we don't care about.
6838 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6839 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6840 KnownZero, KnownOne))
6843 // If the source isn't an instruction or has more than one use then we
6844 // can't do anything more.
6845 Instruction *SrcI = dyn_cast<Instruction>(Src);
6846 if (!SrcI || !Src->hasOneUse())
6849 // Attempt to propagate the cast into the instruction for int->int casts.
6850 int NumCastsRemoved = 0;
6851 if (!isa<BitCastInst>(CI) &&
6852 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6853 CI.getOpcode(), NumCastsRemoved)) {
6854 // If this cast is a truncate, evaluting in a different type always
6855 // eliminates the cast, so it is always a win. If this is a zero-extension,
6856 // we need to do an AND to maintain the clear top-part of the computation,
6857 // so we require that the input have eliminated at least one cast. If this
6858 // is a sign extension, we insert two new casts (to do the extension) so we
6859 // require that two casts have been eliminated.
6861 switch (CI.getOpcode()) {
6863 // All the others use floating point so we shouldn't actually
6864 // get here because of the check above.
6865 assert(0 && "Unknown cast type");
6866 case Instruction::Trunc:
6869 case Instruction::ZExt:
6870 DoXForm = NumCastsRemoved >= 1;
6872 case Instruction::SExt:
6873 DoXForm = NumCastsRemoved >= 2;
6878 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6879 CI.getOpcode() == Instruction::SExt);
6880 assert(Res->getType() == DestTy);
6881 switch (CI.getOpcode()) {
6882 default: assert(0 && "Unknown cast type!");
6883 case Instruction::Trunc:
6884 case Instruction::BitCast:
6885 // Just replace this cast with the result.
6886 return ReplaceInstUsesWith(CI, Res);
6887 case Instruction::ZExt: {
6888 // We need to emit an AND to clear the high bits.
6889 assert(SrcBitSize < DestBitSize && "Not a zext?");
6890 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6892 return BinaryOperator::createAnd(Res, C);
6894 case Instruction::SExt:
6895 // We need to emit a cast to truncate, then a cast to sext.
6896 return CastInst::create(Instruction::SExt,
6897 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6903 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6904 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6906 switch (SrcI->getOpcode()) {
6907 case Instruction::Add:
6908 case Instruction::Mul:
6909 case Instruction::And:
6910 case Instruction::Or:
6911 case Instruction::Xor:
6912 // If we are discarding information, rewrite.
6913 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6914 // Don't insert two casts if they cannot be eliminated. We allow
6915 // two casts to be inserted if the sizes are the same. This could
6916 // only be converting signedness, which is a noop.
6917 if (DestBitSize == SrcBitSize ||
6918 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6919 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6920 Instruction::CastOps opcode = CI.getOpcode();
6921 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6922 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6923 return BinaryOperator::create(
6924 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6928 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6929 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6930 SrcI->getOpcode() == Instruction::Xor &&
6931 Op1 == ConstantInt::getTrue() &&
6932 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6933 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6934 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6937 case Instruction::SDiv:
6938 case Instruction::UDiv:
6939 case Instruction::SRem:
6940 case Instruction::URem:
6941 // If we are just changing the sign, rewrite.
6942 if (DestBitSize == SrcBitSize) {
6943 // Don't insert two casts if they cannot be eliminated. We allow
6944 // two casts to be inserted if the sizes are the same. This could
6945 // only be converting signedness, which is a noop.
6946 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6947 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6948 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6950 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6952 return BinaryOperator::create(
6953 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6958 case Instruction::Shl:
6959 // Allow changing the sign of the source operand. Do not allow
6960 // changing the size of the shift, UNLESS the shift amount is a
6961 // constant. We must not change variable sized shifts to a smaller
6962 // size, because it is undefined to shift more bits out than exist
6964 if (DestBitSize == SrcBitSize ||
6965 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6966 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6967 Instruction::BitCast : Instruction::Trunc);
6968 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6969 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6970 return BinaryOperator::createShl(Op0c, Op1c);
6973 case Instruction::AShr:
6974 // If this is a signed shr, and if all bits shifted in are about to be
6975 // truncated off, turn it into an unsigned shr to allow greater
6977 if (DestBitSize < SrcBitSize &&
6978 isa<ConstantInt>(Op1)) {
6979 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6980 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6981 // Insert the new logical shift right.
6982 return BinaryOperator::createLShr(Op0, Op1);
6990 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6991 if (Instruction *Result = commonIntCastTransforms(CI))
6994 Value *Src = CI.getOperand(0);
6995 const Type *Ty = CI.getType();
6996 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6997 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6999 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7000 switch (SrcI->getOpcode()) {
7002 case Instruction::LShr:
7003 // We can shrink lshr to something smaller if we know the bits shifted in
7004 // are already zeros.
7005 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7006 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7008 // Get a mask for the bits shifting in.
7009 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7010 Value* SrcIOp0 = SrcI->getOperand(0);
7011 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7012 if (ShAmt >= DestBitWidth) // All zeros.
7013 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7015 // Okay, we can shrink this. Truncate the input, then return a new
7017 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7018 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7020 return BinaryOperator::createLShr(V1, V2);
7022 } else { // This is a variable shr.
7024 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7025 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7026 // loop-invariant and CSE'd.
7027 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7028 Value *One = ConstantInt::get(SrcI->getType(), 1);
7030 Value *V = InsertNewInstBefore(
7031 BinaryOperator::createShl(One, SrcI->getOperand(1),
7033 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7034 SrcI->getOperand(0),
7036 Value *Zero = Constant::getNullValue(V->getType());
7037 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7047 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7048 // If one of the common conversion will work ..
7049 if (Instruction *Result = commonIntCastTransforms(CI))
7052 Value *Src = CI.getOperand(0);
7054 // If this is a cast of a cast
7055 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7056 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7057 // types and if the sizes are just right we can convert this into a logical
7058 // 'and' which will be much cheaper than the pair of casts.
7059 if (isa<TruncInst>(CSrc)) {
7060 // Get the sizes of the types involved
7061 Value *A = CSrc->getOperand(0);
7062 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7063 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7064 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7065 // If we're actually extending zero bits and the trunc is a no-op
7066 if (MidSize < DstSize && SrcSize == DstSize) {
7067 // Replace both of the casts with an And of the type mask.
7068 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7069 Constant *AndConst = ConstantInt::get(AndValue);
7071 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7072 // Unfortunately, if the type changed, we need to cast it back.
7073 if (And->getType() != CI.getType()) {
7074 And->setName(CSrc->getName()+".mask");
7075 InsertNewInstBefore(And, CI);
7076 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7083 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7084 // If we are just checking for a icmp eq of a single bit and zext'ing it
7085 // to an integer, then shift the bit to the appropriate place and then
7086 // cast to integer to avoid the comparison.
7087 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7088 const APInt &Op1CV = Op1C->getValue();
7090 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7091 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7092 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7093 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7094 Value *In = ICI->getOperand(0);
7095 Value *Sh = ConstantInt::get(In->getType(),
7096 In->getType()->getPrimitiveSizeInBits()-1);
7097 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7098 In->getName()+".lobit"),
7100 if (In->getType() != CI.getType())
7101 In = CastInst::createIntegerCast(In, CI.getType(),
7102 false/*ZExt*/, "tmp", &CI);
7104 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7105 Constant *One = ConstantInt::get(In->getType(), 1);
7106 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7107 In->getName()+".not"),
7111 return ReplaceInstUsesWith(CI, In);
7116 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7117 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7118 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7119 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7120 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7121 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7122 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7123 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7124 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7125 // This only works for EQ and NE
7126 ICI->isEquality()) {
7127 // If Op1C some other power of two, convert:
7128 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7129 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7130 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7131 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7133 APInt KnownZeroMask(~KnownZero);
7134 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7135 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7136 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7137 // (X&4) == 2 --> false
7138 // (X&4) != 2 --> true
7139 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7140 Res = ConstantExpr::getZExt(Res, CI.getType());
7141 return ReplaceInstUsesWith(CI, Res);
7144 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7145 Value *In = ICI->getOperand(0);
7147 // Perform a logical shr by shiftamt.
7148 // Insert the shift to put the result in the low bit.
7149 In = InsertNewInstBefore(
7150 BinaryOperator::createLShr(In,
7151 ConstantInt::get(In->getType(), ShiftAmt),
7152 In->getName()+".lobit"), CI);
7155 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7156 Constant *One = ConstantInt::get(In->getType(), 1);
7157 In = BinaryOperator::createXor(In, One, "tmp");
7158 InsertNewInstBefore(cast<Instruction>(In), CI);
7161 if (CI.getType() == In->getType())
7162 return ReplaceInstUsesWith(CI, In);
7164 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7172 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7173 if (Instruction *I = commonIntCastTransforms(CI))
7176 Value *Src = CI.getOperand(0);
7178 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7179 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7180 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7181 // If we are just checking for a icmp eq of a single bit and zext'ing it
7182 // to an integer, then shift the bit to the appropriate place and then
7183 // cast to integer to avoid the comparison.
7184 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7185 const APInt &Op1CV = Op1C->getValue();
7187 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7188 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7189 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7190 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7191 Value *In = ICI->getOperand(0);
7192 Value *Sh = ConstantInt::get(In->getType(),
7193 In->getType()->getPrimitiveSizeInBits()-1);
7194 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7195 In->getName()+".lobit"),
7197 if (In->getType() != CI.getType())
7198 In = CastInst::createIntegerCast(In, CI.getType(),
7199 true/*SExt*/, "tmp", &CI);
7201 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7202 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7203 In->getName()+".not"), CI);
7205 return ReplaceInstUsesWith(CI, In);
7213 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7214 /// in the specified FP type without changing its value.
7215 static Constant *FitsInFPType(ConstantFP *CFP, const Type *FPTy,
7216 const fltSemantics &Sem) {
7217 APFloat F = CFP->getValueAPF();
7218 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7219 return ConstantFP::get(FPTy, F);
7223 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7224 /// through it until we get the source value.
7225 static Value *LookThroughFPExtensions(Value *V) {
7226 if (Instruction *I = dyn_cast<Instruction>(V))
7227 if (I->getOpcode() == Instruction::FPExt)
7228 return LookThroughFPExtensions(I->getOperand(0));
7230 // If this value is a constant, return the constant in the smallest FP type
7231 // that can accurately represent it. This allows us to turn
7232 // (float)((double)X+2.0) into x+2.0f.
7233 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7234 if (CFP->getType() == Type::PPC_FP128Ty)
7235 return V; // No constant folding of this.
7236 // See if the value can be truncated to float and then reextended.
7237 if (Value *V = FitsInFPType(CFP, Type::FloatTy, APFloat::IEEEsingle))
7239 if (CFP->getType() == Type::DoubleTy)
7240 return V; // Won't shrink.
7241 if (Value *V = FitsInFPType(CFP, Type::DoubleTy, APFloat::IEEEdouble))
7243 // Don't try to shrink to various long double types.
7249 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7250 if (Instruction *I = commonCastTransforms(CI))
7253 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7254 // smaller than the destination type, we can eliminate the truncate by doing
7255 // the add as the smaller type. This applies to add/sub/mul/div as well as
7256 // many builtins (sqrt, etc).
7257 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7258 if (OpI && OpI->hasOneUse()) {
7259 switch (OpI->getOpcode()) {
7261 case Instruction::Add:
7262 case Instruction::Sub:
7263 case Instruction::Mul:
7264 case Instruction::FDiv:
7265 case Instruction::FRem:
7266 const Type *SrcTy = OpI->getType();
7267 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7268 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7269 if (LHSTrunc->getType() != SrcTy &&
7270 RHSTrunc->getType() != SrcTy) {
7271 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7272 // If the source types were both smaller than the destination type of
7273 // the cast, do this xform.
7274 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7275 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7276 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7278 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7280 return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7289 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7290 return commonCastTransforms(CI);
7293 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7294 return commonCastTransforms(CI);
7297 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7298 return commonCastTransforms(CI);
7301 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7302 return commonCastTransforms(CI);
7305 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7306 return commonCastTransforms(CI);
7309 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7310 return commonPointerCastTransforms(CI);
7313 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7314 if (Instruction *I = commonCastTransforms(CI))
7317 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7318 if (!DestPointee->isSized()) return 0;
7320 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7323 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7324 m_ConstantInt(Cst)))) {
7325 // If the source and destination operands have the same type, see if this
7326 // is a single-index GEP.
7327 if (X->getType() == CI.getType()) {
7328 // Get the size of the pointee type.
7329 uint64_t Size = TD->getABITypeSizeInBits(DestPointee);
7331 // Convert the constant to intptr type.
7332 APInt Offset = Cst->getValue();
7333 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7335 // If Offset is evenly divisible by Size, we can do this xform.
7336 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7337 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7338 return new GetElementPtrInst(X, ConstantInt::get(Offset));
7341 // TODO: Could handle other cases, e.g. where add is indexing into field of
7343 } else if (CI.getOperand(0)->hasOneUse() &&
7344 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7345 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7346 // "inttoptr+GEP" instead of "add+intptr".
7348 // Get the size of the pointee type.
7349 uint64_t Size = TD->getABITypeSize(DestPointee);
7351 // Convert the constant to intptr type.
7352 APInt Offset = Cst->getValue();
7353 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7355 // If Offset is evenly divisible by Size, we can do this xform.
7356 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7357 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7359 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7361 return new GetElementPtrInst(P, ConstantInt::get(Offset), "tmp");
7367 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7368 // If the operands are integer typed then apply the integer transforms,
7369 // otherwise just apply the common ones.
7370 Value *Src = CI.getOperand(0);
7371 const Type *SrcTy = Src->getType();
7372 const Type *DestTy = CI.getType();
7374 if (SrcTy->isInteger() && DestTy->isInteger()) {
7375 if (Instruction *Result = commonIntCastTransforms(CI))
7377 } else if (isa<PointerType>(SrcTy)) {
7378 if (Instruction *I = commonPointerCastTransforms(CI))
7381 if (Instruction *Result = commonCastTransforms(CI))
7386 // Get rid of casts from one type to the same type. These are useless and can
7387 // be replaced by the operand.
7388 if (DestTy == Src->getType())
7389 return ReplaceInstUsesWith(CI, Src);
7391 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7392 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7393 const Type *DstElTy = DstPTy->getElementType();
7394 const Type *SrcElTy = SrcPTy->getElementType();
7396 // If we are casting a malloc or alloca to a pointer to a type of the same
7397 // size, rewrite the allocation instruction to allocate the "right" type.
7398 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7399 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7402 // If the source and destination are pointers, and this cast is equivalent
7403 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7404 // This can enhance SROA and other transforms that want type-safe pointers.
7405 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7406 unsigned NumZeros = 0;
7407 while (SrcElTy != DstElTy &&
7408 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7409 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7410 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7414 // If we found a path from the src to dest, create the getelementptr now.
7415 if (SrcElTy == DstElTy) {
7416 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7417 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7418 ((Instruction*) NULL));
7422 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7423 if (SVI->hasOneUse()) {
7424 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7425 // a bitconvert to a vector with the same # elts.
7426 if (isa<VectorType>(DestTy) &&
7427 cast<VectorType>(DestTy)->getNumElements() ==
7428 SVI->getType()->getNumElements()) {
7430 // If either of the operands is a cast from CI.getType(), then
7431 // evaluating the shuffle in the casted destination's type will allow
7432 // us to eliminate at least one cast.
7433 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7434 Tmp->getOperand(0)->getType() == DestTy) ||
7435 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7436 Tmp->getOperand(0)->getType() == DestTy)) {
7437 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7438 SVI->getOperand(0), DestTy, &CI);
7439 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7440 SVI->getOperand(1), DestTy, &CI);
7441 // Return a new shuffle vector. Use the same element ID's, as we
7442 // know the vector types match #elts.
7443 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7451 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7453 /// %D = select %cond, %C, %A
7455 /// %C = select %cond, %B, 0
7458 /// Assuming that the specified instruction is an operand to the select, return
7459 /// a bitmask indicating which operands of this instruction are foldable if they
7460 /// equal the other incoming value of the select.
7462 static unsigned GetSelectFoldableOperands(Instruction *I) {
7463 switch (I->getOpcode()) {
7464 case Instruction::Add:
7465 case Instruction::Mul:
7466 case Instruction::And:
7467 case Instruction::Or:
7468 case Instruction::Xor:
7469 return 3; // Can fold through either operand.
7470 case Instruction::Sub: // Can only fold on the amount subtracted.
7471 case Instruction::Shl: // Can only fold on the shift amount.
7472 case Instruction::LShr:
7473 case Instruction::AShr:
7476 return 0; // Cannot fold
7480 /// GetSelectFoldableConstant - For the same transformation as the previous
7481 /// function, return the identity constant that goes into the select.
7482 static Constant *GetSelectFoldableConstant(Instruction *I) {
7483 switch (I->getOpcode()) {
7484 default: assert(0 && "This cannot happen!"); abort();
7485 case Instruction::Add:
7486 case Instruction::Sub:
7487 case Instruction::Or:
7488 case Instruction::Xor:
7489 case Instruction::Shl:
7490 case Instruction::LShr:
7491 case Instruction::AShr:
7492 return Constant::getNullValue(I->getType());
7493 case Instruction::And:
7494 return Constant::getAllOnesValue(I->getType());
7495 case Instruction::Mul:
7496 return ConstantInt::get(I->getType(), 1);
7500 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7501 /// have the same opcode and only one use each. Try to simplify this.
7502 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7504 if (TI->getNumOperands() == 1) {
7505 // If this is a non-volatile load or a cast from the same type,
7508 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7511 return 0; // unknown unary op.
7514 // Fold this by inserting a select from the input values.
7515 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7516 FI->getOperand(0), SI.getName()+".v");
7517 InsertNewInstBefore(NewSI, SI);
7518 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7522 // Only handle binary operators here.
7523 if (!isa<BinaryOperator>(TI))
7526 // Figure out if the operations have any operands in common.
7527 Value *MatchOp, *OtherOpT, *OtherOpF;
7529 if (TI->getOperand(0) == FI->getOperand(0)) {
7530 MatchOp = TI->getOperand(0);
7531 OtherOpT = TI->getOperand(1);
7532 OtherOpF = FI->getOperand(1);
7533 MatchIsOpZero = true;
7534 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7535 MatchOp = TI->getOperand(1);
7536 OtherOpT = TI->getOperand(0);
7537 OtherOpF = FI->getOperand(0);
7538 MatchIsOpZero = false;
7539 } else if (!TI->isCommutative()) {
7541 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7542 MatchOp = TI->getOperand(0);
7543 OtherOpT = TI->getOperand(1);
7544 OtherOpF = FI->getOperand(0);
7545 MatchIsOpZero = true;
7546 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7547 MatchOp = TI->getOperand(1);
7548 OtherOpT = TI->getOperand(0);
7549 OtherOpF = FI->getOperand(1);
7550 MatchIsOpZero = true;
7555 // If we reach here, they do have operations in common.
7556 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7557 OtherOpF, SI.getName()+".v");
7558 InsertNewInstBefore(NewSI, SI);
7560 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7562 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7564 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7566 assert(0 && "Shouldn't get here");
7570 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7571 Value *CondVal = SI.getCondition();
7572 Value *TrueVal = SI.getTrueValue();
7573 Value *FalseVal = SI.getFalseValue();
7575 // select true, X, Y -> X
7576 // select false, X, Y -> Y
7577 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7578 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7580 // select C, X, X -> X
7581 if (TrueVal == FalseVal)
7582 return ReplaceInstUsesWith(SI, TrueVal);
7584 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7585 return ReplaceInstUsesWith(SI, FalseVal);
7586 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7587 return ReplaceInstUsesWith(SI, TrueVal);
7588 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7589 if (isa<Constant>(TrueVal))
7590 return ReplaceInstUsesWith(SI, TrueVal);
7592 return ReplaceInstUsesWith(SI, FalseVal);
7595 if (SI.getType() == Type::Int1Ty) {
7596 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7597 if (C->getZExtValue()) {
7598 // Change: A = select B, true, C --> A = or B, C
7599 return BinaryOperator::createOr(CondVal, FalseVal);
7601 // Change: A = select B, false, C --> A = and !B, C
7603 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7604 "not."+CondVal->getName()), SI);
7605 return BinaryOperator::createAnd(NotCond, FalseVal);
7607 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7608 if (C->getZExtValue() == false) {
7609 // Change: A = select B, C, false --> A = and B, C
7610 return BinaryOperator::createAnd(CondVal, TrueVal);
7612 // Change: A = select B, C, true --> A = or !B, C
7614 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7615 "not."+CondVal->getName()), SI);
7616 return BinaryOperator::createOr(NotCond, TrueVal);
7620 // select a, b, a -> a&b
7621 // select a, a, b -> a|b
7622 if (CondVal == TrueVal)
7623 return BinaryOperator::createOr(CondVal, FalseVal);
7624 else if (CondVal == FalseVal)
7625 return BinaryOperator::createAnd(CondVal, TrueVal);
7628 // Selecting between two integer constants?
7629 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7630 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7631 // select C, 1, 0 -> zext C to int
7632 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7633 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7634 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7635 // select C, 0, 1 -> zext !C to int
7637 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7638 "not."+CondVal->getName()), SI);
7639 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7642 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7644 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7646 // (x <s 0) ? -1 : 0 -> ashr x, 31
7647 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7648 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7649 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7650 // The comparison constant and the result are not neccessarily the
7651 // same width. Make an all-ones value by inserting a AShr.
7652 Value *X = IC->getOperand(0);
7653 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7654 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7655 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7657 InsertNewInstBefore(SRA, SI);
7659 // Finally, convert to the type of the select RHS. We figure out
7660 // if this requires a SExt, Trunc or BitCast based on the sizes.
7661 Instruction::CastOps opc = Instruction::BitCast;
7662 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7663 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7664 if (SRASize < SISize)
7665 opc = Instruction::SExt;
7666 else if (SRASize > SISize)
7667 opc = Instruction::Trunc;
7668 return CastInst::create(opc, SRA, SI.getType());
7673 // If one of the constants is zero (we know they can't both be) and we
7674 // have an icmp instruction with zero, and we have an 'and' with the
7675 // non-constant value, eliminate this whole mess. This corresponds to
7676 // cases like this: ((X & 27) ? 27 : 0)
7677 if (TrueValC->isZero() || FalseValC->isZero())
7678 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7679 cast<Constant>(IC->getOperand(1))->isNullValue())
7680 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7681 if (ICA->getOpcode() == Instruction::And &&
7682 isa<ConstantInt>(ICA->getOperand(1)) &&
7683 (ICA->getOperand(1) == TrueValC ||
7684 ICA->getOperand(1) == FalseValC) &&
7685 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7686 // Okay, now we know that everything is set up, we just don't
7687 // know whether we have a icmp_ne or icmp_eq and whether the
7688 // true or false val is the zero.
7689 bool ShouldNotVal = !TrueValC->isZero();
7690 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7693 V = InsertNewInstBefore(BinaryOperator::create(
7694 Instruction::Xor, V, ICA->getOperand(1)), SI);
7695 return ReplaceInstUsesWith(SI, V);
7700 // See if we are selecting two values based on a comparison of the two values.
7701 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7702 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7703 // Transform (X == Y) ? X : Y -> Y
7704 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7705 // This is not safe in general for floating point:
7706 // consider X== -0, Y== +0.
7707 // It becomes safe if either operand is a nonzero constant.
7708 ConstantFP *CFPt, *CFPf;
7709 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7710 !CFPt->getValueAPF().isZero()) ||
7711 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7712 !CFPf->getValueAPF().isZero()))
7713 return ReplaceInstUsesWith(SI, FalseVal);
7715 // Transform (X != Y) ? X : Y -> X
7716 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7717 return ReplaceInstUsesWith(SI, TrueVal);
7718 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7720 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7721 // Transform (X == Y) ? Y : X -> X
7722 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7723 // This is not safe in general for floating point:
7724 // consider X== -0, Y== +0.
7725 // It becomes safe if either operand is a nonzero constant.
7726 ConstantFP *CFPt, *CFPf;
7727 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7728 !CFPt->getValueAPF().isZero()) ||
7729 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7730 !CFPf->getValueAPF().isZero()))
7731 return ReplaceInstUsesWith(SI, FalseVal);
7733 // Transform (X != Y) ? Y : X -> Y
7734 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7735 return ReplaceInstUsesWith(SI, TrueVal);
7736 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7740 // See if we are selecting two values based on a comparison of the two values.
7741 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7742 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7743 // Transform (X == Y) ? X : Y -> Y
7744 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7745 return ReplaceInstUsesWith(SI, FalseVal);
7746 // Transform (X != Y) ? X : Y -> X
7747 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7748 return ReplaceInstUsesWith(SI, TrueVal);
7749 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7751 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7752 // Transform (X == Y) ? Y : X -> X
7753 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7754 return ReplaceInstUsesWith(SI, FalseVal);
7755 // Transform (X != Y) ? Y : X -> Y
7756 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7757 return ReplaceInstUsesWith(SI, TrueVal);
7758 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7762 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7763 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7764 if (TI->hasOneUse() && FI->hasOneUse()) {
7765 Instruction *AddOp = 0, *SubOp = 0;
7767 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7768 if (TI->getOpcode() == FI->getOpcode())
7769 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7772 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7773 // even legal for FP.
7774 if (TI->getOpcode() == Instruction::Sub &&
7775 FI->getOpcode() == Instruction::Add) {
7776 AddOp = FI; SubOp = TI;
7777 } else if (FI->getOpcode() == Instruction::Sub &&
7778 TI->getOpcode() == Instruction::Add) {
7779 AddOp = TI; SubOp = FI;
7783 Value *OtherAddOp = 0;
7784 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7785 OtherAddOp = AddOp->getOperand(1);
7786 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7787 OtherAddOp = AddOp->getOperand(0);
7791 // So at this point we know we have (Y -> OtherAddOp):
7792 // select C, (add X, Y), (sub X, Z)
7793 Value *NegVal; // Compute -Z
7794 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7795 NegVal = ConstantExpr::getNeg(C);
7797 NegVal = InsertNewInstBefore(
7798 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7801 Value *NewTrueOp = OtherAddOp;
7802 Value *NewFalseOp = NegVal;
7804 std::swap(NewTrueOp, NewFalseOp);
7805 Instruction *NewSel =
7806 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7808 NewSel = InsertNewInstBefore(NewSel, SI);
7809 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7814 // See if we can fold the select into one of our operands.
7815 if (SI.getType()->isInteger()) {
7816 // See the comment above GetSelectFoldableOperands for a description of the
7817 // transformation we are doing here.
7818 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7819 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7820 !isa<Constant>(FalseVal))
7821 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7822 unsigned OpToFold = 0;
7823 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7825 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7830 Constant *C = GetSelectFoldableConstant(TVI);
7831 Instruction *NewSel =
7832 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7833 InsertNewInstBefore(NewSel, SI);
7834 NewSel->takeName(TVI);
7835 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7836 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7838 assert(0 && "Unknown instruction!!");
7843 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7844 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7845 !isa<Constant>(TrueVal))
7846 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7847 unsigned OpToFold = 0;
7848 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7850 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7855 Constant *C = GetSelectFoldableConstant(FVI);
7856 Instruction *NewSel =
7857 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7858 InsertNewInstBefore(NewSel, SI);
7859 NewSel->takeName(FVI);
7860 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7861 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7863 assert(0 && "Unknown instruction!!");
7868 if (BinaryOperator::isNot(CondVal)) {
7869 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7870 SI.setOperand(1, FalseVal);
7871 SI.setOperand(2, TrueVal);
7878 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7879 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7880 /// and it is more than the alignment of the ultimate object, see if we can
7881 /// increase the alignment of the ultimate object, making this check succeed.
7882 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7883 unsigned PrefAlign = 0) {
7884 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7885 unsigned Align = GV->getAlignment();
7886 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7887 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7889 // If there is a large requested alignment and we can, bump up the alignment
7891 if (PrefAlign > Align && GV->hasInitializer()) {
7892 GV->setAlignment(PrefAlign);
7896 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7897 unsigned Align = AI->getAlignment();
7898 if (Align == 0 && TD) {
7899 if (isa<AllocaInst>(AI))
7900 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7901 else if (isa<MallocInst>(AI)) {
7902 // Malloc returns maximally aligned memory.
7903 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7906 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7909 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7913 // If there is a requested alignment and if this is an alloca, round up. We
7914 // don't do this for malloc, because some systems can't respect the request.
7915 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7916 AI->setAlignment(PrefAlign);
7920 } else if (isa<BitCastInst>(V) ||
7921 (isa<ConstantExpr>(V) &&
7922 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7923 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7925 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7926 // If all indexes are zero, it is just the alignment of the base pointer.
7927 bool AllZeroOperands = true;
7928 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7929 if (!isa<Constant>(GEPI->getOperand(i)) ||
7930 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7931 AllZeroOperands = false;
7935 if (AllZeroOperands) {
7936 // Treat this like a bitcast.
7937 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7940 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7941 if (BaseAlignment == 0) return 0;
7943 // Otherwise, if the base alignment is >= the alignment we expect for the
7944 // base pointer type, then we know that the resultant pointer is aligned at
7945 // least as much as its type requires.
7948 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7949 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7950 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7951 if (Align <= BaseAlignment) {
7952 const Type *GEPTy = GEPI->getType();
7953 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7954 Align = std::min(Align, (unsigned)
7955 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7963 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
7964 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7965 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7966 unsigned MinAlign = std::min(DstAlign, SrcAlign);
7967 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
7969 if (CopyAlign < MinAlign) {
7970 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
7974 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
7976 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
7977 if (MemOpLength == 0) return 0;
7979 // Source and destination pointer types are always "i8*" for intrinsic. See
7980 // if the size is something we can handle with a single primitive load/store.
7981 // A single load+store correctly handles overlapping memory in the memmove
7983 unsigned Size = MemOpLength->getZExtValue();
7984 if (Size == 0 || Size > 8 || (Size&(Size-1)))
7985 return 0; // If not 1/2/4/8 bytes, exit.
7987 // Use an integer load+store unless we can find something better.
7988 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
7990 // Memcpy forces the use of i8* for the source and destination. That means
7991 // that if you're using memcpy to move one double around, you'll get a cast
7992 // from double* to i8*. We'd much rather use a double load+store rather than
7993 // an i64 load+store, here because this improves the odds that the source or
7994 // dest address will be promotable. See if we can find a better type than the
7995 // integer datatype.
7996 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
7997 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
7998 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
7999 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8000 // down through these levels if so.
8001 while (!SrcETy->isFirstClassType()) {
8002 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8003 if (STy->getNumElements() == 1)
8004 SrcETy = STy->getElementType(0);
8007 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8008 if (ATy->getNumElements() == 1)
8009 SrcETy = ATy->getElementType();
8016 if (SrcETy->isFirstClassType())
8017 NewPtrTy = PointerType::getUnqual(SrcETy);
8022 // If the memcpy/memmove provides better alignment info than we can
8024 SrcAlign = std::max(SrcAlign, CopyAlign);
8025 DstAlign = std::max(DstAlign, CopyAlign);
8027 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8028 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8029 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8030 InsertNewInstBefore(L, *MI);
8031 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8033 // Set the size of the copy to 0, it will be deleted on the next iteration.
8034 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8038 /// visitCallInst - CallInst simplification. This mostly only handles folding
8039 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8040 /// the heavy lifting.
8042 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8043 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8044 if (!II) return visitCallSite(&CI);
8046 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8048 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8049 bool Changed = false;
8051 // memmove/cpy/set of zero bytes is a noop.
8052 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8053 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8055 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8056 if (CI->getZExtValue() == 1) {
8057 // Replace the instruction with just byte operations. We would
8058 // transform other cases to loads/stores, but we don't know if
8059 // alignment is sufficient.
8063 // If we have a memmove and the source operation is a constant global,
8064 // then the source and dest pointers can't alias, so we can change this
8065 // into a call to memcpy.
8066 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8067 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8068 if (GVSrc->isConstant()) {
8069 Module *M = CI.getParent()->getParent()->getParent();
8070 Intrinsic::ID MemCpyID;
8071 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8072 MemCpyID = Intrinsic::memcpy_i32;
8074 MemCpyID = Intrinsic::memcpy_i64;
8075 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8080 // If we can determine a pointer alignment that is bigger than currently
8081 // set, update the alignment.
8082 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8083 if (Instruction *I = SimplifyMemTransfer(MI))
8085 } else if (isa<MemSetInst>(MI)) {
8086 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
8087 if (MI->getAlignment()->getZExtValue() < Alignment) {
8088 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8093 if (Changed) return II;
8095 switch (II->getIntrinsicID()) {
8097 case Intrinsic::ppc_altivec_lvx:
8098 case Intrinsic::ppc_altivec_lvxl:
8099 case Intrinsic::x86_sse_loadu_ps:
8100 case Intrinsic::x86_sse2_loadu_pd:
8101 case Intrinsic::x86_sse2_loadu_dq:
8102 // Turn PPC lvx -> load if the pointer is known aligned.
8103 // Turn X86 loadups -> load if the pointer is known aligned.
8104 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8105 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8106 PointerType::getUnqual(II->getType()),
8108 return new LoadInst(Ptr);
8111 case Intrinsic::ppc_altivec_stvx:
8112 case Intrinsic::ppc_altivec_stvxl:
8113 // Turn stvx -> store if the pointer is known aligned.
8114 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
8115 const Type *OpPtrTy =
8116 PointerType::getUnqual(II->getOperand(1)->getType());
8117 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8118 return new StoreInst(II->getOperand(1), Ptr);
8121 case Intrinsic::x86_sse_storeu_ps:
8122 case Intrinsic::x86_sse2_storeu_pd:
8123 case Intrinsic::x86_sse2_storeu_dq:
8124 case Intrinsic::x86_sse2_storel_dq:
8125 // Turn X86 storeu -> store if the pointer is known aligned.
8126 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8127 const Type *OpPtrTy =
8128 PointerType::getUnqual(II->getOperand(2)->getType());
8129 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8130 return new StoreInst(II->getOperand(2), Ptr);
8134 case Intrinsic::x86_sse_cvttss2si: {
8135 // These intrinsics only demands the 0th element of its input vector. If
8136 // we can simplify the input based on that, do so now.
8138 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8140 II->setOperand(1, V);
8146 case Intrinsic::ppc_altivec_vperm:
8147 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8148 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8149 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8151 // Check that all of the elements are integer constants or undefs.
8152 bool AllEltsOk = true;
8153 for (unsigned i = 0; i != 16; ++i) {
8154 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8155 !isa<UndefValue>(Mask->getOperand(i))) {
8162 // Cast the input vectors to byte vectors.
8163 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8164 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8165 Value *Result = UndefValue::get(Op0->getType());
8167 // Only extract each element once.
8168 Value *ExtractedElts[32];
8169 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8171 for (unsigned i = 0; i != 16; ++i) {
8172 if (isa<UndefValue>(Mask->getOperand(i)))
8174 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8175 Idx &= 31; // Match the hardware behavior.
8177 if (ExtractedElts[Idx] == 0) {
8179 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8180 InsertNewInstBefore(Elt, CI);
8181 ExtractedElts[Idx] = Elt;
8184 // Insert this value into the result vector.
8185 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
8186 InsertNewInstBefore(cast<Instruction>(Result), CI);
8188 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8193 case Intrinsic::stackrestore: {
8194 // If the save is right next to the restore, remove the restore. This can
8195 // happen when variable allocas are DCE'd.
8196 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8197 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8198 BasicBlock::iterator BI = SS;
8200 return EraseInstFromFunction(CI);
8204 // Scan down this block to see if there is another stack restore in the
8205 // same block without an intervening call/alloca.
8206 BasicBlock::iterator BI = II;
8207 TerminatorInst *TI = II->getParent()->getTerminator();
8208 bool CannotRemove = false;
8209 for (++BI; &*BI != TI; ++BI) {
8210 if (isa<AllocaInst>(BI)) {
8211 CannotRemove = true;
8214 if (isa<CallInst>(BI)) {
8215 if (!isa<IntrinsicInst>(BI)) {
8216 CannotRemove = true;
8219 // If there is a stackrestore below this one, remove this one.
8220 return EraseInstFromFunction(CI);
8224 // If the stack restore is in a return/unwind block and if there are no
8225 // allocas or calls between the restore and the return, nuke the restore.
8226 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8227 return EraseInstFromFunction(CI);
8233 return visitCallSite(II);
8236 // InvokeInst simplification
8238 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8239 return visitCallSite(&II);
8242 // visitCallSite - Improvements for call and invoke instructions.
8244 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8245 bool Changed = false;
8247 // If the callee is a constexpr cast of a function, attempt to move the cast
8248 // to the arguments of the call/invoke.
8249 if (transformConstExprCastCall(CS)) return 0;
8251 Value *Callee = CS.getCalledValue();
8253 if (Function *CalleeF = dyn_cast<Function>(Callee))
8254 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8255 Instruction *OldCall = CS.getInstruction();
8256 // If the call and callee calling conventions don't match, this call must
8257 // be unreachable, as the call is undefined.
8258 new StoreInst(ConstantInt::getTrue(),
8259 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8261 if (!OldCall->use_empty())
8262 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8263 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8264 return EraseInstFromFunction(*OldCall);
8268 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8269 // This instruction is not reachable, just remove it. We insert a store to
8270 // undef so that we know that this code is not reachable, despite the fact
8271 // that we can't modify the CFG here.
8272 new StoreInst(ConstantInt::getTrue(),
8273 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8274 CS.getInstruction());
8276 if (!CS.getInstruction()->use_empty())
8277 CS.getInstruction()->
8278 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8280 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8281 // Don't break the CFG, insert a dummy cond branch.
8282 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8283 ConstantInt::getTrue(), II);
8285 return EraseInstFromFunction(*CS.getInstruction());
8288 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8289 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8290 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8291 return transformCallThroughTrampoline(CS);
8293 const PointerType *PTy = cast<PointerType>(Callee->getType());
8294 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8295 if (FTy->isVarArg()) {
8296 // See if we can optimize any arguments passed through the varargs area of
8298 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8299 E = CS.arg_end(); I != E; ++I)
8300 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8301 // If this cast does not effect the value passed through the varargs
8302 // area, we can eliminate the use of the cast.
8303 Value *Op = CI->getOperand(0);
8304 if (CI->isLosslessCast()) {
8311 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8312 // Inline asm calls cannot throw - mark them 'nounwind'.
8313 CS.setDoesNotThrow();
8317 return Changed ? CS.getInstruction() : 0;
8320 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8321 // attempt to move the cast to the arguments of the call/invoke.
8323 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8324 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8325 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8326 if (CE->getOpcode() != Instruction::BitCast ||
8327 !isa<Function>(CE->getOperand(0)))
8329 Function *Callee = cast<Function>(CE->getOperand(0));
8330 Instruction *Caller = CS.getInstruction();
8331 const ParamAttrsList* CallerPAL = CS.getParamAttrs();
8333 // Okay, this is a cast from a function to a different type. Unless doing so
8334 // would cause a type conversion of one of our arguments, change this call to
8335 // be a direct call with arguments casted to the appropriate types.
8337 const FunctionType *FT = Callee->getFunctionType();
8338 const Type *OldRetTy = Caller->getType();
8340 // Check to see if we are changing the return type...
8341 if (OldRetTy != FT->getReturnType()) {
8342 if (Callee->isDeclaration() && !Caller->use_empty() &&
8343 // Conversion is ok if changing from pointer to int of same size.
8344 !(isa<PointerType>(FT->getReturnType()) &&
8345 TD->getIntPtrType() == OldRetTy))
8346 return false; // Cannot transform this return value.
8348 if (!Caller->use_empty() &&
8349 // void -> non-void is handled specially
8350 FT->getReturnType() != Type::VoidTy &&
8351 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8352 return false; // Cannot transform this return value.
8354 if (CallerPAL && !Caller->use_empty()) {
8355 uint16_t RAttrs = CallerPAL->getParamAttrs(0);
8356 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8357 return false; // Attribute not compatible with transformed value.
8360 // If the callsite is an invoke instruction, and the return value is used by
8361 // a PHI node in a successor, we cannot change the return type of the call
8362 // because there is no place to put the cast instruction (without breaking
8363 // the critical edge). Bail out in this case.
8364 if (!Caller->use_empty())
8365 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8366 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8368 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8369 if (PN->getParent() == II->getNormalDest() ||
8370 PN->getParent() == II->getUnwindDest())
8374 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8375 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8377 CallSite::arg_iterator AI = CS.arg_begin();
8378 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8379 const Type *ParamTy = FT->getParamType(i);
8380 const Type *ActTy = (*AI)->getType();
8382 if (!CastInst::isCastable(ActTy, ParamTy))
8383 return false; // Cannot transform this parameter value.
8386 uint16_t PAttrs = CallerPAL->getParamAttrs(i + 1);
8387 if (PAttrs & ParamAttr::typeIncompatible(ParamTy))
8388 return false; // Attribute not compatible with transformed value.
8391 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8392 // Some conversions are safe even if we do not have a body.
8393 // Either we can cast directly, or we can upconvert the argument
8394 bool isConvertible = ActTy == ParamTy ||
8395 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8396 (ParamTy->isInteger() && ActTy->isInteger() &&
8397 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8398 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8399 && c->getValue().isStrictlyPositive());
8400 if (Callee->isDeclaration() && !isConvertible) return false;
8403 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8404 Callee->isDeclaration())
8405 return false; // Do not delete arguments unless we have a function body...
8407 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && CallerPAL)
8408 // In this case we have more arguments than the new function type, but we
8409 // won't be dropping them. Check that these extra arguments have attributes
8410 // that are compatible with being a vararg call argument.
8411 for (unsigned i = CallerPAL->size(); i; --i) {
8412 if (CallerPAL->getParamIndex(i - 1) <= FT->getNumParams())
8414 uint16_t PAttrs = CallerPAL->getParamAttrsAtIndex(i - 1);
8415 if (PAttrs & ParamAttr::VarArgsIncompatible)
8419 // Okay, we decided that this is a safe thing to do: go ahead and start
8420 // inserting cast instructions as necessary...
8421 std::vector<Value*> Args;
8422 Args.reserve(NumActualArgs);
8423 ParamAttrsVector attrVec;
8424 attrVec.reserve(NumCommonArgs);
8426 // Get any return attributes.
8427 uint16_t RAttrs = CallerPAL ? CallerPAL->getParamAttrs(0) : 0;
8429 // If the return value is not being used, the type may not be compatible
8430 // with the existing attributes. Wipe out any problematic attributes.
8431 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
8433 // Add the new return attributes.
8435 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8437 AI = CS.arg_begin();
8438 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8439 const Type *ParamTy = FT->getParamType(i);
8440 if ((*AI)->getType() == ParamTy) {
8441 Args.push_back(*AI);
8443 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8444 false, ParamTy, false);
8445 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8446 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8449 // Add any parameter attributes.
8450 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8452 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8455 // If the function takes more arguments than the call was taking, add them
8457 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8458 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8460 // If we are removing arguments to the function, emit an obnoxious warning...
8461 if (FT->getNumParams() < NumActualArgs)
8462 if (!FT->isVarArg()) {
8463 cerr << "WARNING: While resolving call to function '"
8464 << Callee->getName() << "' arguments were dropped!\n";
8466 // Add all of the arguments in their promoted form to the arg list...
8467 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8468 const Type *PTy = getPromotedType((*AI)->getType());
8469 if (PTy != (*AI)->getType()) {
8470 // Must promote to pass through va_arg area!
8471 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8473 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8474 InsertNewInstBefore(Cast, *Caller);
8475 Args.push_back(Cast);
8477 Args.push_back(*AI);
8480 // Add any parameter attributes.
8481 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8483 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8487 if (FT->getReturnType() == Type::VoidTy)
8488 Caller->setName(""); // Void type should not have a name.
8490 const ParamAttrsList* NewCallerPAL = ParamAttrsList::get(attrVec);
8493 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8494 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8495 Args.begin(), Args.end(), Caller->getName(), Caller);
8496 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8497 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
8499 NC = new CallInst(Callee, Args.begin(), Args.end(),
8500 Caller->getName(), Caller);
8501 CallInst *CI = cast<CallInst>(Caller);
8502 if (CI->isTailCall())
8503 cast<CallInst>(NC)->setTailCall();
8504 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8505 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
8508 // Insert a cast of the return type as necessary.
8510 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8511 if (NV->getType() != Type::VoidTy) {
8512 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8514 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8516 // If this is an invoke instruction, we should insert it after the first
8517 // non-phi, instruction in the normal successor block.
8518 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8519 BasicBlock::iterator I = II->getNormalDest()->begin();
8520 while (isa<PHINode>(I)) ++I;
8521 InsertNewInstBefore(NC, *I);
8523 // Otherwise, it's a call, just insert cast right after the call instr
8524 InsertNewInstBefore(NC, *Caller);
8526 AddUsersToWorkList(*Caller);
8528 NV = UndefValue::get(Caller->getType());
8532 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8533 Caller->replaceAllUsesWith(NV);
8534 Caller->eraseFromParent();
8535 RemoveFromWorkList(Caller);
8539 // transformCallThroughTrampoline - Turn a call to a function created by the
8540 // init_trampoline intrinsic into a direct call to the underlying function.
8542 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8543 Value *Callee = CS.getCalledValue();
8544 const PointerType *PTy = cast<PointerType>(Callee->getType());
8545 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8546 const ParamAttrsList *Attrs = CS.getParamAttrs();
8548 // If the call already has the 'nest' attribute somewhere then give up -
8549 // otherwise 'nest' would occur twice after splicing in the chain.
8550 if (Attrs && Attrs->hasAttrSomewhere(ParamAttr::Nest))
8553 IntrinsicInst *Tramp =
8554 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8557 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8558 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8559 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8561 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8562 unsigned NestIdx = 1;
8563 const Type *NestTy = 0;
8564 uint16_t NestAttr = 0;
8566 // Look for a parameter marked with the 'nest' attribute.
8567 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8568 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8569 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8570 // Record the parameter type and any other attributes.
8572 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8577 Instruction *Caller = CS.getInstruction();
8578 std::vector<Value*> NewArgs;
8579 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8581 ParamAttrsVector NewAttrs;
8582 NewAttrs.reserve(Attrs ? Attrs->size() + 1 : 1);
8584 // Insert the nest argument into the call argument list, which may
8585 // mean appending it. Likewise for attributes.
8587 // Add any function result attributes.
8588 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
8590 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8594 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8596 if (Idx == NestIdx) {
8597 // Add the chain argument and attributes.
8598 Value *NestVal = Tramp->getOperand(3);
8599 if (NestVal->getType() != NestTy)
8600 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8601 NewArgs.push_back(NestVal);
8602 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8608 // Add the original argument and attributes.
8609 NewArgs.push_back(*I);
8610 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8613 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8619 // The trampoline may have been bitcast to a bogus type (FTy).
8620 // Handle this by synthesizing a new function type, equal to FTy
8621 // with the chain parameter inserted.
8623 std::vector<const Type*> NewTypes;
8624 NewTypes.reserve(FTy->getNumParams()+1);
8626 // Insert the chain's type into the list of parameter types, which may
8627 // mean appending it.
8630 FunctionType::param_iterator I = FTy->param_begin(),
8631 E = FTy->param_end();
8635 // Add the chain's type.
8636 NewTypes.push_back(NestTy);
8641 // Add the original type.
8642 NewTypes.push_back(*I);
8648 // Replace the trampoline call with a direct call. Let the generic
8649 // code sort out any function type mismatches.
8650 FunctionType *NewFTy =
8651 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8652 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8653 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8654 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8656 Instruction *NewCaller;
8657 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8658 NewCaller = new InvokeInst(NewCallee,
8659 II->getNormalDest(), II->getUnwindDest(),
8660 NewArgs.begin(), NewArgs.end(),
8661 Caller->getName(), Caller);
8662 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8663 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8665 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8666 Caller->getName(), Caller);
8667 if (cast<CallInst>(Caller)->isTailCall())
8668 cast<CallInst>(NewCaller)->setTailCall();
8669 cast<CallInst>(NewCaller)->
8670 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8671 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8673 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8674 Caller->replaceAllUsesWith(NewCaller);
8675 Caller->eraseFromParent();
8676 RemoveFromWorkList(Caller);
8681 // Replace the trampoline call with a direct call. Since there is no 'nest'
8682 // parameter, there is no need to adjust the argument list. Let the generic
8683 // code sort out any function type mismatches.
8684 Constant *NewCallee =
8685 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8686 CS.setCalledFunction(NewCallee);
8687 return CS.getInstruction();
8690 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8691 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8692 /// and a single binop.
8693 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8694 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8695 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8696 isa<CmpInst>(FirstInst));
8697 unsigned Opc = FirstInst->getOpcode();
8698 Value *LHSVal = FirstInst->getOperand(0);
8699 Value *RHSVal = FirstInst->getOperand(1);
8701 const Type *LHSType = LHSVal->getType();
8702 const Type *RHSType = RHSVal->getType();
8704 // Scan to see if all operands are the same opcode, all have one use, and all
8705 // kill their operands (i.e. the operands have one use).
8706 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8707 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8708 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8709 // Verify type of the LHS matches so we don't fold cmp's of different
8710 // types or GEP's with different index types.
8711 I->getOperand(0)->getType() != LHSType ||
8712 I->getOperand(1)->getType() != RHSType)
8715 // If they are CmpInst instructions, check their predicates
8716 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8717 if (cast<CmpInst>(I)->getPredicate() !=
8718 cast<CmpInst>(FirstInst)->getPredicate())
8721 // Keep track of which operand needs a phi node.
8722 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8723 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8726 // Otherwise, this is safe to transform, determine if it is profitable.
8728 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8729 // Indexes are often folded into load/store instructions, so we don't want to
8730 // hide them behind a phi.
8731 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8734 Value *InLHS = FirstInst->getOperand(0);
8735 Value *InRHS = FirstInst->getOperand(1);
8736 PHINode *NewLHS = 0, *NewRHS = 0;
8738 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8739 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8740 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8741 InsertNewInstBefore(NewLHS, PN);
8746 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8747 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8748 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8749 InsertNewInstBefore(NewRHS, PN);
8753 // Add all operands to the new PHIs.
8754 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8756 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8757 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8760 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8761 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8765 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8766 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8767 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8768 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8771 assert(isa<GetElementPtrInst>(FirstInst));
8772 return new GetElementPtrInst(LHSVal, RHSVal);
8776 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8777 /// of the block that defines it. This means that it must be obvious the value
8778 /// of the load is not changed from the point of the load to the end of the
8781 /// Finally, it is safe, but not profitable, to sink a load targetting a
8782 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8784 static bool isSafeToSinkLoad(LoadInst *L) {
8785 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8787 for (++BBI; BBI != E; ++BBI)
8788 if (BBI->mayWriteToMemory())
8791 // Check for non-address taken alloca. If not address-taken already, it isn't
8792 // profitable to do this xform.
8793 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8794 bool isAddressTaken = false;
8795 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8797 if (isa<LoadInst>(UI)) continue;
8798 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8799 // If storing TO the alloca, then the address isn't taken.
8800 if (SI->getOperand(1) == AI) continue;
8802 isAddressTaken = true;
8806 if (!isAddressTaken)
8814 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8815 // operator and they all are only used by the PHI, PHI together their
8816 // inputs, and do the operation once, to the result of the PHI.
8817 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8818 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8820 // Scan the instruction, looking for input operations that can be folded away.
8821 // If all input operands to the phi are the same instruction (e.g. a cast from
8822 // the same type or "+42") we can pull the operation through the PHI, reducing
8823 // code size and simplifying code.
8824 Constant *ConstantOp = 0;
8825 const Type *CastSrcTy = 0;
8826 bool isVolatile = false;
8827 if (isa<CastInst>(FirstInst)) {
8828 CastSrcTy = FirstInst->getOperand(0)->getType();
8829 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8830 // Can fold binop, compare or shift here if the RHS is a constant,
8831 // otherwise call FoldPHIArgBinOpIntoPHI.
8832 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8833 if (ConstantOp == 0)
8834 return FoldPHIArgBinOpIntoPHI(PN);
8835 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8836 isVolatile = LI->isVolatile();
8837 // We can't sink the load if the loaded value could be modified between the
8838 // load and the PHI.
8839 if (LI->getParent() != PN.getIncomingBlock(0) ||
8840 !isSafeToSinkLoad(LI))
8842 } else if (isa<GetElementPtrInst>(FirstInst)) {
8843 if (FirstInst->getNumOperands() == 2)
8844 return FoldPHIArgBinOpIntoPHI(PN);
8845 // Can't handle general GEPs yet.
8848 return 0; // Cannot fold this operation.
8851 // Check to see if all arguments are the same operation.
8852 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8853 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8854 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8855 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8858 if (I->getOperand(0)->getType() != CastSrcTy)
8859 return 0; // Cast operation must match.
8860 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8861 // We can't sink the load if the loaded value could be modified between
8862 // the load and the PHI.
8863 if (LI->isVolatile() != isVolatile ||
8864 LI->getParent() != PN.getIncomingBlock(i) ||
8865 !isSafeToSinkLoad(LI))
8867 } else if (I->getOperand(1) != ConstantOp) {
8872 // Okay, they are all the same operation. Create a new PHI node of the
8873 // correct type, and PHI together all of the LHS's of the instructions.
8874 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8875 PN.getName()+".in");
8876 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8878 Value *InVal = FirstInst->getOperand(0);
8879 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8881 // Add all operands to the new PHI.
8882 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8883 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8884 if (NewInVal != InVal)
8886 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8891 // The new PHI unions all of the same values together. This is really
8892 // common, so we handle it intelligently here for compile-time speed.
8896 InsertNewInstBefore(NewPN, PN);
8900 // Insert and return the new operation.
8901 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8902 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8903 else if (isa<LoadInst>(FirstInst))
8904 return new LoadInst(PhiVal, "", isVolatile);
8905 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8906 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8907 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8908 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8909 PhiVal, ConstantOp);
8911 assert(0 && "Unknown operation");
8915 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8917 static bool DeadPHICycle(PHINode *PN,
8918 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8919 if (PN->use_empty()) return true;
8920 if (!PN->hasOneUse()) return false;
8922 // Remember this node, and if we find the cycle, return.
8923 if (!PotentiallyDeadPHIs.insert(PN))
8926 // Don't scan crazily complex things.
8927 if (PotentiallyDeadPHIs.size() == 16)
8930 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8931 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8936 /// PHIsEqualValue - Return true if this phi node is always equal to
8937 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8938 /// z = some value; x = phi (y, z); y = phi (x, z)
8939 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8940 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8941 // See if we already saw this PHI node.
8942 if (!ValueEqualPHIs.insert(PN))
8945 // Don't scan crazily complex things.
8946 if (ValueEqualPHIs.size() == 16)
8949 // Scan the operands to see if they are either phi nodes or are equal to
8951 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8952 Value *Op = PN->getIncomingValue(i);
8953 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8954 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8956 } else if (Op != NonPhiInVal)
8964 // PHINode simplification
8966 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8967 // If LCSSA is around, don't mess with Phi nodes
8968 if (MustPreserveLCSSA) return 0;
8970 if (Value *V = PN.hasConstantValue())
8971 return ReplaceInstUsesWith(PN, V);
8973 // If all PHI operands are the same operation, pull them through the PHI,
8974 // reducing code size.
8975 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8976 PN.getIncomingValue(0)->hasOneUse())
8977 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8980 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8981 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8982 // PHI)... break the cycle.
8983 if (PN.hasOneUse()) {
8984 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8985 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8986 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8987 PotentiallyDeadPHIs.insert(&PN);
8988 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8989 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8992 // If this phi has a single use, and if that use just computes a value for
8993 // the next iteration of a loop, delete the phi. This occurs with unused
8994 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8995 // common case here is good because the only other things that catch this
8996 // are induction variable analysis (sometimes) and ADCE, which is only run
8998 if (PHIUser->hasOneUse() &&
8999 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9000 PHIUser->use_back() == &PN) {
9001 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9005 // We sometimes end up with phi cycles that non-obviously end up being the
9006 // same value, for example:
9007 // z = some value; x = phi (y, z); y = phi (x, z)
9008 // where the phi nodes don't necessarily need to be in the same block. Do a
9009 // quick check to see if the PHI node only contains a single non-phi value, if
9010 // so, scan to see if the phi cycle is actually equal to that value.
9012 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9013 // Scan for the first non-phi operand.
9014 while (InValNo != NumOperandVals &&
9015 isa<PHINode>(PN.getIncomingValue(InValNo)))
9018 if (InValNo != NumOperandVals) {
9019 Value *NonPhiInVal = PN.getOperand(InValNo);
9021 // Scan the rest of the operands to see if there are any conflicts, if so
9022 // there is no need to recursively scan other phis.
9023 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9024 Value *OpVal = PN.getIncomingValue(InValNo);
9025 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9029 // If we scanned over all operands, then we have one unique value plus
9030 // phi values. Scan PHI nodes to see if they all merge in each other or
9032 if (InValNo == NumOperandVals) {
9033 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9034 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9035 return ReplaceInstUsesWith(PN, NonPhiInVal);
9042 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9043 Instruction *InsertPoint,
9045 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9046 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9047 // We must cast correctly to the pointer type. Ensure that we
9048 // sign extend the integer value if it is smaller as this is
9049 // used for address computation.
9050 Instruction::CastOps opcode =
9051 (VTySize < PtrSize ? Instruction::SExt :
9052 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9053 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9057 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9058 Value *PtrOp = GEP.getOperand(0);
9059 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9060 // If so, eliminate the noop.
9061 if (GEP.getNumOperands() == 1)
9062 return ReplaceInstUsesWith(GEP, PtrOp);
9064 if (isa<UndefValue>(GEP.getOperand(0)))
9065 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9067 bool HasZeroPointerIndex = false;
9068 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9069 HasZeroPointerIndex = C->isNullValue();
9071 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9072 return ReplaceInstUsesWith(GEP, PtrOp);
9074 // Eliminate unneeded casts for indices.
9075 bool MadeChange = false;
9077 gep_type_iterator GTI = gep_type_begin(GEP);
9078 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
9079 if (isa<SequentialType>(*GTI)) {
9080 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
9081 if (CI->getOpcode() == Instruction::ZExt ||
9082 CI->getOpcode() == Instruction::SExt) {
9083 const Type *SrcTy = CI->getOperand(0)->getType();
9084 // We can eliminate a cast from i32 to i64 iff the target
9085 // is a 32-bit pointer target.
9086 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9088 GEP.setOperand(i, CI->getOperand(0));
9092 // If we are using a wider index than needed for this platform, shrink it
9093 // to what we need. If the incoming value needs a cast instruction,
9094 // insert it. This explicit cast can make subsequent optimizations more
9096 Value *Op = GEP.getOperand(i);
9097 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits())
9098 if (Constant *C = dyn_cast<Constant>(Op)) {
9099 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
9102 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9104 GEP.setOperand(i, Op);
9109 if (MadeChange) return &GEP;
9111 // If this GEP instruction doesn't move the pointer, and if the input operand
9112 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9113 // real input to the dest type.
9114 if (GEP.hasAllZeroIndices()) {
9115 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9116 // If the bitcast is of an allocation, and the allocation will be
9117 // converted to match the type of the cast, don't touch this.
9118 if (isa<AllocationInst>(BCI->getOperand(0))) {
9119 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9120 if (Instruction *I = visitBitCast(*BCI)) {
9123 BCI->getParent()->getInstList().insert(BCI, I);
9124 ReplaceInstUsesWith(*BCI, I);
9129 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9133 // Combine Indices - If the source pointer to this getelementptr instruction
9134 // is a getelementptr instruction, combine the indices of the two
9135 // getelementptr instructions into a single instruction.
9137 SmallVector<Value*, 8> SrcGEPOperands;
9138 if (User *Src = dyn_castGetElementPtr(PtrOp))
9139 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9141 if (!SrcGEPOperands.empty()) {
9142 // Note that if our source is a gep chain itself that we wait for that
9143 // chain to be resolved before we perform this transformation. This
9144 // avoids us creating a TON of code in some cases.
9146 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9147 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9148 return 0; // Wait until our source is folded to completion.
9150 SmallVector<Value*, 8> Indices;
9152 // Find out whether the last index in the source GEP is a sequential idx.
9153 bool EndsWithSequential = false;
9154 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9155 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9156 EndsWithSequential = !isa<StructType>(*I);
9158 // Can we combine the two pointer arithmetics offsets?
9159 if (EndsWithSequential) {
9160 // Replace: gep (gep %P, long B), long A, ...
9161 // With: T = long A+B; gep %P, T, ...
9163 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9164 if (SO1 == Constant::getNullValue(SO1->getType())) {
9166 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9169 // If they aren't the same type, convert both to an integer of the
9170 // target's pointer size.
9171 if (SO1->getType() != GO1->getType()) {
9172 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9173 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9174 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9175 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9177 unsigned PS = TD->getPointerSizeInBits();
9178 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9179 // Convert GO1 to SO1's type.
9180 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9182 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9183 // Convert SO1 to GO1's type.
9184 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9186 const Type *PT = TD->getIntPtrType();
9187 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9188 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9192 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9193 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9195 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
9196 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9200 // Recycle the GEP we already have if possible.
9201 if (SrcGEPOperands.size() == 2) {
9202 GEP.setOperand(0, SrcGEPOperands[0]);
9203 GEP.setOperand(1, Sum);
9206 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9207 SrcGEPOperands.end()-1);
9208 Indices.push_back(Sum);
9209 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9211 } else if (isa<Constant>(*GEP.idx_begin()) &&
9212 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9213 SrcGEPOperands.size() != 1) {
9214 // Otherwise we can do the fold if the first index of the GEP is a zero
9215 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9216 SrcGEPOperands.end());
9217 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9220 if (!Indices.empty())
9221 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
9222 Indices.end(), GEP.getName());
9224 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9225 // GEP of global variable. If all of the indices for this GEP are
9226 // constants, we can promote this to a constexpr instead of an instruction.
9228 // Scan for nonconstants...
9229 SmallVector<Constant*, 8> Indices;
9230 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9231 for (; I != E && isa<Constant>(*I); ++I)
9232 Indices.push_back(cast<Constant>(*I));
9234 if (I == E) { // If they are all constants...
9235 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9236 &Indices[0],Indices.size());
9238 // Replace all uses of the GEP with the new constexpr...
9239 return ReplaceInstUsesWith(GEP, CE);
9241 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9242 if (!isa<PointerType>(X->getType())) {
9243 // Not interesting. Source pointer must be a cast from pointer.
9244 } else if (HasZeroPointerIndex) {
9245 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9246 // into : GEP [10 x i8]* X, i32 0, ...
9248 // This occurs when the program declares an array extern like "int X[];"
9250 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9251 const PointerType *XTy = cast<PointerType>(X->getType());
9252 if (const ArrayType *XATy =
9253 dyn_cast<ArrayType>(XTy->getElementType()))
9254 if (const ArrayType *CATy =
9255 dyn_cast<ArrayType>(CPTy->getElementType()))
9256 if (CATy->getElementType() == XATy->getElementType()) {
9257 // At this point, we know that the cast source type is a pointer
9258 // to an array of the same type as the destination pointer
9259 // array. Because the array type is never stepped over (there
9260 // is a leading zero) we can fold the cast into this GEP.
9261 GEP.setOperand(0, X);
9264 } else if (GEP.getNumOperands() == 2) {
9265 // Transform things like:
9266 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9267 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9268 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9269 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9270 if (isa<ArrayType>(SrcElTy) &&
9271 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9272 TD->getABITypeSize(ResElTy)) {
9274 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9275 Idx[1] = GEP.getOperand(1);
9276 Value *V = InsertNewInstBefore(
9277 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
9278 // V and GEP are both pointer types --> BitCast
9279 return new BitCastInst(V, GEP.getType());
9282 // Transform things like:
9283 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9284 // (where tmp = 8*tmp2) into:
9285 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9287 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9288 uint64_t ArrayEltSize =
9289 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9291 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9292 // allow either a mul, shift, or constant here.
9294 ConstantInt *Scale = 0;
9295 if (ArrayEltSize == 1) {
9296 NewIdx = GEP.getOperand(1);
9297 Scale = ConstantInt::get(NewIdx->getType(), 1);
9298 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9299 NewIdx = ConstantInt::get(CI->getType(), 1);
9301 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9302 if (Inst->getOpcode() == Instruction::Shl &&
9303 isa<ConstantInt>(Inst->getOperand(1))) {
9304 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9305 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9306 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9307 NewIdx = Inst->getOperand(0);
9308 } else if (Inst->getOpcode() == Instruction::Mul &&
9309 isa<ConstantInt>(Inst->getOperand(1))) {
9310 Scale = cast<ConstantInt>(Inst->getOperand(1));
9311 NewIdx = Inst->getOperand(0);
9315 // If the index will be to exactly the right offset with the scale taken
9316 // out, perform the transformation. Note, we don't know whether Scale is
9317 // signed or not. We'll use unsigned version of division/modulo
9318 // operation after making sure Scale doesn't have the sign bit set.
9319 if (Scale && Scale->getSExtValue() >= 0LL &&
9320 Scale->getZExtValue() % ArrayEltSize == 0) {
9321 Scale = ConstantInt::get(Scale->getType(),
9322 Scale->getZExtValue() / ArrayEltSize);
9323 if (Scale->getZExtValue() != 1) {
9324 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9326 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9327 NewIdx = InsertNewInstBefore(Sc, GEP);
9330 // Insert the new GEP instruction.
9332 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9334 Instruction *NewGEP =
9335 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
9336 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9337 // The NewGEP must be pointer typed, so must the old one -> BitCast
9338 return new BitCastInst(NewGEP, GEP.getType());
9347 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9348 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9349 if (AI.isArrayAllocation()) // Check C != 1
9350 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9352 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9353 AllocationInst *New = 0;
9355 // Create and insert the replacement instruction...
9356 if (isa<MallocInst>(AI))
9357 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9359 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9360 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9363 InsertNewInstBefore(New, AI);
9365 // Scan to the end of the allocation instructions, to skip over a block of
9366 // allocas if possible...
9368 BasicBlock::iterator It = New;
9369 while (isa<AllocationInst>(*It)) ++It;
9371 // Now that I is pointing to the first non-allocation-inst in the block,
9372 // insert our getelementptr instruction...
9374 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9378 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9379 New->getName()+".sub", It);
9381 // Now make everything use the getelementptr instead of the original
9383 return ReplaceInstUsesWith(AI, V);
9384 } else if (isa<UndefValue>(AI.getArraySize())) {
9385 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9388 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9389 // Note that we only do this for alloca's, because malloc should allocate and
9390 // return a unique pointer, even for a zero byte allocation.
9391 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9392 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9393 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9398 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9399 Value *Op = FI.getOperand(0);
9401 // free undef -> unreachable.
9402 if (isa<UndefValue>(Op)) {
9403 // Insert a new store to null because we cannot modify the CFG here.
9404 new StoreInst(ConstantInt::getTrue(),
9405 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9406 return EraseInstFromFunction(FI);
9409 // If we have 'free null' delete the instruction. This can happen in stl code
9410 // when lots of inlining happens.
9411 if (isa<ConstantPointerNull>(Op))
9412 return EraseInstFromFunction(FI);
9414 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9415 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9416 FI.setOperand(0, CI->getOperand(0));
9420 // Change free (gep X, 0,0,0,0) into free(X)
9421 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9422 if (GEPI->hasAllZeroIndices()) {
9423 AddToWorkList(GEPI);
9424 FI.setOperand(0, GEPI->getOperand(0));
9429 // Change free(malloc) into nothing, if the malloc has a single use.
9430 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9431 if (MI->hasOneUse()) {
9432 EraseInstFromFunction(FI);
9433 return EraseInstFromFunction(*MI);
9440 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9441 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9442 const TargetData *TD) {
9443 User *CI = cast<User>(LI.getOperand(0));
9444 Value *CastOp = CI->getOperand(0);
9446 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9447 // Instead of loading constant c string, use corresponding integer value
9448 // directly if string length is small enough.
9449 const std::string &Str = CE->getOperand(0)->getStringValue();
9451 unsigned len = Str.length();
9452 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9453 unsigned numBits = Ty->getPrimitiveSizeInBits();
9454 // Replace LI with immediate integer store.
9455 if ((numBits >> 3) == len + 1) {
9456 APInt StrVal(numBits, 0);
9457 APInt SingleChar(numBits, 0);
9458 if (TD->isLittleEndian()) {
9459 for (signed i = len-1; i >= 0; i--) {
9460 SingleChar = (uint64_t) Str[i];
9461 StrVal = (StrVal << 8) | SingleChar;
9464 for (unsigned i = 0; i < len; i++) {
9465 SingleChar = (uint64_t) Str[i];
9466 StrVal = (StrVal << 8) | SingleChar;
9468 // Append NULL at the end.
9470 StrVal = (StrVal << 8) | SingleChar;
9472 Value *NL = ConstantInt::get(StrVal);
9473 return IC.ReplaceInstUsesWith(LI, NL);
9478 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9479 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9480 const Type *SrcPTy = SrcTy->getElementType();
9482 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9483 isa<VectorType>(DestPTy)) {
9484 // If the source is an array, the code below will not succeed. Check to
9485 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9487 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9488 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9489 if (ASrcTy->getNumElements() != 0) {
9491 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9492 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9493 SrcTy = cast<PointerType>(CastOp->getType());
9494 SrcPTy = SrcTy->getElementType();
9497 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9498 isa<VectorType>(SrcPTy)) &&
9499 // Do not allow turning this into a load of an integer, which is then
9500 // casted to a pointer, this pessimizes pointer analysis a lot.
9501 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9502 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9503 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9505 // Okay, we are casting from one integer or pointer type to another of
9506 // the same size. Instead of casting the pointer before the load, cast
9507 // the result of the loaded value.
9508 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9510 LI.isVolatile()),LI);
9511 // Now cast the result of the load.
9512 return new BitCastInst(NewLoad, LI.getType());
9519 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9520 /// from this value cannot trap. If it is not obviously safe to load from the
9521 /// specified pointer, we do a quick local scan of the basic block containing
9522 /// ScanFrom, to determine if the address is already accessed.
9523 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9524 // If it is an alloca it is always safe to load from.
9525 if (isa<AllocaInst>(V)) return true;
9527 // If it is a global variable it is mostly safe to load from.
9528 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9529 // Don't try to evaluate aliases. External weak GV can be null.
9530 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9532 // Otherwise, be a little bit agressive by scanning the local block where we
9533 // want to check to see if the pointer is already being loaded or stored
9534 // from/to. If so, the previous load or store would have already trapped,
9535 // so there is no harm doing an extra load (also, CSE will later eliminate
9536 // the load entirely).
9537 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9542 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9543 if (LI->getOperand(0) == V) return true;
9544 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9545 if (SI->getOperand(1) == V) return true;
9551 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9552 /// until we find the underlying object a pointer is referring to or something
9553 /// we don't understand. Note that the returned pointer may be offset from the
9554 /// input, because we ignore GEP indices.
9555 static Value *GetUnderlyingObject(Value *Ptr) {
9557 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9558 if (CE->getOpcode() == Instruction::BitCast ||
9559 CE->getOpcode() == Instruction::GetElementPtr)
9560 Ptr = CE->getOperand(0);
9563 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9564 Ptr = BCI->getOperand(0);
9565 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9566 Ptr = GEP->getOperand(0);
9573 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9574 Value *Op = LI.getOperand(0);
9576 // Attempt to improve the alignment.
9577 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9578 if (KnownAlign > LI.getAlignment())
9579 LI.setAlignment(KnownAlign);
9581 // load (cast X) --> cast (load X) iff safe
9582 if (isa<CastInst>(Op))
9583 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9586 // None of the following transforms are legal for volatile loads.
9587 if (LI.isVolatile()) return 0;
9589 if (&LI.getParent()->front() != &LI) {
9590 BasicBlock::iterator BBI = &LI; --BBI;
9591 // If the instruction immediately before this is a store to the same
9592 // address, do a simple form of store->load forwarding.
9593 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9594 if (SI->getOperand(1) == LI.getOperand(0))
9595 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9596 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9597 if (LIB->getOperand(0) == LI.getOperand(0))
9598 return ReplaceInstUsesWith(LI, LIB);
9601 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9602 const Value *GEPI0 = GEPI->getOperand(0);
9603 // TODO: Consider a target hook for valid address spaces for this xform.
9604 if (isa<ConstantPointerNull>(GEPI0) &&
9605 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
9606 // Insert a new store to null instruction before the load to indicate
9607 // that this code is not reachable. We do this instead of inserting
9608 // an unreachable instruction directly because we cannot modify the
9610 new StoreInst(UndefValue::get(LI.getType()),
9611 Constant::getNullValue(Op->getType()), &LI);
9612 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9616 if (Constant *C = dyn_cast<Constant>(Op)) {
9617 // load null/undef -> undef
9618 // TODO: Consider a target hook for valid address spaces for this xform.
9619 if (isa<UndefValue>(C) || (C->isNullValue() &&
9620 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
9621 // Insert a new store to null instruction before the load to indicate that
9622 // this code is not reachable. We do this instead of inserting an
9623 // unreachable instruction directly because we cannot modify the CFG.
9624 new StoreInst(UndefValue::get(LI.getType()),
9625 Constant::getNullValue(Op->getType()), &LI);
9626 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9629 // Instcombine load (constant global) into the value loaded.
9630 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9631 if (GV->isConstant() && !GV->isDeclaration())
9632 return ReplaceInstUsesWith(LI, GV->getInitializer());
9634 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9635 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9636 if (CE->getOpcode() == Instruction::GetElementPtr) {
9637 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9638 if (GV->isConstant() && !GV->isDeclaration())
9640 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9641 return ReplaceInstUsesWith(LI, V);
9642 if (CE->getOperand(0)->isNullValue()) {
9643 // Insert a new store to null instruction before the load to indicate
9644 // that this code is not reachable. We do this instead of inserting
9645 // an unreachable instruction directly because we cannot modify the
9647 new StoreInst(UndefValue::get(LI.getType()),
9648 Constant::getNullValue(Op->getType()), &LI);
9649 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9652 } else if (CE->isCast()) {
9653 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9658 // If this load comes from anywhere in a constant global, and if the global
9659 // is all undef or zero, we know what it loads.
9660 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9661 if (GV->isConstant() && GV->hasInitializer()) {
9662 if (GV->getInitializer()->isNullValue())
9663 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9664 else if (isa<UndefValue>(GV->getInitializer()))
9665 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9669 if (Op->hasOneUse()) {
9670 // Change select and PHI nodes to select values instead of addresses: this
9671 // helps alias analysis out a lot, allows many others simplifications, and
9672 // exposes redundancy in the code.
9674 // Note that we cannot do the transformation unless we know that the
9675 // introduced loads cannot trap! Something like this is valid as long as
9676 // the condition is always false: load (select bool %C, int* null, int* %G),
9677 // but it would not be valid if we transformed it to load from null
9680 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9681 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9682 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9683 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9684 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9685 SI->getOperand(1)->getName()+".val"), LI);
9686 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9687 SI->getOperand(2)->getName()+".val"), LI);
9688 return new SelectInst(SI->getCondition(), V1, V2);
9691 // load (select (cond, null, P)) -> load P
9692 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9693 if (C->isNullValue()) {
9694 LI.setOperand(0, SI->getOperand(2));
9698 // load (select (cond, P, null)) -> load P
9699 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9700 if (C->isNullValue()) {
9701 LI.setOperand(0, SI->getOperand(1));
9709 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9711 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9712 User *CI = cast<User>(SI.getOperand(1));
9713 Value *CastOp = CI->getOperand(0);
9715 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9716 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9717 const Type *SrcPTy = SrcTy->getElementType();
9719 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9720 // If the source is an array, the code below will not succeed. Check to
9721 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9723 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9724 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9725 if (ASrcTy->getNumElements() != 0) {
9727 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9728 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9729 SrcTy = cast<PointerType>(CastOp->getType());
9730 SrcPTy = SrcTy->getElementType();
9733 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9734 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9735 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9737 // Okay, we are casting from one integer or pointer type to another of
9738 // the same size. Instead of casting the pointer before
9739 // the store, cast the value to be stored.
9741 Value *SIOp0 = SI.getOperand(0);
9742 Instruction::CastOps opcode = Instruction::BitCast;
9743 const Type* CastSrcTy = SIOp0->getType();
9744 const Type* CastDstTy = SrcPTy;
9745 if (isa<PointerType>(CastDstTy)) {
9746 if (CastSrcTy->isInteger())
9747 opcode = Instruction::IntToPtr;
9748 } else if (isa<IntegerType>(CastDstTy)) {
9749 if (isa<PointerType>(SIOp0->getType()))
9750 opcode = Instruction::PtrToInt;
9752 if (Constant *C = dyn_cast<Constant>(SIOp0))
9753 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9755 NewCast = IC.InsertNewInstBefore(
9756 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9758 return new StoreInst(NewCast, CastOp);
9765 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9766 Value *Val = SI.getOperand(0);
9767 Value *Ptr = SI.getOperand(1);
9769 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9770 EraseInstFromFunction(SI);
9775 // If the RHS is an alloca with a single use, zapify the store, making the
9777 if (Ptr->hasOneUse()) {
9778 if (isa<AllocaInst>(Ptr)) {
9779 EraseInstFromFunction(SI);
9784 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9785 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9786 GEP->getOperand(0)->hasOneUse()) {
9787 EraseInstFromFunction(SI);
9793 // Attempt to improve the alignment.
9794 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9795 if (KnownAlign > SI.getAlignment())
9796 SI.setAlignment(KnownAlign);
9798 // Do really simple DSE, to catch cases where there are several consequtive
9799 // stores to the same location, separated by a few arithmetic operations. This
9800 // situation often occurs with bitfield accesses.
9801 BasicBlock::iterator BBI = &SI;
9802 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9806 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9807 // Prev store isn't volatile, and stores to the same location?
9808 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9811 EraseInstFromFunction(*PrevSI);
9817 // If this is a load, we have to stop. However, if the loaded value is from
9818 // the pointer we're loading and is producing the pointer we're storing,
9819 // then *this* store is dead (X = load P; store X -> P).
9820 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9821 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9822 EraseInstFromFunction(SI);
9826 // Otherwise, this is a load from some other location. Stores before it
9831 // Don't skip over loads or things that can modify memory.
9832 if (BBI->mayWriteToMemory())
9837 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9839 // store X, null -> turns into 'unreachable' in SimplifyCFG
9840 if (isa<ConstantPointerNull>(Ptr)) {
9841 if (!isa<UndefValue>(Val)) {
9842 SI.setOperand(0, UndefValue::get(Val->getType()));
9843 if (Instruction *U = dyn_cast<Instruction>(Val))
9844 AddToWorkList(U); // Dropped a use.
9847 return 0; // Do not modify these!
9850 // store undef, Ptr -> noop
9851 if (isa<UndefValue>(Val)) {
9852 EraseInstFromFunction(SI);
9857 // If the pointer destination is a cast, see if we can fold the cast into the
9859 if (isa<CastInst>(Ptr))
9860 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9862 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9864 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9868 // If this store is the last instruction in the basic block, and if the block
9869 // ends with an unconditional branch, try to move it to the successor block.
9871 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9872 if (BI->isUnconditional())
9873 if (SimplifyStoreAtEndOfBlock(SI))
9874 return 0; // xform done!
9879 /// SimplifyStoreAtEndOfBlock - Turn things like:
9880 /// if () { *P = v1; } else { *P = v2 }
9881 /// into a phi node with a store in the successor.
9883 /// Simplify things like:
9884 /// *P = v1; if () { *P = v2; }
9885 /// into a phi node with a store in the successor.
9887 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9888 BasicBlock *StoreBB = SI.getParent();
9890 // Check to see if the successor block has exactly two incoming edges. If
9891 // so, see if the other predecessor contains a store to the same location.
9892 // if so, insert a PHI node (if needed) and move the stores down.
9893 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9895 // Determine whether Dest has exactly two predecessors and, if so, compute
9896 // the other predecessor.
9897 pred_iterator PI = pred_begin(DestBB);
9898 BasicBlock *OtherBB = 0;
9902 if (PI == pred_end(DestBB))
9905 if (*PI != StoreBB) {
9910 if (++PI != pred_end(DestBB))
9914 // Verify that the other block ends in a branch and is not otherwise empty.
9915 BasicBlock::iterator BBI = OtherBB->getTerminator();
9916 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9917 if (!OtherBr || BBI == OtherBB->begin())
9920 // If the other block ends in an unconditional branch, check for the 'if then
9921 // else' case. there is an instruction before the branch.
9922 StoreInst *OtherStore = 0;
9923 if (OtherBr->isUnconditional()) {
9924 // If this isn't a store, or isn't a store to the same location, bail out.
9926 OtherStore = dyn_cast<StoreInst>(BBI);
9927 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9930 // Otherwise, the other block ended with a conditional branch. If one of the
9931 // destinations is StoreBB, then we have the if/then case.
9932 if (OtherBr->getSuccessor(0) != StoreBB &&
9933 OtherBr->getSuccessor(1) != StoreBB)
9936 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9937 // if/then triangle. See if there is a store to the same ptr as SI that
9938 // lives in OtherBB.
9940 // Check to see if we find the matching store.
9941 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9942 if (OtherStore->getOperand(1) != SI.getOperand(1))
9946 // If we find something that may be using the stored value, or if we run
9947 // out of instructions, we can't do the xform.
9948 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9949 BBI == OtherBB->begin())
9953 // In order to eliminate the store in OtherBr, we have to
9954 // make sure nothing reads the stored value in StoreBB.
9955 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9956 // FIXME: This should really be AA driven.
9957 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9962 // Insert a PHI node now if we need it.
9963 Value *MergedVal = OtherStore->getOperand(0);
9964 if (MergedVal != SI.getOperand(0)) {
9965 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9966 PN->reserveOperandSpace(2);
9967 PN->addIncoming(SI.getOperand(0), SI.getParent());
9968 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9969 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9972 // Advance to a place where it is safe to insert the new store and
9974 BBI = DestBB->begin();
9975 while (isa<PHINode>(BBI)) ++BBI;
9976 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9977 OtherStore->isVolatile()), *BBI);
9979 // Nuke the old stores.
9980 EraseInstFromFunction(SI);
9981 EraseInstFromFunction(*OtherStore);
9987 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9988 // Change br (not X), label True, label False to: br X, label False, True
9990 BasicBlock *TrueDest;
9991 BasicBlock *FalseDest;
9992 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9993 !isa<Constant>(X)) {
9994 // Swap Destinations and condition...
9996 BI.setSuccessor(0, FalseDest);
9997 BI.setSuccessor(1, TrueDest);
10001 // Cannonicalize fcmp_one -> fcmp_oeq
10002 FCmpInst::Predicate FPred; Value *Y;
10003 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10004 TrueDest, FalseDest)))
10005 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10006 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10007 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10008 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10009 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10010 NewSCC->takeName(I);
10011 // Swap Destinations and condition...
10012 BI.setCondition(NewSCC);
10013 BI.setSuccessor(0, FalseDest);
10014 BI.setSuccessor(1, TrueDest);
10015 RemoveFromWorkList(I);
10016 I->eraseFromParent();
10017 AddToWorkList(NewSCC);
10021 // Cannonicalize icmp_ne -> icmp_eq
10022 ICmpInst::Predicate IPred;
10023 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10024 TrueDest, FalseDest)))
10025 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10026 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10027 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10028 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10029 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10030 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10031 NewSCC->takeName(I);
10032 // Swap Destinations and condition...
10033 BI.setCondition(NewSCC);
10034 BI.setSuccessor(0, FalseDest);
10035 BI.setSuccessor(1, TrueDest);
10036 RemoveFromWorkList(I);
10037 I->eraseFromParent();;
10038 AddToWorkList(NewSCC);
10045 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10046 Value *Cond = SI.getCondition();
10047 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10048 if (I->getOpcode() == Instruction::Add)
10049 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10050 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10051 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10052 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10054 SI.setOperand(0, I->getOperand(0));
10062 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10063 /// is to leave as a vector operation.
10064 static bool CheapToScalarize(Value *V, bool isConstant) {
10065 if (isa<ConstantAggregateZero>(V))
10067 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10068 if (isConstant) return true;
10069 // If all elts are the same, we can extract.
10070 Constant *Op0 = C->getOperand(0);
10071 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10072 if (C->getOperand(i) != Op0)
10076 Instruction *I = dyn_cast<Instruction>(V);
10077 if (!I) return false;
10079 // Insert element gets simplified to the inserted element or is deleted if
10080 // this is constant idx extract element and its a constant idx insertelt.
10081 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10082 isa<ConstantInt>(I->getOperand(2)))
10084 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10086 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10087 if (BO->hasOneUse() &&
10088 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10089 CheapToScalarize(BO->getOperand(1), isConstant)))
10091 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10092 if (CI->hasOneUse() &&
10093 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10094 CheapToScalarize(CI->getOperand(1), isConstant)))
10100 /// Read and decode a shufflevector mask.
10102 /// It turns undef elements into values that are larger than the number of
10103 /// elements in the input.
10104 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10105 unsigned NElts = SVI->getType()->getNumElements();
10106 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10107 return std::vector<unsigned>(NElts, 0);
10108 if (isa<UndefValue>(SVI->getOperand(2)))
10109 return std::vector<unsigned>(NElts, 2*NElts);
10111 std::vector<unsigned> Result;
10112 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10113 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
10114 if (isa<UndefValue>(CP->getOperand(i)))
10115 Result.push_back(NElts*2); // undef -> 8
10117 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
10121 /// FindScalarElement - Given a vector and an element number, see if the scalar
10122 /// value is already around as a register, for example if it were inserted then
10123 /// extracted from the vector.
10124 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10125 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10126 const VectorType *PTy = cast<VectorType>(V->getType());
10127 unsigned Width = PTy->getNumElements();
10128 if (EltNo >= Width) // Out of range access.
10129 return UndefValue::get(PTy->getElementType());
10131 if (isa<UndefValue>(V))
10132 return UndefValue::get(PTy->getElementType());
10133 else if (isa<ConstantAggregateZero>(V))
10134 return Constant::getNullValue(PTy->getElementType());
10135 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10136 return CP->getOperand(EltNo);
10137 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10138 // If this is an insert to a variable element, we don't know what it is.
10139 if (!isa<ConstantInt>(III->getOperand(2)))
10141 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10143 // If this is an insert to the element we are looking for, return the
10145 if (EltNo == IIElt)
10146 return III->getOperand(1);
10148 // Otherwise, the insertelement doesn't modify the value, recurse on its
10150 return FindScalarElement(III->getOperand(0), EltNo);
10151 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10152 unsigned InEl = getShuffleMask(SVI)[EltNo];
10154 return FindScalarElement(SVI->getOperand(0), InEl);
10155 else if (InEl < Width*2)
10156 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10158 return UndefValue::get(PTy->getElementType());
10161 // Otherwise, we don't know.
10165 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10167 // If vector val is undef, replace extract with scalar undef.
10168 if (isa<UndefValue>(EI.getOperand(0)))
10169 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10171 // If vector val is constant 0, replace extract with scalar 0.
10172 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10173 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10175 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10176 // If vector val is constant with uniform operands, replace EI
10177 // with that operand
10178 Constant *op0 = C->getOperand(0);
10179 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10180 if (C->getOperand(i) != op0) {
10185 return ReplaceInstUsesWith(EI, op0);
10188 // If extracting a specified index from the vector, see if we can recursively
10189 // find a previously computed scalar that was inserted into the vector.
10190 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10191 unsigned IndexVal = IdxC->getZExtValue();
10192 unsigned VectorWidth =
10193 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10195 // If this is extracting an invalid index, turn this into undef, to avoid
10196 // crashing the code below.
10197 if (IndexVal >= VectorWidth)
10198 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10200 // This instruction only demands the single element from the input vector.
10201 // If the input vector has a single use, simplify it based on this use
10203 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10204 uint64_t UndefElts;
10205 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10208 EI.setOperand(0, V);
10213 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10214 return ReplaceInstUsesWith(EI, Elt);
10216 // If the this extractelement is directly using a bitcast from a vector of
10217 // the same number of elements, see if we can find the source element from
10218 // it. In this case, we will end up needing to bitcast the scalars.
10219 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10220 if (const VectorType *VT =
10221 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10222 if (VT->getNumElements() == VectorWidth)
10223 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10224 return new BitCastInst(Elt, EI.getType());
10228 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10229 if (I->hasOneUse()) {
10230 // Push extractelement into predecessor operation if legal and
10231 // profitable to do so
10232 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10233 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10234 if (CheapToScalarize(BO, isConstantElt)) {
10235 ExtractElementInst *newEI0 =
10236 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10237 EI.getName()+".lhs");
10238 ExtractElementInst *newEI1 =
10239 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10240 EI.getName()+".rhs");
10241 InsertNewInstBefore(newEI0, EI);
10242 InsertNewInstBefore(newEI1, EI);
10243 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10245 } else if (isa<LoadInst>(I)) {
10247 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10248 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10249 PointerType::get(EI.getType(), AS),EI);
10250 GetElementPtrInst *GEP =
10251 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
10252 InsertNewInstBefore(GEP, EI);
10253 return new LoadInst(GEP);
10256 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10257 // Extracting the inserted element?
10258 if (IE->getOperand(2) == EI.getOperand(1))
10259 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10260 // If the inserted and extracted elements are constants, they must not
10261 // be the same value, extract from the pre-inserted value instead.
10262 if (isa<Constant>(IE->getOperand(2)) &&
10263 isa<Constant>(EI.getOperand(1))) {
10264 AddUsesToWorkList(EI);
10265 EI.setOperand(0, IE->getOperand(0));
10268 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10269 // If this is extracting an element from a shufflevector, figure out where
10270 // it came from and extract from the appropriate input element instead.
10271 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10272 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10274 if (SrcIdx < SVI->getType()->getNumElements())
10275 Src = SVI->getOperand(0);
10276 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10277 SrcIdx -= SVI->getType()->getNumElements();
10278 Src = SVI->getOperand(1);
10280 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10282 return new ExtractElementInst(Src, SrcIdx);
10289 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10290 /// elements from either LHS or RHS, return the shuffle mask and true.
10291 /// Otherwise, return false.
10292 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10293 std::vector<Constant*> &Mask) {
10294 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10295 "Invalid CollectSingleShuffleElements");
10296 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10298 if (isa<UndefValue>(V)) {
10299 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10301 } else if (V == LHS) {
10302 for (unsigned i = 0; i != NumElts; ++i)
10303 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10305 } else if (V == RHS) {
10306 for (unsigned i = 0; i != NumElts; ++i)
10307 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10309 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10310 // If this is an insert of an extract from some other vector, include it.
10311 Value *VecOp = IEI->getOperand(0);
10312 Value *ScalarOp = IEI->getOperand(1);
10313 Value *IdxOp = IEI->getOperand(2);
10315 if (!isa<ConstantInt>(IdxOp))
10317 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10319 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10320 // Okay, we can handle this if the vector we are insertinting into is
10321 // transitively ok.
10322 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10323 // If so, update the mask to reflect the inserted undef.
10324 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10327 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10328 if (isa<ConstantInt>(EI->getOperand(1)) &&
10329 EI->getOperand(0)->getType() == V->getType()) {
10330 unsigned ExtractedIdx =
10331 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10333 // This must be extracting from either LHS or RHS.
10334 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10335 // Okay, we can handle this if the vector we are insertinting into is
10336 // transitively ok.
10337 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10338 // If so, update the mask to reflect the inserted value.
10339 if (EI->getOperand(0) == LHS) {
10340 Mask[InsertedIdx & (NumElts-1)] =
10341 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10343 assert(EI->getOperand(0) == RHS);
10344 Mask[InsertedIdx & (NumElts-1)] =
10345 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10354 // TODO: Handle shufflevector here!
10359 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10360 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10361 /// that computes V and the LHS value of the shuffle.
10362 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10364 assert(isa<VectorType>(V->getType()) &&
10365 (RHS == 0 || V->getType() == RHS->getType()) &&
10366 "Invalid shuffle!");
10367 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10369 if (isa<UndefValue>(V)) {
10370 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10372 } else if (isa<ConstantAggregateZero>(V)) {
10373 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10375 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10376 // If this is an insert of an extract from some other vector, include it.
10377 Value *VecOp = IEI->getOperand(0);
10378 Value *ScalarOp = IEI->getOperand(1);
10379 Value *IdxOp = IEI->getOperand(2);
10381 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10382 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10383 EI->getOperand(0)->getType() == V->getType()) {
10384 unsigned ExtractedIdx =
10385 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10386 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10388 // Either the extracted from or inserted into vector must be RHSVec,
10389 // otherwise we'd end up with a shuffle of three inputs.
10390 if (EI->getOperand(0) == RHS || RHS == 0) {
10391 RHS = EI->getOperand(0);
10392 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10393 Mask[InsertedIdx & (NumElts-1)] =
10394 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10398 if (VecOp == RHS) {
10399 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10400 // Everything but the extracted element is replaced with the RHS.
10401 for (unsigned i = 0; i != NumElts; ++i) {
10402 if (i != InsertedIdx)
10403 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10408 // If this insertelement is a chain that comes from exactly these two
10409 // vectors, return the vector and the effective shuffle.
10410 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10411 return EI->getOperand(0);
10416 // TODO: Handle shufflevector here!
10418 // Otherwise, can't do anything fancy. Return an identity vector.
10419 for (unsigned i = 0; i != NumElts; ++i)
10420 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10424 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10425 Value *VecOp = IE.getOperand(0);
10426 Value *ScalarOp = IE.getOperand(1);
10427 Value *IdxOp = IE.getOperand(2);
10429 // Inserting an undef or into an undefined place, remove this.
10430 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10431 ReplaceInstUsesWith(IE, VecOp);
10433 // If the inserted element was extracted from some other vector, and if the
10434 // indexes are constant, try to turn this into a shufflevector operation.
10435 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10436 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10437 EI->getOperand(0)->getType() == IE.getType()) {
10438 unsigned NumVectorElts = IE.getType()->getNumElements();
10439 unsigned ExtractedIdx =
10440 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10441 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10443 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10444 return ReplaceInstUsesWith(IE, VecOp);
10446 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10447 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10449 // If we are extracting a value from a vector, then inserting it right
10450 // back into the same place, just use the input vector.
10451 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10452 return ReplaceInstUsesWith(IE, VecOp);
10454 // We could theoretically do this for ANY input. However, doing so could
10455 // turn chains of insertelement instructions into a chain of shufflevector
10456 // instructions, and right now we do not merge shufflevectors. As such,
10457 // only do this in a situation where it is clear that there is benefit.
10458 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10459 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10460 // the values of VecOp, except then one read from EIOp0.
10461 // Build a new shuffle mask.
10462 std::vector<Constant*> Mask;
10463 if (isa<UndefValue>(VecOp))
10464 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10466 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10467 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10470 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10471 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10472 ConstantVector::get(Mask));
10475 // If this insertelement isn't used by some other insertelement, turn it
10476 // (and any insertelements it points to), into one big shuffle.
10477 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10478 std::vector<Constant*> Mask;
10480 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10481 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10482 // We now have a shuffle of LHS, RHS, Mask.
10483 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10492 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10493 Value *LHS = SVI.getOperand(0);
10494 Value *RHS = SVI.getOperand(1);
10495 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10497 bool MadeChange = false;
10499 // Undefined shuffle mask -> undefined value.
10500 if (isa<UndefValue>(SVI.getOperand(2)))
10501 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10503 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10504 // the undef, change them to undefs.
10505 if (isa<UndefValue>(SVI.getOperand(1))) {
10506 // Scan to see if there are any references to the RHS. If so, replace them
10507 // with undef element refs and set MadeChange to true.
10508 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10509 if (Mask[i] >= e && Mask[i] != 2*e) {
10516 // Remap any references to RHS to use LHS.
10517 std::vector<Constant*> Elts;
10518 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10519 if (Mask[i] == 2*e)
10520 Elts.push_back(UndefValue::get(Type::Int32Ty));
10522 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10524 SVI.setOperand(2, ConstantVector::get(Elts));
10528 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10529 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10530 if (LHS == RHS || isa<UndefValue>(LHS)) {
10531 if (isa<UndefValue>(LHS) && LHS == RHS) {
10532 // shuffle(undef,undef,mask) -> undef.
10533 return ReplaceInstUsesWith(SVI, LHS);
10536 // Remap any references to RHS to use LHS.
10537 std::vector<Constant*> Elts;
10538 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10539 if (Mask[i] >= 2*e)
10540 Elts.push_back(UndefValue::get(Type::Int32Ty));
10542 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10543 (Mask[i] < e && isa<UndefValue>(LHS)))
10544 Mask[i] = 2*e; // Turn into undef.
10546 Mask[i] &= (e-1); // Force to LHS.
10547 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10550 SVI.setOperand(0, SVI.getOperand(1));
10551 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10552 SVI.setOperand(2, ConstantVector::get(Elts));
10553 LHS = SVI.getOperand(0);
10554 RHS = SVI.getOperand(1);
10558 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10559 bool isLHSID = true, isRHSID = true;
10561 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10562 if (Mask[i] >= e*2) continue; // Ignore undef values.
10563 // Is this an identity shuffle of the LHS value?
10564 isLHSID &= (Mask[i] == i);
10566 // Is this an identity shuffle of the RHS value?
10567 isRHSID &= (Mask[i]-e == i);
10570 // Eliminate identity shuffles.
10571 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10572 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10574 // If the LHS is a shufflevector itself, see if we can combine it with this
10575 // one without producing an unusual shuffle. Here we are really conservative:
10576 // we are absolutely afraid of producing a shuffle mask not in the input
10577 // program, because the code gen may not be smart enough to turn a merged
10578 // shuffle into two specific shuffles: it may produce worse code. As such,
10579 // we only merge two shuffles if the result is one of the two input shuffle
10580 // masks. In this case, merging the shuffles just removes one instruction,
10581 // which we know is safe. This is good for things like turning:
10582 // (splat(splat)) -> splat.
10583 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10584 if (isa<UndefValue>(RHS)) {
10585 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10587 std::vector<unsigned> NewMask;
10588 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10589 if (Mask[i] >= 2*e)
10590 NewMask.push_back(2*e);
10592 NewMask.push_back(LHSMask[Mask[i]]);
10594 // If the result mask is equal to the src shuffle or this shuffle mask, do
10595 // the replacement.
10596 if (NewMask == LHSMask || NewMask == Mask) {
10597 std::vector<Constant*> Elts;
10598 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10599 if (NewMask[i] >= e*2) {
10600 Elts.push_back(UndefValue::get(Type::Int32Ty));
10602 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10605 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10606 LHSSVI->getOperand(1),
10607 ConstantVector::get(Elts));
10612 return MadeChange ? &SVI : 0;
10618 /// TryToSinkInstruction - Try to move the specified instruction from its
10619 /// current block into the beginning of DestBlock, which can only happen if it's
10620 /// safe to move the instruction past all of the instructions between it and the
10621 /// end of its block.
10622 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10623 assert(I->hasOneUse() && "Invariants didn't hold!");
10625 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10626 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10628 // Do not sink alloca instructions out of the entry block.
10629 if (isa<AllocaInst>(I) && I->getParent() ==
10630 &DestBlock->getParent()->getEntryBlock())
10633 // We can only sink load instructions if there is nothing between the load and
10634 // the end of block that could change the value.
10635 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10636 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10638 if (Scan->mayWriteToMemory())
10642 BasicBlock::iterator InsertPos = DestBlock->begin();
10643 while (isa<PHINode>(InsertPos)) ++InsertPos;
10645 I->moveBefore(InsertPos);
10651 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10652 /// all reachable code to the worklist.
10654 /// This has a couple of tricks to make the code faster and more powerful. In
10655 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10656 /// them to the worklist (this significantly speeds up instcombine on code where
10657 /// many instructions are dead or constant). Additionally, if we find a branch
10658 /// whose condition is a known constant, we only visit the reachable successors.
10660 static void AddReachableCodeToWorklist(BasicBlock *BB,
10661 SmallPtrSet<BasicBlock*, 64> &Visited,
10663 const TargetData *TD) {
10664 std::vector<BasicBlock*> Worklist;
10665 Worklist.push_back(BB);
10667 while (!Worklist.empty()) {
10668 BB = Worklist.back();
10669 Worklist.pop_back();
10671 // We have now visited this block! If we've already been here, ignore it.
10672 if (!Visited.insert(BB)) continue;
10674 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10675 Instruction *Inst = BBI++;
10677 // DCE instruction if trivially dead.
10678 if (isInstructionTriviallyDead(Inst)) {
10680 DOUT << "IC: DCE: " << *Inst;
10681 Inst->eraseFromParent();
10685 // ConstantProp instruction if trivially constant.
10686 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10687 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10688 Inst->replaceAllUsesWith(C);
10690 Inst->eraseFromParent();
10694 IC.AddToWorkList(Inst);
10697 // Recursively visit successors. If this is a branch or switch on a
10698 // constant, only visit the reachable successor.
10699 TerminatorInst *TI = BB->getTerminator();
10700 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10701 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10702 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10703 Worklist.push_back(BI->getSuccessor(!CondVal));
10706 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10707 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10708 // See if this is an explicit destination.
10709 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10710 if (SI->getCaseValue(i) == Cond) {
10711 Worklist.push_back(SI->getSuccessor(i));
10715 // Otherwise it is the default destination.
10716 Worklist.push_back(SI->getSuccessor(0));
10721 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10722 Worklist.push_back(TI->getSuccessor(i));
10726 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10727 bool Changed = false;
10728 TD = &getAnalysis<TargetData>();
10730 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10731 << F.getNameStr() << "\n");
10734 // Do a depth-first traversal of the function, populate the worklist with
10735 // the reachable instructions. Ignore blocks that are not reachable. Keep
10736 // track of which blocks we visit.
10737 SmallPtrSet<BasicBlock*, 64> Visited;
10738 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10740 // Do a quick scan over the function. If we find any blocks that are
10741 // unreachable, remove any instructions inside of them. This prevents
10742 // the instcombine code from having to deal with some bad special cases.
10743 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10744 if (!Visited.count(BB)) {
10745 Instruction *Term = BB->getTerminator();
10746 while (Term != BB->begin()) { // Remove instrs bottom-up
10747 BasicBlock::iterator I = Term; --I;
10749 DOUT << "IC: DCE: " << *I;
10752 if (!I->use_empty())
10753 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10754 I->eraseFromParent();
10759 while (!Worklist.empty()) {
10760 Instruction *I = RemoveOneFromWorkList();
10761 if (I == 0) continue; // skip null values.
10763 // Check to see if we can DCE the instruction.
10764 if (isInstructionTriviallyDead(I)) {
10765 // Add operands to the worklist.
10766 if (I->getNumOperands() < 4)
10767 AddUsesToWorkList(*I);
10770 DOUT << "IC: DCE: " << *I;
10772 I->eraseFromParent();
10773 RemoveFromWorkList(I);
10777 // Instruction isn't dead, see if we can constant propagate it.
10778 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10779 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10781 // Add operands to the worklist.
10782 AddUsesToWorkList(*I);
10783 ReplaceInstUsesWith(*I, C);
10786 I->eraseFromParent();
10787 RemoveFromWorkList(I);
10791 // See if we can trivially sink this instruction to a successor basic block.
10792 if (I->hasOneUse()) {
10793 BasicBlock *BB = I->getParent();
10794 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10795 if (UserParent != BB) {
10796 bool UserIsSuccessor = false;
10797 // See if the user is one of our successors.
10798 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10799 if (*SI == UserParent) {
10800 UserIsSuccessor = true;
10804 // If the user is one of our immediate successors, and if that successor
10805 // only has us as a predecessors (we'd have to split the critical edge
10806 // otherwise), we can keep going.
10807 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10808 next(pred_begin(UserParent)) == pred_end(UserParent))
10809 // Okay, the CFG is simple enough, try to sink this instruction.
10810 Changed |= TryToSinkInstruction(I, UserParent);
10814 // Now that we have an instruction, try combining it to simplify it...
10818 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10819 if (Instruction *Result = visit(*I)) {
10821 // Should we replace the old instruction with a new one?
10823 DOUT << "IC: Old = " << *I
10824 << " New = " << *Result;
10826 // Everything uses the new instruction now.
10827 I->replaceAllUsesWith(Result);
10829 // Push the new instruction and any users onto the worklist.
10830 AddToWorkList(Result);
10831 AddUsersToWorkList(*Result);
10833 // Move the name to the new instruction first.
10834 Result->takeName(I);
10836 // Insert the new instruction into the basic block...
10837 BasicBlock *InstParent = I->getParent();
10838 BasicBlock::iterator InsertPos = I;
10840 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10841 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10844 InstParent->getInstList().insert(InsertPos, Result);
10846 // Make sure that we reprocess all operands now that we reduced their
10848 AddUsesToWorkList(*I);
10850 // Instructions can end up on the worklist more than once. Make sure
10851 // we do not process an instruction that has been deleted.
10852 RemoveFromWorkList(I);
10854 // Erase the old instruction.
10855 InstParent->getInstList().erase(I);
10858 DOUT << "IC: Mod = " << OrigI
10859 << " New = " << *I;
10862 // If the instruction was modified, it's possible that it is now dead.
10863 // if so, remove it.
10864 if (isInstructionTriviallyDead(I)) {
10865 // Make sure we process all operands now that we are reducing their
10867 AddUsesToWorkList(*I);
10869 // Instructions may end up in the worklist more than once. Erase all
10870 // occurrences of this instruction.
10871 RemoveFromWorkList(I);
10872 I->eraseFromParent();
10875 AddUsersToWorkList(*I);
10882 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10884 // Do an explicit clear, this shrinks the map if needed.
10885 WorklistMap.clear();
10890 bool InstCombiner::runOnFunction(Function &F) {
10891 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10893 bool EverMadeChange = false;
10895 // Iterate while there is work to do.
10896 unsigned Iteration = 0;
10897 while (DoOneIteration(F, Iteration++))
10898 EverMadeChange = true;
10899 return EverMadeChange;
10902 FunctionPass *llvm::createInstructionCombiningPass() {
10903 return new InstCombiner();