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 if (Value *V = dyn_castNegVal(LHS))
2094 return BinaryOperator::createSub(RHS, V);
2097 if (!isa<Constant>(RHS))
2098 if (Value *V = dyn_castNegVal(RHS))
2099 return BinaryOperator::createSub(LHS, V);
2103 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2104 if (X == RHS) // X*C + X --> X * (C+1)
2105 return BinaryOperator::createMul(RHS, AddOne(C2));
2107 // X*C1 + X*C2 --> X * (C1+C2)
2109 if (X == dyn_castFoldableMul(RHS, C1))
2110 return BinaryOperator::createMul(X, Add(C1, C2));
2113 // X + X*C --> X * (C+1)
2114 if (dyn_castFoldableMul(RHS, C2) == LHS)
2115 return BinaryOperator::createMul(LHS, AddOne(C2));
2117 // X + ~X --> -1 since ~X = -X-1
2118 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2119 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2122 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2123 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2124 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2127 // W*X + Y*Z --> W * (X+Z) iff W == Y
2128 if (I.getType()->isIntOrIntVector()) {
2129 Value *W, *X, *Y, *Z;
2130 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2131 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2135 } else if (Y == X) {
2137 } else if (X == Z) {
2144 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, Z,
2145 LHS->getName()), I);
2146 return BinaryOperator::createMul(W, NewAdd);
2151 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2153 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2154 return BinaryOperator::createSub(SubOne(CRHS), X);
2156 // (X & FF00) + xx00 -> (X+xx00) & FF00
2157 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2158 Constant *Anded = And(CRHS, C2);
2159 if (Anded == CRHS) {
2160 // See if all bits from the first bit set in the Add RHS up are included
2161 // in the mask. First, get the rightmost bit.
2162 const APInt& AddRHSV = CRHS->getValue();
2164 // Form a mask of all bits from the lowest bit added through the top.
2165 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2167 // See if the and mask includes all of these bits.
2168 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2170 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2171 // Okay, the xform is safe. Insert the new add pronto.
2172 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2173 LHS->getName()), I);
2174 return BinaryOperator::createAnd(NewAdd, C2);
2179 // Try to fold constant add into select arguments.
2180 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2181 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2185 // add (cast *A to intptrtype) B ->
2186 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2188 CastInst *CI = dyn_cast<CastInst>(LHS);
2191 CI = dyn_cast<CastInst>(RHS);
2194 if (CI && CI->getType()->isSized() &&
2195 (CI->getType()->getPrimitiveSizeInBits() ==
2196 TD->getIntPtrType()->getPrimitiveSizeInBits())
2197 && isa<PointerType>(CI->getOperand(0)->getType())) {
2199 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2200 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2201 PointerType::get(Type::Int8Ty, AS), I);
2202 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2203 return new PtrToIntInst(I2, CI->getType());
2207 // add (select X 0 (sub n A)) A --> select X A n
2209 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2212 SI = dyn_cast<SelectInst>(RHS);
2215 if (SI && SI->hasOneUse()) {
2216 Value *TV = SI->getTrueValue();
2217 Value *FV = SI->getFalseValue();
2220 // Can we fold the add into the argument of the select?
2221 // We check both true and false select arguments for a matching subtract.
2222 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2223 A == Other) // Fold the add into the true select value.
2224 return new SelectInst(SI->getCondition(), N, A);
2225 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2226 A == Other) // Fold the add into the false select value.
2227 return new SelectInst(SI->getCondition(), A, N);
2231 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2232 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2233 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2234 return ReplaceInstUsesWith(I, LHS);
2236 return Changed ? &I : 0;
2239 // isSignBit - Return true if the value represented by the constant only has the
2240 // highest order bit set.
2241 static bool isSignBit(ConstantInt *CI) {
2242 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2243 return CI->getValue() == APInt::getSignBit(NumBits);
2246 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2247 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2249 if (Op0 == Op1) // sub X, X -> 0
2250 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2252 // If this is a 'B = x-(-A)', change to B = x+A...
2253 if (Value *V = dyn_castNegVal(Op1))
2254 return BinaryOperator::createAdd(Op0, V);
2256 if (isa<UndefValue>(Op0))
2257 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2258 if (isa<UndefValue>(Op1))
2259 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2261 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2262 // Replace (-1 - A) with (~A)...
2263 if (C->isAllOnesValue())
2264 return BinaryOperator::createNot(Op1);
2266 // C - ~X == X + (1+C)
2268 if (match(Op1, m_Not(m_Value(X))))
2269 return BinaryOperator::createAdd(X, AddOne(C));
2271 // -(X >>u 31) -> (X >>s 31)
2272 // -(X >>s 31) -> (X >>u 31)
2274 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2275 if (SI->getOpcode() == Instruction::LShr) {
2276 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2277 // Check to see if we are shifting out everything but the sign bit.
2278 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2279 SI->getType()->getPrimitiveSizeInBits()-1) {
2280 // Ok, the transformation is safe. Insert AShr.
2281 return BinaryOperator::create(Instruction::AShr,
2282 SI->getOperand(0), CU, SI->getName());
2286 else if (SI->getOpcode() == Instruction::AShr) {
2287 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2288 // Check to see if we are shifting out everything but the sign bit.
2289 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2290 SI->getType()->getPrimitiveSizeInBits()-1) {
2291 // Ok, the transformation is safe. Insert LShr.
2292 return BinaryOperator::createLShr(
2293 SI->getOperand(0), CU, SI->getName());
2299 // Try to fold constant sub into select arguments.
2300 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2301 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2304 if (isa<PHINode>(Op0))
2305 if (Instruction *NV = FoldOpIntoPhi(I))
2309 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2310 if (Op1I->getOpcode() == Instruction::Add &&
2311 !Op0->getType()->isFPOrFPVector()) {
2312 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2313 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2314 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2315 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2316 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2317 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2318 // C1-(X+C2) --> (C1-C2)-X
2319 return BinaryOperator::createSub(Subtract(CI1, CI2),
2320 Op1I->getOperand(0));
2324 if (Op1I->hasOneUse()) {
2325 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2326 // is not used by anyone else...
2328 if (Op1I->getOpcode() == Instruction::Sub &&
2329 !Op1I->getType()->isFPOrFPVector()) {
2330 // Swap the two operands of the subexpr...
2331 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2332 Op1I->setOperand(0, IIOp1);
2333 Op1I->setOperand(1, IIOp0);
2335 // Create the new top level add instruction...
2336 return BinaryOperator::createAdd(Op0, Op1);
2339 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2341 if (Op1I->getOpcode() == Instruction::And &&
2342 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2343 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2346 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2347 return BinaryOperator::createAnd(Op0, NewNot);
2350 // 0 - (X sdiv C) -> (X sdiv -C)
2351 if (Op1I->getOpcode() == Instruction::SDiv)
2352 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2354 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2355 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2356 ConstantExpr::getNeg(DivRHS));
2358 // X - X*C --> X * (1-C)
2359 ConstantInt *C2 = 0;
2360 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2361 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2362 return BinaryOperator::createMul(Op0, CP1);
2365 // X - ((X / Y) * Y) --> X % Y
2366 if (Op1I->getOpcode() == Instruction::Mul)
2367 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2368 if (Op0 == I->getOperand(0) &&
2369 Op1I->getOperand(1) == I->getOperand(1)) {
2370 if (I->getOpcode() == Instruction::SDiv)
2371 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2372 if (I->getOpcode() == Instruction::UDiv)
2373 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2378 if (!Op0->getType()->isFPOrFPVector())
2379 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2380 if (Op0I->getOpcode() == Instruction::Add) {
2381 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2382 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2383 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2384 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2385 } else if (Op0I->getOpcode() == Instruction::Sub) {
2386 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2387 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2391 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2392 if (X == Op1) // X*C - X --> X * (C-1)
2393 return BinaryOperator::createMul(Op1, SubOne(C1));
2395 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2396 if (X == dyn_castFoldableMul(Op1, C2))
2397 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2402 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2403 /// comparison only checks the sign bit. If it only checks the sign bit, set
2404 /// TrueIfSigned if the result of the comparison is true when the input value is
2406 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2407 bool &TrueIfSigned) {
2409 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2410 TrueIfSigned = true;
2411 return RHS->isZero();
2412 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2413 TrueIfSigned = true;
2414 return RHS->isAllOnesValue();
2415 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2416 TrueIfSigned = false;
2417 return RHS->isAllOnesValue();
2418 case ICmpInst::ICMP_UGT:
2419 // True if LHS u> RHS and RHS == high-bit-mask - 1
2420 TrueIfSigned = true;
2421 return RHS->getValue() ==
2422 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2423 case ICmpInst::ICMP_UGE:
2424 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2425 TrueIfSigned = true;
2426 return RHS->getValue() ==
2427 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2433 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2434 bool Changed = SimplifyCommutative(I);
2435 Value *Op0 = I.getOperand(0);
2437 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2438 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2440 // Simplify mul instructions with a constant RHS...
2441 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2442 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2444 // ((X << C1)*C2) == (X * (C2 << C1))
2445 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2446 if (SI->getOpcode() == Instruction::Shl)
2447 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2448 return BinaryOperator::createMul(SI->getOperand(0),
2449 ConstantExpr::getShl(CI, ShOp));
2452 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2453 if (CI->equalsInt(1)) // X * 1 == X
2454 return ReplaceInstUsesWith(I, Op0);
2455 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2456 return BinaryOperator::createNeg(Op0, I.getName());
2458 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2459 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2460 return BinaryOperator::createShl(Op0,
2461 ConstantInt::get(Op0->getType(), Val.logBase2()));
2463 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2464 if (Op1F->isNullValue())
2465 return ReplaceInstUsesWith(I, Op1);
2467 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2468 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2469 // We need a better interface for long double here.
2470 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2471 if (Op1F->isExactlyValue(1.0))
2472 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2475 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2476 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2477 isa<ConstantInt>(Op0I->getOperand(1))) {
2478 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2479 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2481 InsertNewInstBefore(Add, I);
2482 Value *C1C2 = ConstantExpr::getMul(Op1,
2483 cast<Constant>(Op0I->getOperand(1)));
2484 return BinaryOperator::createAdd(Add, C1C2);
2488 // Try to fold constant mul into select arguments.
2489 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2490 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2493 if (isa<PHINode>(Op0))
2494 if (Instruction *NV = FoldOpIntoPhi(I))
2498 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2499 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2500 return BinaryOperator::createMul(Op0v, Op1v);
2502 // If one of the operands of the multiply is a cast from a boolean value, then
2503 // we know the bool is either zero or one, so this is a 'masking' multiply.
2504 // See if we can simplify things based on how the boolean was originally
2506 CastInst *BoolCast = 0;
2507 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2508 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2511 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2512 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2515 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2516 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2517 const Type *SCOpTy = SCIOp0->getType();
2520 // If the icmp is true iff the sign bit of X is set, then convert this
2521 // multiply into a shift/and combination.
2522 if (isa<ConstantInt>(SCIOp1) &&
2523 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2525 // Shift the X value right to turn it into "all signbits".
2526 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2527 SCOpTy->getPrimitiveSizeInBits()-1);
2529 InsertNewInstBefore(
2530 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2531 BoolCast->getOperand(0)->getName()+
2534 // If the multiply type is not the same as the source type, sign extend
2535 // or truncate to the multiply type.
2536 if (I.getType() != V->getType()) {
2537 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2538 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2539 Instruction::CastOps opcode =
2540 (SrcBits == DstBits ? Instruction::BitCast :
2541 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2542 V = InsertCastBefore(opcode, V, I.getType(), I);
2545 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2546 return BinaryOperator::createAnd(V, OtherOp);
2551 return Changed ? &I : 0;
2554 /// This function implements the transforms on div instructions that work
2555 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2556 /// used by the visitors to those instructions.
2557 /// @brief Transforms common to all three div instructions
2558 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2559 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2562 if (isa<UndefValue>(Op0))
2563 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2565 // X / undef -> undef
2566 if (isa<UndefValue>(Op1))
2567 return ReplaceInstUsesWith(I, Op1);
2569 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2570 // This does not apply for fdiv.
2571 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2572 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2573 // the same basic block, then we replace the select with Y, and the
2574 // condition of the select with false (if the cond value is in the same BB).
2575 // If the select has uses other than the div, this allows them to be
2576 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2577 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2578 if (ST->isNullValue()) {
2579 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2580 if (CondI && CondI->getParent() == I.getParent())
2581 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2582 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2583 I.setOperand(1, SI->getOperand(2));
2585 UpdateValueUsesWith(SI, SI->getOperand(2));
2589 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2590 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2591 if (ST->isNullValue()) {
2592 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2593 if (CondI && CondI->getParent() == I.getParent())
2594 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2595 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2596 I.setOperand(1, SI->getOperand(1));
2598 UpdateValueUsesWith(SI, SI->getOperand(1));
2606 /// This function implements the transforms common to both integer division
2607 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2608 /// division instructions.
2609 /// @brief Common integer divide transforms
2610 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2611 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2613 if (Instruction *Common = commonDivTransforms(I))
2616 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2618 if (RHS->equalsInt(1))
2619 return ReplaceInstUsesWith(I, Op0);
2621 // (X / C1) / C2 -> X / (C1*C2)
2622 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2623 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2624 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2625 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2626 Multiply(RHS, LHSRHS));
2629 if (!RHS->isZero()) { // avoid X udiv 0
2630 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2631 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2633 if (isa<PHINode>(Op0))
2634 if (Instruction *NV = FoldOpIntoPhi(I))
2639 // 0 / X == 0, we don't need to preserve faults!
2640 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2641 if (LHS->equalsInt(0))
2642 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2647 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2648 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2650 // Handle the integer div common cases
2651 if (Instruction *Common = commonIDivTransforms(I))
2654 // X udiv C^2 -> X >> C
2655 // Check to see if this is an unsigned division with an exact power of 2,
2656 // if so, convert to a right shift.
2657 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2658 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2659 return BinaryOperator::createLShr(Op0,
2660 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2663 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2664 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2665 if (RHSI->getOpcode() == Instruction::Shl &&
2666 isa<ConstantInt>(RHSI->getOperand(0))) {
2667 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2668 if (C1.isPowerOf2()) {
2669 Value *N = RHSI->getOperand(1);
2670 const Type *NTy = N->getType();
2671 if (uint32_t C2 = C1.logBase2()) {
2672 Constant *C2V = ConstantInt::get(NTy, C2);
2673 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2675 return BinaryOperator::createLShr(Op0, N);
2680 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2681 // where C1&C2 are powers of two.
2682 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2683 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2684 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2685 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2686 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2687 // Compute the shift amounts
2688 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2689 // Construct the "on true" case of the select
2690 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2691 Instruction *TSI = BinaryOperator::createLShr(
2692 Op0, TC, SI->getName()+".t");
2693 TSI = InsertNewInstBefore(TSI, I);
2695 // Construct the "on false" case of the select
2696 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2697 Instruction *FSI = BinaryOperator::createLShr(
2698 Op0, FC, SI->getName()+".f");
2699 FSI = InsertNewInstBefore(FSI, I);
2701 // construct the select instruction and return it.
2702 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2708 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2709 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2711 // Handle the integer div common cases
2712 if (Instruction *Common = commonIDivTransforms(I))
2715 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2717 if (RHS->isAllOnesValue())
2718 return BinaryOperator::createNeg(Op0);
2721 if (Value *LHSNeg = dyn_castNegVal(Op0))
2722 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2725 // If the sign bits of both operands are zero (i.e. we can prove they are
2726 // unsigned inputs), turn this into a udiv.
2727 if (I.getType()->isInteger()) {
2728 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2729 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2730 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2731 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2738 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2739 return commonDivTransforms(I);
2742 /// GetFactor - If we can prove that the specified value is at least a multiple
2743 /// of some factor, return that factor.
2744 static Constant *GetFactor(Value *V) {
2745 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2748 // Unless we can be tricky, we know this is a multiple of 1.
2749 Constant *Result = ConstantInt::get(V->getType(), 1);
2751 Instruction *I = dyn_cast<Instruction>(V);
2752 if (!I) return Result;
2754 if (I->getOpcode() == Instruction::Mul) {
2755 // Handle multiplies by a constant, etc.
2756 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2757 GetFactor(I->getOperand(1)));
2758 } else if (I->getOpcode() == Instruction::Shl) {
2759 // (X<<C) -> X * (1 << C)
2760 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2761 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2762 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2764 } else if (I->getOpcode() == Instruction::And) {
2765 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2766 // X & 0xFFF0 is known to be a multiple of 16.
2767 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2768 if (Zeros != V->getType()->getPrimitiveSizeInBits())// don't shift by "32"
2769 return ConstantExpr::getShl(Result,
2770 ConstantInt::get(Result->getType(), Zeros));
2772 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2773 // Only handle int->int casts.
2774 if (!CI->isIntegerCast())
2776 Value *Op = CI->getOperand(0);
2777 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2782 /// This function implements the transforms on rem instructions that work
2783 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2784 /// is used by the visitors to those instructions.
2785 /// @brief Transforms common to all three rem instructions
2786 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2787 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2789 // 0 % X == 0, we don't need to preserve faults!
2790 if (Constant *LHS = dyn_cast<Constant>(Op0))
2791 if (LHS->isNullValue())
2792 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2794 if (isa<UndefValue>(Op0)) // undef % X -> 0
2795 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2796 if (isa<UndefValue>(Op1))
2797 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2799 // Handle cases involving: rem X, (select Cond, Y, Z)
2800 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2801 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2802 // the same basic block, then we replace the select with Y, and the
2803 // condition of the select with false (if the cond value is in the same
2804 // BB). If the select has uses other than the div, this allows them to be
2806 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2807 if (ST->isNullValue()) {
2808 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2809 if (CondI && CondI->getParent() == I.getParent())
2810 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2811 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2812 I.setOperand(1, SI->getOperand(2));
2814 UpdateValueUsesWith(SI, SI->getOperand(2));
2817 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2818 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2819 if (ST->isNullValue()) {
2820 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2821 if (CondI && CondI->getParent() == I.getParent())
2822 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2823 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2824 I.setOperand(1, SI->getOperand(1));
2826 UpdateValueUsesWith(SI, SI->getOperand(1));
2834 /// This function implements the transforms common to both integer remainder
2835 /// instructions (urem and srem). It is called by the visitors to those integer
2836 /// remainder instructions.
2837 /// @brief Common integer remainder transforms
2838 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2839 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2841 if (Instruction *common = commonRemTransforms(I))
2844 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2845 // X % 0 == undef, we don't need to preserve faults!
2846 if (RHS->equalsInt(0))
2847 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2849 if (RHS->equalsInt(1)) // X % 1 == 0
2850 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2852 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2853 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2854 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2856 } else if (isa<PHINode>(Op0I)) {
2857 if (Instruction *NV = FoldOpIntoPhi(I))
2860 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2861 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2862 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2869 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2870 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2872 if (Instruction *common = commonIRemTransforms(I))
2875 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2876 // X urem C^2 -> X and C
2877 // Check to see if this is an unsigned remainder with an exact power of 2,
2878 // if so, convert to a bitwise and.
2879 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2880 if (C->getValue().isPowerOf2())
2881 return BinaryOperator::createAnd(Op0, SubOne(C));
2884 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2885 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2886 if (RHSI->getOpcode() == Instruction::Shl &&
2887 isa<ConstantInt>(RHSI->getOperand(0))) {
2888 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2889 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2890 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2892 return BinaryOperator::createAnd(Op0, Add);
2897 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2898 // where C1&C2 are powers of two.
2899 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2900 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2901 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2902 // STO == 0 and SFO == 0 handled above.
2903 if ((STO->getValue().isPowerOf2()) &&
2904 (SFO->getValue().isPowerOf2())) {
2905 Value *TrueAnd = InsertNewInstBefore(
2906 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2907 Value *FalseAnd = InsertNewInstBefore(
2908 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2909 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2917 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2918 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2920 // Handle the integer rem common cases
2921 if (Instruction *common = commonIRemTransforms(I))
2924 if (Value *RHSNeg = dyn_castNegVal(Op1))
2925 if (!isa<ConstantInt>(RHSNeg) ||
2926 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2928 AddUsesToWorkList(I);
2929 I.setOperand(1, RHSNeg);
2933 // If the sign bits of both operands are zero (i.e. we can prove they are
2934 // unsigned inputs), turn this into a urem.
2935 if (I.getType()->isInteger()) {
2936 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2937 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2938 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2939 return BinaryOperator::createURem(Op0, Op1, I.getName());
2946 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2947 return commonRemTransforms(I);
2950 // isMaxValueMinusOne - return true if this is Max-1
2951 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2952 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2954 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2955 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2958 // isMinValuePlusOne - return true if this is Min+1
2959 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2961 return C->getValue() == 1; // unsigned
2963 // Calculate 1111111111000000000000
2964 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2965 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2968 // isOneBitSet - Return true if there is exactly one bit set in the specified
2970 static bool isOneBitSet(const ConstantInt *CI) {
2971 return CI->getValue().isPowerOf2();
2974 // isHighOnes - Return true if the constant is of the form 1+0+.
2975 // This is the same as lowones(~X).
2976 static bool isHighOnes(const ConstantInt *CI) {
2977 return (~CI->getValue() + 1).isPowerOf2();
2980 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2981 /// are carefully arranged to allow folding of expressions such as:
2983 /// (A < B) | (A > B) --> (A != B)
2985 /// Note that this is only valid if the first and second predicates have the
2986 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2988 /// Three bits are used to represent the condition, as follows:
2993 /// <=> Value Definition
2994 /// 000 0 Always false
3001 /// 111 7 Always true
3003 static unsigned getICmpCode(const ICmpInst *ICI) {
3004 switch (ICI->getPredicate()) {
3006 case ICmpInst::ICMP_UGT: return 1; // 001
3007 case ICmpInst::ICMP_SGT: return 1; // 001
3008 case ICmpInst::ICMP_EQ: return 2; // 010
3009 case ICmpInst::ICMP_UGE: return 3; // 011
3010 case ICmpInst::ICMP_SGE: return 3; // 011
3011 case ICmpInst::ICMP_ULT: return 4; // 100
3012 case ICmpInst::ICMP_SLT: return 4; // 100
3013 case ICmpInst::ICMP_NE: return 5; // 101
3014 case ICmpInst::ICMP_ULE: return 6; // 110
3015 case ICmpInst::ICMP_SLE: return 6; // 110
3018 assert(0 && "Invalid ICmp predicate!");
3023 /// getICmpValue - This is the complement of getICmpCode, which turns an
3024 /// opcode and two operands into either a constant true or false, or a brand
3025 /// new ICmp instruction. The sign is passed in to determine which kind
3026 /// of predicate to use in new icmp instructions.
3027 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3029 default: assert(0 && "Illegal ICmp code!");
3030 case 0: return ConstantInt::getFalse();
3033 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3035 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3036 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3039 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3041 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3044 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3046 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3047 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3050 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3052 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3053 case 7: return ConstantInt::getTrue();
3057 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3058 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3059 (ICmpInst::isSignedPredicate(p1) &&
3060 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3061 (ICmpInst::isSignedPredicate(p2) &&
3062 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3066 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3067 struct FoldICmpLogical {
3070 ICmpInst::Predicate pred;
3071 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3072 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3073 pred(ICI->getPredicate()) {}
3074 bool shouldApply(Value *V) const {
3075 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3076 if (PredicatesFoldable(pred, ICI->getPredicate()))
3077 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
3078 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
3081 Instruction *apply(Instruction &Log) const {
3082 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3083 if (ICI->getOperand(0) != LHS) {
3084 assert(ICI->getOperand(1) == LHS);
3085 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3088 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3089 unsigned LHSCode = getICmpCode(ICI);
3090 unsigned RHSCode = getICmpCode(RHSICI);
3092 switch (Log.getOpcode()) {
3093 case Instruction::And: Code = LHSCode & RHSCode; break;
3094 case Instruction::Or: Code = LHSCode | RHSCode; break;
3095 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3096 default: assert(0 && "Illegal logical opcode!"); return 0;
3099 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3100 ICmpInst::isSignedPredicate(ICI->getPredicate());
3102 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3103 if (Instruction *I = dyn_cast<Instruction>(RV))
3105 // Otherwise, it's a constant boolean value...
3106 return IC.ReplaceInstUsesWith(Log, RV);
3109 } // end anonymous namespace
3111 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3112 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3113 // guaranteed to be a binary operator.
3114 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3116 ConstantInt *AndRHS,
3117 BinaryOperator &TheAnd) {
3118 Value *X = Op->getOperand(0);
3119 Constant *Together = 0;
3121 Together = And(AndRHS, OpRHS);
3123 switch (Op->getOpcode()) {
3124 case Instruction::Xor:
3125 if (Op->hasOneUse()) {
3126 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3127 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3128 InsertNewInstBefore(And, TheAnd);
3130 return BinaryOperator::createXor(And, Together);
3133 case Instruction::Or:
3134 if (Together == AndRHS) // (X | C) & C --> C
3135 return ReplaceInstUsesWith(TheAnd, AndRHS);
3137 if (Op->hasOneUse() && Together != OpRHS) {
3138 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3139 Instruction *Or = BinaryOperator::createOr(X, Together);
3140 InsertNewInstBefore(Or, TheAnd);
3142 return BinaryOperator::createAnd(Or, AndRHS);
3145 case Instruction::Add:
3146 if (Op->hasOneUse()) {
3147 // Adding a one to a single bit bit-field should be turned into an XOR
3148 // of the bit. First thing to check is to see if this AND is with a
3149 // single bit constant.
3150 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3152 // If there is only one bit set...
3153 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3154 // Ok, at this point, we know that we are masking the result of the
3155 // ADD down to exactly one bit. If the constant we are adding has
3156 // no bits set below this bit, then we can eliminate the ADD.
3157 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3159 // Check to see if any bits below the one bit set in AndRHSV are set.
3160 if ((AddRHS & (AndRHSV-1)) == 0) {
3161 // If not, the only thing that can effect the output of the AND is
3162 // the bit specified by AndRHSV. If that bit is set, the effect of
3163 // the XOR is to toggle the bit. If it is clear, then the ADD has
3165 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3166 TheAnd.setOperand(0, X);
3169 // Pull the XOR out of the AND.
3170 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3171 InsertNewInstBefore(NewAnd, TheAnd);
3172 NewAnd->takeName(Op);
3173 return BinaryOperator::createXor(NewAnd, AndRHS);
3180 case Instruction::Shl: {
3181 // We know that the AND will not produce any of the bits shifted in, so if
3182 // the anded constant includes them, clear them now!
3184 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3185 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3186 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3187 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3189 if (CI->getValue() == ShlMask) {
3190 // Masking out bits that the shift already masks
3191 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3192 } else if (CI != AndRHS) { // Reducing bits set in and.
3193 TheAnd.setOperand(1, CI);
3198 case Instruction::LShr:
3200 // We know that the AND will not produce any of the bits shifted in, so if
3201 // the anded constant includes them, clear them now! This only applies to
3202 // unsigned shifts, because a signed shr may bring in set bits!
3204 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3205 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3206 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3207 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3209 if (CI->getValue() == ShrMask) {
3210 // Masking out bits that the shift already masks.
3211 return ReplaceInstUsesWith(TheAnd, Op);
3212 } else if (CI != AndRHS) {
3213 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3218 case Instruction::AShr:
3220 // See if this is shifting in some sign extension, then masking it out
3222 if (Op->hasOneUse()) {
3223 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3224 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3225 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3226 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3227 if (C == AndRHS) { // Masking out bits shifted in.
3228 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3229 // Make the argument unsigned.
3230 Value *ShVal = Op->getOperand(0);
3231 ShVal = InsertNewInstBefore(
3232 BinaryOperator::createLShr(ShVal, OpRHS,
3233 Op->getName()), TheAnd);
3234 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3243 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3244 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3245 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3246 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3247 /// insert new instructions.
3248 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3249 bool isSigned, bool Inside,
3251 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3252 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3253 "Lo is not <= Hi in range emission code!");
3256 if (Lo == Hi) // Trivially false.
3257 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3259 // V >= Min && V < Hi --> V < Hi
3260 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3261 ICmpInst::Predicate pred = (isSigned ?
3262 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3263 return new ICmpInst(pred, V, Hi);
3266 // Emit V-Lo <u Hi-Lo
3267 Constant *NegLo = ConstantExpr::getNeg(Lo);
3268 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3269 InsertNewInstBefore(Add, IB);
3270 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3271 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3274 if (Lo == Hi) // Trivially true.
3275 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3277 // V < Min || V >= Hi -> V > Hi-1
3278 Hi = SubOne(cast<ConstantInt>(Hi));
3279 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3280 ICmpInst::Predicate pred = (isSigned ?
3281 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3282 return new ICmpInst(pred, V, Hi);
3285 // Emit V-Lo >u Hi-1-Lo
3286 // Note that Hi has already had one subtracted from it, above.
3287 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3288 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3289 InsertNewInstBefore(Add, IB);
3290 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3291 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3294 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3295 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3296 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3297 // not, since all 1s are not contiguous.
3298 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3299 const APInt& V = Val->getValue();
3300 uint32_t BitWidth = Val->getType()->getBitWidth();
3301 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3303 // look for the first zero bit after the run of ones
3304 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3305 // look for the first non-zero bit
3306 ME = V.getActiveBits();
3310 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3311 /// where isSub determines whether the operator is a sub. If we can fold one of
3312 /// the following xforms:
3314 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3315 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3316 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3318 /// return (A +/- B).
3320 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3321 ConstantInt *Mask, bool isSub,
3323 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3324 if (!LHSI || LHSI->getNumOperands() != 2 ||
3325 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3327 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3329 switch (LHSI->getOpcode()) {
3331 case Instruction::And:
3332 if (And(N, Mask) == Mask) {
3333 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3334 if ((Mask->getValue().countLeadingZeros() +
3335 Mask->getValue().countPopulation()) ==
3336 Mask->getValue().getBitWidth())
3339 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3340 // part, we don't need any explicit masks to take them out of A. If that
3341 // is all N is, ignore it.
3342 uint32_t MB = 0, ME = 0;
3343 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3344 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3345 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3346 if (MaskedValueIsZero(RHS, Mask))
3351 case Instruction::Or:
3352 case Instruction::Xor:
3353 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3354 if ((Mask->getValue().countLeadingZeros() +
3355 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3356 && And(N, Mask)->isZero())
3363 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3365 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3366 return InsertNewInstBefore(New, I);
3369 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3370 bool Changed = SimplifyCommutative(I);
3371 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3373 if (isa<UndefValue>(Op1)) // X & undef -> 0
3374 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3378 return ReplaceInstUsesWith(I, Op1);
3380 // See if we can simplify any instructions used by the instruction whose sole
3381 // purpose is to compute bits we don't care about.
3382 if (!isa<VectorType>(I.getType())) {
3383 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3384 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3385 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3386 KnownZero, KnownOne))
3389 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3390 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3391 return ReplaceInstUsesWith(I, I.getOperand(0));
3392 } else if (isa<ConstantAggregateZero>(Op1)) {
3393 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3397 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3398 const APInt& AndRHSMask = AndRHS->getValue();
3399 APInt NotAndRHS(~AndRHSMask);
3401 // Optimize a variety of ((val OP C1) & C2) combinations...
3402 if (isa<BinaryOperator>(Op0)) {
3403 Instruction *Op0I = cast<Instruction>(Op0);
3404 Value *Op0LHS = Op0I->getOperand(0);
3405 Value *Op0RHS = Op0I->getOperand(1);
3406 switch (Op0I->getOpcode()) {
3407 case Instruction::Xor:
3408 case Instruction::Or:
3409 // If the mask is only needed on one incoming arm, push it up.
3410 if (Op0I->hasOneUse()) {
3411 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3412 // Not masking anything out for the LHS, move to RHS.
3413 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3414 Op0RHS->getName()+".masked");
3415 InsertNewInstBefore(NewRHS, I);
3416 return BinaryOperator::create(
3417 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3419 if (!isa<Constant>(Op0RHS) &&
3420 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3421 // Not masking anything out for the RHS, move to LHS.
3422 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3423 Op0LHS->getName()+".masked");
3424 InsertNewInstBefore(NewLHS, I);
3425 return BinaryOperator::create(
3426 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3431 case Instruction::Add:
3432 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3433 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3434 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3435 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3436 return BinaryOperator::createAnd(V, AndRHS);
3437 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3438 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3441 case Instruction::Sub:
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, true, I))
3446 return BinaryOperator::createAnd(V, AndRHS);
3450 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3451 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3453 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3454 // If this is an integer truncation or change from signed-to-unsigned, and
3455 // if the source is an and/or with immediate, transform it. This
3456 // frequently occurs for bitfield accesses.
3457 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3458 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3459 CastOp->getNumOperands() == 2)
3460 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3461 if (CastOp->getOpcode() == Instruction::And) {
3462 // Change: and (cast (and X, C1) to T), C2
3463 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3464 // This will fold the two constants together, which may allow
3465 // other simplifications.
3466 Instruction *NewCast = CastInst::createTruncOrBitCast(
3467 CastOp->getOperand(0), I.getType(),
3468 CastOp->getName()+".shrunk");
3469 NewCast = InsertNewInstBefore(NewCast, I);
3470 // trunc_or_bitcast(C1)&C2
3471 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3472 C3 = ConstantExpr::getAnd(C3, AndRHS);
3473 return BinaryOperator::createAnd(NewCast, C3);
3474 } else if (CastOp->getOpcode() == Instruction::Or) {
3475 // Change: and (cast (or X, C1) to T), C2
3476 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3477 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3478 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3479 return ReplaceInstUsesWith(I, AndRHS);
3484 // Try to fold constant and into select arguments.
3485 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3486 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3488 if (isa<PHINode>(Op0))
3489 if (Instruction *NV = FoldOpIntoPhi(I))
3493 Value *Op0NotVal = dyn_castNotVal(Op0);
3494 Value *Op1NotVal = dyn_castNotVal(Op1);
3496 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3497 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3499 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3500 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3501 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3502 I.getName()+".demorgan");
3503 InsertNewInstBefore(Or, I);
3504 return BinaryOperator::createNot(Or);
3508 Value *A = 0, *B = 0, *C = 0, *D = 0;
3509 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3510 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3511 return ReplaceInstUsesWith(I, Op1);
3513 // (A|B) & ~(A&B) -> A^B
3514 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3515 if ((A == C && B == D) || (A == D && B == C))
3516 return BinaryOperator::createXor(A, B);
3520 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3521 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3522 return ReplaceInstUsesWith(I, Op0);
3524 // ~(A&B) & (A|B) -> A^B
3525 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3526 if ((A == C && B == D) || (A == D && B == C))
3527 return BinaryOperator::createXor(A, B);
3531 if (Op0->hasOneUse() &&
3532 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3533 if (A == Op1) { // (A^B)&A -> A&(A^B)
3534 I.swapOperands(); // Simplify below
3535 std::swap(Op0, Op1);
3536 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3537 cast<BinaryOperator>(Op0)->swapOperands();
3538 I.swapOperands(); // Simplify below
3539 std::swap(Op0, Op1);
3542 if (Op1->hasOneUse() &&
3543 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3544 if (B == Op0) { // B&(A^B) -> B&(B^A)
3545 cast<BinaryOperator>(Op1)->swapOperands();
3548 if (A == Op0) { // A&(A^B) -> A & ~B
3549 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3550 InsertNewInstBefore(NotB, I);
3551 return BinaryOperator::createAnd(A, NotB);
3556 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3557 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3558 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3561 Value *LHSVal, *RHSVal;
3562 ConstantInt *LHSCst, *RHSCst;
3563 ICmpInst::Predicate LHSCC, RHSCC;
3564 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3565 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3566 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3567 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3568 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3569 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3570 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3571 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3573 // Don't try to fold ICMP_SLT + ICMP_ULT.
3574 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3575 ICmpInst::isSignedPredicate(LHSCC) ==
3576 ICmpInst::isSignedPredicate(RHSCC))) {
3577 // Ensure that the larger constant is on the RHS.
3578 ICmpInst::Predicate GT;
3579 if (ICmpInst::isSignedPredicate(LHSCC) ||
3580 (ICmpInst::isEquality(LHSCC) &&
3581 ICmpInst::isSignedPredicate(RHSCC)))
3582 GT = ICmpInst::ICMP_SGT;
3584 GT = ICmpInst::ICMP_UGT;
3586 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3587 ICmpInst *LHS = cast<ICmpInst>(Op0);
3588 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3589 std::swap(LHS, RHS);
3590 std::swap(LHSCst, RHSCst);
3591 std::swap(LHSCC, RHSCC);
3594 // At this point, we know we have have two icmp instructions
3595 // comparing a value against two constants and and'ing the result
3596 // together. Because of the above check, we know that we only have
3597 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3598 // (from the FoldICmpLogical check above), that the two constants
3599 // are not equal and that the larger constant is on the RHS
3600 assert(LHSCst != RHSCst && "Compares not folded above?");
3603 default: assert(0 && "Unknown integer condition code!");
3604 case ICmpInst::ICMP_EQ:
3606 default: assert(0 && "Unknown integer condition code!");
3607 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3608 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3609 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3610 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3611 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3612 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3613 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3614 return ReplaceInstUsesWith(I, LHS);
3616 case ICmpInst::ICMP_NE:
3618 default: assert(0 && "Unknown integer condition code!");
3619 case ICmpInst::ICMP_ULT:
3620 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3621 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3622 break; // (X != 13 & X u< 15) -> no change
3623 case ICmpInst::ICMP_SLT:
3624 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3625 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3626 break; // (X != 13 & X s< 15) -> no change
3627 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3628 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3629 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3630 return ReplaceInstUsesWith(I, RHS);
3631 case ICmpInst::ICMP_NE:
3632 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3633 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3634 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3635 LHSVal->getName()+".off");
3636 InsertNewInstBefore(Add, I);
3637 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3638 ConstantInt::get(Add->getType(), 1));
3640 break; // (X != 13 & X != 15) -> no change
3643 case ICmpInst::ICMP_ULT:
3645 default: assert(0 && "Unknown integer condition code!");
3646 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3647 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3648 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3649 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3651 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3652 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3653 return ReplaceInstUsesWith(I, LHS);
3654 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3658 case ICmpInst::ICMP_SLT:
3660 default: assert(0 && "Unknown integer condition code!");
3661 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3662 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3663 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3664 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3666 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3667 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3668 return ReplaceInstUsesWith(I, LHS);
3669 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3673 case ICmpInst::ICMP_UGT:
3675 default: assert(0 && "Unknown integer condition code!");
3676 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3677 return ReplaceInstUsesWith(I, LHS);
3678 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3679 return ReplaceInstUsesWith(I, RHS);
3680 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3682 case ICmpInst::ICMP_NE:
3683 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3684 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3685 break; // (X u> 13 & X != 15) -> no change
3686 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3687 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3689 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3693 case ICmpInst::ICMP_SGT:
3695 default: assert(0 && "Unknown integer condition code!");
3696 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3697 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3698 return ReplaceInstUsesWith(I, RHS);
3699 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3701 case ICmpInst::ICMP_NE:
3702 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3703 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3704 break; // (X s> 13 & X != 15) -> no change
3705 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3706 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3708 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3716 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3717 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3718 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3719 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3720 const Type *SrcTy = Op0C->getOperand(0)->getType();
3721 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3722 // Only do this if the casts both really cause code to be generated.
3723 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3725 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3727 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3728 Op1C->getOperand(0),
3730 InsertNewInstBefore(NewOp, I);
3731 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3735 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3736 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3737 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3738 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3739 SI0->getOperand(1) == SI1->getOperand(1) &&
3740 (SI0->hasOneUse() || SI1->hasOneUse())) {
3741 Instruction *NewOp =
3742 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3744 SI0->getName()), I);
3745 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3746 SI1->getOperand(1));
3750 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3751 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3752 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3753 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3754 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3755 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3756 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3757 // If either of the constants are nans, then the whole thing returns
3759 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3760 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3761 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3762 RHS->getOperand(0));
3767 return Changed ? &I : 0;
3770 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3771 /// in the result. If it does, and if the specified byte hasn't been filled in
3772 /// yet, fill it in and return false.
3773 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3774 Instruction *I = dyn_cast<Instruction>(V);
3775 if (I == 0) return true;
3777 // If this is an or instruction, it is an inner node of the bswap.
3778 if (I->getOpcode() == Instruction::Or)
3779 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3780 CollectBSwapParts(I->getOperand(1), ByteValues);
3782 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3783 // If this is a shift by a constant int, and it is "24", then its operand
3784 // defines a byte. We only handle unsigned types here.
3785 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3786 // Not shifting the entire input by N-1 bytes?
3787 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3788 8*(ByteValues.size()-1))
3792 if (I->getOpcode() == Instruction::Shl) {
3793 // X << 24 defines the top byte with the lowest of the input bytes.
3794 DestNo = ByteValues.size()-1;
3796 // X >>u 24 defines the low byte with the highest of the input bytes.
3800 // If the destination byte value is already defined, the values are or'd
3801 // together, which isn't a bswap (unless it's an or of the same bits).
3802 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3804 ByteValues[DestNo] = I->getOperand(0);
3808 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3810 Value *Shift = 0, *ShiftLHS = 0;
3811 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3812 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3813 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3815 Instruction *SI = cast<Instruction>(Shift);
3817 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3818 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3819 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3822 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3824 if (AndAmt->getValue().getActiveBits() > 64)
3826 uint64_t AndAmtVal = AndAmt->getZExtValue();
3827 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3828 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3830 // Unknown mask for bswap.
3831 if (DestByte == ByteValues.size()) return true;
3833 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3835 if (SI->getOpcode() == Instruction::Shl)
3836 SrcByte = DestByte - ShiftBytes;
3838 SrcByte = DestByte + ShiftBytes;
3840 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3841 if (SrcByte != ByteValues.size()-DestByte-1)
3844 // If the destination byte value is already defined, the values are or'd
3845 // together, which isn't a bswap (unless it's an or of the same bits).
3846 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3848 ByteValues[DestByte] = SI->getOperand(0);
3852 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3853 /// If so, insert the new bswap intrinsic and return it.
3854 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3855 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3856 if (!ITy || ITy->getBitWidth() % 16)
3857 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3859 /// ByteValues - For each byte of the result, we keep track of which value
3860 /// defines each byte.
3861 SmallVector<Value*, 8> ByteValues;
3862 ByteValues.resize(ITy->getBitWidth()/8);
3864 // Try to find all the pieces corresponding to the bswap.
3865 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3866 CollectBSwapParts(I.getOperand(1), ByteValues))
3869 // Check to see if all of the bytes come from the same value.
3870 Value *V = ByteValues[0];
3871 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3873 // Check to make sure that all of the bytes come from the same value.
3874 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3875 if (ByteValues[i] != V)
3877 const Type *Tys[] = { ITy };
3878 Module *M = I.getParent()->getParent()->getParent();
3879 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3880 return new CallInst(F, V);
3884 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3885 bool Changed = SimplifyCommutative(I);
3886 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3888 if (isa<UndefValue>(Op1)) // X | undef -> -1
3889 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3893 return ReplaceInstUsesWith(I, Op0);
3895 // See if we can simplify any instructions used by the instruction whose sole
3896 // purpose is to compute bits we don't care about.
3897 if (!isa<VectorType>(I.getType())) {
3898 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3899 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3900 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3901 KnownZero, KnownOne))
3903 } else if (isa<ConstantAggregateZero>(Op1)) {
3904 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3905 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3906 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3907 return ReplaceInstUsesWith(I, I.getOperand(1));
3913 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3914 ConstantInt *C1 = 0; Value *X = 0;
3915 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3916 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3917 Instruction *Or = BinaryOperator::createOr(X, RHS);
3918 InsertNewInstBefore(Or, I);
3920 return BinaryOperator::createAnd(Or,
3921 ConstantInt::get(RHS->getValue() | C1->getValue()));
3924 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3925 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3926 Instruction *Or = BinaryOperator::createOr(X, RHS);
3927 InsertNewInstBefore(Or, I);
3929 return BinaryOperator::createXor(Or,
3930 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3933 // Try to fold constant and into select arguments.
3934 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3935 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3937 if (isa<PHINode>(Op0))
3938 if (Instruction *NV = FoldOpIntoPhi(I))
3942 Value *A = 0, *B = 0;
3943 ConstantInt *C1 = 0, *C2 = 0;
3945 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3946 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3947 return ReplaceInstUsesWith(I, Op1);
3948 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3949 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3950 return ReplaceInstUsesWith(I, Op0);
3952 // (A | B) | C and A | (B | C) -> bswap if possible.
3953 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3954 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3955 match(Op1, m_Or(m_Value(), m_Value())) ||
3956 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3957 match(Op1, m_Shift(m_Value(), m_Value())))) {
3958 if (Instruction *BSwap = MatchBSwap(I))
3962 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3963 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3964 MaskedValueIsZero(Op1, C1->getValue())) {
3965 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3966 InsertNewInstBefore(NOr, I);
3968 return BinaryOperator::createXor(NOr, C1);
3971 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3972 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3973 MaskedValueIsZero(Op0, C1->getValue())) {
3974 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3975 InsertNewInstBefore(NOr, I);
3977 return BinaryOperator::createXor(NOr, C1);
3981 Value *C = 0, *D = 0;
3982 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3983 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3984 Value *V1 = 0, *V2 = 0, *V3 = 0;
3985 C1 = dyn_cast<ConstantInt>(C);
3986 C2 = dyn_cast<ConstantInt>(D);
3987 if (C1 && C2) { // (A & C1)|(B & C2)
3988 // If we have: ((V + N) & C1) | (V & C2)
3989 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3990 // replace with V+N.
3991 if (C1->getValue() == ~C2->getValue()) {
3992 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3993 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3994 // Add commutes, try both ways.
3995 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3996 return ReplaceInstUsesWith(I, A);
3997 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3998 return ReplaceInstUsesWith(I, A);
4000 // Or commutes, try both ways.
4001 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4002 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4003 // Add commutes, try both ways.
4004 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4005 return ReplaceInstUsesWith(I, B);
4006 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4007 return ReplaceInstUsesWith(I, B);
4010 V1 = 0; V2 = 0; V3 = 0;
4013 // Check to see if we have any common things being and'ed. If so, find the
4014 // terms for V1 & (V2|V3).
4015 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4016 if (A == B) // (A & C)|(A & D) == A & (C|D)
4017 V1 = A, V2 = C, V3 = D;
4018 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4019 V1 = A, V2 = B, V3 = C;
4020 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4021 V1 = C, V2 = A, V3 = D;
4022 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4023 V1 = C, V2 = A, V3 = B;
4027 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
4028 return BinaryOperator::createAnd(V1, Or);
4033 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4034 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4035 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4036 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4037 SI0->getOperand(1) == SI1->getOperand(1) &&
4038 (SI0->hasOneUse() || SI1->hasOneUse())) {
4039 Instruction *NewOp =
4040 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4042 SI0->getName()), I);
4043 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4044 SI1->getOperand(1));
4048 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4049 if (A == Op1) // ~A | A == -1
4050 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4054 // Note, A is still live here!
4055 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4057 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4059 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4060 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4061 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4062 I.getName()+".demorgan"), I);
4063 return BinaryOperator::createNot(And);
4067 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4068 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4069 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4072 Value *LHSVal, *RHSVal;
4073 ConstantInt *LHSCst, *RHSCst;
4074 ICmpInst::Predicate LHSCC, RHSCC;
4075 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4076 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4077 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4078 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4079 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4080 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4081 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4082 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4083 // We can't fold (ugt x, C) | (sgt x, C2).
4084 PredicatesFoldable(LHSCC, RHSCC)) {
4085 // Ensure that the larger constant is on the RHS.
4086 ICmpInst *LHS = cast<ICmpInst>(Op0);
4088 if (ICmpInst::isSignedPredicate(LHSCC))
4089 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4091 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4094 std::swap(LHS, RHS);
4095 std::swap(LHSCst, RHSCst);
4096 std::swap(LHSCC, RHSCC);
4099 // At this point, we know we have have two icmp instructions
4100 // comparing a value against two constants and or'ing the result
4101 // together. Because of the above check, we know that we only have
4102 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4103 // FoldICmpLogical check above), that the two constants are not
4105 assert(LHSCst != RHSCst && "Compares not folded above?");
4108 default: assert(0 && "Unknown integer condition code!");
4109 case ICmpInst::ICMP_EQ:
4111 default: assert(0 && "Unknown integer condition code!");
4112 case ICmpInst::ICMP_EQ:
4113 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4114 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4115 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4116 LHSVal->getName()+".off");
4117 InsertNewInstBefore(Add, I);
4118 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4119 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4121 break; // (X == 13 | X == 15) -> no change
4122 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4123 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4125 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4126 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4127 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4128 return ReplaceInstUsesWith(I, RHS);
4131 case ICmpInst::ICMP_NE:
4133 default: assert(0 && "Unknown integer condition code!");
4134 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4135 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4136 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4137 return ReplaceInstUsesWith(I, LHS);
4138 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4139 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4140 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4141 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4144 case ICmpInst::ICMP_ULT:
4146 default: assert(0 && "Unknown integer condition code!");
4147 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4149 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4150 // If RHSCst is [us]MAXINT, it is always false. Not handling
4151 // this can cause overflow.
4152 if (RHSCst->isMaxValue(false))
4153 return ReplaceInstUsesWith(I, LHS);
4154 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4156 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4158 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4159 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4160 return ReplaceInstUsesWith(I, RHS);
4161 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4165 case ICmpInst::ICMP_SLT:
4167 default: assert(0 && "Unknown integer condition code!");
4168 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4170 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4171 // If RHSCst is [us]MAXINT, it is always false. Not handling
4172 // this can cause overflow.
4173 if (RHSCst->isMaxValue(true))
4174 return ReplaceInstUsesWith(I, LHS);
4175 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4177 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4179 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4180 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4181 return ReplaceInstUsesWith(I, RHS);
4182 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4186 case ICmpInst::ICMP_UGT:
4188 default: assert(0 && "Unknown integer condition code!");
4189 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4190 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4191 return ReplaceInstUsesWith(I, LHS);
4192 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4194 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4195 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4196 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4197 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4201 case ICmpInst::ICMP_SGT:
4203 default: assert(0 && "Unknown integer condition code!");
4204 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4205 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4206 return ReplaceInstUsesWith(I, LHS);
4207 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4209 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4210 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4211 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4212 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4220 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4221 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4222 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4223 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4224 const Type *SrcTy = Op0C->getOperand(0)->getType();
4225 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4226 // Only do this if the casts both really cause code to be generated.
4227 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4229 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4231 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4232 Op1C->getOperand(0),
4234 InsertNewInstBefore(NewOp, I);
4235 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4241 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4242 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4243 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4244 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4245 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4246 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4247 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4248 // If either of the constants are nans, then the whole thing returns
4250 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4251 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4253 // Otherwise, no need to compare the two constants, compare the
4255 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4256 RHS->getOperand(0));
4261 return Changed ? &I : 0;
4264 // XorSelf - Implements: X ^ X --> 0
4267 XorSelf(Value *rhs) : RHS(rhs) {}
4268 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4269 Instruction *apply(BinaryOperator &Xor) const {
4275 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4276 bool Changed = SimplifyCommutative(I);
4277 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4279 if (isa<UndefValue>(Op1))
4280 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4282 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4283 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4284 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4285 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4288 // See if we can simplify any instructions used by the instruction whose sole
4289 // purpose is to compute bits we don't care about.
4290 if (!isa<VectorType>(I.getType())) {
4291 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4292 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4293 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4294 KnownZero, KnownOne))
4296 } else if (isa<ConstantAggregateZero>(Op1)) {
4297 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4300 // Is this a ~ operation?
4301 if (Value *NotOp = dyn_castNotVal(&I)) {
4302 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4303 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4304 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4305 if (Op0I->getOpcode() == Instruction::And ||
4306 Op0I->getOpcode() == Instruction::Or) {
4307 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4308 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4310 BinaryOperator::createNot(Op0I->getOperand(1),
4311 Op0I->getOperand(1)->getName()+".not");
4312 InsertNewInstBefore(NotY, I);
4313 if (Op0I->getOpcode() == Instruction::And)
4314 return BinaryOperator::createOr(Op0NotVal, NotY);
4316 return BinaryOperator::createAnd(Op0NotVal, NotY);
4323 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4324 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4325 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4326 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4327 return new ICmpInst(ICI->getInversePredicate(),
4328 ICI->getOperand(0), ICI->getOperand(1));
4330 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4331 return new FCmpInst(FCI->getInversePredicate(),
4332 FCI->getOperand(0), FCI->getOperand(1));
4335 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4336 // ~(c-X) == X-c-1 == X+(-c-1)
4337 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4338 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4339 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4340 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4341 ConstantInt::get(I.getType(), 1));
4342 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4345 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4346 if (Op0I->getOpcode() == Instruction::Add) {
4347 // ~(X-c) --> (-c-1)-X
4348 if (RHS->isAllOnesValue()) {
4349 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4350 return BinaryOperator::createSub(
4351 ConstantExpr::getSub(NegOp0CI,
4352 ConstantInt::get(I.getType(), 1)),
4353 Op0I->getOperand(0));
4354 } else if (RHS->getValue().isSignBit()) {
4355 // (X + C) ^ signbit -> (X + C + signbit)
4356 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4357 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4360 } else if (Op0I->getOpcode() == Instruction::Or) {
4361 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4362 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4363 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4364 // Anything in both C1 and C2 is known to be zero, remove it from
4366 Constant *CommonBits = And(Op0CI, RHS);
4367 NewRHS = ConstantExpr::getAnd(NewRHS,
4368 ConstantExpr::getNot(CommonBits));
4369 AddToWorkList(Op0I);
4370 I.setOperand(0, Op0I->getOperand(0));
4371 I.setOperand(1, NewRHS);
4377 // Try to fold constant and into select arguments.
4378 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4379 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4381 if (isa<PHINode>(Op0))
4382 if (Instruction *NV = FoldOpIntoPhi(I))
4386 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4388 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4390 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4392 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4395 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4398 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4399 if (A == Op0) { // B^(B|A) == (A|B)^B
4400 Op1I->swapOperands();
4402 std::swap(Op0, Op1);
4403 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4404 I.swapOperands(); // Simplified below.
4405 std::swap(Op0, Op1);
4407 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4408 if (Op0 == A) // A^(A^B) == B
4409 return ReplaceInstUsesWith(I, B);
4410 else if (Op0 == B) // A^(B^A) == B
4411 return ReplaceInstUsesWith(I, A);
4412 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4413 if (A == Op0) { // A^(A&B) -> A^(B&A)
4414 Op1I->swapOperands();
4417 if (B == Op0) { // A^(B&A) -> (B&A)^A
4418 I.swapOperands(); // Simplified below.
4419 std::swap(Op0, Op1);
4424 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4427 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4428 if (A == Op1) // (B|A)^B == (A|B)^B
4430 if (B == Op1) { // (A|B)^B == A & ~B
4432 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4433 return BinaryOperator::createAnd(A, NotB);
4435 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4436 if (Op1 == A) // (A^B)^A == B
4437 return ReplaceInstUsesWith(I, B);
4438 else if (Op1 == B) // (B^A)^A == B
4439 return ReplaceInstUsesWith(I, A);
4440 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4441 if (A == Op1) // (A&B)^A -> (B&A)^A
4443 if (B == Op1 && // (B&A)^A == ~B & A
4444 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4446 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4447 return BinaryOperator::createAnd(N, Op1);
4452 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4453 if (Op0I && Op1I && Op0I->isShift() &&
4454 Op0I->getOpcode() == Op1I->getOpcode() &&
4455 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4456 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4457 Instruction *NewOp =
4458 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4459 Op1I->getOperand(0),
4460 Op0I->getName()), I);
4461 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4462 Op1I->getOperand(1));
4466 Value *A, *B, *C, *D;
4467 // (A & B)^(A | B) -> A ^ B
4468 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4469 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4470 if ((A == C && B == D) || (A == D && B == C))
4471 return BinaryOperator::createXor(A, B);
4473 // (A | B)^(A & B) -> A ^ B
4474 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4475 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4476 if ((A == C && B == D) || (A == D && B == C))
4477 return BinaryOperator::createXor(A, B);
4481 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4482 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4483 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4484 // (X & Y)^(X & Y) -> (Y^Z) & X
4485 Value *X = 0, *Y = 0, *Z = 0;
4487 X = A, Y = B, Z = D;
4489 X = A, Y = B, Z = C;
4491 X = B, Y = A, Z = D;
4493 X = B, Y = A, Z = C;
4496 Instruction *NewOp =
4497 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4498 return BinaryOperator::createAnd(NewOp, X);
4503 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4504 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4505 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4508 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4509 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4510 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4511 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4512 const Type *SrcTy = Op0C->getOperand(0)->getType();
4513 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4514 // Only do this if the casts both really cause code to be generated.
4515 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4517 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4519 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4520 Op1C->getOperand(0),
4522 InsertNewInstBefore(NewOp, I);
4523 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4527 return Changed ? &I : 0;
4530 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4531 /// overflowed for this type.
4532 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4533 ConstantInt *In2, bool IsSigned = false) {
4534 Result = cast<ConstantInt>(Add(In1, In2));
4537 if (In2->getValue().isNegative())
4538 return Result->getValue().sgt(In1->getValue());
4540 return Result->getValue().slt(In1->getValue());
4542 return Result->getValue().ult(In1->getValue());
4545 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4546 /// code necessary to compute the offset from the base pointer (without adding
4547 /// in the base pointer). Return the result as a signed integer of intptr size.
4548 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4549 TargetData &TD = IC.getTargetData();
4550 gep_type_iterator GTI = gep_type_begin(GEP);
4551 const Type *IntPtrTy = TD.getIntPtrType();
4552 Value *Result = Constant::getNullValue(IntPtrTy);
4554 // Build a mask for high order bits.
4555 unsigned IntPtrWidth = TD.getPointerSize()*8;
4556 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4558 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4559 Value *Op = GEP->getOperand(i);
4560 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4561 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4562 if (OpC->isZero()) continue;
4564 // Handle a struct index, which adds its field offset to the pointer.
4565 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4566 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4568 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4569 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4571 Result = IC.InsertNewInstBefore(
4572 BinaryOperator::createAdd(Result,
4573 ConstantInt::get(IntPtrTy, Size),
4574 GEP->getName()+".offs"), I);
4578 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4579 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4580 Scale = ConstantExpr::getMul(OC, Scale);
4581 if (Constant *RC = dyn_cast<Constant>(Result))
4582 Result = ConstantExpr::getAdd(RC, Scale);
4584 // Emit an add instruction.
4585 Result = IC.InsertNewInstBefore(
4586 BinaryOperator::createAdd(Result, Scale,
4587 GEP->getName()+".offs"), I);
4591 // Convert to correct type.
4592 if (Op->getType() != IntPtrTy) {
4593 if (Constant *OpC = dyn_cast<Constant>(Op))
4594 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4596 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4597 Op->getName()+".c"), I);
4600 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4601 if (Constant *OpC = dyn_cast<Constant>(Op))
4602 Op = ConstantExpr::getMul(OpC, Scale);
4603 else // We'll let instcombine(mul) convert this to a shl if possible.
4604 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4605 GEP->getName()+".idx"), I);
4608 // Emit an add instruction.
4609 if (isa<Constant>(Op) && isa<Constant>(Result))
4610 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4611 cast<Constant>(Result));
4613 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4614 GEP->getName()+".offs"), I);
4619 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4620 /// else. At this point we know that the GEP is on the LHS of the comparison.
4621 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4622 ICmpInst::Predicate Cond,
4624 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4626 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4627 if (isa<PointerType>(CI->getOperand(0)->getType()))
4628 RHS = CI->getOperand(0);
4630 Value *PtrBase = GEPLHS->getOperand(0);
4631 if (PtrBase == RHS) {
4632 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4633 // This transformation is valid because we know pointers can't overflow.
4634 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4635 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4636 Constant::getNullValue(Offset->getType()));
4637 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4638 // If the base pointers are different, but the indices are the same, just
4639 // compare the base pointer.
4640 if (PtrBase != GEPRHS->getOperand(0)) {
4641 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4642 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4643 GEPRHS->getOperand(0)->getType();
4645 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4646 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4647 IndicesTheSame = false;
4651 // If all indices are the same, just compare the base pointers.
4653 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4654 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4656 // Otherwise, the base pointers are different and the indices are
4657 // different, bail out.
4661 // If one of the GEPs has all zero indices, recurse.
4662 bool AllZeros = true;
4663 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4664 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4665 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4670 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4671 ICmpInst::getSwappedPredicate(Cond), I);
4673 // If the other GEP has all zero indices, recurse.
4675 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4676 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4677 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4682 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4684 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4685 // If the GEPs only differ by one index, compare it.
4686 unsigned NumDifferences = 0; // Keep track of # differences.
4687 unsigned DiffOperand = 0; // The operand that differs.
4688 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4689 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4690 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4691 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4692 // Irreconcilable differences.
4696 if (NumDifferences++) break;
4701 if (NumDifferences == 0) // SAME GEP?
4702 return ReplaceInstUsesWith(I, // No comparison is needed here.
4703 ConstantInt::get(Type::Int1Ty,
4704 isTrueWhenEqual(Cond)));
4706 else if (NumDifferences == 1) {
4707 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4708 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4709 // Make sure we do a signed comparison here.
4710 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4714 // Only lower this if the icmp is the only user of the GEP or if we expect
4715 // the result to fold to a constant!
4716 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4717 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4718 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4719 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4720 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4721 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4727 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4728 bool Changed = SimplifyCompare(I);
4729 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4731 // Fold trivial predicates.
4732 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4733 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4734 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4735 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4737 // Simplify 'fcmp pred X, X'
4739 switch (I.getPredicate()) {
4740 default: assert(0 && "Unknown predicate!");
4741 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4742 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4743 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4744 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4745 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4746 case FCmpInst::FCMP_OLT: // True if ordered and less than
4747 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4748 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4750 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4751 case FCmpInst::FCMP_ULT: // True if unordered or less than
4752 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4753 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4754 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4755 I.setPredicate(FCmpInst::FCMP_UNO);
4756 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4759 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4760 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4761 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4762 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4763 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4764 I.setPredicate(FCmpInst::FCMP_ORD);
4765 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4770 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4771 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4773 // Handle fcmp with constant RHS
4774 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4775 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4776 switch (LHSI->getOpcode()) {
4777 case Instruction::PHI:
4778 if (Instruction *NV = FoldOpIntoPhi(I))
4781 case Instruction::Select:
4782 // If either operand of the select is a constant, we can fold the
4783 // comparison into the select arms, which will cause one to be
4784 // constant folded and the select turned into a bitwise or.
4785 Value *Op1 = 0, *Op2 = 0;
4786 if (LHSI->hasOneUse()) {
4787 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4788 // Fold the known value into the constant operand.
4789 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4790 // Insert a new FCmp of the other select operand.
4791 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4792 LHSI->getOperand(2), RHSC,
4794 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4795 // Fold the known value into the constant operand.
4796 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4797 // Insert a new FCmp of the other select operand.
4798 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4799 LHSI->getOperand(1), RHSC,
4805 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4810 return Changed ? &I : 0;
4813 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4814 bool Changed = SimplifyCompare(I);
4815 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4816 const Type *Ty = Op0->getType();
4820 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4821 isTrueWhenEqual(I)));
4823 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4824 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4826 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4827 // addresses never equal each other! We already know that Op0 != Op1.
4828 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4829 isa<ConstantPointerNull>(Op0)) &&
4830 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4831 isa<ConstantPointerNull>(Op1)))
4832 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4833 !isTrueWhenEqual(I)));
4835 // icmp's with boolean values can always be turned into bitwise operations
4836 if (Ty == Type::Int1Ty) {
4837 switch (I.getPredicate()) {
4838 default: assert(0 && "Invalid icmp instruction!");
4839 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4840 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4841 InsertNewInstBefore(Xor, I);
4842 return BinaryOperator::createNot(Xor);
4844 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4845 return BinaryOperator::createXor(Op0, Op1);
4847 case ICmpInst::ICMP_UGT:
4848 case ICmpInst::ICMP_SGT:
4849 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4851 case ICmpInst::ICMP_ULT:
4852 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4853 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4854 InsertNewInstBefore(Not, I);
4855 return BinaryOperator::createAnd(Not, Op1);
4857 case ICmpInst::ICMP_UGE:
4858 case ICmpInst::ICMP_SGE:
4859 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4861 case ICmpInst::ICMP_ULE:
4862 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4863 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4864 InsertNewInstBefore(Not, I);
4865 return BinaryOperator::createOr(Not, Op1);
4870 // See if we are doing a comparison between a constant and an instruction that
4871 // can be folded into the comparison.
4872 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4875 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
4876 if (I.isEquality() && CI->isNullValue() &&
4877 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
4878 // (icmp cond A B) if cond is equality
4879 return new ICmpInst(I.getPredicate(), A, B);
4882 switch (I.getPredicate()) {
4884 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4885 if (CI->isMinValue(false))
4886 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4887 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4888 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4889 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4890 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4891 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4892 if (CI->isMinValue(true))
4893 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4894 ConstantInt::getAllOnesValue(Op0->getType()));
4898 case ICmpInst::ICMP_SLT:
4899 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4900 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4901 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4902 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4903 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4904 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4907 case ICmpInst::ICMP_UGT:
4908 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4909 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4910 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4911 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4912 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4913 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4915 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4916 if (CI->isMaxValue(true))
4917 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4918 ConstantInt::getNullValue(Op0->getType()));
4921 case ICmpInst::ICMP_SGT:
4922 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4923 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4924 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4925 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4926 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4927 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4930 case ICmpInst::ICMP_ULE:
4931 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4932 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4933 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4934 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4935 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4936 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4939 case ICmpInst::ICMP_SLE:
4940 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4941 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4942 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4943 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4944 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4945 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4948 case ICmpInst::ICMP_UGE:
4949 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4950 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4951 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4952 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4953 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4954 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4957 case ICmpInst::ICMP_SGE:
4958 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4959 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4960 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4961 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4962 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4963 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4967 // If we still have a icmp le or icmp ge instruction, turn it into the
4968 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4969 // already been handled above, this requires little checking.
4971 switch (I.getPredicate()) {
4973 case ICmpInst::ICMP_ULE:
4974 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4975 case ICmpInst::ICMP_SLE:
4976 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4977 case ICmpInst::ICMP_UGE:
4978 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4979 case ICmpInst::ICMP_SGE:
4980 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4983 // See if we can fold the comparison based on bits known to be zero or one
4984 // in the input. If this comparison is a normal comparison, it demands all
4985 // bits, if it is a sign bit comparison, it only demands the sign bit.
4988 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4990 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4991 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4992 if (SimplifyDemandedBits(Op0,
4993 isSignBit ? APInt::getSignBit(BitWidth)
4994 : APInt::getAllOnesValue(BitWidth),
4995 KnownZero, KnownOne, 0))
4998 // Given the known and unknown bits, compute a range that the LHS could be
5000 if ((KnownOne | KnownZero) != 0) {
5001 // Compute the Min, Max and RHS values based on the known bits. For the
5002 // EQ and NE we use unsigned values.
5003 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5004 const APInt& RHSVal = CI->getValue();
5005 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5006 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5009 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5012 switch (I.getPredicate()) { // LE/GE have been folded already.
5013 default: assert(0 && "Unknown icmp opcode!");
5014 case ICmpInst::ICMP_EQ:
5015 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5016 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5018 case ICmpInst::ICMP_NE:
5019 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5020 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5022 case ICmpInst::ICMP_ULT:
5023 if (Max.ult(RHSVal))
5024 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5025 if (Min.uge(RHSVal))
5026 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5028 case ICmpInst::ICMP_UGT:
5029 if (Min.ugt(RHSVal))
5030 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5031 if (Max.ule(RHSVal))
5032 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5034 case ICmpInst::ICMP_SLT:
5035 if (Max.slt(RHSVal))
5036 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5037 if (Min.sgt(RHSVal))
5038 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5040 case ICmpInst::ICMP_SGT:
5041 if (Min.sgt(RHSVal))
5042 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5043 if (Max.sle(RHSVal))
5044 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5049 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5050 // instruction, see if that instruction also has constants so that the
5051 // instruction can be folded into the icmp
5052 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5053 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5057 // Handle icmp with constant (but not simple integer constant) RHS
5058 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5059 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5060 switch (LHSI->getOpcode()) {
5061 case Instruction::GetElementPtr:
5062 if (RHSC->isNullValue()) {
5063 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5064 bool isAllZeros = true;
5065 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5066 if (!isa<Constant>(LHSI->getOperand(i)) ||
5067 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5072 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5073 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5077 case Instruction::PHI:
5078 if (Instruction *NV = FoldOpIntoPhi(I))
5081 case Instruction::Select: {
5082 // If either operand of the select is a constant, we can fold the
5083 // comparison into the select arms, which will cause one to be
5084 // constant folded and the select turned into a bitwise or.
5085 Value *Op1 = 0, *Op2 = 0;
5086 if (LHSI->hasOneUse()) {
5087 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5088 // Fold the known value into the constant operand.
5089 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5090 // Insert a new ICmp of the other select operand.
5091 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5092 LHSI->getOperand(2), RHSC,
5094 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5095 // Fold the known value into the constant operand.
5096 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5097 // Insert a new ICmp of the other select operand.
5098 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5099 LHSI->getOperand(1), RHSC,
5105 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5108 case Instruction::Malloc:
5109 // If we have (malloc != null), and if the malloc has a single use, we
5110 // can assume it is successful and remove the malloc.
5111 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5112 AddToWorkList(LHSI);
5113 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5114 !isTrueWhenEqual(I)));
5120 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5121 if (User *GEP = dyn_castGetElementPtr(Op0))
5122 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5124 if (User *GEP = dyn_castGetElementPtr(Op1))
5125 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5126 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5129 // Test to see if the operands of the icmp are casted versions of other
5130 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5132 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5133 if (isa<PointerType>(Op0->getType()) &&
5134 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5135 // We keep moving the cast from the left operand over to the right
5136 // operand, where it can often be eliminated completely.
5137 Op0 = CI->getOperand(0);
5139 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5140 // so eliminate it as well.
5141 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5142 Op1 = CI2->getOperand(0);
5144 // If Op1 is a constant, we can fold the cast into the constant.
5145 if (Op0->getType() != Op1->getType())
5146 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5147 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5149 // Otherwise, cast the RHS right before the icmp
5150 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5152 return new ICmpInst(I.getPredicate(), Op0, Op1);
5156 if (isa<CastInst>(Op0)) {
5157 // Handle the special case of: icmp (cast bool to X), <cst>
5158 // This comes up when you have code like
5161 // For generality, we handle any zero-extension of any operand comparison
5162 // with a constant or another cast from the same type.
5163 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5164 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5168 if (I.isEquality()) {
5169 Value *A, *B, *C, *D;
5170 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5171 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5172 Value *OtherVal = A == Op1 ? B : A;
5173 return new ICmpInst(I.getPredicate(), OtherVal,
5174 Constant::getNullValue(A->getType()));
5177 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5178 // A^c1 == C^c2 --> A == C^(c1^c2)
5179 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5180 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5181 if (Op1->hasOneUse()) {
5182 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5183 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5184 return new ICmpInst(I.getPredicate(), A,
5185 InsertNewInstBefore(Xor, I));
5188 // A^B == A^D -> B == D
5189 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5190 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5191 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5192 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5196 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5197 (A == Op0 || B == Op0)) {
5198 // A == (A^B) -> B == 0
5199 Value *OtherVal = A == Op0 ? B : A;
5200 return new ICmpInst(I.getPredicate(), OtherVal,
5201 Constant::getNullValue(A->getType()));
5203 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5204 // (A-B) == A -> B == 0
5205 return new ICmpInst(I.getPredicate(), B,
5206 Constant::getNullValue(B->getType()));
5208 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5209 // A == (A-B) -> B == 0
5210 return new ICmpInst(I.getPredicate(), B,
5211 Constant::getNullValue(B->getType()));
5214 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5215 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5216 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5217 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5218 Value *X = 0, *Y = 0, *Z = 0;
5221 X = B; Y = D; Z = A;
5222 } else if (A == D) {
5223 X = B; Y = C; Z = A;
5224 } else if (B == C) {
5225 X = A; Y = D; Z = B;
5226 } else if (B == D) {
5227 X = A; Y = C; Z = B;
5230 if (X) { // Build (X^Y) & Z
5231 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5232 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5233 I.setOperand(0, Op1);
5234 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5239 return Changed ? &I : 0;
5243 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5244 /// and CmpRHS are both known to be integer constants.
5245 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5246 ConstantInt *DivRHS) {
5247 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5248 const APInt &CmpRHSV = CmpRHS->getValue();
5250 // FIXME: If the operand types don't match the type of the divide
5251 // then don't attempt this transform. The code below doesn't have the
5252 // logic to deal with a signed divide and an unsigned compare (and
5253 // vice versa). This is because (x /s C1) <s C2 produces different
5254 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5255 // (x /u C1) <u C2. Simply casting the operands and result won't
5256 // work. :( The if statement below tests that condition and bails
5258 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5259 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5261 if (DivRHS->isZero())
5262 return 0; // The ProdOV computation fails on divide by zero.
5264 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5265 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5266 // C2 (CI). By solving for X we can turn this into a range check
5267 // instead of computing a divide.
5268 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5270 // Determine if the product overflows by seeing if the product is
5271 // not equal to the divide. Make sure we do the same kind of divide
5272 // as in the LHS instruction that we're folding.
5273 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5274 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5276 // Get the ICmp opcode
5277 ICmpInst::Predicate Pred = ICI.getPredicate();
5279 // Figure out the interval that is being checked. For example, a comparison
5280 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5281 // Compute this interval based on the constants involved and the signedness of
5282 // the compare/divide. This computes a half-open interval, keeping track of
5283 // whether either value in the interval overflows. After analysis each
5284 // overflow variable is set to 0 if it's corresponding bound variable is valid
5285 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5286 int LoOverflow = 0, HiOverflow = 0;
5287 ConstantInt *LoBound = 0, *HiBound = 0;
5290 if (!DivIsSigned) { // udiv
5291 // e.g. X/5 op 3 --> [15, 20)
5293 HiOverflow = LoOverflow = ProdOV;
5295 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5296 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5297 if (CmpRHSV == 0) { // (X / pos) op 0
5298 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5299 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5301 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5302 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5303 HiOverflow = LoOverflow = ProdOV;
5305 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5306 } else { // (X / pos) op neg
5307 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5308 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5309 LoOverflow = AddWithOverflow(LoBound, Prod,
5310 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5311 HiBound = AddOne(Prod);
5312 HiOverflow = ProdOV ? -1 : 0;
5314 } else { // Divisor is < 0.
5315 if (CmpRHSV == 0) { // (X / neg) op 0
5316 // e.g. X/-5 op 0 --> [-4, 5)
5317 LoBound = AddOne(DivRHS);
5318 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5319 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5320 HiOverflow = 1; // [INTMIN+1, overflow)
5321 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5323 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5324 // e.g. X/-5 op 3 --> [-19, -14)
5325 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5327 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5328 HiBound = AddOne(Prod);
5329 } else { // (X / neg) op neg
5330 // e.g. X/-5 op -3 --> [15, 20)
5332 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5333 HiBound = Subtract(Prod, DivRHS);
5336 // Dividing by a negative swaps the condition. LT <-> GT
5337 Pred = ICmpInst::getSwappedPredicate(Pred);
5340 Value *X = DivI->getOperand(0);
5342 default: assert(0 && "Unhandled icmp opcode!");
5343 case ICmpInst::ICMP_EQ:
5344 if (LoOverflow && HiOverflow)
5345 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5346 else if (HiOverflow)
5347 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5348 ICmpInst::ICMP_UGE, X, LoBound);
5349 else if (LoOverflow)
5350 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5351 ICmpInst::ICMP_ULT, X, HiBound);
5353 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5354 case ICmpInst::ICMP_NE:
5355 if (LoOverflow && HiOverflow)
5356 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5357 else if (HiOverflow)
5358 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5359 ICmpInst::ICMP_ULT, X, LoBound);
5360 else if (LoOverflow)
5361 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5362 ICmpInst::ICMP_UGE, X, HiBound);
5364 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5365 case ICmpInst::ICMP_ULT:
5366 case ICmpInst::ICMP_SLT:
5367 if (LoOverflow == +1) // Low bound is greater than input range.
5368 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5369 if (LoOverflow == -1) // Low bound is less than input range.
5370 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5371 return new ICmpInst(Pred, X, LoBound);
5372 case ICmpInst::ICMP_UGT:
5373 case ICmpInst::ICMP_SGT:
5374 if (HiOverflow == +1) // High bound greater than input range.
5375 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5376 else if (HiOverflow == -1) // High bound less than input range.
5377 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5378 if (Pred == ICmpInst::ICMP_UGT)
5379 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5381 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5386 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5388 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5391 const APInt &RHSV = RHS->getValue();
5393 switch (LHSI->getOpcode()) {
5394 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5395 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5396 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5398 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5399 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5400 Value *CompareVal = LHSI->getOperand(0);
5402 // If the sign bit of the XorCST is not set, there is no change to
5403 // the operation, just stop using the Xor.
5404 if (!XorCST->getValue().isNegative()) {
5405 ICI.setOperand(0, CompareVal);
5406 AddToWorkList(LHSI);
5410 // Was the old condition true if the operand is positive?
5411 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5413 // If so, the new one isn't.
5414 isTrueIfPositive ^= true;
5416 if (isTrueIfPositive)
5417 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5419 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5423 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5424 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5425 LHSI->getOperand(0)->hasOneUse()) {
5426 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5428 // If the LHS is an AND of a truncating cast, we can widen the
5429 // and/compare to be the input width without changing the value
5430 // produced, eliminating a cast.
5431 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5432 // We can do this transformation if either the AND constant does not
5433 // have its sign bit set or if it is an equality comparison.
5434 // Extending a relational comparison when we're checking the sign
5435 // bit would not work.
5436 if (Cast->hasOneUse() &&
5437 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5438 RHSV.isPositive())) {
5440 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5441 APInt NewCST = AndCST->getValue();
5442 NewCST.zext(BitWidth);
5444 NewCI.zext(BitWidth);
5445 Instruction *NewAnd =
5446 BinaryOperator::createAnd(Cast->getOperand(0),
5447 ConstantInt::get(NewCST),LHSI->getName());
5448 InsertNewInstBefore(NewAnd, ICI);
5449 return new ICmpInst(ICI.getPredicate(), NewAnd,
5450 ConstantInt::get(NewCI));
5454 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5455 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5456 // happens a LOT in code produced by the C front-end, for bitfield
5458 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5459 if (Shift && !Shift->isShift())
5463 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5464 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5465 const Type *AndTy = AndCST->getType(); // Type of the and.
5467 // We can fold this as long as we can't shift unknown bits
5468 // into the mask. This can only happen with signed shift
5469 // rights, as they sign-extend.
5471 bool CanFold = Shift->isLogicalShift();
5473 // To test for the bad case of the signed shr, see if any
5474 // of the bits shifted in could be tested after the mask.
5475 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5476 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5478 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5479 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5480 AndCST->getValue()) == 0)
5486 if (Shift->getOpcode() == Instruction::Shl)
5487 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5489 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5491 // Check to see if we are shifting out any of the bits being
5493 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5494 // If we shifted bits out, the fold is not going to work out.
5495 // As a special case, check to see if this means that the
5496 // result is always true or false now.
5497 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5498 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5499 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5500 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5502 ICI.setOperand(1, NewCst);
5503 Constant *NewAndCST;
5504 if (Shift->getOpcode() == Instruction::Shl)
5505 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5507 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5508 LHSI->setOperand(1, NewAndCST);
5509 LHSI->setOperand(0, Shift->getOperand(0));
5510 AddToWorkList(Shift); // Shift is dead.
5511 AddUsesToWorkList(ICI);
5517 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5518 // preferable because it allows the C<<Y expression to be hoisted out
5519 // of a loop if Y is invariant and X is not.
5520 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5521 ICI.isEquality() && !Shift->isArithmeticShift() &&
5522 isa<Instruction>(Shift->getOperand(0))) {
5525 if (Shift->getOpcode() == Instruction::LShr) {
5526 NS = BinaryOperator::createShl(AndCST,
5527 Shift->getOperand(1), "tmp");
5529 // Insert a logical shift.
5530 NS = BinaryOperator::createLShr(AndCST,
5531 Shift->getOperand(1), "tmp");
5533 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5535 // Compute X & (C << Y).
5536 Instruction *NewAnd =
5537 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5538 InsertNewInstBefore(NewAnd, ICI);
5540 ICI.setOperand(0, NewAnd);
5546 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5547 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5550 uint32_t TypeBits = RHSV.getBitWidth();
5552 // Check that the shift amount is in range. If not, don't perform
5553 // undefined shifts. When the shift is visited it will be
5555 if (ShAmt->uge(TypeBits))
5558 if (ICI.isEquality()) {
5559 // If we are comparing against bits always shifted out, the
5560 // comparison cannot succeed.
5562 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5563 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5564 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5565 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5566 return ReplaceInstUsesWith(ICI, Cst);
5569 if (LHSI->hasOneUse()) {
5570 // Otherwise strength reduce the shift into an and.
5571 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5573 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5576 BinaryOperator::createAnd(LHSI->getOperand(0),
5577 Mask, LHSI->getName()+".mask");
5578 Value *And = InsertNewInstBefore(AndI, ICI);
5579 return new ICmpInst(ICI.getPredicate(), And,
5580 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5584 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5585 bool TrueIfSigned = false;
5586 if (LHSI->hasOneUse() &&
5587 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5588 // (X << 31) <s 0 --> (X&1) != 0
5589 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5590 (TypeBits-ShAmt->getZExtValue()-1));
5592 BinaryOperator::createAnd(LHSI->getOperand(0),
5593 Mask, LHSI->getName()+".mask");
5594 Value *And = InsertNewInstBefore(AndI, ICI);
5596 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5597 And, Constant::getNullValue(And->getType()));
5602 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5603 case Instruction::AShr: {
5604 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5607 if (ICI.isEquality()) {
5608 // Check that the shift amount is in range. If not, don't perform
5609 // undefined shifts. When the shift is visited it will be
5611 uint32_t TypeBits = RHSV.getBitWidth();
5612 if (ShAmt->uge(TypeBits))
5614 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5616 // If we are comparing against bits always shifted out, the
5617 // comparison cannot succeed.
5618 APInt Comp = RHSV << ShAmtVal;
5619 if (LHSI->getOpcode() == Instruction::LShr)
5620 Comp = Comp.lshr(ShAmtVal);
5622 Comp = Comp.ashr(ShAmtVal);
5624 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5625 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5626 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5627 return ReplaceInstUsesWith(ICI, Cst);
5630 if (LHSI->hasOneUse() || RHSV == 0) {
5631 // Otherwise strength reduce the shift into an and.
5632 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5633 Constant *Mask = ConstantInt::get(Val);
5636 BinaryOperator::createAnd(LHSI->getOperand(0),
5637 Mask, LHSI->getName()+".mask");
5638 Value *And = InsertNewInstBefore(AndI, ICI);
5639 return new ICmpInst(ICI.getPredicate(), And,
5640 ConstantExpr::getShl(RHS, ShAmt));
5646 case Instruction::SDiv:
5647 case Instruction::UDiv:
5648 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5649 // Fold this div into the comparison, producing a range check.
5650 // Determine, based on the divide type, what the range is being
5651 // checked. If there is an overflow on the low or high side, remember
5652 // it, otherwise compute the range [low, hi) bounding the new value.
5653 // See: InsertRangeTest above for the kinds of replacements possible.
5654 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5655 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5660 case Instruction::Add:
5661 // Fold: icmp pred (add, X, C1), C2
5663 if (!ICI.isEquality()) {
5664 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5666 const APInt &LHSV = LHSC->getValue();
5668 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
5671 if (ICI.isSignedPredicate()) {
5672 if (CR.getLower().isSignBit()) {
5673 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
5674 ConstantInt::get(CR.getUpper()));
5675 } else if (CR.getUpper().isSignBit()) {
5676 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
5677 ConstantInt::get(CR.getLower()));
5680 if (CR.getLower().isMinValue()) {
5681 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
5682 ConstantInt::get(CR.getUpper()));
5683 } else if (CR.getUpper().isMinValue()) {
5684 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
5685 ConstantInt::get(CR.getLower()));
5692 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5693 if (ICI.isEquality()) {
5694 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5696 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5697 // the second operand is a constant, simplify a bit.
5698 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5699 switch (BO->getOpcode()) {
5700 case Instruction::SRem:
5701 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5702 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5703 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5704 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5705 Instruction *NewRem =
5706 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5708 InsertNewInstBefore(NewRem, ICI);
5709 return new ICmpInst(ICI.getPredicate(), NewRem,
5710 Constant::getNullValue(BO->getType()));
5714 case Instruction::Add:
5715 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5716 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5717 if (BO->hasOneUse())
5718 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5719 Subtract(RHS, BOp1C));
5720 } else if (RHSV == 0) {
5721 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5722 // efficiently invertible, or if the add has just this one use.
5723 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5725 if (Value *NegVal = dyn_castNegVal(BOp1))
5726 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5727 else if (Value *NegVal = dyn_castNegVal(BOp0))
5728 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5729 else if (BO->hasOneUse()) {
5730 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5731 InsertNewInstBefore(Neg, ICI);
5733 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5737 case Instruction::Xor:
5738 // For the xor case, we can xor two constants together, eliminating
5739 // the explicit xor.
5740 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5741 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5742 ConstantExpr::getXor(RHS, BOC));
5745 case Instruction::Sub:
5746 // Replace (([sub|xor] A, B) != 0) with (A != B)
5748 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5752 case Instruction::Or:
5753 // If bits are being or'd in that are not present in the constant we
5754 // are comparing against, then the comparison could never succeed!
5755 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5756 Constant *NotCI = ConstantExpr::getNot(RHS);
5757 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5758 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5763 case Instruction::And:
5764 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5765 // If bits are being compared against that are and'd out, then the
5766 // comparison can never succeed!
5767 if ((RHSV & ~BOC->getValue()) != 0)
5768 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5771 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5772 if (RHS == BOC && RHSV.isPowerOf2())
5773 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5774 ICmpInst::ICMP_NE, LHSI,
5775 Constant::getNullValue(RHS->getType()));
5777 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5778 if (isSignBit(BOC)) {
5779 Value *X = BO->getOperand(0);
5780 Constant *Zero = Constant::getNullValue(X->getType());
5781 ICmpInst::Predicate pred = isICMP_NE ?
5782 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5783 return new ICmpInst(pred, X, Zero);
5786 // ((X & ~7) == 0) --> X < 8
5787 if (RHSV == 0 && isHighOnes(BOC)) {
5788 Value *X = BO->getOperand(0);
5789 Constant *NegX = ConstantExpr::getNeg(BOC);
5790 ICmpInst::Predicate pred = isICMP_NE ?
5791 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5792 return new ICmpInst(pred, X, NegX);
5797 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5798 // Handle icmp {eq|ne} <intrinsic>, intcst.
5799 if (II->getIntrinsicID() == Intrinsic::bswap) {
5801 ICI.setOperand(0, II->getOperand(1));
5802 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5806 } else { // Not a ICMP_EQ/ICMP_NE
5807 // If the LHS is a cast from an integral value of the same size,
5808 // then since we know the RHS is a constant, try to simlify.
5809 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5810 Value *CastOp = Cast->getOperand(0);
5811 const Type *SrcTy = CastOp->getType();
5812 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5813 if (SrcTy->isInteger() &&
5814 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5815 // If this is an unsigned comparison, try to make the comparison use
5816 // smaller constant values.
5817 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5818 // X u< 128 => X s> -1
5819 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5820 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5821 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5822 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5823 // X u> 127 => X s< 0
5824 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5825 Constant::getNullValue(SrcTy));
5833 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5834 /// We only handle extending casts so far.
5836 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5837 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5838 Value *LHSCIOp = LHSCI->getOperand(0);
5839 const Type *SrcTy = LHSCIOp->getType();
5840 const Type *DestTy = LHSCI->getType();
5843 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5844 // integer type is the same size as the pointer type.
5845 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5846 getTargetData().getPointerSizeInBits() ==
5847 cast<IntegerType>(DestTy)->getBitWidth()) {
5849 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5850 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5851 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5852 RHSOp = RHSC->getOperand(0);
5853 // If the pointer types don't match, insert a bitcast.
5854 if (LHSCIOp->getType() != RHSOp->getType())
5855 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
5859 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5862 // The code below only handles extension cast instructions, so far.
5864 if (LHSCI->getOpcode() != Instruction::ZExt &&
5865 LHSCI->getOpcode() != Instruction::SExt)
5868 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5869 bool isSignedCmp = ICI.isSignedPredicate();
5871 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5872 // Not an extension from the same type?
5873 RHSCIOp = CI->getOperand(0);
5874 if (RHSCIOp->getType() != LHSCIOp->getType())
5877 // If the signedness of the two casts doesn't agree (i.e. one is a sext
5878 // and the other is a zext), then we can't handle this.
5879 if (CI->getOpcode() != LHSCI->getOpcode())
5882 // Deal with equality cases early.
5883 if (ICI.isEquality())
5884 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5886 // A signed comparison of sign extended values simplifies into a
5887 // signed comparison.
5888 if (isSignedCmp && isSignedExt)
5889 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5891 // The other three cases all fold into an unsigned comparison.
5892 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
5895 // If we aren't dealing with a constant on the RHS, exit early
5896 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5900 // Compute the constant that would happen if we truncated to SrcTy then
5901 // reextended to DestTy.
5902 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5903 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5905 // If the re-extended constant didn't change...
5907 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5908 // For example, we might have:
5909 // %A = sext short %X to uint
5910 // %B = icmp ugt uint %A, 1330
5911 // It is incorrect to transform this into
5912 // %B = icmp ugt short %X, 1330
5913 // because %A may have negative value.
5915 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5916 // OR operation is EQ/NE.
5917 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5918 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5923 // The re-extended constant changed so the constant cannot be represented
5924 // in the shorter type. Consequently, we cannot emit a simple comparison.
5926 // First, handle some easy cases. We know the result cannot be equal at this
5927 // point so handle the ICI.isEquality() cases
5928 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5929 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5930 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5931 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5933 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5934 // should have been folded away previously and not enter in here.
5937 // We're performing a signed comparison.
5938 if (cast<ConstantInt>(CI)->getValue().isNegative())
5939 Result = ConstantInt::getFalse(); // X < (small) --> false
5941 Result = ConstantInt::getTrue(); // X < (large) --> true
5943 // We're performing an unsigned comparison.
5945 // We're performing an unsigned comp with a sign extended value.
5946 // This is true if the input is >= 0. [aka >s -1]
5947 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5948 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5949 NegOne, ICI.getName()), ICI);
5951 // Unsigned extend & unsigned compare -> always true.
5952 Result = ConstantInt::getTrue();
5956 // Finally, return the value computed.
5957 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5958 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5959 return ReplaceInstUsesWith(ICI, Result);
5961 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5962 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5963 "ICmp should be folded!");
5964 if (Constant *CI = dyn_cast<Constant>(Result))
5965 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5967 return BinaryOperator::createNot(Result);
5971 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5972 return commonShiftTransforms(I);
5975 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5976 return commonShiftTransforms(I);
5979 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5980 if (Instruction *R = commonShiftTransforms(I))
5983 Value *Op0 = I.getOperand(0);
5985 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5986 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5987 if (CSI->isAllOnesValue())
5988 return ReplaceInstUsesWith(I, CSI);
5990 // See if we can turn a signed shr into an unsigned shr.
5991 if (MaskedValueIsZero(Op0,
5992 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
5993 return BinaryOperator::createLShr(Op0, I.getOperand(1));
5998 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5999 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6000 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6002 // shl X, 0 == X and shr X, 0 == X
6003 // shl 0, X == 0 and shr 0, X == 0
6004 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6005 Op0 == Constant::getNullValue(Op0->getType()))
6006 return ReplaceInstUsesWith(I, Op0);
6008 if (isa<UndefValue>(Op0)) {
6009 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6010 return ReplaceInstUsesWith(I, Op0);
6011 else // undef << X -> 0, undef >>u X -> 0
6012 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6014 if (isa<UndefValue>(Op1)) {
6015 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6016 return ReplaceInstUsesWith(I, Op0);
6017 else // X << undef, X >>u undef -> 0
6018 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6021 // Try to fold constant and into select arguments.
6022 if (isa<Constant>(Op0))
6023 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6024 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6027 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6028 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6033 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6034 BinaryOperator &I) {
6035 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6037 // See if we can simplify any instructions used by the instruction whose sole
6038 // purpose is to compute bits we don't care about.
6039 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6040 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6041 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6042 KnownZero, KnownOne))
6045 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6046 // of a signed value.
6048 if (Op1->uge(TypeBits)) {
6049 if (I.getOpcode() != Instruction::AShr)
6050 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6052 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6057 // ((X*C1) << C2) == (X * (C1 << C2))
6058 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6059 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6060 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6061 return BinaryOperator::createMul(BO->getOperand(0),
6062 ConstantExpr::getShl(BOOp, Op1));
6064 // Try to fold constant and into select arguments.
6065 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6066 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6068 if (isa<PHINode>(Op0))
6069 if (Instruction *NV = FoldOpIntoPhi(I))
6072 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6073 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6074 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6075 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6076 // place. Don't try to do this transformation in this case. Also, we
6077 // require that the input operand is a shift-by-constant so that we have
6078 // confidence that the shifts will get folded together. We could do this
6079 // xform in more cases, but it is unlikely to be profitable.
6080 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6081 isa<ConstantInt>(TrOp->getOperand(1))) {
6082 // Okay, we'll do this xform. Make the shift of shift.
6083 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6084 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6086 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6088 // For logical shifts, the truncation has the effect of making the high
6089 // part of the register be zeros. Emulate this by inserting an AND to
6090 // clear the top bits as needed. This 'and' will usually be zapped by
6091 // other xforms later if dead.
6092 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6093 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6094 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6096 // The mask we constructed says what the trunc would do if occurring
6097 // between the shifts. We want to know the effect *after* the second
6098 // shift. We know that it is a logical shift by a constant, so adjust the
6099 // mask as appropriate.
6100 if (I.getOpcode() == Instruction::Shl)
6101 MaskV <<= Op1->getZExtValue();
6103 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6104 MaskV = MaskV.lshr(Op1->getZExtValue());
6107 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6109 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6111 // Return the value truncated to the interesting size.
6112 return new TruncInst(And, I.getType());
6116 if (Op0->hasOneUse()) {
6117 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6118 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6121 switch (Op0BO->getOpcode()) {
6123 case Instruction::Add:
6124 case Instruction::And:
6125 case Instruction::Or:
6126 case Instruction::Xor: {
6127 // These operators commute.
6128 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6129 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6130 match(Op0BO->getOperand(1),
6131 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6132 Instruction *YS = BinaryOperator::createShl(
6133 Op0BO->getOperand(0), Op1,
6135 InsertNewInstBefore(YS, I); // (Y << C)
6137 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6138 Op0BO->getOperand(1)->getName());
6139 InsertNewInstBefore(X, I); // (X + (Y << C))
6140 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6141 return BinaryOperator::createAnd(X, ConstantInt::get(
6142 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6145 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6146 Value *Op0BOOp1 = Op0BO->getOperand(1);
6147 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6149 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6150 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6152 Instruction *YS = BinaryOperator::createShl(
6153 Op0BO->getOperand(0), Op1,
6155 InsertNewInstBefore(YS, I); // (Y << C)
6157 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6158 V1->getName()+".mask");
6159 InsertNewInstBefore(XM, I); // X & (CC << C)
6161 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6166 case Instruction::Sub: {
6167 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6168 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6169 match(Op0BO->getOperand(0),
6170 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6171 Instruction *YS = BinaryOperator::createShl(
6172 Op0BO->getOperand(1), Op1,
6174 InsertNewInstBefore(YS, I); // (Y << C)
6176 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6177 Op0BO->getOperand(0)->getName());
6178 InsertNewInstBefore(X, I); // (X + (Y << C))
6179 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6180 return BinaryOperator::createAnd(X, ConstantInt::get(
6181 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6184 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6185 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6186 match(Op0BO->getOperand(0),
6187 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6188 m_ConstantInt(CC))) && V2 == Op1 &&
6189 cast<BinaryOperator>(Op0BO->getOperand(0))
6190 ->getOperand(0)->hasOneUse()) {
6191 Instruction *YS = BinaryOperator::createShl(
6192 Op0BO->getOperand(1), Op1,
6194 InsertNewInstBefore(YS, I); // (Y << C)
6196 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6197 V1->getName()+".mask");
6198 InsertNewInstBefore(XM, I); // X & (CC << C)
6200 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6208 // If the operand is an bitwise operator with a constant RHS, and the
6209 // shift is the only use, we can pull it out of the shift.
6210 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6211 bool isValid = true; // Valid only for And, Or, Xor
6212 bool highBitSet = false; // Transform if high bit of constant set?
6214 switch (Op0BO->getOpcode()) {
6215 default: isValid = false; break; // Do not perform transform!
6216 case Instruction::Add:
6217 isValid = isLeftShift;
6219 case Instruction::Or:
6220 case Instruction::Xor:
6223 case Instruction::And:
6228 // If this is a signed shift right, and the high bit is modified
6229 // by the logical operation, do not perform the transformation.
6230 // The highBitSet boolean indicates the value of the high bit of
6231 // the constant which would cause it to be modified for this
6234 if (isValid && I.getOpcode() == Instruction::AShr)
6235 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6238 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6240 Instruction *NewShift =
6241 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6242 InsertNewInstBefore(NewShift, I);
6243 NewShift->takeName(Op0BO);
6245 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6252 // Find out if this is a shift of a shift by a constant.
6253 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6254 if (ShiftOp && !ShiftOp->isShift())
6257 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6258 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6259 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6260 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6261 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6262 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6263 Value *X = ShiftOp->getOperand(0);
6265 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6266 if (AmtSum > TypeBits)
6269 const IntegerType *Ty = cast<IntegerType>(I.getType());
6271 // Check for (X << c1) << c2 and (X >> c1) >> c2
6272 if (I.getOpcode() == ShiftOp->getOpcode()) {
6273 return BinaryOperator::create(I.getOpcode(), X,
6274 ConstantInt::get(Ty, AmtSum));
6275 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6276 I.getOpcode() == Instruction::AShr) {
6277 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6278 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6279 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6280 I.getOpcode() == Instruction::LShr) {
6281 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6282 Instruction *Shift =
6283 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6284 InsertNewInstBefore(Shift, I);
6286 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6287 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6290 // Okay, if we get here, one shift must be left, and the other shift must be
6291 // right. See if the amounts are equal.
6292 if (ShiftAmt1 == ShiftAmt2) {
6293 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6294 if (I.getOpcode() == Instruction::Shl) {
6295 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6296 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6298 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6299 if (I.getOpcode() == Instruction::LShr) {
6300 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6301 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6303 // We can simplify ((X << C) >>s C) into a trunc + sext.
6304 // NOTE: we could do this for any C, but that would make 'unusual' integer
6305 // types. For now, just stick to ones well-supported by the code
6307 const Type *SExtType = 0;
6308 switch (Ty->getBitWidth() - ShiftAmt1) {
6315 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6320 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6321 InsertNewInstBefore(NewTrunc, I);
6322 return new SExtInst(NewTrunc, Ty);
6324 // Otherwise, we can't handle it yet.
6325 } else if (ShiftAmt1 < ShiftAmt2) {
6326 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6328 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6329 if (I.getOpcode() == Instruction::Shl) {
6330 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6331 ShiftOp->getOpcode() == Instruction::AShr);
6332 Instruction *Shift =
6333 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6334 InsertNewInstBefore(Shift, I);
6336 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6337 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6340 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6341 if (I.getOpcode() == Instruction::LShr) {
6342 assert(ShiftOp->getOpcode() == Instruction::Shl);
6343 Instruction *Shift =
6344 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6345 InsertNewInstBefore(Shift, I);
6347 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6348 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6351 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6353 assert(ShiftAmt2 < ShiftAmt1);
6354 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6356 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6357 if (I.getOpcode() == Instruction::Shl) {
6358 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6359 ShiftOp->getOpcode() == Instruction::AShr);
6360 Instruction *Shift =
6361 BinaryOperator::create(ShiftOp->getOpcode(), X,
6362 ConstantInt::get(Ty, ShiftDiff));
6363 InsertNewInstBefore(Shift, I);
6365 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6366 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6369 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6370 if (I.getOpcode() == Instruction::LShr) {
6371 assert(ShiftOp->getOpcode() == Instruction::Shl);
6372 Instruction *Shift =
6373 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6374 InsertNewInstBefore(Shift, I);
6376 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6377 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6380 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6387 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6388 /// expression. If so, decompose it, returning some value X, such that Val is
6391 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6393 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6394 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6395 Offset = CI->getZExtValue();
6397 return ConstantInt::get(Type::Int32Ty, 0);
6398 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6399 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6400 if (I->getOpcode() == Instruction::Shl) {
6401 // This is a value scaled by '1 << the shift amt'.
6402 Scale = 1U << RHS->getZExtValue();
6404 return I->getOperand(0);
6405 } else if (I->getOpcode() == Instruction::Mul) {
6406 // This value is scaled by 'RHS'.
6407 Scale = RHS->getZExtValue();
6409 return I->getOperand(0);
6410 } else if (I->getOpcode() == Instruction::Add) {
6411 // We have X+C. Check to see if we really have (X*C2)+C1,
6412 // where C1 is divisible by C2.
6415 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6416 Offset += RHS->getZExtValue();
6423 // Otherwise, we can't look past this.
6430 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6431 /// try to eliminate the cast by moving the type information into the alloc.
6432 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6433 AllocationInst &AI) {
6434 const PointerType *PTy = cast<PointerType>(CI.getType());
6436 // Remove any uses of AI that are dead.
6437 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6439 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6440 Instruction *User = cast<Instruction>(*UI++);
6441 if (isInstructionTriviallyDead(User)) {
6442 while (UI != E && *UI == User)
6443 ++UI; // If this instruction uses AI more than once, don't break UI.
6446 DOUT << "IC: DCE: " << *User;
6447 EraseInstFromFunction(*User);
6451 // Get the type really allocated and the type casted to.
6452 const Type *AllocElTy = AI.getAllocatedType();
6453 const Type *CastElTy = PTy->getElementType();
6454 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6456 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6457 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6458 if (CastElTyAlign < AllocElTyAlign) return 0;
6460 // If the allocation has multiple uses, only promote it if we are strictly
6461 // increasing the alignment of the resultant allocation. If we keep it the
6462 // same, we open the door to infinite loops of various kinds.
6463 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6465 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6466 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6467 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6469 // See if we can satisfy the modulus by pulling a scale out of the array
6471 unsigned ArraySizeScale;
6473 Value *NumElements = // See if the array size is a decomposable linear expr.
6474 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6476 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6478 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6479 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6481 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6486 // If the allocation size is constant, form a constant mul expression
6487 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6488 if (isa<ConstantInt>(NumElements))
6489 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6490 // otherwise multiply the amount and the number of elements
6491 else if (Scale != 1) {
6492 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6493 Amt = InsertNewInstBefore(Tmp, AI);
6497 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6498 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6499 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6500 Amt = InsertNewInstBefore(Tmp, AI);
6503 AllocationInst *New;
6504 if (isa<MallocInst>(AI))
6505 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6507 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6508 InsertNewInstBefore(New, AI);
6511 // If the allocation has multiple uses, insert a cast and change all things
6512 // that used it to use the new cast. This will also hack on CI, but it will
6514 if (!AI.hasOneUse()) {
6515 AddUsesToWorkList(AI);
6516 // New is the allocation instruction, pointer typed. AI is the original
6517 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6518 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6519 InsertNewInstBefore(NewCast, AI);
6520 AI.replaceAllUsesWith(NewCast);
6522 return ReplaceInstUsesWith(CI, New);
6525 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6526 /// and return it as type Ty without inserting any new casts and without
6527 /// changing the computed value. This is used by code that tries to decide
6528 /// whether promoting or shrinking integer operations to wider or smaller types
6529 /// will allow us to eliminate a truncate or extend.
6531 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6532 /// extension operation if Ty is larger.
6533 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6534 unsigned CastOpc, int &NumCastsRemoved) {
6535 // We can always evaluate constants in another type.
6536 if (isa<ConstantInt>(V))
6539 Instruction *I = dyn_cast<Instruction>(V);
6540 if (!I) return false;
6542 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6544 // If this is an extension or truncate, we can often eliminate it.
6545 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6546 // If this is a cast from the destination type, we can trivially eliminate
6547 // it, and this will remove a cast overall.
6548 if (I->getOperand(0)->getType() == Ty) {
6549 // If the first operand is itself a cast, and is eliminable, do not count
6550 // this as an eliminable cast. We would prefer to eliminate those two
6552 if (!isa<CastInst>(I->getOperand(0)))
6558 // We can't extend or shrink something that has multiple uses: doing so would
6559 // require duplicating the instruction in general, which isn't profitable.
6560 if (!I->hasOneUse()) return false;
6562 switch (I->getOpcode()) {
6563 case Instruction::Add:
6564 case Instruction::Sub:
6565 case Instruction::And:
6566 case Instruction::Or:
6567 case Instruction::Xor:
6568 // These operators can all arbitrarily be extended or truncated.
6569 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6571 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6574 case Instruction::Mul:
6575 // A multiply can be truncated by truncating its operands.
6576 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6577 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6579 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6582 case Instruction::Shl:
6583 // If we are truncating the result of this SHL, and if it's a shift of a
6584 // constant amount, we can always perform a SHL in a smaller type.
6585 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6586 uint32_t BitWidth = Ty->getBitWidth();
6587 if (BitWidth < OrigTy->getBitWidth() &&
6588 CI->getLimitedValue(BitWidth) < BitWidth)
6589 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6593 case Instruction::LShr:
6594 // If this is a truncate of a logical shr, we can truncate it to a smaller
6595 // lshr iff we know that the bits we would otherwise be shifting in are
6597 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6598 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6599 uint32_t BitWidth = Ty->getBitWidth();
6600 if (BitWidth < OrigBitWidth &&
6601 MaskedValueIsZero(I->getOperand(0),
6602 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6603 CI->getLimitedValue(BitWidth) < BitWidth) {
6604 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6609 case Instruction::ZExt:
6610 case Instruction::SExt:
6611 case Instruction::Trunc:
6612 // If this is the same kind of case as our original (e.g. zext+zext), we
6613 // can safely replace it. Note that replacing it does not reduce the number
6614 // of casts in the input.
6615 if (I->getOpcode() == CastOpc)
6620 // TODO: Can handle more cases here.
6627 /// EvaluateInDifferentType - Given an expression that
6628 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6629 /// evaluate the expression.
6630 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6632 if (Constant *C = dyn_cast<Constant>(V))
6633 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6635 // Otherwise, it must be an instruction.
6636 Instruction *I = cast<Instruction>(V);
6637 Instruction *Res = 0;
6638 switch (I->getOpcode()) {
6639 case Instruction::Add:
6640 case Instruction::Sub:
6641 case Instruction::Mul:
6642 case Instruction::And:
6643 case Instruction::Or:
6644 case Instruction::Xor:
6645 case Instruction::AShr:
6646 case Instruction::LShr:
6647 case Instruction::Shl: {
6648 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6649 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6650 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6651 LHS, RHS, I->getName());
6654 case Instruction::Trunc:
6655 case Instruction::ZExt:
6656 case Instruction::SExt:
6657 // If the source type of the cast is the type we're trying for then we can
6658 // just return the source. There's no need to insert it because it is not
6660 if (I->getOperand(0)->getType() == Ty)
6661 return I->getOperand(0);
6663 // Otherwise, must be the same type of case, so just reinsert a new one.
6664 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6668 // TODO: Can handle more cases here.
6669 assert(0 && "Unreachable!");
6673 return InsertNewInstBefore(Res, *I);
6676 /// @brief Implement the transforms common to all CastInst visitors.
6677 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6678 Value *Src = CI.getOperand(0);
6680 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6681 // eliminate it now.
6682 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6683 if (Instruction::CastOps opc =
6684 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6685 // The first cast (CSrc) is eliminable so we need to fix up or replace
6686 // the second cast (CI). CSrc will then have a good chance of being dead.
6687 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6691 // If we are casting a select then fold the cast into the select
6692 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6693 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6696 // If we are casting a PHI then fold the cast into the PHI
6697 if (isa<PHINode>(Src))
6698 if (Instruction *NV = FoldOpIntoPhi(CI))
6704 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6705 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6706 Value *Src = CI.getOperand(0);
6708 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6709 // If casting the result of a getelementptr instruction with no offset, turn
6710 // this into a cast of the original pointer!
6711 if (GEP->hasAllZeroIndices()) {
6712 // Changing the cast operand is usually not a good idea but it is safe
6713 // here because the pointer operand is being replaced with another
6714 // pointer operand so the opcode doesn't need to change.
6716 CI.setOperand(0, GEP->getOperand(0));
6720 // If the GEP has a single use, and the base pointer is a bitcast, and the
6721 // GEP computes a constant offset, see if we can convert these three
6722 // instructions into fewer. This typically happens with unions and other
6723 // non-type-safe code.
6724 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6725 if (GEP->hasAllConstantIndices()) {
6726 // We are guaranteed to get a constant from EmitGEPOffset.
6727 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6728 int64_t Offset = OffsetV->getSExtValue();
6730 // Get the base pointer input of the bitcast, and the type it points to.
6731 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6732 const Type *GEPIdxTy =
6733 cast<PointerType>(OrigBase->getType())->getElementType();
6734 if (GEPIdxTy->isSized()) {
6735 SmallVector<Value*, 8> NewIndices;
6737 // Start with the index over the outer type. Note that the type size
6738 // might be zero (even if the offset isn't zero) if the indexed type
6739 // is something like [0 x {int, int}]
6740 const Type *IntPtrTy = TD->getIntPtrType();
6741 int64_t FirstIdx = 0;
6742 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6743 FirstIdx = Offset/TySize;
6746 // Handle silly modulus not returning values values [0..TySize).
6750 assert(Offset >= 0);
6752 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6755 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6757 // Index into the types. If we fail, set OrigBase to null.
6759 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6760 const StructLayout *SL = TD->getStructLayout(STy);
6761 if (Offset < (int64_t)SL->getSizeInBytes()) {
6762 unsigned Elt = SL->getElementContainingOffset(Offset);
6763 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6765 Offset -= SL->getElementOffset(Elt);
6766 GEPIdxTy = STy->getElementType(Elt);
6768 // Otherwise, we can't index into this, bail out.
6772 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6773 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6774 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6775 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6778 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6780 GEPIdxTy = STy->getElementType();
6782 // Otherwise, we can't index into this, bail out.
6788 // If we were able to index down into an element, create the GEP
6789 // and bitcast the result. This eliminates one bitcast, potentially
6791 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6793 NewIndices.end(), "");
6794 InsertNewInstBefore(NGEP, CI);
6795 NGEP->takeName(GEP);
6797 if (isa<BitCastInst>(CI))
6798 return new BitCastInst(NGEP, CI.getType());
6799 assert(isa<PtrToIntInst>(CI));
6800 return new PtrToIntInst(NGEP, CI.getType());
6807 return commonCastTransforms(CI);
6812 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6813 /// integer types. This function implements the common transforms for all those
6815 /// @brief Implement the transforms common to CastInst with integer operands
6816 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6817 if (Instruction *Result = commonCastTransforms(CI))
6820 Value *Src = CI.getOperand(0);
6821 const Type *SrcTy = Src->getType();
6822 const Type *DestTy = CI.getType();
6823 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6824 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6826 // See if we can simplify any instructions used by the LHS whose sole
6827 // purpose is to compute bits we don't care about.
6828 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6829 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6830 KnownZero, KnownOne))
6833 // If the source isn't an instruction or has more than one use then we
6834 // can't do anything more.
6835 Instruction *SrcI = dyn_cast<Instruction>(Src);
6836 if (!SrcI || !Src->hasOneUse())
6839 // Attempt to propagate the cast into the instruction for int->int casts.
6840 int NumCastsRemoved = 0;
6841 if (!isa<BitCastInst>(CI) &&
6842 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6843 CI.getOpcode(), NumCastsRemoved)) {
6844 // If this cast is a truncate, evaluting in a different type always
6845 // eliminates the cast, so it is always a win. If this is a zero-extension,
6846 // we need to do an AND to maintain the clear top-part of the computation,
6847 // so we require that the input have eliminated at least one cast. If this
6848 // is a sign extension, we insert two new casts (to do the extension) so we
6849 // require that two casts have been eliminated.
6851 switch (CI.getOpcode()) {
6853 // All the others use floating point so we shouldn't actually
6854 // get here because of the check above.
6855 assert(0 && "Unknown cast type");
6856 case Instruction::Trunc:
6859 case Instruction::ZExt:
6860 DoXForm = NumCastsRemoved >= 1;
6862 case Instruction::SExt:
6863 DoXForm = NumCastsRemoved >= 2;
6868 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6869 CI.getOpcode() == Instruction::SExt);
6870 assert(Res->getType() == DestTy);
6871 switch (CI.getOpcode()) {
6872 default: assert(0 && "Unknown cast type!");
6873 case Instruction::Trunc:
6874 case Instruction::BitCast:
6875 // Just replace this cast with the result.
6876 return ReplaceInstUsesWith(CI, Res);
6877 case Instruction::ZExt: {
6878 // We need to emit an AND to clear the high bits.
6879 assert(SrcBitSize < DestBitSize && "Not a zext?");
6880 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6882 return BinaryOperator::createAnd(Res, C);
6884 case Instruction::SExt:
6885 // We need to emit a cast to truncate, then a cast to sext.
6886 return CastInst::create(Instruction::SExt,
6887 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6893 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6894 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6896 switch (SrcI->getOpcode()) {
6897 case Instruction::Add:
6898 case Instruction::Mul:
6899 case Instruction::And:
6900 case Instruction::Or:
6901 case Instruction::Xor:
6902 // If we are discarding information, rewrite.
6903 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6904 // Don't insert two casts if they cannot be eliminated. We allow
6905 // two casts to be inserted if the sizes are the same. This could
6906 // only be converting signedness, which is a noop.
6907 if (DestBitSize == SrcBitSize ||
6908 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6909 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6910 Instruction::CastOps opcode = CI.getOpcode();
6911 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6912 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6913 return BinaryOperator::create(
6914 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6918 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6919 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6920 SrcI->getOpcode() == Instruction::Xor &&
6921 Op1 == ConstantInt::getTrue() &&
6922 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6923 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6924 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6927 case Instruction::SDiv:
6928 case Instruction::UDiv:
6929 case Instruction::SRem:
6930 case Instruction::URem:
6931 // If we are just changing the sign, rewrite.
6932 if (DestBitSize == SrcBitSize) {
6933 // Don't insert two casts if they cannot be eliminated. We allow
6934 // two casts to be inserted if the sizes are the same. This could
6935 // only be converting signedness, which is a noop.
6936 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6937 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6938 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6940 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6942 return BinaryOperator::create(
6943 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6948 case Instruction::Shl:
6949 // Allow changing the sign of the source operand. Do not allow
6950 // changing the size of the shift, UNLESS the shift amount is a
6951 // constant. We must not change variable sized shifts to a smaller
6952 // size, because it is undefined to shift more bits out than exist
6954 if (DestBitSize == SrcBitSize ||
6955 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6956 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6957 Instruction::BitCast : Instruction::Trunc);
6958 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6959 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6960 return BinaryOperator::createShl(Op0c, Op1c);
6963 case Instruction::AShr:
6964 // If this is a signed shr, and if all bits shifted in are about to be
6965 // truncated off, turn it into an unsigned shr to allow greater
6967 if (DestBitSize < SrcBitSize &&
6968 isa<ConstantInt>(Op1)) {
6969 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6970 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6971 // Insert the new logical shift right.
6972 return BinaryOperator::createLShr(Op0, Op1);
6980 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6981 if (Instruction *Result = commonIntCastTransforms(CI))
6984 Value *Src = CI.getOperand(0);
6985 const Type *Ty = CI.getType();
6986 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6987 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6989 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6990 switch (SrcI->getOpcode()) {
6992 case Instruction::LShr:
6993 // We can shrink lshr to something smaller if we know the bits shifted in
6994 // are already zeros.
6995 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6996 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6998 // Get a mask for the bits shifting in.
6999 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7000 Value* SrcIOp0 = SrcI->getOperand(0);
7001 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7002 if (ShAmt >= DestBitWidth) // All zeros.
7003 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7005 // Okay, we can shrink this. Truncate the input, then return a new
7007 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7008 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7010 return BinaryOperator::createLShr(V1, V2);
7012 } else { // This is a variable shr.
7014 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7015 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7016 // loop-invariant and CSE'd.
7017 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7018 Value *One = ConstantInt::get(SrcI->getType(), 1);
7020 Value *V = InsertNewInstBefore(
7021 BinaryOperator::createShl(One, SrcI->getOperand(1),
7023 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7024 SrcI->getOperand(0),
7026 Value *Zero = Constant::getNullValue(V->getType());
7027 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7037 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7038 // If one of the common conversion will work ..
7039 if (Instruction *Result = commonIntCastTransforms(CI))
7042 Value *Src = CI.getOperand(0);
7044 // If this is a cast of a cast
7045 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7046 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7047 // types and if the sizes are just right we can convert this into a logical
7048 // 'and' which will be much cheaper than the pair of casts.
7049 if (isa<TruncInst>(CSrc)) {
7050 // Get the sizes of the types involved
7051 Value *A = CSrc->getOperand(0);
7052 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7053 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7054 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7055 // If we're actually extending zero bits and the trunc is a no-op
7056 if (MidSize < DstSize && SrcSize == DstSize) {
7057 // Replace both of the casts with an And of the type mask.
7058 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7059 Constant *AndConst = ConstantInt::get(AndValue);
7061 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7062 // Unfortunately, if the type changed, we need to cast it back.
7063 if (And->getType() != CI.getType()) {
7064 And->setName(CSrc->getName()+".mask");
7065 InsertNewInstBefore(And, CI);
7066 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7073 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7074 // If we are just checking for a icmp eq of a single bit and zext'ing it
7075 // to an integer, then shift the bit to the appropriate place and then
7076 // cast to integer to avoid the comparison.
7077 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7078 const APInt &Op1CV = Op1C->getValue();
7080 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7081 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7082 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7083 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7084 Value *In = ICI->getOperand(0);
7085 Value *Sh = ConstantInt::get(In->getType(),
7086 In->getType()->getPrimitiveSizeInBits()-1);
7087 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7088 In->getName()+".lobit"),
7090 if (In->getType() != CI.getType())
7091 In = CastInst::createIntegerCast(In, CI.getType(),
7092 false/*ZExt*/, "tmp", &CI);
7094 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7095 Constant *One = ConstantInt::get(In->getType(), 1);
7096 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7097 In->getName()+".not"),
7101 return ReplaceInstUsesWith(CI, In);
7106 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7107 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7108 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7109 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7110 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7111 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7112 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7113 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7114 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7115 // This only works for EQ and NE
7116 ICI->isEquality()) {
7117 // If Op1C some other power of two, convert:
7118 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7119 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7120 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7121 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7123 APInt KnownZeroMask(~KnownZero);
7124 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7125 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7126 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7127 // (X&4) == 2 --> false
7128 // (X&4) != 2 --> true
7129 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7130 Res = ConstantExpr::getZExt(Res, CI.getType());
7131 return ReplaceInstUsesWith(CI, Res);
7134 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7135 Value *In = ICI->getOperand(0);
7137 // Perform a logical shr by shiftamt.
7138 // Insert the shift to put the result in the low bit.
7139 In = InsertNewInstBefore(
7140 BinaryOperator::createLShr(In,
7141 ConstantInt::get(In->getType(), ShiftAmt),
7142 In->getName()+".lobit"), CI);
7145 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7146 Constant *One = ConstantInt::get(In->getType(), 1);
7147 In = BinaryOperator::createXor(In, One, "tmp");
7148 InsertNewInstBefore(cast<Instruction>(In), CI);
7151 if (CI.getType() == In->getType())
7152 return ReplaceInstUsesWith(CI, In);
7154 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7162 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7163 if (Instruction *I = commonIntCastTransforms(CI))
7166 Value *Src = CI.getOperand(0);
7168 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7169 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7170 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7171 // If we are just checking for a icmp eq of a single bit and zext'ing it
7172 // to an integer, then shift the bit to the appropriate place and then
7173 // cast to integer to avoid the comparison.
7174 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7175 const APInt &Op1CV = Op1C->getValue();
7177 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7178 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7179 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7180 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7181 Value *In = ICI->getOperand(0);
7182 Value *Sh = ConstantInt::get(In->getType(),
7183 In->getType()->getPrimitiveSizeInBits()-1);
7184 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7185 In->getName()+".lobit"),
7187 if (In->getType() != CI.getType())
7188 In = CastInst::createIntegerCast(In, CI.getType(),
7189 true/*SExt*/, "tmp", &CI);
7191 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7192 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7193 In->getName()+".not"), CI);
7195 return ReplaceInstUsesWith(CI, In);
7203 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7204 /// in the specified FP type without changing its value.
7205 static Constant *FitsInFPType(ConstantFP *CFP, const Type *FPTy,
7206 const fltSemantics &Sem) {
7207 APFloat F = CFP->getValueAPF();
7208 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7209 return ConstantFP::get(FPTy, F);
7213 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7214 /// through it until we get the source value.
7215 static Value *LookThroughFPExtensions(Value *V) {
7216 if (Instruction *I = dyn_cast<Instruction>(V))
7217 if (I->getOpcode() == Instruction::FPExt)
7218 return LookThroughFPExtensions(I->getOperand(0));
7220 // If this value is a constant, return the constant in the smallest FP type
7221 // that can accurately represent it. This allows us to turn
7222 // (float)((double)X+2.0) into x+2.0f.
7223 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7224 if (CFP->getType() == Type::PPC_FP128Ty)
7225 return V; // No constant folding of this.
7226 // See if the value can be truncated to float and then reextended.
7227 if (Value *V = FitsInFPType(CFP, Type::FloatTy, APFloat::IEEEsingle))
7229 if (CFP->getType() == Type::DoubleTy)
7230 return V; // Won't shrink.
7231 if (Value *V = FitsInFPType(CFP, Type::DoubleTy, APFloat::IEEEdouble))
7233 // Don't try to shrink to various long double types.
7239 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7240 if (Instruction *I = commonCastTransforms(CI))
7243 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7244 // smaller than the destination type, we can eliminate the truncate by doing
7245 // the add as the smaller type. This applies to add/sub/mul/div as well as
7246 // many builtins (sqrt, etc).
7247 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7248 if (OpI && OpI->hasOneUse()) {
7249 switch (OpI->getOpcode()) {
7251 case Instruction::Add:
7252 case Instruction::Sub:
7253 case Instruction::Mul:
7254 case Instruction::FDiv:
7255 case Instruction::FRem:
7256 const Type *SrcTy = OpI->getType();
7257 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7258 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7259 if (LHSTrunc->getType() != SrcTy &&
7260 RHSTrunc->getType() != SrcTy) {
7261 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7262 // If the source types were both smaller than the destination type of
7263 // the cast, do this xform.
7264 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7265 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7266 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7268 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7270 return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7279 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7280 return commonCastTransforms(CI);
7283 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7284 return commonCastTransforms(CI);
7287 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7288 return commonCastTransforms(CI);
7291 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7292 return commonCastTransforms(CI);
7295 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7296 return commonCastTransforms(CI);
7299 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7300 return commonPointerCastTransforms(CI);
7303 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7304 if (Instruction *I = commonCastTransforms(CI))
7307 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7308 if (!DestPointee->isSized()) return 0;
7310 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7313 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7314 m_ConstantInt(Cst)))) {
7315 // If the source and destination operands have the same type, see if this
7316 // is a single-index GEP.
7317 if (X->getType() == CI.getType()) {
7318 // Get the size of the pointee type.
7319 uint64_t Size = TD->getABITypeSizeInBits(DestPointee);
7321 // Convert the constant to intptr type.
7322 APInt Offset = Cst->getValue();
7323 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7325 // If Offset is evenly divisible by Size, we can do this xform.
7326 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7327 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7328 return new GetElementPtrInst(X, ConstantInt::get(Offset));
7331 // TODO: Could handle other cases, e.g. where add is indexing into field of
7333 } else if (CI.getOperand(0)->hasOneUse() &&
7334 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7335 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7336 // "inttoptr+GEP" instead of "add+intptr".
7338 // Get the size of the pointee type.
7339 uint64_t Size = TD->getABITypeSize(DestPointee);
7341 // Convert the constant to intptr type.
7342 APInt Offset = Cst->getValue();
7343 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7345 // If Offset is evenly divisible by Size, we can do this xform.
7346 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7347 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7349 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7351 return new GetElementPtrInst(P, ConstantInt::get(Offset), "tmp");
7357 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7358 // If the operands are integer typed then apply the integer transforms,
7359 // otherwise just apply the common ones.
7360 Value *Src = CI.getOperand(0);
7361 const Type *SrcTy = Src->getType();
7362 const Type *DestTy = CI.getType();
7364 if (SrcTy->isInteger() && DestTy->isInteger()) {
7365 if (Instruction *Result = commonIntCastTransforms(CI))
7367 } else if (isa<PointerType>(SrcTy)) {
7368 if (Instruction *I = commonPointerCastTransforms(CI))
7371 if (Instruction *Result = commonCastTransforms(CI))
7376 // Get rid of casts from one type to the same type. These are useless and can
7377 // be replaced by the operand.
7378 if (DestTy == Src->getType())
7379 return ReplaceInstUsesWith(CI, Src);
7381 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7382 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7383 const Type *DstElTy = DstPTy->getElementType();
7384 const Type *SrcElTy = SrcPTy->getElementType();
7386 // If we are casting a malloc or alloca to a pointer to a type of the same
7387 // size, rewrite the allocation instruction to allocate the "right" type.
7388 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7389 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7392 // If the source and destination are pointers, and this cast is equivalent
7393 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7394 // This can enhance SROA and other transforms that want type-safe pointers.
7395 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7396 unsigned NumZeros = 0;
7397 while (SrcElTy != DstElTy &&
7398 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7399 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7400 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7404 // If we found a path from the src to dest, create the getelementptr now.
7405 if (SrcElTy == DstElTy) {
7406 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7407 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7408 ((Instruction*) NULL));
7412 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7413 if (SVI->hasOneUse()) {
7414 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7415 // a bitconvert to a vector with the same # elts.
7416 if (isa<VectorType>(DestTy) &&
7417 cast<VectorType>(DestTy)->getNumElements() ==
7418 SVI->getType()->getNumElements()) {
7420 // If either of the operands is a cast from CI.getType(), then
7421 // evaluating the shuffle in the casted destination's type will allow
7422 // us to eliminate at least one cast.
7423 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7424 Tmp->getOperand(0)->getType() == DestTy) ||
7425 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7426 Tmp->getOperand(0)->getType() == DestTy)) {
7427 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7428 SVI->getOperand(0), DestTy, &CI);
7429 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7430 SVI->getOperand(1), DestTy, &CI);
7431 // Return a new shuffle vector. Use the same element ID's, as we
7432 // know the vector types match #elts.
7433 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7441 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7443 /// %D = select %cond, %C, %A
7445 /// %C = select %cond, %B, 0
7448 /// Assuming that the specified instruction is an operand to the select, return
7449 /// a bitmask indicating which operands of this instruction are foldable if they
7450 /// equal the other incoming value of the select.
7452 static unsigned GetSelectFoldableOperands(Instruction *I) {
7453 switch (I->getOpcode()) {
7454 case Instruction::Add:
7455 case Instruction::Mul:
7456 case Instruction::And:
7457 case Instruction::Or:
7458 case Instruction::Xor:
7459 return 3; // Can fold through either operand.
7460 case Instruction::Sub: // Can only fold on the amount subtracted.
7461 case Instruction::Shl: // Can only fold on the shift amount.
7462 case Instruction::LShr:
7463 case Instruction::AShr:
7466 return 0; // Cannot fold
7470 /// GetSelectFoldableConstant - For the same transformation as the previous
7471 /// function, return the identity constant that goes into the select.
7472 static Constant *GetSelectFoldableConstant(Instruction *I) {
7473 switch (I->getOpcode()) {
7474 default: assert(0 && "This cannot happen!"); abort();
7475 case Instruction::Add:
7476 case Instruction::Sub:
7477 case Instruction::Or:
7478 case Instruction::Xor:
7479 case Instruction::Shl:
7480 case Instruction::LShr:
7481 case Instruction::AShr:
7482 return Constant::getNullValue(I->getType());
7483 case Instruction::And:
7484 return Constant::getAllOnesValue(I->getType());
7485 case Instruction::Mul:
7486 return ConstantInt::get(I->getType(), 1);
7490 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7491 /// have the same opcode and only one use each. Try to simplify this.
7492 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7494 if (TI->getNumOperands() == 1) {
7495 // If this is a non-volatile load or a cast from the same type,
7498 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7501 return 0; // unknown unary op.
7504 // Fold this by inserting a select from the input values.
7505 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7506 FI->getOperand(0), SI.getName()+".v");
7507 InsertNewInstBefore(NewSI, SI);
7508 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7512 // Only handle binary operators here.
7513 if (!isa<BinaryOperator>(TI))
7516 // Figure out if the operations have any operands in common.
7517 Value *MatchOp, *OtherOpT, *OtherOpF;
7519 if (TI->getOperand(0) == FI->getOperand(0)) {
7520 MatchOp = TI->getOperand(0);
7521 OtherOpT = TI->getOperand(1);
7522 OtherOpF = FI->getOperand(1);
7523 MatchIsOpZero = true;
7524 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7525 MatchOp = TI->getOperand(1);
7526 OtherOpT = TI->getOperand(0);
7527 OtherOpF = FI->getOperand(0);
7528 MatchIsOpZero = false;
7529 } else if (!TI->isCommutative()) {
7531 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7532 MatchOp = TI->getOperand(0);
7533 OtherOpT = TI->getOperand(1);
7534 OtherOpF = FI->getOperand(0);
7535 MatchIsOpZero = true;
7536 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7537 MatchOp = TI->getOperand(1);
7538 OtherOpT = TI->getOperand(0);
7539 OtherOpF = FI->getOperand(1);
7540 MatchIsOpZero = true;
7545 // If we reach here, they do have operations in common.
7546 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7547 OtherOpF, SI.getName()+".v");
7548 InsertNewInstBefore(NewSI, SI);
7550 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7552 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7554 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7556 assert(0 && "Shouldn't get here");
7560 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7561 Value *CondVal = SI.getCondition();
7562 Value *TrueVal = SI.getTrueValue();
7563 Value *FalseVal = SI.getFalseValue();
7565 // select true, X, Y -> X
7566 // select false, X, Y -> Y
7567 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7568 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7570 // select C, X, X -> X
7571 if (TrueVal == FalseVal)
7572 return ReplaceInstUsesWith(SI, TrueVal);
7574 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7575 return ReplaceInstUsesWith(SI, FalseVal);
7576 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7577 return ReplaceInstUsesWith(SI, TrueVal);
7578 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7579 if (isa<Constant>(TrueVal))
7580 return ReplaceInstUsesWith(SI, TrueVal);
7582 return ReplaceInstUsesWith(SI, FalseVal);
7585 if (SI.getType() == Type::Int1Ty) {
7586 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7587 if (C->getZExtValue()) {
7588 // Change: A = select B, true, C --> A = or B, C
7589 return BinaryOperator::createOr(CondVal, FalseVal);
7591 // Change: A = select B, false, C --> A = and !B, C
7593 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7594 "not."+CondVal->getName()), SI);
7595 return BinaryOperator::createAnd(NotCond, FalseVal);
7597 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7598 if (C->getZExtValue() == false) {
7599 // Change: A = select B, C, false --> A = and B, C
7600 return BinaryOperator::createAnd(CondVal, TrueVal);
7602 // Change: A = select B, C, true --> A = or !B, C
7604 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7605 "not."+CondVal->getName()), SI);
7606 return BinaryOperator::createOr(NotCond, TrueVal);
7610 // select a, b, a -> a&b
7611 // select a, a, b -> a|b
7612 if (CondVal == TrueVal)
7613 return BinaryOperator::createOr(CondVal, FalseVal);
7614 else if (CondVal == FalseVal)
7615 return BinaryOperator::createAnd(CondVal, TrueVal);
7618 // Selecting between two integer constants?
7619 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7620 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7621 // select C, 1, 0 -> zext C to int
7622 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7623 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7624 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7625 // select C, 0, 1 -> zext !C to int
7627 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7628 "not."+CondVal->getName()), SI);
7629 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7632 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7634 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7636 // (x <s 0) ? -1 : 0 -> ashr x, 31
7637 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7638 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7639 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7640 // The comparison constant and the result are not neccessarily the
7641 // same width. Make an all-ones value by inserting a AShr.
7642 Value *X = IC->getOperand(0);
7643 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7644 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7645 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7647 InsertNewInstBefore(SRA, SI);
7649 // Finally, convert to the type of the select RHS. We figure out
7650 // if this requires a SExt, Trunc or BitCast based on the sizes.
7651 Instruction::CastOps opc = Instruction::BitCast;
7652 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7653 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7654 if (SRASize < SISize)
7655 opc = Instruction::SExt;
7656 else if (SRASize > SISize)
7657 opc = Instruction::Trunc;
7658 return CastInst::create(opc, SRA, SI.getType());
7663 // If one of the constants is zero (we know they can't both be) and we
7664 // have an icmp instruction with zero, and we have an 'and' with the
7665 // non-constant value, eliminate this whole mess. This corresponds to
7666 // cases like this: ((X & 27) ? 27 : 0)
7667 if (TrueValC->isZero() || FalseValC->isZero())
7668 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7669 cast<Constant>(IC->getOperand(1))->isNullValue())
7670 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7671 if (ICA->getOpcode() == Instruction::And &&
7672 isa<ConstantInt>(ICA->getOperand(1)) &&
7673 (ICA->getOperand(1) == TrueValC ||
7674 ICA->getOperand(1) == FalseValC) &&
7675 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7676 // Okay, now we know that everything is set up, we just don't
7677 // know whether we have a icmp_ne or icmp_eq and whether the
7678 // true or false val is the zero.
7679 bool ShouldNotVal = !TrueValC->isZero();
7680 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7683 V = InsertNewInstBefore(BinaryOperator::create(
7684 Instruction::Xor, V, ICA->getOperand(1)), SI);
7685 return ReplaceInstUsesWith(SI, V);
7690 // See if we are selecting two values based on a comparison of the two values.
7691 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7692 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7693 // Transform (X == Y) ? X : Y -> Y
7694 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7695 // This is not safe in general for floating point:
7696 // consider X== -0, Y== +0.
7697 // It becomes safe if either operand is a nonzero constant.
7698 ConstantFP *CFPt, *CFPf;
7699 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7700 !CFPt->getValueAPF().isZero()) ||
7701 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7702 !CFPf->getValueAPF().isZero()))
7703 return ReplaceInstUsesWith(SI, FalseVal);
7705 // Transform (X != Y) ? X : Y -> X
7706 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7707 return ReplaceInstUsesWith(SI, TrueVal);
7708 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7710 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7711 // Transform (X == Y) ? Y : X -> X
7712 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7713 // This is not safe in general for floating point:
7714 // consider X== -0, Y== +0.
7715 // It becomes safe if either operand is a nonzero constant.
7716 ConstantFP *CFPt, *CFPf;
7717 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7718 !CFPt->getValueAPF().isZero()) ||
7719 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7720 !CFPf->getValueAPF().isZero()))
7721 return ReplaceInstUsesWith(SI, FalseVal);
7723 // Transform (X != Y) ? Y : X -> Y
7724 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7725 return ReplaceInstUsesWith(SI, TrueVal);
7726 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7730 // See if we are selecting two values based on a comparison of the two values.
7731 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7732 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7733 // Transform (X == Y) ? X : Y -> Y
7734 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7735 return ReplaceInstUsesWith(SI, FalseVal);
7736 // Transform (X != Y) ? X : Y -> X
7737 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7738 return ReplaceInstUsesWith(SI, TrueVal);
7739 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7741 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7742 // Transform (X == Y) ? Y : X -> X
7743 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7744 return ReplaceInstUsesWith(SI, FalseVal);
7745 // Transform (X != Y) ? Y : X -> Y
7746 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7747 return ReplaceInstUsesWith(SI, TrueVal);
7748 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7752 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7753 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7754 if (TI->hasOneUse() && FI->hasOneUse()) {
7755 Instruction *AddOp = 0, *SubOp = 0;
7757 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7758 if (TI->getOpcode() == FI->getOpcode())
7759 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7762 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7763 // even legal for FP.
7764 if (TI->getOpcode() == Instruction::Sub &&
7765 FI->getOpcode() == Instruction::Add) {
7766 AddOp = FI; SubOp = TI;
7767 } else if (FI->getOpcode() == Instruction::Sub &&
7768 TI->getOpcode() == Instruction::Add) {
7769 AddOp = TI; SubOp = FI;
7773 Value *OtherAddOp = 0;
7774 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7775 OtherAddOp = AddOp->getOperand(1);
7776 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7777 OtherAddOp = AddOp->getOperand(0);
7781 // So at this point we know we have (Y -> OtherAddOp):
7782 // select C, (add X, Y), (sub X, Z)
7783 Value *NegVal; // Compute -Z
7784 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7785 NegVal = ConstantExpr::getNeg(C);
7787 NegVal = InsertNewInstBefore(
7788 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7791 Value *NewTrueOp = OtherAddOp;
7792 Value *NewFalseOp = NegVal;
7794 std::swap(NewTrueOp, NewFalseOp);
7795 Instruction *NewSel =
7796 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7798 NewSel = InsertNewInstBefore(NewSel, SI);
7799 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7804 // See if we can fold the select into one of our operands.
7805 if (SI.getType()->isInteger()) {
7806 // See the comment above GetSelectFoldableOperands for a description of the
7807 // transformation we are doing here.
7808 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7809 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7810 !isa<Constant>(FalseVal))
7811 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7812 unsigned OpToFold = 0;
7813 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7815 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7820 Constant *C = GetSelectFoldableConstant(TVI);
7821 Instruction *NewSel =
7822 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7823 InsertNewInstBefore(NewSel, SI);
7824 NewSel->takeName(TVI);
7825 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7826 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7828 assert(0 && "Unknown instruction!!");
7833 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7834 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7835 !isa<Constant>(TrueVal))
7836 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7837 unsigned OpToFold = 0;
7838 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7840 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7845 Constant *C = GetSelectFoldableConstant(FVI);
7846 Instruction *NewSel =
7847 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7848 InsertNewInstBefore(NewSel, SI);
7849 NewSel->takeName(FVI);
7850 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7851 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7853 assert(0 && "Unknown instruction!!");
7858 if (BinaryOperator::isNot(CondVal)) {
7859 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7860 SI.setOperand(1, FalseVal);
7861 SI.setOperand(2, TrueVal);
7868 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7869 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7870 /// and it is more than the alignment of the ultimate object, see if we can
7871 /// increase the alignment of the ultimate object, making this check succeed.
7872 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7873 unsigned PrefAlign = 0) {
7874 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7875 unsigned Align = GV->getAlignment();
7876 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7877 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7879 // If there is a large requested alignment and we can, bump up the alignment
7881 if (PrefAlign > Align && GV->hasInitializer()) {
7882 GV->setAlignment(PrefAlign);
7886 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7887 unsigned Align = AI->getAlignment();
7888 if (Align == 0 && TD) {
7889 if (isa<AllocaInst>(AI))
7890 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7891 else if (isa<MallocInst>(AI)) {
7892 // Malloc returns maximally aligned memory.
7893 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7896 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7899 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7903 // If there is a requested alignment and if this is an alloca, round up. We
7904 // don't do this for malloc, because some systems can't respect the request.
7905 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7906 AI->setAlignment(PrefAlign);
7910 } else if (isa<BitCastInst>(V) ||
7911 (isa<ConstantExpr>(V) &&
7912 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7913 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7915 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7916 // If all indexes are zero, it is just the alignment of the base pointer.
7917 bool AllZeroOperands = true;
7918 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7919 if (!isa<Constant>(GEPI->getOperand(i)) ||
7920 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7921 AllZeroOperands = false;
7925 if (AllZeroOperands) {
7926 // Treat this like a bitcast.
7927 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7930 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7931 if (BaseAlignment == 0) return 0;
7933 // Otherwise, if the base alignment is >= the alignment we expect for the
7934 // base pointer type, then we know that the resultant pointer is aligned at
7935 // least as much as its type requires.
7938 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7939 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7940 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7941 if (Align <= BaseAlignment) {
7942 const Type *GEPTy = GEPI->getType();
7943 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7944 Align = std::min(Align, (unsigned)
7945 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7953 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
7954 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7955 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7956 unsigned MinAlign = std::min(DstAlign, SrcAlign);
7957 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
7959 if (CopyAlign < MinAlign) {
7960 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
7964 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
7966 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
7967 if (MemOpLength == 0) return 0;
7969 // Source and destination pointer types are always "i8*" for intrinsic. See
7970 // if the size is something we can handle with a single primitive load/store.
7971 // A single load+store correctly handles overlapping memory in the memmove
7973 unsigned Size = MemOpLength->getZExtValue();
7974 if (Size == 0 || Size > 8 || (Size&(Size-1)))
7975 return 0; // If not 1/2/4/8 bytes, exit.
7977 // Use an integer load+store unless we can find something better.
7978 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
7980 // Memcpy forces the use of i8* for the source and destination. That means
7981 // that if you're using memcpy to move one double around, you'll get a cast
7982 // from double* to i8*. We'd much rather use a double load+store rather than
7983 // an i64 load+store, here because this improves the odds that the source or
7984 // dest address will be promotable. See if we can find a better type than the
7985 // integer datatype.
7986 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
7987 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
7988 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
7989 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
7990 // down through these levels if so.
7991 while (!SrcETy->isFirstClassType()) {
7992 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
7993 if (STy->getNumElements() == 1)
7994 SrcETy = STy->getElementType(0);
7997 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
7998 if (ATy->getNumElements() == 1)
7999 SrcETy = ATy->getElementType();
8006 if (SrcETy->isFirstClassType())
8007 NewPtrTy = PointerType::getUnqual(SrcETy);
8012 // If the memcpy/memmove provides better alignment info than we can
8014 SrcAlign = std::max(SrcAlign, CopyAlign);
8015 DstAlign = std::max(DstAlign, CopyAlign);
8017 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8018 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8019 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8020 InsertNewInstBefore(L, *MI);
8021 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8023 // Set the size of the copy to 0, it will be deleted on the next iteration.
8024 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8028 /// visitCallInst - CallInst simplification. This mostly only handles folding
8029 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8030 /// the heavy lifting.
8032 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8033 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8034 if (!II) return visitCallSite(&CI);
8036 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8038 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8039 bool Changed = false;
8041 // memmove/cpy/set of zero bytes is a noop.
8042 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8043 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8045 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8046 if (CI->getZExtValue() == 1) {
8047 // Replace the instruction with just byte operations. We would
8048 // transform other cases to loads/stores, but we don't know if
8049 // alignment is sufficient.
8053 // If we have a memmove and the source operation is a constant global,
8054 // then the source and dest pointers can't alias, so we can change this
8055 // into a call to memcpy.
8056 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8057 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8058 if (GVSrc->isConstant()) {
8059 Module *M = CI.getParent()->getParent()->getParent();
8060 Intrinsic::ID MemCpyID;
8061 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8062 MemCpyID = Intrinsic::memcpy_i32;
8064 MemCpyID = Intrinsic::memcpy_i64;
8065 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8070 // If we can determine a pointer alignment that is bigger than currently
8071 // set, update the alignment.
8072 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8073 if (Instruction *I = SimplifyMemTransfer(MI))
8075 } else if (isa<MemSetInst>(MI)) {
8076 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
8077 if (MI->getAlignment()->getZExtValue() < Alignment) {
8078 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8083 if (Changed) return II;
8085 switch (II->getIntrinsicID()) {
8087 case Intrinsic::ppc_altivec_lvx:
8088 case Intrinsic::ppc_altivec_lvxl:
8089 case Intrinsic::x86_sse_loadu_ps:
8090 case Intrinsic::x86_sse2_loadu_pd:
8091 case Intrinsic::x86_sse2_loadu_dq:
8092 // Turn PPC lvx -> load if the pointer is known aligned.
8093 // Turn X86 loadups -> load if the pointer is known aligned.
8094 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8095 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8096 PointerType::getUnqual(II->getType()),
8098 return new LoadInst(Ptr);
8101 case Intrinsic::ppc_altivec_stvx:
8102 case Intrinsic::ppc_altivec_stvxl:
8103 // Turn stvx -> store if the pointer is known aligned.
8104 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
8105 const Type *OpPtrTy =
8106 PointerType::getUnqual(II->getOperand(1)->getType());
8107 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8108 return new StoreInst(II->getOperand(1), Ptr);
8111 case Intrinsic::x86_sse_storeu_ps:
8112 case Intrinsic::x86_sse2_storeu_pd:
8113 case Intrinsic::x86_sse2_storeu_dq:
8114 case Intrinsic::x86_sse2_storel_dq:
8115 // Turn X86 storeu -> store if the pointer is known aligned.
8116 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8117 const Type *OpPtrTy =
8118 PointerType::getUnqual(II->getOperand(2)->getType());
8119 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8120 return new StoreInst(II->getOperand(2), Ptr);
8124 case Intrinsic::x86_sse_cvttss2si: {
8125 // These intrinsics only demands the 0th element of its input vector. If
8126 // we can simplify the input based on that, do so now.
8128 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8130 II->setOperand(1, V);
8136 case Intrinsic::ppc_altivec_vperm:
8137 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8138 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8139 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8141 // Check that all of the elements are integer constants or undefs.
8142 bool AllEltsOk = true;
8143 for (unsigned i = 0; i != 16; ++i) {
8144 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8145 !isa<UndefValue>(Mask->getOperand(i))) {
8152 // Cast the input vectors to byte vectors.
8153 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8154 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8155 Value *Result = UndefValue::get(Op0->getType());
8157 // Only extract each element once.
8158 Value *ExtractedElts[32];
8159 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8161 for (unsigned i = 0; i != 16; ++i) {
8162 if (isa<UndefValue>(Mask->getOperand(i)))
8164 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8165 Idx &= 31; // Match the hardware behavior.
8167 if (ExtractedElts[Idx] == 0) {
8169 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8170 InsertNewInstBefore(Elt, CI);
8171 ExtractedElts[Idx] = Elt;
8174 // Insert this value into the result vector.
8175 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
8176 InsertNewInstBefore(cast<Instruction>(Result), CI);
8178 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8183 case Intrinsic::stackrestore: {
8184 // If the save is right next to the restore, remove the restore. This can
8185 // happen when variable allocas are DCE'd.
8186 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8187 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8188 BasicBlock::iterator BI = SS;
8190 return EraseInstFromFunction(CI);
8194 // If the stack restore is in a return/unwind block and if there are no
8195 // allocas or calls between the restore and the return, nuke the restore.
8196 TerminatorInst *TI = II->getParent()->getTerminator();
8197 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
8198 BasicBlock::iterator BI = II;
8199 bool CannotRemove = false;
8200 for (++BI; &*BI != TI; ++BI) {
8201 if (isa<AllocaInst>(BI) ||
8202 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
8203 CannotRemove = true;
8208 return EraseInstFromFunction(CI);
8215 return visitCallSite(II);
8218 // InvokeInst simplification
8220 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8221 return visitCallSite(&II);
8224 // visitCallSite - Improvements for call and invoke instructions.
8226 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8227 bool Changed = false;
8229 // If the callee is a constexpr cast of a function, attempt to move the cast
8230 // to the arguments of the call/invoke.
8231 if (transformConstExprCastCall(CS)) return 0;
8233 Value *Callee = CS.getCalledValue();
8235 if (Function *CalleeF = dyn_cast<Function>(Callee))
8236 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8237 Instruction *OldCall = CS.getInstruction();
8238 // If the call and callee calling conventions don't match, this call must
8239 // be unreachable, as the call is undefined.
8240 new StoreInst(ConstantInt::getTrue(),
8241 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8243 if (!OldCall->use_empty())
8244 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8245 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8246 return EraseInstFromFunction(*OldCall);
8250 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8251 // This instruction is not reachable, just remove it. We insert a store to
8252 // undef so that we know that this code is not reachable, despite the fact
8253 // that we can't modify the CFG here.
8254 new StoreInst(ConstantInt::getTrue(),
8255 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8256 CS.getInstruction());
8258 if (!CS.getInstruction()->use_empty())
8259 CS.getInstruction()->
8260 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8262 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8263 // Don't break the CFG, insert a dummy cond branch.
8264 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8265 ConstantInt::getTrue(), II);
8267 return EraseInstFromFunction(*CS.getInstruction());
8270 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8271 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8272 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8273 return transformCallThroughTrampoline(CS);
8275 const PointerType *PTy = cast<PointerType>(Callee->getType());
8276 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8277 if (FTy->isVarArg()) {
8278 // See if we can optimize any arguments passed through the varargs area of
8280 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8281 E = CS.arg_end(); I != E; ++I)
8282 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8283 // If this cast does not effect the value passed through the varargs
8284 // area, we can eliminate the use of the cast.
8285 Value *Op = CI->getOperand(0);
8286 if (CI->isLosslessCast()) {
8293 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8294 // Inline asm calls cannot throw - mark them 'nounwind'.
8295 CS.setDoesNotThrow();
8299 return Changed ? CS.getInstruction() : 0;
8302 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8303 // attempt to move the cast to the arguments of the call/invoke.
8305 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8306 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8307 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8308 if (CE->getOpcode() != Instruction::BitCast ||
8309 !isa<Function>(CE->getOperand(0)))
8311 Function *Callee = cast<Function>(CE->getOperand(0));
8312 Instruction *Caller = CS.getInstruction();
8313 const ParamAttrsList* CallerPAL = CS.getParamAttrs();
8315 // Okay, this is a cast from a function to a different type. Unless doing so
8316 // would cause a type conversion of one of our arguments, change this call to
8317 // be a direct call with arguments casted to the appropriate types.
8319 const FunctionType *FT = Callee->getFunctionType();
8320 const Type *OldRetTy = Caller->getType();
8322 // Check to see if we are changing the return type...
8323 if (OldRetTy != FT->getReturnType()) {
8324 if (Callee->isDeclaration() && !Caller->use_empty() &&
8325 // Conversion is ok if changing from pointer to int of same size.
8326 !(isa<PointerType>(FT->getReturnType()) &&
8327 TD->getIntPtrType() == OldRetTy))
8328 return false; // Cannot transform this return value.
8330 if (!Caller->use_empty() &&
8331 // void -> non-void is handled specially
8332 FT->getReturnType() != Type::VoidTy &&
8333 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8334 return false; // Cannot transform this return value.
8336 if (CallerPAL && !Caller->use_empty()) {
8337 uint16_t RAttrs = CallerPAL->getParamAttrs(0);
8338 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8339 return false; // Attribute not compatible with transformed value.
8342 // If the callsite is an invoke instruction, and the return value is used by
8343 // a PHI node in a successor, we cannot change the return type of the call
8344 // because there is no place to put the cast instruction (without breaking
8345 // the critical edge). Bail out in this case.
8346 if (!Caller->use_empty())
8347 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8348 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8350 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8351 if (PN->getParent() == II->getNormalDest() ||
8352 PN->getParent() == II->getUnwindDest())
8356 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8357 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8359 CallSite::arg_iterator AI = CS.arg_begin();
8360 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8361 const Type *ParamTy = FT->getParamType(i);
8362 const Type *ActTy = (*AI)->getType();
8364 if (!CastInst::isCastable(ActTy, ParamTy))
8365 return false; // Cannot transform this parameter value.
8368 uint16_t PAttrs = CallerPAL->getParamAttrs(i + 1);
8369 if (PAttrs & ParamAttr::typeIncompatible(ParamTy))
8370 return false; // Attribute not compatible with transformed value.
8373 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8374 // Some conversions are safe even if we do not have a body.
8375 // Either we can cast directly, or we can upconvert the argument
8376 bool isConvertible = ActTy == ParamTy ||
8377 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8378 (ParamTy->isInteger() && ActTy->isInteger() &&
8379 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8380 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8381 && c->getValue().isStrictlyPositive());
8382 if (Callee->isDeclaration() && !isConvertible) return false;
8385 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8386 Callee->isDeclaration())
8387 return false; // Do not delete arguments unless we have a function body...
8389 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && CallerPAL)
8390 // In this case we have more arguments than the new function type, but we
8391 // won't be dropping them. Check that these extra arguments have attributes
8392 // that are compatible with being a vararg call argument.
8393 for (unsigned i = CallerPAL->size(); i; --i) {
8394 if (CallerPAL->getParamIndex(i - 1) <= FT->getNumParams())
8396 uint16_t PAttrs = CallerPAL->getParamAttrsAtIndex(i - 1);
8397 if (PAttrs & ParamAttr::VarArgsIncompatible)
8401 // Okay, we decided that this is a safe thing to do: go ahead and start
8402 // inserting cast instructions as necessary...
8403 std::vector<Value*> Args;
8404 Args.reserve(NumActualArgs);
8405 ParamAttrsVector attrVec;
8406 attrVec.reserve(NumCommonArgs);
8408 // Get any return attributes.
8409 uint16_t RAttrs = CallerPAL ? CallerPAL->getParamAttrs(0) : 0;
8411 // If the return value is not being used, the type may not be compatible
8412 // with the existing attributes. Wipe out any problematic attributes.
8413 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
8415 // Add the new return attributes.
8417 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8419 AI = CS.arg_begin();
8420 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8421 const Type *ParamTy = FT->getParamType(i);
8422 if ((*AI)->getType() == ParamTy) {
8423 Args.push_back(*AI);
8425 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8426 false, ParamTy, false);
8427 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8428 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8431 // Add any parameter attributes.
8432 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8434 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8437 // If the function takes more arguments than the call was taking, add them
8439 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8440 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8442 // If we are removing arguments to the function, emit an obnoxious warning...
8443 if (FT->getNumParams() < NumActualArgs)
8444 if (!FT->isVarArg()) {
8445 cerr << "WARNING: While resolving call to function '"
8446 << Callee->getName() << "' arguments were dropped!\n";
8448 // Add all of the arguments in their promoted form to the arg list...
8449 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8450 const Type *PTy = getPromotedType((*AI)->getType());
8451 if (PTy != (*AI)->getType()) {
8452 // Must promote to pass through va_arg area!
8453 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8455 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8456 InsertNewInstBefore(Cast, *Caller);
8457 Args.push_back(Cast);
8459 Args.push_back(*AI);
8462 // Add any parameter attributes.
8463 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8465 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8469 if (FT->getReturnType() == Type::VoidTy)
8470 Caller->setName(""); // Void type should not have a name.
8472 const ParamAttrsList* NewCallerPAL = ParamAttrsList::get(attrVec);
8475 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8476 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8477 Args.begin(), Args.end(), Caller->getName(), Caller);
8478 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8479 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
8481 NC = new CallInst(Callee, Args.begin(), Args.end(),
8482 Caller->getName(), Caller);
8483 CallInst *CI = cast<CallInst>(Caller);
8484 if (CI->isTailCall())
8485 cast<CallInst>(NC)->setTailCall();
8486 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8487 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
8490 // Insert a cast of the return type as necessary.
8492 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8493 if (NV->getType() != Type::VoidTy) {
8494 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8496 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8498 // If this is an invoke instruction, we should insert it after the first
8499 // non-phi, instruction in the normal successor block.
8500 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8501 BasicBlock::iterator I = II->getNormalDest()->begin();
8502 while (isa<PHINode>(I)) ++I;
8503 InsertNewInstBefore(NC, *I);
8505 // Otherwise, it's a call, just insert cast right after the call instr
8506 InsertNewInstBefore(NC, *Caller);
8508 AddUsersToWorkList(*Caller);
8510 NV = UndefValue::get(Caller->getType());
8514 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8515 Caller->replaceAllUsesWith(NV);
8516 Caller->eraseFromParent();
8517 RemoveFromWorkList(Caller);
8521 // transformCallThroughTrampoline - Turn a call to a function created by the
8522 // init_trampoline intrinsic into a direct call to the underlying function.
8524 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8525 Value *Callee = CS.getCalledValue();
8526 const PointerType *PTy = cast<PointerType>(Callee->getType());
8527 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8528 const ParamAttrsList *Attrs = CS.getParamAttrs();
8530 // If the call already has the 'nest' attribute somewhere then give up -
8531 // otherwise 'nest' would occur twice after splicing in the chain.
8532 if (Attrs && Attrs->hasAttrSomewhere(ParamAttr::Nest))
8535 IntrinsicInst *Tramp =
8536 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8539 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8540 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8541 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8543 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8544 unsigned NestIdx = 1;
8545 const Type *NestTy = 0;
8546 uint16_t NestAttr = 0;
8548 // Look for a parameter marked with the 'nest' attribute.
8549 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8550 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8551 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8552 // Record the parameter type and any other attributes.
8554 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8559 Instruction *Caller = CS.getInstruction();
8560 std::vector<Value*> NewArgs;
8561 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8563 ParamAttrsVector NewAttrs;
8564 NewAttrs.reserve(Attrs ? Attrs->size() + 1 : 1);
8566 // Insert the nest argument into the call argument list, which may
8567 // mean appending it. Likewise for attributes.
8569 // Add any function result attributes.
8570 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
8572 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8576 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8578 if (Idx == NestIdx) {
8579 // Add the chain argument and attributes.
8580 Value *NestVal = Tramp->getOperand(3);
8581 if (NestVal->getType() != NestTy)
8582 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8583 NewArgs.push_back(NestVal);
8584 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8590 // Add the original argument and attributes.
8591 NewArgs.push_back(*I);
8592 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8595 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8601 // The trampoline may have been bitcast to a bogus type (FTy).
8602 // Handle this by synthesizing a new function type, equal to FTy
8603 // with the chain parameter inserted.
8605 std::vector<const Type*> NewTypes;
8606 NewTypes.reserve(FTy->getNumParams()+1);
8608 // Insert the chain's type into the list of parameter types, which may
8609 // mean appending it.
8612 FunctionType::param_iterator I = FTy->param_begin(),
8613 E = FTy->param_end();
8617 // Add the chain's type.
8618 NewTypes.push_back(NestTy);
8623 // Add the original type.
8624 NewTypes.push_back(*I);
8630 // Replace the trampoline call with a direct call. Let the generic
8631 // code sort out any function type mismatches.
8632 FunctionType *NewFTy =
8633 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8634 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8635 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8636 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8638 Instruction *NewCaller;
8639 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8640 NewCaller = new InvokeInst(NewCallee,
8641 II->getNormalDest(), II->getUnwindDest(),
8642 NewArgs.begin(), NewArgs.end(),
8643 Caller->getName(), Caller);
8644 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8645 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8647 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8648 Caller->getName(), Caller);
8649 if (cast<CallInst>(Caller)->isTailCall())
8650 cast<CallInst>(NewCaller)->setTailCall();
8651 cast<CallInst>(NewCaller)->
8652 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8653 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8655 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8656 Caller->replaceAllUsesWith(NewCaller);
8657 Caller->eraseFromParent();
8658 RemoveFromWorkList(Caller);
8663 // Replace the trampoline call with a direct call. Since there is no 'nest'
8664 // parameter, there is no need to adjust the argument list. Let the generic
8665 // code sort out any function type mismatches.
8666 Constant *NewCallee =
8667 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8668 CS.setCalledFunction(NewCallee);
8669 return CS.getInstruction();
8672 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8673 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8674 /// and a single binop.
8675 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8676 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8677 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8678 isa<CmpInst>(FirstInst));
8679 unsigned Opc = FirstInst->getOpcode();
8680 Value *LHSVal = FirstInst->getOperand(0);
8681 Value *RHSVal = FirstInst->getOperand(1);
8683 const Type *LHSType = LHSVal->getType();
8684 const Type *RHSType = RHSVal->getType();
8686 // Scan to see if all operands are the same opcode, all have one use, and all
8687 // kill their operands (i.e. the operands have one use).
8688 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8689 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8690 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8691 // Verify type of the LHS matches so we don't fold cmp's of different
8692 // types or GEP's with different index types.
8693 I->getOperand(0)->getType() != LHSType ||
8694 I->getOperand(1)->getType() != RHSType)
8697 // If they are CmpInst instructions, check their predicates
8698 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8699 if (cast<CmpInst>(I)->getPredicate() !=
8700 cast<CmpInst>(FirstInst)->getPredicate())
8703 // Keep track of which operand needs a phi node.
8704 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8705 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8708 // Otherwise, this is safe to transform, determine if it is profitable.
8710 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8711 // Indexes are often folded into load/store instructions, so we don't want to
8712 // hide them behind a phi.
8713 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8716 Value *InLHS = FirstInst->getOperand(0);
8717 Value *InRHS = FirstInst->getOperand(1);
8718 PHINode *NewLHS = 0, *NewRHS = 0;
8720 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8721 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8722 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8723 InsertNewInstBefore(NewLHS, PN);
8728 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8729 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8730 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8731 InsertNewInstBefore(NewRHS, PN);
8735 // Add all operands to the new PHIs.
8736 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8738 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8739 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8742 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8743 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8747 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8748 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8749 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8750 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8753 assert(isa<GetElementPtrInst>(FirstInst));
8754 return new GetElementPtrInst(LHSVal, RHSVal);
8758 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8759 /// of the block that defines it. This means that it must be obvious the value
8760 /// of the load is not changed from the point of the load to the end of the
8763 /// Finally, it is safe, but not profitable, to sink a load targetting a
8764 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8766 static bool isSafeToSinkLoad(LoadInst *L) {
8767 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8769 for (++BBI; BBI != E; ++BBI)
8770 if (BBI->mayWriteToMemory())
8773 // Check for non-address taken alloca. If not address-taken already, it isn't
8774 // profitable to do this xform.
8775 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8776 bool isAddressTaken = false;
8777 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8779 if (isa<LoadInst>(UI)) continue;
8780 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8781 // If storing TO the alloca, then the address isn't taken.
8782 if (SI->getOperand(1) == AI) continue;
8784 isAddressTaken = true;
8788 if (!isAddressTaken)
8796 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8797 // operator and they all are only used by the PHI, PHI together their
8798 // inputs, and do the operation once, to the result of the PHI.
8799 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8800 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8802 // Scan the instruction, looking for input operations that can be folded away.
8803 // If all input operands to the phi are the same instruction (e.g. a cast from
8804 // the same type or "+42") we can pull the operation through the PHI, reducing
8805 // code size and simplifying code.
8806 Constant *ConstantOp = 0;
8807 const Type *CastSrcTy = 0;
8808 bool isVolatile = false;
8809 if (isa<CastInst>(FirstInst)) {
8810 CastSrcTy = FirstInst->getOperand(0)->getType();
8811 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8812 // Can fold binop, compare or shift here if the RHS is a constant,
8813 // otherwise call FoldPHIArgBinOpIntoPHI.
8814 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8815 if (ConstantOp == 0)
8816 return FoldPHIArgBinOpIntoPHI(PN);
8817 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8818 isVolatile = LI->isVolatile();
8819 // We can't sink the load if the loaded value could be modified between the
8820 // load and the PHI.
8821 if (LI->getParent() != PN.getIncomingBlock(0) ||
8822 !isSafeToSinkLoad(LI))
8824 } else if (isa<GetElementPtrInst>(FirstInst)) {
8825 if (FirstInst->getNumOperands() == 2)
8826 return FoldPHIArgBinOpIntoPHI(PN);
8827 // Can't handle general GEPs yet.
8830 return 0; // Cannot fold this operation.
8833 // Check to see if all arguments are the same operation.
8834 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8835 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8836 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8837 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8840 if (I->getOperand(0)->getType() != CastSrcTy)
8841 return 0; // Cast operation must match.
8842 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8843 // We can't sink the load if the loaded value could be modified between
8844 // the load and the PHI.
8845 if (LI->isVolatile() != isVolatile ||
8846 LI->getParent() != PN.getIncomingBlock(i) ||
8847 !isSafeToSinkLoad(LI))
8849 } else if (I->getOperand(1) != ConstantOp) {
8854 // Okay, they are all the same operation. Create a new PHI node of the
8855 // correct type, and PHI together all of the LHS's of the instructions.
8856 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8857 PN.getName()+".in");
8858 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8860 Value *InVal = FirstInst->getOperand(0);
8861 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8863 // Add all operands to the new PHI.
8864 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8865 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8866 if (NewInVal != InVal)
8868 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8873 // The new PHI unions all of the same values together. This is really
8874 // common, so we handle it intelligently here for compile-time speed.
8878 InsertNewInstBefore(NewPN, PN);
8882 // Insert and return the new operation.
8883 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8884 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8885 else if (isa<LoadInst>(FirstInst))
8886 return new LoadInst(PhiVal, "", isVolatile);
8887 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8888 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8889 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8890 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8891 PhiVal, ConstantOp);
8893 assert(0 && "Unknown operation");
8897 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8899 static bool DeadPHICycle(PHINode *PN,
8900 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8901 if (PN->use_empty()) return true;
8902 if (!PN->hasOneUse()) return false;
8904 // Remember this node, and if we find the cycle, return.
8905 if (!PotentiallyDeadPHIs.insert(PN))
8908 // Don't scan crazily complex things.
8909 if (PotentiallyDeadPHIs.size() == 16)
8912 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8913 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8918 /// PHIsEqualValue - Return true if this phi node is always equal to
8919 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8920 /// z = some value; x = phi (y, z); y = phi (x, z)
8921 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8922 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8923 // See if we already saw this PHI node.
8924 if (!ValueEqualPHIs.insert(PN))
8927 // Don't scan crazily complex things.
8928 if (ValueEqualPHIs.size() == 16)
8931 // Scan the operands to see if they are either phi nodes or are equal to
8933 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8934 Value *Op = PN->getIncomingValue(i);
8935 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8936 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8938 } else if (Op != NonPhiInVal)
8946 // PHINode simplification
8948 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8949 // If LCSSA is around, don't mess with Phi nodes
8950 if (MustPreserveLCSSA) return 0;
8952 if (Value *V = PN.hasConstantValue())
8953 return ReplaceInstUsesWith(PN, V);
8955 // If all PHI operands are the same operation, pull them through the PHI,
8956 // reducing code size.
8957 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8958 PN.getIncomingValue(0)->hasOneUse())
8959 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8962 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8963 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8964 // PHI)... break the cycle.
8965 if (PN.hasOneUse()) {
8966 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8967 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8968 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8969 PotentiallyDeadPHIs.insert(&PN);
8970 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8971 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8974 // If this phi has a single use, and if that use just computes a value for
8975 // the next iteration of a loop, delete the phi. This occurs with unused
8976 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8977 // common case here is good because the only other things that catch this
8978 // are induction variable analysis (sometimes) and ADCE, which is only run
8980 if (PHIUser->hasOneUse() &&
8981 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8982 PHIUser->use_back() == &PN) {
8983 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8987 // We sometimes end up with phi cycles that non-obviously end up being the
8988 // same value, for example:
8989 // z = some value; x = phi (y, z); y = phi (x, z)
8990 // where the phi nodes don't necessarily need to be in the same block. Do a
8991 // quick check to see if the PHI node only contains a single non-phi value, if
8992 // so, scan to see if the phi cycle is actually equal to that value.
8994 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
8995 // Scan for the first non-phi operand.
8996 while (InValNo != NumOperandVals &&
8997 isa<PHINode>(PN.getIncomingValue(InValNo)))
9000 if (InValNo != NumOperandVals) {
9001 Value *NonPhiInVal = PN.getOperand(InValNo);
9003 // Scan the rest of the operands to see if there are any conflicts, if so
9004 // there is no need to recursively scan other phis.
9005 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9006 Value *OpVal = PN.getIncomingValue(InValNo);
9007 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9011 // If we scanned over all operands, then we have one unique value plus
9012 // phi values. Scan PHI nodes to see if they all merge in each other or
9014 if (InValNo == NumOperandVals) {
9015 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9016 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9017 return ReplaceInstUsesWith(PN, NonPhiInVal);
9024 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9025 Instruction *InsertPoint,
9027 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9028 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9029 // We must cast correctly to the pointer type. Ensure that we
9030 // sign extend the integer value if it is smaller as this is
9031 // used for address computation.
9032 Instruction::CastOps opcode =
9033 (VTySize < PtrSize ? Instruction::SExt :
9034 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9035 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9039 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9040 Value *PtrOp = GEP.getOperand(0);
9041 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9042 // If so, eliminate the noop.
9043 if (GEP.getNumOperands() == 1)
9044 return ReplaceInstUsesWith(GEP, PtrOp);
9046 if (isa<UndefValue>(GEP.getOperand(0)))
9047 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9049 bool HasZeroPointerIndex = false;
9050 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9051 HasZeroPointerIndex = C->isNullValue();
9053 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9054 return ReplaceInstUsesWith(GEP, PtrOp);
9056 // Eliminate unneeded casts for indices.
9057 bool MadeChange = false;
9059 gep_type_iterator GTI = gep_type_begin(GEP);
9060 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
9061 if (isa<SequentialType>(*GTI)) {
9062 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
9063 if (CI->getOpcode() == Instruction::ZExt ||
9064 CI->getOpcode() == Instruction::SExt) {
9065 const Type *SrcTy = CI->getOperand(0)->getType();
9066 // We can eliminate a cast from i32 to i64 iff the target
9067 // is a 32-bit pointer target.
9068 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9070 GEP.setOperand(i, CI->getOperand(0));
9074 // If we are using a wider index than needed for this platform, shrink it
9075 // to what we need. If the incoming value needs a cast instruction,
9076 // insert it. This explicit cast can make subsequent optimizations more
9078 Value *Op = GEP.getOperand(i);
9079 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits())
9080 if (Constant *C = dyn_cast<Constant>(Op)) {
9081 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
9084 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9086 GEP.setOperand(i, Op);
9091 if (MadeChange) return &GEP;
9093 // If this GEP instruction doesn't move the pointer, and if the input operand
9094 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9095 // real input to the dest type.
9096 if (GEP.hasAllZeroIndices()) {
9097 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9098 // If the bitcast is of an allocation, and the allocation will be
9099 // converted to match the type of the cast, don't touch this.
9100 if (isa<AllocationInst>(BCI->getOperand(0))) {
9101 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9102 if (Instruction *I = visitBitCast(*BCI)) {
9105 BCI->getParent()->getInstList().insert(BCI, I);
9106 ReplaceInstUsesWith(*BCI, I);
9111 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9115 // Combine Indices - If the source pointer to this getelementptr instruction
9116 // is a getelementptr instruction, combine the indices of the two
9117 // getelementptr instructions into a single instruction.
9119 SmallVector<Value*, 8> SrcGEPOperands;
9120 if (User *Src = dyn_castGetElementPtr(PtrOp))
9121 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9123 if (!SrcGEPOperands.empty()) {
9124 // Note that if our source is a gep chain itself that we wait for that
9125 // chain to be resolved before we perform this transformation. This
9126 // avoids us creating a TON of code in some cases.
9128 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9129 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9130 return 0; // Wait until our source is folded to completion.
9132 SmallVector<Value*, 8> Indices;
9134 // Find out whether the last index in the source GEP is a sequential idx.
9135 bool EndsWithSequential = false;
9136 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9137 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9138 EndsWithSequential = !isa<StructType>(*I);
9140 // Can we combine the two pointer arithmetics offsets?
9141 if (EndsWithSequential) {
9142 // Replace: gep (gep %P, long B), long A, ...
9143 // With: T = long A+B; gep %P, T, ...
9145 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9146 if (SO1 == Constant::getNullValue(SO1->getType())) {
9148 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9151 // If they aren't the same type, convert both to an integer of the
9152 // target's pointer size.
9153 if (SO1->getType() != GO1->getType()) {
9154 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9155 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9156 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9157 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9159 unsigned PS = TD->getPointerSizeInBits();
9160 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9161 // Convert GO1 to SO1's type.
9162 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9164 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9165 // Convert SO1 to GO1's type.
9166 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9168 const Type *PT = TD->getIntPtrType();
9169 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9170 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9174 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9175 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9177 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
9178 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9182 // Recycle the GEP we already have if possible.
9183 if (SrcGEPOperands.size() == 2) {
9184 GEP.setOperand(0, SrcGEPOperands[0]);
9185 GEP.setOperand(1, Sum);
9188 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9189 SrcGEPOperands.end()-1);
9190 Indices.push_back(Sum);
9191 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9193 } else if (isa<Constant>(*GEP.idx_begin()) &&
9194 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9195 SrcGEPOperands.size() != 1) {
9196 // Otherwise we can do the fold if the first index of the GEP is a zero
9197 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9198 SrcGEPOperands.end());
9199 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9202 if (!Indices.empty())
9203 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
9204 Indices.end(), GEP.getName());
9206 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9207 // GEP of global variable. If all of the indices for this GEP are
9208 // constants, we can promote this to a constexpr instead of an instruction.
9210 // Scan for nonconstants...
9211 SmallVector<Constant*, 8> Indices;
9212 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9213 for (; I != E && isa<Constant>(*I); ++I)
9214 Indices.push_back(cast<Constant>(*I));
9216 if (I == E) { // If they are all constants...
9217 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9218 &Indices[0],Indices.size());
9220 // Replace all uses of the GEP with the new constexpr...
9221 return ReplaceInstUsesWith(GEP, CE);
9223 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9224 if (!isa<PointerType>(X->getType())) {
9225 // Not interesting. Source pointer must be a cast from pointer.
9226 } else if (HasZeroPointerIndex) {
9227 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9228 // into : GEP [10 x i8]* X, i32 0, ...
9230 // This occurs when the program declares an array extern like "int X[];"
9232 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9233 const PointerType *XTy = cast<PointerType>(X->getType());
9234 if (const ArrayType *XATy =
9235 dyn_cast<ArrayType>(XTy->getElementType()))
9236 if (const ArrayType *CATy =
9237 dyn_cast<ArrayType>(CPTy->getElementType()))
9238 if (CATy->getElementType() == XATy->getElementType()) {
9239 // At this point, we know that the cast source type is a pointer
9240 // to an array of the same type as the destination pointer
9241 // array. Because the array type is never stepped over (there
9242 // is a leading zero) we can fold the cast into this GEP.
9243 GEP.setOperand(0, X);
9246 } else if (GEP.getNumOperands() == 2) {
9247 // Transform things like:
9248 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9249 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9250 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9251 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9252 if (isa<ArrayType>(SrcElTy) &&
9253 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9254 TD->getABITypeSize(ResElTy)) {
9256 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9257 Idx[1] = GEP.getOperand(1);
9258 Value *V = InsertNewInstBefore(
9259 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
9260 // V and GEP are both pointer types --> BitCast
9261 return new BitCastInst(V, GEP.getType());
9264 // Transform things like:
9265 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9266 // (where tmp = 8*tmp2) into:
9267 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9269 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9270 uint64_t ArrayEltSize =
9271 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9273 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9274 // allow either a mul, shift, or constant here.
9276 ConstantInt *Scale = 0;
9277 if (ArrayEltSize == 1) {
9278 NewIdx = GEP.getOperand(1);
9279 Scale = ConstantInt::get(NewIdx->getType(), 1);
9280 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9281 NewIdx = ConstantInt::get(CI->getType(), 1);
9283 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9284 if (Inst->getOpcode() == Instruction::Shl &&
9285 isa<ConstantInt>(Inst->getOperand(1))) {
9286 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9287 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9288 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9289 NewIdx = Inst->getOperand(0);
9290 } else if (Inst->getOpcode() == Instruction::Mul &&
9291 isa<ConstantInt>(Inst->getOperand(1))) {
9292 Scale = cast<ConstantInt>(Inst->getOperand(1));
9293 NewIdx = Inst->getOperand(0);
9297 // If the index will be to exactly the right offset with the scale taken
9298 // out, perform the transformation. Note, we don't know whether Scale is
9299 // signed or not. We'll use unsigned version of division/modulo
9300 // operation after making sure Scale doesn't have the sign bit set.
9301 if (Scale && Scale->getSExtValue() >= 0LL &&
9302 Scale->getZExtValue() % ArrayEltSize == 0) {
9303 Scale = ConstantInt::get(Scale->getType(),
9304 Scale->getZExtValue() / ArrayEltSize);
9305 if (Scale->getZExtValue() != 1) {
9306 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9308 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9309 NewIdx = InsertNewInstBefore(Sc, GEP);
9312 // Insert the new GEP instruction.
9314 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9316 Instruction *NewGEP =
9317 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
9318 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9319 // The NewGEP must be pointer typed, so must the old one -> BitCast
9320 return new BitCastInst(NewGEP, GEP.getType());
9329 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9330 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9331 if (AI.isArrayAllocation()) // Check C != 1
9332 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9334 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9335 AllocationInst *New = 0;
9337 // Create and insert the replacement instruction...
9338 if (isa<MallocInst>(AI))
9339 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9341 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9342 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9345 InsertNewInstBefore(New, AI);
9347 // Scan to the end of the allocation instructions, to skip over a block of
9348 // allocas if possible...
9350 BasicBlock::iterator It = New;
9351 while (isa<AllocationInst>(*It)) ++It;
9353 // Now that I is pointing to the first non-allocation-inst in the block,
9354 // insert our getelementptr instruction...
9356 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9360 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9361 New->getName()+".sub", It);
9363 // Now make everything use the getelementptr instead of the original
9365 return ReplaceInstUsesWith(AI, V);
9366 } else if (isa<UndefValue>(AI.getArraySize())) {
9367 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9370 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9371 // Note that we only do this for alloca's, because malloc should allocate and
9372 // return a unique pointer, even for a zero byte allocation.
9373 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9374 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9375 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9380 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9381 Value *Op = FI.getOperand(0);
9383 // free undef -> unreachable.
9384 if (isa<UndefValue>(Op)) {
9385 // Insert a new store to null because we cannot modify the CFG here.
9386 new StoreInst(ConstantInt::getTrue(),
9387 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9388 return EraseInstFromFunction(FI);
9391 // If we have 'free null' delete the instruction. This can happen in stl code
9392 // when lots of inlining happens.
9393 if (isa<ConstantPointerNull>(Op))
9394 return EraseInstFromFunction(FI);
9396 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9397 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9398 FI.setOperand(0, CI->getOperand(0));
9402 // Change free (gep X, 0,0,0,0) into free(X)
9403 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9404 if (GEPI->hasAllZeroIndices()) {
9405 AddToWorkList(GEPI);
9406 FI.setOperand(0, GEPI->getOperand(0));
9411 // Change free(malloc) into nothing, if the malloc has a single use.
9412 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9413 if (MI->hasOneUse()) {
9414 EraseInstFromFunction(FI);
9415 return EraseInstFromFunction(*MI);
9422 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9423 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9424 const TargetData *TD) {
9425 User *CI = cast<User>(LI.getOperand(0));
9426 Value *CastOp = CI->getOperand(0);
9428 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9429 // Instead of loading constant c string, use corresponding integer value
9430 // directly if string length is small enough.
9431 const std::string &Str = CE->getOperand(0)->getStringValue();
9433 unsigned len = Str.length();
9434 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9435 unsigned numBits = Ty->getPrimitiveSizeInBits();
9436 // Replace LI with immediate integer store.
9437 if ((numBits >> 3) == len + 1) {
9438 APInt StrVal(numBits, 0);
9439 APInt SingleChar(numBits, 0);
9440 if (TD->isLittleEndian()) {
9441 for (signed i = len-1; i >= 0; i--) {
9442 SingleChar = (uint64_t) Str[i];
9443 StrVal = (StrVal << 8) | SingleChar;
9446 for (unsigned i = 0; i < len; i++) {
9447 SingleChar = (uint64_t) Str[i];
9448 StrVal = (StrVal << 8) | SingleChar;
9450 // Append NULL at the end.
9452 StrVal = (StrVal << 8) | SingleChar;
9454 Value *NL = ConstantInt::get(StrVal);
9455 return IC.ReplaceInstUsesWith(LI, NL);
9460 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9461 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9462 const Type *SrcPTy = SrcTy->getElementType();
9464 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9465 isa<VectorType>(DestPTy)) {
9466 // If the source is an array, the code below will not succeed. Check to
9467 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9469 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9470 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9471 if (ASrcTy->getNumElements() != 0) {
9473 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9474 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9475 SrcTy = cast<PointerType>(CastOp->getType());
9476 SrcPTy = SrcTy->getElementType();
9479 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9480 isa<VectorType>(SrcPTy)) &&
9481 // Do not allow turning this into a load of an integer, which is then
9482 // casted to a pointer, this pessimizes pointer analysis a lot.
9483 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9484 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9485 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9487 // Okay, we are casting from one integer or pointer type to another of
9488 // the same size. Instead of casting the pointer before the load, cast
9489 // the result of the loaded value.
9490 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9492 LI.isVolatile()),LI);
9493 // Now cast the result of the load.
9494 return new BitCastInst(NewLoad, LI.getType());
9501 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9502 /// from this value cannot trap. If it is not obviously safe to load from the
9503 /// specified pointer, we do a quick local scan of the basic block containing
9504 /// ScanFrom, to determine if the address is already accessed.
9505 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9506 // If it is an alloca it is always safe to load from.
9507 if (isa<AllocaInst>(V)) return true;
9509 // If it is a global variable it is mostly safe to load from.
9510 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9511 // Don't try to evaluate aliases. External weak GV can be null.
9512 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9514 // Otherwise, be a little bit agressive by scanning the local block where we
9515 // want to check to see if the pointer is already being loaded or stored
9516 // from/to. If so, the previous load or store would have already trapped,
9517 // so there is no harm doing an extra load (also, CSE will later eliminate
9518 // the load entirely).
9519 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9524 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9525 if (LI->getOperand(0) == V) return true;
9526 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9527 if (SI->getOperand(1) == V) return true;
9533 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9534 /// until we find the underlying object a pointer is referring to or something
9535 /// we don't understand. Note that the returned pointer may be offset from the
9536 /// input, because we ignore GEP indices.
9537 static Value *GetUnderlyingObject(Value *Ptr) {
9539 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9540 if (CE->getOpcode() == Instruction::BitCast ||
9541 CE->getOpcode() == Instruction::GetElementPtr)
9542 Ptr = CE->getOperand(0);
9545 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9546 Ptr = BCI->getOperand(0);
9547 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9548 Ptr = GEP->getOperand(0);
9555 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9556 Value *Op = LI.getOperand(0);
9558 // Attempt to improve the alignment.
9559 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9560 if (KnownAlign > LI.getAlignment())
9561 LI.setAlignment(KnownAlign);
9563 // load (cast X) --> cast (load X) iff safe
9564 if (isa<CastInst>(Op))
9565 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9568 // None of the following transforms are legal for volatile loads.
9569 if (LI.isVolatile()) return 0;
9571 if (&LI.getParent()->front() != &LI) {
9572 BasicBlock::iterator BBI = &LI; --BBI;
9573 // If the instruction immediately before this is a store to the same
9574 // address, do a simple form of store->load forwarding.
9575 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9576 if (SI->getOperand(1) == LI.getOperand(0))
9577 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9578 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9579 if (LIB->getOperand(0) == LI.getOperand(0))
9580 return ReplaceInstUsesWith(LI, LIB);
9583 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9584 const Value *GEPI0 = GEPI->getOperand(0);
9585 // TODO: Consider a target hook for valid address spaces for this xform.
9586 if (isa<ConstantPointerNull>(GEPI0) &&
9587 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
9588 // Insert a new store to null instruction before the load to indicate
9589 // that this code is not reachable. We do this instead of inserting
9590 // an unreachable instruction directly because we cannot modify the
9592 new StoreInst(UndefValue::get(LI.getType()),
9593 Constant::getNullValue(Op->getType()), &LI);
9594 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9598 if (Constant *C = dyn_cast<Constant>(Op)) {
9599 // load null/undef -> undef
9600 // TODO: Consider a target hook for valid address spaces for this xform.
9601 if (isa<UndefValue>(C) || (C->isNullValue() &&
9602 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
9603 // Insert a new store to null instruction before the load to indicate that
9604 // this code is not reachable. We do this instead of inserting an
9605 // unreachable instruction directly because we cannot modify the CFG.
9606 new StoreInst(UndefValue::get(LI.getType()),
9607 Constant::getNullValue(Op->getType()), &LI);
9608 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9611 // Instcombine load (constant global) into the value loaded.
9612 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9613 if (GV->isConstant() && !GV->isDeclaration())
9614 return ReplaceInstUsesWith(LI, GV->getInitializer());
9616 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9617 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9618 if (CE->getOpcode() == Instruction::GetElementPtr) {
9619 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9620 if (GV->isConstant() && !GV->isDeclaration())
9622 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9623 return ReplaceInstUsesWith(LI, V);
9624 if (CE->getOperand(0)->isNullValue()) {
9625 // Insert a new store to null instruction before the load to indicate
9626 // that this code is not reachable. We do this instead of inserting
9627 // an unreachable instruction directly because we cannot modify the
9629 new StoreInst(UndefValue::get(LI.getType()),
9630 Constant::getNullValue(Op->getType()), &LI);
9631 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9634 } else if (CE->isCast()) {
9635 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9640 // If this load comes from anywhere in a constant global, and if the global
9641 // is all undef or zero, we know what it loads.
9642 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9643 if (GV->isConstant() && GV->hasInitializer()) {
9644 if (GV->getInitializer()->isNullValue())
9645 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9646 else if (isa<UndefValue>(GV->getInitializer()))
9647 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9651 if (Op->hasOneUse()) {
9652 // Change select and PHI nodes to select values instead of addresses: this
9653 // helps alias analysis out a lot, allows many others simplifications, and
9654 // exposes redundancy in the code.
9656 // Note that we cannot do the transformation unless we know that the
9657 // introduced loads cannot trap! Something like this is valid as long as
9658 // the condition is always false: load (select bool %C, int* null, int* %G),
9659 // but it would not be valid if we transformed it to load from null
9662 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9663 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9664 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9665 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9666 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9667 SI->getOperand(1)->getName()+".val"), LI);
9668 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9669 SI->getOperand(2)->getName()+".val"), LI);
9670 return new SelectInst(SI->getCondition(), V1, V2);
9673 // load (select (cond, null, P)) -> load P
9674 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9675 if (C->isNullValue()) {
9676 LI.setOperand(0, SI->getOperand(2));
9680 // load (select (cond, P, null)) -> load P
9681 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9682 if (C->isNullValue()) {
9683 LI.setOperand(0, SI->getOperand(1));
9691 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9693 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9694 User *CI = cast<User>(SI.getOperand(1));
9695 Value *CastOp = CI->getOperand(0);
9697 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9698 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9699 const Type *SrcPTy = SrcTy->getElementType();
9701 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9702 // If the source is an array, the code below will not succeed. Check to
9703 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9705 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9706 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9707 if (ASrcTy->getNumElements() != 0) {
9709 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9710 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9711 SrcTy = cast<PointerType>(CastOp->getType());
9712 SrcPTy = SrcTy->getElementType();
9715 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9716 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9717 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9719 // Okay, we are casting from one integer or pointer type to another of
9720 // the same size. Instead of casting the pointer before
9721 // the store, cast the value to be stored.
9723 Value *SIOp0 = SI.getOperand(0);
9724 Instruction::CastOps opcode = Instruction::BitCast;
9725 const Type* CastSrcTy = SIOp0->getType();
9726 const Type* CastDstTy = SrcPTy;
9727 if (isa<PointerType>(CastDstTy)) {
9728 if (CastSrcTy->isInteger())
9729 opcode = Instruction::IntToPtr;
9730 } else if (isa<IntegerType>(CastDstTy)) {
9731 if (isa<PointerType>(SIOp0->getType()))
9732 opcode = Instruction::PtrToInt;
9734 if (Constant *C = dyn_cast<Constant>(SIOp0))
9735 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9737 NewCast = IC.InsertNewInstBefore(
9738 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9740 return new StoreInst(NewCast, CastOp);
9747 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9748 Value *Val = SI.getOperand(0);
9749 Value *Ptr = SI.getOperand(1);
9751 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9752 EraseInstFromFunction(SI);
9757 // If the RHS is an alloca with a single use, zapify the store, making the
9759 if (Ptr->hasOneUse()) {
9760 if (isa<AllocaInst>(Ptr)) {
9761 EraseInstFromFunction(SI);
9766 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9767 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9768 GEP->getOperand(0)->hasOneUse()) {
9769 EraseInstFromFunction(SI);
9775 // Attempt to improve the alignment.
9776 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9777 if (KnownAlign > SI.getAlignment())
9778 SI.setAlignment(KnownAlign);
9780 // Do really simple DSE, to catch cases where there are several consequtive
9781 // stores to the same location, separated by a few arithmetic operations. This
9782 // situation often occurs with bitfield accesses.
9783 BasicBlock::iterator BBI = &SI;
9784 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9788 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9789 // Prev store isn't volatile, and stores to the same location?
9790 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9793 EraseInstFromFunction(*PrevSI);
9799 // If this is a load, we have to stop. However, if the loaded value is from
9800 // the pointer we're loading and is producing the pointer we're storing,
9801 // then *this* store is dead (X = load P; store X -> P).
9802 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9803 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9804 EraseInstFromFunction(SI);
9808 // Otherwise, this is a load from some other location. Stores before it
9813 // Don't skip over loads or things that can modify memory.
9814 if (BBI->mayWriteToMemory())
9819 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9821 // store X, null -> turns into 'unreachable' in SimplifyCFG
9822 if (isa<ConstantPointerNull>(Ptr)) {
9823 if (!isa<UndefValue>(Val)) {
9824 SI.setOperand(0, UndefValue::get(Val->getType()));
9825 if (Instruction *U = dyn_cast<Instruction>(Val))
9826 AddToWorkList(U); // Dropped a use.
9829 return 0; // Do not modify these!
9832 // store undef, Ptr -> noop
9833 if (isa<UndefValue>(Val)) {
9834 EraseInstFromFunction(SI);
9839 // If the pointer destination is a cast, see if we can fold the cast into the
9841 if (isa<CastInst>(Ptr))
9842 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9844 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9846 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9850 // If this store is the last instruction in the basic block, and if the block
9851 // ends with an unconditional branch, try to move it to the successor block.
9853 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9854 if (BI->isUnconditional())
9855 if (SimplifyStoreAtEndOfBlock(SI))
9856 return 0; // xform done!
9861 /// SimplifyStoreAtEndOfBlock - Turn things like:
9862 /// if () { *P = v1; } else { *P = v2 }
9863 /// into a phi node with a store in the successor.
9865 /// Simplify things like:
9866 /// *P = v1; if () { *P = v2; }
9867 /// into a phi node with a store in the successor.
9869 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9870 BasicBlock *StoreBB = SI.getParent();
9872 // Check to see if the successor block has exactly two incoming edges. If
9873 // so, see if the other predecessor contains a store to the same location.
9874 // if so, insert a PHI node (if needed) and move the stores down.
9875 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9877 // Determine whether Dest has exactly two predecessors and, if so, compute
9878 // the other predecessor.
9879 pred_iterator PI = pred_begin(DestBB);
9880 BasicBlock *OtherBB = 0;
9884 if (PI == pred_end(DestBB))
9887 if (*PI != StoreBB) {
9892 if (++PI != pred_end(DestBB))
9896 // Verify that the other block ends in a branch and is not otherwise empty.
9897 BasicBlock::iterator BBI = OtherBB->getTerminator();
9898 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9899 if (!OtherBr || BBI == OtherBB->begin())
9902 // If the other block ends in an unconditional branch, check for the 'if then
9903 // else' case. there is an instruction before the branch.
9904 StoreInst *OtherStore = 0;
9905 if (OtherBr->isUnconditional()) {
9906 // If this isn't a store, or isn't a store to the same location, bail out.
9908 OtherStore = dyn_cast<StoreInst>(BBI);
9909 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9912 // Otherwise, the other block ended with a conditional branch. If one of the
9913 // destinations is StoreBB, then we have the if/then case.
9914 if (OtherBr->getSuccessor(0) != StoreBB &&
9915 OtherBr->getSuccessor(1) != StoreBB)
9918 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9919 // if/then triangle. See if there is a store to the same ptr as SI that
9920 // lives in OtherBB.
9922 // Check to see if we find the matching store.
9923 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9924 if (OtherStore->getOperand(1) != SI.getOperand(1))
9928 // If we find something that may be using the stored value, or if we run
9929 // out of instructions, we can't do the xform.
9930 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9931 BBI == OtherBB->begin())
9935 // In order to eliminate the store in OtherBr, we have to
9936 // make sure nothing reads the stored value in StoreBB.
9937 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9938 // FIXME: This should really be AA driven.
9939 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9944 // Insert a PHI node now if we need it.
9945 Value *MergedVal = OtherStore->getOperand(0);
9946 if (MergedVal != SI.getOperand(0)) {
9947 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9948 PN->reserveOperandSpace(2);
9949 PN->addIncoming(SI.getOperand(0), SI.getParent());
9950 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9951 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9954 // Advance to a place where it is safe to insert the new store and
9956 BBI = DestBB->begin();
9957 while (isa<PHINode>(BBI)) ++BBI;
9958 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9959 OtherStore->isVolatile()), *BBI);
9961 // Nuke the old stores.
9962 EraseInstFromFunction(SI);
9963 EraseInstFromFunction(*OtherStore);
9969 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9970 // Change br (not X), label True, label False to: br X, label False, True
9972 BasicBlock *TrueDest;
9973 BasicBlock *FalseDest;
9974 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9975 !isa<Constant>(X)) {
9976 // Swap Destinations and condition...
9978 BI.setSuccessor(0, FalseDest);
9979 BI.setSuccessor(1, TrueDest);
9983 // Cannonicalize fcmp_one -> fcmp_oeq
9984 FCmpInst::Predicate FPred; Value *Y;
9985 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9986 TrueDest, FalseDest)))
9987 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9988 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9989 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9990 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9991 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9992 NewSCC->takeName(I);
9993 // Swap Destinations and condition...
9994 BI.setCondition(NewSCC);
9995 BI.setSuccessor(0, FalseDest);
9996 BI.setSuccessor(1, TrueDest);
9997 RemoveFromWorkList(I);
9998 I->eraseFromParent();
9999 AddToWorkList(NewSCC);
10003 // Cannonicalize icmp_ne -> icmp_eq
10004 ICmpInst::Predicate IPred;
10005 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10006 TrueDest, FalseDest)))
10007 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10008 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10009 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10010 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10011 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10012 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10013 NewSCC->takeName(I);
10014 // Swap Destinations and condition...
10015 BI.setCondition(NewSCC);
10016 BI.setSuccessor(0, FalseDest);
10017 BI.setSuccessor(1, TrueDest);
10018 RemoveFromWorkList(I);
10019 I->eraseFromParent();;
10020 AddToWorkList(NewSCC);
10027 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10028 Value *Cond = SI.getCondition();
10029 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10030 if (I->getOpcode() == Instruction::Add)
10031 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10032 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10033 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10034 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10036 SI.setOperand(0, I->getOperand(0));
10044 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10045 /// is to leave as a vector operation.
10046 static bool CheapToScalarize(Value *V, bool isConstant) {
10047 if (isa<ConstantAggregateZero>(V))
10049 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10050 if (isConstant) return true;
10051 // If all elts are the same, we can extract.
10052 Constant *Op0 = C->getOperand(0);
10053 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10054 if (C->getOperand(i) != Op0)
10058 Instruction *I = dyn_cast<Instruction>(V);
10059 if (!I) return false;
10061 // Insert element gets simplified to the inserted element or is deleted if
10062 // this is constant idx extract element and its a constant idx insertelt.
10063 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10064 isa<ConstantInt>(I->getOperand(2)))
10066 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10068 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10069 if (BO->hasOneUse() &&
10070 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10071 CheapToScalarize(BO->getOperand(1), isConstant)))
10073 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10074 if (CI->hasOneUse() &&
10075 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10076 CheapToScalarize(CI->getOperand(1), isConstant)))
10082 /// Read and decode a shufflevector mask.
10084 /// It turns undef elements into values that are larger than the number of
10085 /// elements in the input.
10086 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10087 unsigned NElts = SVI->getType()->getNumElements();
10088 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10089 return std::vector<unsigned>(NElts, 0);
10090 if (isa<UndefValue>(SVI->getOperand(2)))
10091 return std::vector<unsigned>(NElts, 2*NElts);
10093 std::vector<unsigned> Result;
10094 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10095 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
10096 if (isa<UndefValue>(CP->getOperand(i)))
10097 Result.push_back(NElts*2); // undef -> 8
10099 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
10103 /// FindScalarElement - Given a vector and an element number, see if the scalar
10104 /// value is already around as a register, for example if it were inserted then
10105 /// extracted from the vector.
10106 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10107 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10108 const VectorType *PTy = cast<VectorType>(V->getType());
10109 unsigned Width = PTy->getNumElements();
10110 if (EltNo >= Width) // Out of range access.
10111 return UndefValue::get(PTy->getElementType());
10113 if (isa<UndefValue>(V))
10114 return UndefValue::get(PTy->getElementType());
10115 else if (isa<ConstantAggregateZero>(V))
10116 return Constant::getNullValue(PTy->getElementType());
10117 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10118 return CP->getOperand(EltNo);
10119 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10120 // If this is an insert to a variable element, we don't know what it is.
10121 if (!isa<ConstantInt>(III->getOperand(2)))
10123 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10125 // If this is an insert to the element we are looking for, return the
10127 if (EltNo == IIElt)
10128 return III->getOperand(1);
10130 // Otherwise, the insertelement doesn't modify the value, recurse on its
10132 return FindScalarElement(III->getOperand(0), EltNo);
10133 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10134 unsigned InEl = getShuffleMask(SVI)[EltNo];
10136 return FindScalarElement(SVI->getOperand(0), InEl);
10137 else if (InEl < Width*2)
10138 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10140 return UndefValue::get(PTy->getElementType());
10143 // Otherwise, we don't know.
10147 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10149 // If vector val is undef, replace extract with scalar undef.
10150 if (isa<UndefValue>(EI.getOperand(0)))
10151 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10153 // If vector val is constant 0, replace extract with scalar 0.
10154 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10155 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10157 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10158 // If vector val is constant with uniform operands, replace EI
10159 // with that operand
10160 Constant *op0 = C->getOperand(0);
10161 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10162 if (C->getOperand(i) != op0) {
10167 return ReplaceInstUsesWith(EI, op0);
10170 // If extracting a specified index from the vector, see if we can recursively
10171 // find a previously computed scalar that was inserted into the vector.
10172 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10173 unsigned IndexVal = IdxC->getZExtValue();
10174 unsigned VectorWidth =
10175 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10177 // If this is extracting an invalid index, turn this into undef, to avoid
10178 // crashing the code below.
10179 if (IndexVal >= VectorWidth)
10180 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10182 // This instruction only demands the single element from the input vector.
10183 // If the input vector has a single use, simplify it based on this use
10185 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10186 uint64_t UndefElts;
10187 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10190 EI.setOperand(0, V);
10195 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10196 return ReplaceInstUsesWith(EI, Elt);
10198 // If the this extractelement is directly using a bitcast from a vector of
10199 // the same number of elements, see if we can find the source element from
10200 // it. In this case, we will end up needing to bitcast the scalars.
10201 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10202 if (const VectorType *VT =
10203 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10204 if (VT->getNumElements() == VectorWidth)
10205 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10206 return new BitCastInst(Elt, EI.getType());
10210 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10211 if (I->hasOneUse()) {
10212 // Push extractelement into predecessor operation if legal and
10213 // profitable to do so
10214 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10215 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10216 if (CheapToScalarize(BO, isConstantElt)) {
10217 ExtractElementInst *newEI0 =
10218 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10219 EI.getName()+".lhs");
10220 ExtractElementInst *newEI1 =
10221 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10222 EI.getName()+".rhs");
10223 InsertNewInstBefore(newEI0, EI);
10224 InsertNewInstBefore(newEI1, EI);
10225 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10227 } else if (isa<LoadInst>(I)) {
10229 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10230 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10231 PointerType::get(EI.getType(), AS),EI);
10232 GetElementPtrInst *GEP =
10233 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
10234 InsertNewInstBefore(GEP, EI);
10235 return new LoadInst(GEP);
10238 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10239 // Extracting the inserted element?
10240 if (IE->getOperand(2) == EI.getOperand(1))
10241 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10242 // If the inserted and extracted elements are constants, they must not
10243 // be the same value, extract from the pre-inserted value instead.
10244 if (isa<Constant>(IE->getOperand(2)) &&
10245 isa<Constant>(EI.getOperand(1))) {
10246 AddUsesToWorkList(EI);
10247 EI.setOperand(0, IE->getOperand(0));
10250 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10251 // If this is extracting an element from a shufflevector, figure out where
10252 // it came from and extract from the appropriate input element instead.
10253 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10254 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10256 if (SrcIdx < SVI->getType()->getNumElements())
10257 Src = SVI->getOperand(0);
10258 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10259 SrcIdx -= SVI->getType()->getNumElements();
10260 Src = SVI->getOperand(1);
10262 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10264 return new ExtractElementInst(Src, SrcIdx);
10271 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10272 /// elements from either LHS or RHS, return the shuffle mask and true.
10273 /// Otherwise, return false.
10274 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10275 std::vector<Constant*> &Mask) {
10276 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10277 "Invalid CollectSingleShuffleElements");
10278 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10280 if (isa<UndefValue>(V)) {
10281 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10283 } else if (V == LHS) {
10284 for (unsigned i = 0; i != NumElts; ++i)
10285 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10287 } else if (V == RHS) {
10288 for (unsigned i = 0; i != NumElts; ++i)
10289 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10291 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10292 // If this is an insert of an extract from some other vector, include it.
10293 Value *VecOp = IEI->getOperand(0);
10294 Value *ScalarOp = IEI->getOperand(1);
10295 Value *IdxOp = IEI->getOperand(2);
10297 if (!isa<ConstantInt>(IdxOp))
10299 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10301 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10302 // Okay, we can handle this if the vector we are insertinting into is
10303 // transitively ok.
10304 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10305 // If so, update the mask to reflect the inserted undef.
10306 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10309 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10310 if (isa<ConstantInt>(EI->getOperand(1)) &&
10311 EI->getOperand(0)->getType() == V->getType()) {
10312 unsigned ExtractedIdx =
10313 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10315 // This must be extracting from either LHS or RHS.
10316 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10317 // Okay, we can handle this if the vector we are insertinting into is
10318 // transitively ok.
10319 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10320 // If so, update the mask to reflect the inserted value.
10321 if (EI->getOperand(0) == LHS) {
10322 Mask[InsertedIdx & (NumElts-1)] =
10323 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10325 assert(EI->getOperand(0) == RHS);
10326 Mask[InsertedIdx & (NumElts-1)] =
10327 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10336 // TODO: Handle shufflevector here!
10341 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10342 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10343 /// that computes V and the LHS value of the shuffle.
10344 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10346 assert(isa<VectorType>(V->getType()) &&
10347 (RHS == 0 || V->getType() == RHS->getType()) &&
10348 "Invalid shuffle!");
10349 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10351 if (isa<UndefValue>(V)) {
10352 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10354 } else if (isa<ConstantAggregateZero>(V)) {
10355 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10357 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10358 // If this is an insert of an extract from some other vector, include it.
10359 Value *VecOp = IEI->getOperand(0);
10360 Value *ScalarOp = IEI->getOperand(1);
10361 Value *IdxOp = IEI->getOperand(2);
10363 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10364 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10365 EI->getOperand(0)->getType() == V->getType()) {
10366 unsigned ExtractedIdx =
10367 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10368 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10370 // Either the extracted from or inserted into vector must be RHSVec,
10371 // otherwise we'd end up with a shuffle of three inputs.
10372 if (EI->getOperand(0) == RHS || RHS == 0) {
10373 RHS = EI->getOperand(0);
10374 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10375 Mask[InsertedIdx & (NumElts-1)] =
10376 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10380 if (VecOp == RHS) {
10381 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10382 // Everything but the extracted element is replaced with the RHS.
10383 for (unsigned i = 0; i != NumElts; ++i) {
10384 if (i != InsertedIdx)
10385 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10390 // If this insertelement is a chain that comes from exactly these two
10391 // vectors, return the vector and the effective shuffle.
10392 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10393 return EI->getOperand(0);
10398 // TODO: Handle shufflevector here!
10400 // Otherwise, can't do anything fancy. Return an identity vector.
10401 for (unsigned i = 0; i != NumElts; ++i)
10402 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10406 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10407 Value *VecOp = IE.getOperand(0);
10408 Value *ScalarOp = IE.getOperand(1);
10409 Value *IdxOp = IE.getOperand(2);
10411 // Inserting an undef or into an undefined place, remove this.
10412 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10413 ReplaceInstUsesWith(IE, VecOp);
10415 // If the inserted element was extracted from some other vector, and if the
10416 // indexes are constant, try to turn this into a shufflevector operation.
10417 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10418 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10419 EI->getOperand(0)->getType() == IE.getType()) {
10420 unsigned NumVectorElts = IE.getType()->getNumElements();
10421 unsigned ExtractedIdx =
10422 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10423 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10425 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10426 return ReplaceInstUsesWith(IE, VecOp);
10428 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10429 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10431 // If we are extracting a value from a vector, then inserting it right
10432 // back into the same place, just use the input vector.
10433 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10434 return ReplaceInstUsesWith(IE, VecOp);
10436 // We could theoretically do this for ANY input. However, doing so could
10437 // turn chains of insertelement instructions into a chain of shufflevector
10438 // instructions, and right now we do not merge shufflevectors. As such,
10439 // only do this in a situation where it is clear that there is benefit.
10440 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10441 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10442 // the values of VecOp, except then one read from EIOp0.
10443 // Build a new shuffle mask.
10444 std::vector<Constant*> Mask;
10445 if (isa<UndefValue>(VecOp))
10446 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10448 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10449 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10452 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10453 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10454 ConstantVector::get(Mask));
10457 // If this insertelement isn't used by some other insertelement, turn it
10458 // (and any insertelements it points to), into one big shuffle.
10459 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10460 std::vector<Constant*> Mask;
10462 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10463 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10464 // We now have a shuffle of LHS, RHS, Mask.
10465 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10474 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10475 Value *LHS = SVI.getOperand(0);
10476 Value *RHS = SVI.getOperand(1);
10477 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10479 bool MadeChange = false;
10481 // Undefined shuffle mask -> undefined value.
10482 if (isa<UndefValue>(SVI.getOperand(2)))
10483 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10485 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10486 // the undef, change them to undefs.
10487 if (isa<UndefValue>(SVI.getOperand(1))) {
10488 // Scan to see if there are any references to the RHS. If so, replace them
10489 // with undef element refs and set MadeChange to true.
10490 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10491 if (Mask[i] >= e && Mask[i] != 2*e) {
10498 // Remap any references to RHS to use LHS.
10499 std::vector<Constant*> Elts;
10500 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10501 if (Mask[i] == 2*e)
10502 Elts.push_back(UndefValue::get(Type::Int32Ty));
10504 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10506 SVI.setOperand(2, ConstantVector::get(Elts));
10510 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10511 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10512 if (LHS == RHS || isa<UndefValue>(LHS)) {
10513 if (isa<UndefValue>(LHS) && LHS == RHS) {
10514 // shuffle(undef,undef,mask) -> undef.
10515 return ReplaceInstUsesWith(SVI, LHS);
10518 // Remap any references to RHS to use LHS.
10519 std::vector<Constant*> Elts;
10520 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10521 if (Mask[i] >= 2*e)
10522 Elts.push_back(UndefValue::get(Type::Int32Ty));
10524 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10525 (Mask[i] < e && isa<UndefValue>(LHS)))
10526 Mask[i] = 2*e; // Turn into undef.
10528 Mask[i] &= (e-1); // Force to LHS.
10529 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10532 SVI.setOperand(0, SVI.getOperand(1));
10533 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10534 SVI.setOperand(2, ConstantVector::get(Elts));
10535 LHS = SVI.getOperand(0);
10536 RHS = SVI.getOperand(1);
10540 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10541 bool isLHSID = true, isRHSID = true;
10543 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10544 if (Mask[i] >= e*2) continue; // Ignore undef values.
10545 // Is this an identity shuffle of the LHS value?
10546 isLHSID &= (Mask[i] == i);
10548 // Is this an identity shuffle of the RHS value?
10549 isRHSID &= (Mask[i]-e == i);
10552 // Eliminate identity shuffles.
10553 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10554 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10556 // If the LHS is a shufflevector itself, see if we can combine it with this
10557 // one without producing an unusual shuffle. Here we are really conservative:
10558 // we are absolutely afraid of producing a shuffle mask not in the input
10559 // program, because the code gen may not be smart enough to turn a merged
10560 // shuffle into two specific shuffles: it may produce worse code. As such,
10561 // we only merge two shuffles if the result is one of the two input shuffle
10562 // masks. In this case, merging the shuffles just removes one instruction,
10563 // which we know is safe. This is good for things like turning:
10564 // (splat(splat)) -> splat.
10565 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10566 if (isa<UndefValue>(RHS)) {
10567 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10569 std::vector<unsigned> NewMask;
10570 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10571 if (Mask[i] >= 2*e)
10572 NewMask.push_back(2*e);
10574 NewMask.push_back(LHSMask[Mask[i]]);
10576 // If the result mask is equal to the src shuffle or this shuffle mask, do
10577 // the replacement.
10578 if (NewMask == LHSMask || NewMask == Mask) {
10579 std::vector<Constant*> Elts;
10580 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10581 if (NewMask[i] >= e*2) {
10582 Elts.push_back(UndefValue::get(Type::Int32Ty));
10584 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10587 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10588 LHSSVI->getOperand(1),
10589 ConstantVector::get(Elts));
10594 return MadeChange ? &SVI : 0;
10600 /// TryToSinkInstruction - Try to move the specified instruction from its
10601 /// current block into the beginning of DestBlock, which can only happen if it's
10602 /// safe to move the instruction past all of the instructions between it and the
10603 /// end of its block.
10604 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10605 assert(I->hasOneUse() && "Invariants didn't hold!");
10607 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10608 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10610 // Do not sink alloca instructions out of the entry block.
10611 if (isa<AllocaInst>(I) && I->getParent() ==
10612 &DestBlock->getParent()->getEntryBlock())
10615 // We can only sink load instructions if there is nothing between the load and
10616 // the end of block that could change the value.
10617 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10618 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10620 if (Scan->mayWriteToMemory())
10624 BasicBlock::iterator InsertPos = DestBlock->begin();
10625 while (isa<PHINode>(InsertPos)) ++InsertPos;
10627 I->moveBefore(InsertPos);
10633 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10634 /// all reachable code to the worklist.
10636 /// This has a couple of tricks to make the code faster and more powerful. In
10637 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10638 /// them to the worklist (this significantly speeds up instcombine on code where
10639 /// many instructions are dead or constant). Additionally, if we find a branch
10640 /// whose condition is a known constant, we only visit the reachable successors.
10642 static void AddReachableCodeToWorklist(BasicBlock *BB,
10643 SmallPtrSet<BasicBlock*, 64> &Visited,
10645 const TargetData *TD) {
10646 std::vector<BasicBlock*> Worklist;
10647 Worklist.push_back(BB);
10649 while (!Worklist.empty()) {
10650 BB = Worklist.back();
10651 Worklist.pop_back();
10653 // We have now visited this block! If we've already been here, ignore it.
10654 if (!Visited.insert(BB)) continue;
10656 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10657 Instruction *Inst = BBI++;
10659 // DCE instruction if trivially dead.
10660 if (isInstructionTriviallyDead(Inst)) {
10662 DOUT << "IC: DCE: " << *Inst;
10663 Inst->eraseFromParent();
10667 // ConstantProp instruction if trivially constant.
10668 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10669 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10670 Inst->replaceAllUsesWith(C);
10672 Inst->eraseFromParent();
10676 IC.AddToWorkList(Inst);
10679 // Recursively visit successors. If this is a branch or switch on a
10680 // constant, only visit the reachable successor.
10681 TerminatorInst *TI = BB->getTerminator();
10682 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10683 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10684 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10685 Worklist.push_back(BI->getSuccessor(!CondVal));
10688 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10689 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10690 // See if this is an explicit destination.
10691 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10692 if (SI->getCaseValue(i) == Cond) {
10693 Worklist.push_back(SI->getSuccessor(i));
10697 // Otherwise it is the default destination.
10698 Worklist.push_back(SI->getSuccessor(0));
10703 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10704 Worklist.push_back(TI->getSuccessor(i));
10708 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10709 bool Changed = false;
10710 TD = &getAnalysis<TargetData>();
10712 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10713 << F.getNameStr() << "\n");
10716 // Do a depth-first traversal of the function, populate the worklist with
10717 // the reachable instructions. Ignore blocks that are not reachable. Keep
10718 // track of which blocks we visit.
10719 SmallPtrSet<BasicBlock*, 64> Visited;
10720 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10722 // Do a quick scan over the function. If we find any blocks that are
10723 // unreachable, remove any instructions inside of them. This prevents
10724 // the instcombine code from having to deal with some bad special cases.
10725 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10726 if (!Visited.count(BB)) {
10727 Instruction *Term = BB->getTerminator();
10728 while (Term != BB->begin()) { // Remove instrs bottom-up
10729 BasicBlock::iterator I = Term; --I;
10731 DOUT << "IC: DCE: " << *I;
10734 if (!I->use_empty())
10735 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10736 I->eraseFromParent();
10741 while (!Worklist.empty()) {
10742 Instruction *I = RemoveOneFromWorkList();
10743 if (I == 0) continue; // skip null values.
10745 // Check to see if we can DCE the instruction.
10746 if (isInstructionTriviallyDead(I)) {
10747 // Add operands to the worklist.
10748 if (I->getNumOperands() < 4)
10749 AddUsesToWorkList(*I);
10752 DOUT << "IC: DCE: " << *I;
10754 I->eraseFromParent();
10755 RemoveFromWorkList(I);
10759 // Instruction isn't dead, see if we can constant propagate it.
10760 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10761 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10763 // Add operands to the worklist.
10764 AddUsesToWorkList(*I);
10765 ReplaceInstUsesWith(*I, C);
10768 I->eraseFromParent();
10769 RemoveFromWorkList(I);
10773 // See if we can trivially sink this instruction to a successor basic block.
10774 if (I->hasOneUse()) {
10775 BasicBlock *BB = I->getParent();
10776 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10777 if (UserParent != BB) {
10778 bool UserIsSuccessor = false;
10779 // See if the user is one of our successors.
10780 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10781 if (*SI == UserParent) {
10782 UserIsSuccessor = true;
10786 // If the user is one of our immediate successors, and if that successor
10787 // only has us as a predecessors (we'd have to split the critical edge
10788 // otherwise), we can keep going.
10789 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10790 next(pred_begin(UserParent)) == pred_end(UserParent))
10791 // Okay, the CFG is simple enough, try to sink this instruction.
10792 Changed |= TryToSinkInstruction(I, UserParent);
10796 // Now that we have an instruction, try combining it to simplify it...
10800 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10801 if (Instruction *Result = visit(*I)) {
10803 // Should we replace the old instruction with a new one?
10805 DOUT << "IC: Old = " << *I
10806 << " New = " << *Result;
10808 // Everything uses the new instruction now.
10809 I->replaceAllUsesWith(Result);
10811 // Push the new instruction and any users onto the worklist.
10812 AddToWorkList(Result);
10813 AddUsersToWorkList(*Result);
10815 // Move the name to the new instruction first.
10816 Result->takeName(I);
10818 // Insert the new instruction into the basic block...
10819 BasicBlock *InstParent = I->getParent();
10820 BasicBlock::iterator InsertPos = I;
10822 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10823 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10826 InstParent->getInstList().insert(InsertPos, Result);
10828 // Make sure that we reprocess all operands now that we reduced their
10830 AddUsesToWorkList(*I);
10832 // Instructions can end up on the worklist more than once. Make sure
10833 // we do not process an instruction that has been deleted.
10834 RemoveFromWorkList(I);
10836 // Erase the old instruction.
10837 InstParent->getInstList().erase(I);
10840 DOUT << "IC: Mod = " << OrigI
10841 << " New = " << *I;
10844 // If the instruction was modified, it's possible that it is now dead.
10845 // if so, remove it.
10846 if (isInstructionTriviallyDead(I)) {
10847 // Make sure we process all operands now that we are reducing their
10849 AddUsesToWorkList(*I);
10851 // Instructions may end up in the worklist more than once. Erase all
10852 // occurrences of this instruction.
10853 RemoveFromWorkList(I);
10854 I->eraseFromParent();
10857 AddUsersToWorkList(*I);
10864 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10866 // Do an explicit clear, this shrinks the map if needed.
10867 WorklistMap.clear();
10872 bool InstCombiner::runOnFunction(Function &F) {
10873 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10875 bool EverMadeChange = false;
10877 // Iterate while there is work to do.
10878 unsigned Iteration = 0;
10879 while (DoOneIteration(F, Iteration++))
10880 EverMadeChange = true;
10881 return EverMadeChange;
10884 FunctionPass *llvm::createInstructionCombiningPass() {
10885 return new InstCombiner();