1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/GetElementPtrTypeIterator.h"
49 #include "llvm/Support/InstVisitor.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/PatternMatch.h"
52 #include "llvm/Support/Compiler.h"
53 #include "llvm/ADT/DenseMap.h"
54 #include "llvm/ADT/SmallVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/ADT/STLExtras.h"
61 using namespace llvm::PatternMatch;
63 STATISTIC(NumCombined , "Number of insts combined");
64 STATISTIC(NumConstProp, "Number of constant folds");
65 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
66 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
67 STATISTIC(NumSunkInst , "Number of instructions sunk");
70 class VISIBILITY_HIDDEN InstCombiner
71 : public FunctionPass,
72 public InstVisitor<InstCombiner, Instruction*> {
73 // Worklist of all of the instructions that need to be simplified.
74 std::vector<Instruction*> Worklist;
75 DenseMap<Instruction*, unsigned> WorklistMap;
78 /// AddToWorkList - Add the specified instruction to the worklist if it
79 /// isn't already in it.
80 void AddToWorkList(Instruction *I) {
81 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
82 Worklist.push_back(I);
85 // RemoveFromWorkList - remove I from the worklist if it exists.
86 void RemoveFromWorkList(Instruction *I) {
87 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
88 if (It == WorklistMap.end()) return; // Not in worklist.
90 // Don't bother moving everything down, just null out the slot.
91 Worklist[It->second] = 0;
93 WorklistMap.erase(It);
96 Instruction *RemoveOneFromWorkList() {
97 Instruction *I = Worklist.back();
104 /// AddUsersToWorkList - When an instruction is simplified, add all users of
105 /// the instruction to the work lists because they might get more simplified
108 void AddUsersToWorkList(Value &I) {
109 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
111 AddToWorkList(cast<Instruction>(*UI));
114 /// AddUsesToWorkList - When an instruction is simplified, add operands to
115 /// the work lists because they might get more simplified now.
117 void AddUsesToWorkList(Instruction &I) {
118 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
119 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
123 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
124 /// dead. Add all of its operands to the worklist, turning them into
125 /// undef's to reduce the number of uses of those instructions.
127 /// Return the specified operand before it is turned into an undef.
129 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
130 Value *R = I.getOperand(op);
132 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
133 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
135 // Set the operand to undef to drop the use.
136 I.setOperand(i, UndefValue::get(Op->getType()));
143 virtual bool runOnFunction(Function &F);
145 bool DoOneIteration(Function &F, unsigned ItNum);
147 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
148 AU.addRequired<TargetData>();
149 AU.addPreservedID(LCSSAID);
150 AU.setPreservesCFG();
153 TargetData &getTargetData() const { return *TD; }
155 // Visitation implementation - Implement instruction combining for different
156 // instruction types. The semantics are as follows:
158 // null - No change was made
159 // I - Change was made, I is still valid, I may be dead though
160 // otherwise - Change was made, replace I with returned instruction
162 Instruction *visitAdd(BinaryOperator &I);
163 Instruction *visitSub(BinaryOperator &I);
164 Instruction *visitMul(BinaryOperator &I);
165 Instruction *visitURem(BinaryOperator &I);
166 Instruction *visitSRem(BinaryOperator &I);
167 Instruction *visitFRem(BinaryOperator &I);
168 Instruction *commonRemTransforms(BinaryOperator &I);
169 Instruction *commonIRemTransforms(BinaryOperator &I);
170 Instruction *commonDivTransforms(BinaryOperator &I);
171 Instruction *commonIDivTransforms(BinaryOperator &I);
172 Instruction *visitUDiv(BinaryOperator &I);
173 Instruction *visitSDiv(BinaryOperator &I);
174 Instruction *visitFDiv(BinaryOperator &I);
175 Instruction *visitAnd(BinaryOperator &I);
176 Instruction *visitOr (BinaryOperator &I);
177 Instruction *visitXor(BinaryOperator &I);
178 Instruction *visitShl(BinaryOperator &I);
179 Instruction *visitAShr(BinaryOperator &I);
180 Instruction *visitLShr(BinaryOperator &I);
181 Instruction *commonShiftTransforms(BinaryOperator &I);
182 Instruction *visitFCmpInst(FCmpInst &I);
183 Instruction *visitICmpInst(ICmpInst &I);
184 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
186 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
187 ICmpInst::Predicate Cond, Instruction &I);
188 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
190 Instruction *commonCastTransforms(CastInst &CI);
191 Instruction *commonIntCastTransforms(CastInst &CI);
192 Instruction *visitTrunc(CastInst &CI);
193 Instruction *visitZExt(CastInst &CI);
194 Instruction *visitSExt(CastInst &CI);
195 Instruction *visitFPTrunc(CastInst &CI);
196 Instruction *visitFPExt(CastInst &CI);
197 Instruction *visitFPToUI(CastInst &CI);
198 Instruction *visitFPToSI(CastInst &CI);
199 Instruction *visitUIToFP(CastInst &CI);
200 Instruction *visitSIToFP(CastInst &CI);
201 Instruction *visitPtrToInt(CastInst &CI);
202 Instruction *visitIntToPtr(CastInst &CI);
203 Instruction *visitBitCast(CastInst &CI);
204 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
206 Instruction *visitSelectInst(SelectInst &CI);
207 Instruction *visitCallInst(CallInst &CI);
208 Instruction *visitInvokeInst(InvokeInst &II);
209 Instruction *visitPHINode(PHINode &PN);
210 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
211 Instruction *visitAllocationInst(AllocationInst &AI);
212 Instruction *visitFreeInst(FreeInst &FI);
213 Instruction *visitLoadInst(LoadInst &LI);
214 Instruction *visitStoreInst(StoreInst &SI);
215 Instruction *visitBranchInst(BranchInst &BI);
216 Instruction *visitSwitchInst(SwitchInst &SI);
217 Instruction *visitInsertElementInst(InsertElementInst &IE);
218 Instruction *visitExtractElementInst(ExtractElementInst &EI);
219 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
221 // visitInstruction - Specify what to return for unhandled instructions...
222 Instruction *visitInstruction(Instruction &I) { return 0; }
225 Instruction *visitCallSite(CallSite CS);
226 bool transformConstExprCastCall(CallSite CS);
229 // InsertNewInstBefore - insert an instruction New before instruction Old
230 // in the program. Add the new instruction to the worklist.
232 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
233 assert(New && New->getParent() == 0 &&
234 "New instruction already inserted into a basic block!");
235 BasicBlock *BB = Old.getParent();
236 BB->getInstList().insert(&Old, New); // Insert inst
241 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
242 /// This also adds the cast to the worklist. Finally, this returns the
244 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
246 if (V->getType() == Ty) return V;
248 if (Constant *CV = dyn_cast<Constant>(V))
249 return ConstantExpr::getCast(opc, CV, Ty);
251 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
256 // ReplaceInstUsesWith - This method is to be used when an instruction is
257 // found to be dead, replacable with another preexisting expression. Here
258 // we add all uses of I to the worklist, replace all uses of I with the new
259 // value, then return I, so that the inst combiner will know that I was
262 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
263 AddUsersToWorkList(I); // Add all modified instrs to worklist
265 I.replaceAllUsesWith(V);
268 // If we are replacing the instruction with itself, this must be in a
269 // segment of unreachable code, so just clobber the instruction.
270 I.replaceAllUsesWith(UndefValue::get(I.getType()));
275 // UpdateValueUsesWith - This method is to be used when an value is
276 // found to be replacable with another preexisting expression or was
277 // updated. Here we add all uses of I to the worklist, replace all uses of
278 // I with the new value (unless the instruction was just updated), then
279 // return true, so that the inst combiner will know that I was modified.
281 bool UpdateValueUsesWith(Value *Old, Value *New) {
282 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
284 Old->replaceAllUsesWith(New);
285 if (Instruction *I = dyn_cast<Instruction>(Old))
287 if (Instruction *I = dyn_cast<Instruction>(New))
292 // EraseInstFromFunction - When dealing with an instruction that has side
293 // effects or produces a void value, we can't rely on DCE to delete the
294 // instruction. Instead, visit methods should return the value returned by
296 Instruction *EraseInstFromFunction(Instruction &I) {
297 assert(I.use_empty() && "Cannot erase instruction that is used!");
298 AddUsesToWorkList(I);
299 RemoveFromWorkList(&I);
301 return 0; // Don't do anything with FI
305 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
306 /// InsertBefore instruction. This is specialized a bit to avoid inserting
307 /// casts that are known to not do anything...
309 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
310 Value *V, const Type *DestTy,
311 Instruction *InsertBefore);
313 /// SimplifyCommutative - This performs a few simplifications for
314 /// commutative operators.
315 bool SimplifyCommutative(BinaryOperator &I);
317 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
318 /// most-complex to least-complex order.
319 bool SimplifyCompare(CmpInst &I);
321 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
322 uint64_t &KnownZero, uint64_t &KnownOne,
325 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
326 uint64_t &UndefElts, unsigned Depth = 0);
328 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
329 // PHI node as operand #0, see if we can fold the instruction into the PHI
330 // (which is only possible if all operands to the PHI are constants).
331 Instruction *FoldOpIntoPhi(Instruction &I);
333 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
334 // operator and they all are only used by the PHI, PHI together their
335 // inputs, and do the operation once, to the result of the PHI.
336 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
337 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
340 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
341 ConstantInt *AndRHS, BinaryOperator &TheAnd);
343 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
344 bool isSub, Instruction &I);
345 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
346 bool isSigned, bool Inside, Instruction &IB);
347 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
348 Instruction *MatchBSwap(BinaryOperator &I);
350 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
353 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
356 // getComplexity: Assign a complexity or rank value to LLVM Values...
357 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
358 static unsigned getComplexity(Value *V) {
359 if (isa<Instruction>(V)) {
360 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
364 if (isa<Argument>(V)) return 3;
365 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
368 // isOnlyUse - Return true if this instruction will be deleted if we stop using
370 static bool isOnlyUse(Value *V) {
371 return V->hasOneUse() || isa<Constant>(V);
374 // getPromotedType - Return the specified type promoted as it would be to pass
375 // though a va_arg area...
376 static const Type *getPromotedType(const Type *Ty) {
377 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
378 if (ITy->getBitWidth() < 32)
379 return Type::Int32Ty;
380 } else if (Ty == Type::FloatTy)
381 return Type::DoubleTy;
385 /// getBitCastOperand - If the specified operand is a CastInst or a constant
386 /// expression bitcast, return the operand value, otherwise return null.
387 static Value *getBitCastOperand(Value *V) {
388 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
389 return I->getOperand(0);
390 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
391 if (CE->getOpcode() == Instruction::BitCast)
392 return CE->getOperand(0);
396 /// This function is a wrapper around CastInst::isEliminableCastPair. It
397 /// simply extracts arguments and returns what that function returns.
398 static Instruction::CastOps
399 isEliminableCastPair(
400 const CastInst *CI, ///< The first cast instruction
401 unsigned opcode, ///< The opcode of the second cast instruction
402 const Type *DstTy, ///< The target type for the second cast instruction
403 TargetData *TD ///< The target data for pointer size
406 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
407 const Type *MidTy = CI->getType(); // B from above
409 // Get the opcodes of the two Cast instructions
410 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
411 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
413 return Instruction::CastOps(
414 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
415 DstTy, TD->getIntPtrType()));
418 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
419 /// in any code being generated. It does not require codegen if V is simple
420 /// enough or if the cast can be folded into other casts.
421 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
422 const Type *Ty, TargetData *TD) {
423 if (V->getType() == Ty || isa<Constant>(V)) return false;
425 // If this is another cast that can be eliminated, it isn't codegen either.
426 if (const CastInst *CI = dyn_cast<CastInst>(V))
427 if (isEliminableCastPair(CI, opcode, Ty, TD))
432 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
433 /// InsertBefore instruction. This is specialized a bit to avoid inserting
434 /// casts that are known to not do anything...
436 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
437 Value *V, const Type *DestTy,
438 Instruction *InsertBefore) {
439 if (V->getType() == DestTy) return V;
440 if (Constant *C = dyn_cast<Constant>(V))
441 return ConstantExpr::getCast(opcode, C, DestTy);
443 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
446 // SimplifyCommutative - This performs a few simplifications for commutative
449 // 1. Order operands such that they are listed from right (least complex) to
450 // left (most complex). This puts constants before unary operators before
453 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
454 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
456 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
457 bool Changed = false;
458 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
459 Changed = !I.swapOperands();
461 if (!I.isAssociative()) return Changed;
462 Instruction::BinaryOps Opcode = I.getOpcode();
463 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
464 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
465 if (isa<Constant>(I.getOperand(1))) {
466 Constant *Folded = ConstantExpr::get(I.getOpcode(),
467 cast<Constant>(I.getOperand(1)),
468 cast<Constant>(Op->getOperand(1)));
469 I.setOperand(0, Op->getOperand(0));
470 I.setOperand(1, Folded);
472 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
473 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
474 isOnlyUse(Op) && isOnlyUse(Op1)) {
475 Constant *C1 = cast<Constant>(Op->getOperand(1));
476 Constant *C2 = cast<Constant>(Op1->getOperand(1));
478 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
479 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
480 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
484 I.setOperand(0, New);
485 I.setOperand(1, Folded);
492 /// SimplifyCompare - For a CmpInst this function just orders the operands
493 /// so that theyare listed from right (least complex) to left (most complex).
494 /// This puts constants before unary operators before binary operators.
495 bool InstCombiner::SimplifyCompare(CmpInst &I) {
496 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
499 // Compare instructions are not associative so there's nothing else we can do.
503 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
504 // if the LHS is a constant zero (which is the 'negate' form).
506 static inline Value *dyn_castNegVal(Value *V) {
507 if (BinaryOperator::isNeg(V))
508 return BinaryOperator::getNegArgument(V);
510 // Constants can be considered to be negated values if they can be folded.
511 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
512 return ConstantExpr::getNeg(C);
516 static inline Value *dyn_castNotVal(Value *V) {
517 if (BinaryOperator::isNot(V))
518 return BinaryOperator::getNotArgument(V);
520 // Constants can be considered to be not'ed values...
521 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
522 return ConstantExpr::getNot(C);
526 // dyn_castFoldableMul - If this value is a multiply that can be folded into
527 // other computations (because it has a constant operand), return the
528 // non-constant operand of the multiply, and set CST to point to the multiplier.
529 // Otherwise, return null.
531 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
532 if (V->hasOneUse() && V->getType()->isInteger())
533 if (Instruction *I = dyn_cast<Instruction>(V)) {
534 if (I->getOpcode() == Instruction::Mul)
535 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
536 return I->getOperand(0);
537 if (I->getOpcode() == Instruction::Shl)
538 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
539 // The multiplier is really 1 << CST.
540 Constant *One = ConstantInt::get(V->getType(), 1);
541 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
542 return I->getOperand(0);
548 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
549 /// expression, return it.
550 static User *dyn_castGetElementPtr(Value *V) {
551 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
552 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
553 if (CE->getOpcode() == Instruction::GetElementPtr)
554 return cast<User>(V);
558 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
559 static ConstantInt *AddOne(ConstantInt *C) {
560 return cast<ConstantInt>(ConstantExpr::getAdd(C,
561 ConstantInt::get(C->getType(), 1)));
563 static ConstantInt *SubOne(ConstantInt *C) {
564 return cast<ConstantInt>(ConstantExpr::getSub(C,
565 ConstantInt::get(C->getType(), 1)));
568 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
569 /// known to be either zero or one and return them in the KnownZero/KnownOne
570 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
572 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
573 uint64_t &KnownOne, unsigned Depth = 0) {
574 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
575 // we cannot optimize based on the assumption that it is zero without changing
576 // it to be an explicit zero. If we don't change it to zero, other code could
577 // optimized based on the contradictory assumption that it is non-zero.
578 // Because instcombine aggressively folds operations with undef args anyway,
579 // this won't lose us code quality.
580 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
581 // We know all of the bits for a constant!
582 KnownOne = CI->getZExtValue() & Mask;
583 KnownZero = ~KnownOne & Mask;
587 KnownZero = KnownOne = 0; // Don't know anything.
588 if (Depth == 6 || Mask == 0)
589 return; // Limit search depth.
591 uint64_t KnownZero2, KnownOne2;
592 Instruction *I = dyn_cast<Instruction>(V);
595 Mask &= cast<IntegerType>(V->getType())->getBitMask();
597 switch (I->getOpcode()) {
598 case Instruction::And:
599 // If either the LHS or the RHS are Zero, the result is zero.
600 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
602 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
603 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
604 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
606 // Output known-1 bits are only known if set in both the LHS & RHS.
607 KnownOne &= KnownOne2;
608 // Output known-0 are known to be clear if zero in either the LHS | RHS.
609 KnownZero |= KnownZero2;
611 case Instruction::Or:
612 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
614 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
615 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
616 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
618 // Output known-0 bits are only known if clear in both the LHS & RHS.
619 KnownZero &= KnownZero2;
620 // Output known-1 are known to be set if set in either the LHS | RHS.
621 KnownOne |= KnownOne2;
623 case Instruction::Xor: {
624 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
625 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
626 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
627 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
629 // Output known-0 bits are known if clear or set in both the LHS & RHS.
630 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
631 // Output known-1 are known to be set if set in only one of the LHS, RHS.
632 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
633 KnownZero = KnownZeroOut;
636 case Instruction::Select:
637 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
638 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
639 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
640 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
642 // Only known if known in both the LHS and RHS.
643 KnownOne &= KnownOne2;
644 KnownZero &= KnownZero2;
646 case Instruction::FPTrunc:
647 case Instruction::FPExt:
648 case Instruction::FPToUI:
649 case Instruction::FPToSI:
650 case Instruction::SIToFP:
651 case Instruction::PtrToInt:
652 case Instruction::UIToFP:
653 case Instruction::IntToPtr:
654 return; // Can't work with floating point or pointers
655 case Instruction::Trunc:
656 // All these have integer operands
657 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
659 case Instruction::BitCast: {
660 const Type *SrcTy = I->getOperand(0)->getType();
661 if (SrcTy->isInteger()) {
662 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
667 case Instruction::ZExt: {
668 // Compute the bits in the result that are not present in the input.
669 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
670 uint64_t NotIn = ~SrcTy->getBitMask();
671 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
673 Mask &= SrcTy->getBitMask();
674 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
675 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
676 // The top bits are known to be zero.
677 KnownZero |= NewBits;
680 case Instruction::SExt: {
681 // Compute the bits in the result that are not present in the input.
682 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
683 uint64_t NotIn = ~SrcTy->getBitMask();
684 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
686 Mask &= SrcTy->getBitMask();
687 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
688 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
690 // If the sign bit of the input is known set or clear, then we know the
691 // top bits of the result.
692 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
693 if (KnownZero & InSignBit) { // Input sign bit known zero
694 KnownZero |= NewBits;
695 KnownOne &= ~NewBits;
696 } else if (KnownOne & InSignBit) { // Input sign bit known set
698 KnownZero &= ~NewBits;
699 } else { // Input sign bit unknown
700 KnownZero &= ~NewBits;
701 KnownOne &= ~NewBits;
705 case Instruction::Shl:
706 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
707 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
708 uint64_t ShiftAmt = SA->getZExtValue();
710 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
711 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
712 KnownZero <<= ShiftAmt;
713 KnownOne <<= ShiftAmt;
714 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
718 case Instruction::LShr:
719 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
720 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
721 // Compute the new bits that are at the top now.
722 uint64_t ShiftAmt = SA->getZExtValue();
723 uint64_t HighBits = (1ULL << ShiftAmt)-1;
724 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
726 // Unsigned shift right.
728 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
729 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
730 KnownZero >>= ShiftAmt;
731 KnownOne >>= ShiftAmt;
732 KnownZero |= HighBits; // high bits known zero.
736 case Instruction::AShr:
737 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
738 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
739 // Compute the new bits that are at the top now.
740 uint64_t ShiftAmt = SA->getZExtValue();
741 uint64_t HighBits = (1ULL << ShiftAmt)-1;
742 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
744 // Signed shift right.
746 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
747 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
748 KnownZero >>= ShiftAmt;
749 KnownOne >>= ShiftAmt;
751 // Handle the sign bits.
752 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
753 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
755 if (KnownZero & SignBit) { // New bits are known zero.
756 KnownZero |= HighBits;
757 } else if (KnownOne & SignBit) { // New bits are known one.
758 KnownOne |= HighBits;
766 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
767 /// this predicate to simplify operations downstream. Mask is known to be zero
768 /// for bits that V cannot have.
769 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
770 uint64_t KnownZero, KnownOne;
771 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
772 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
773 return (KnownZero & Mask) == Mask;
776 /// ShrinkDemandedConstant - Check to see if the specified operand of the
777 /// specified instruction is a constant integer. If so, check to see if there
778 /// are any bits set in the constant that are not demanded. If so, shrink the
779 /// constant and return true.
780 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
782 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
783 if (!OpC) return false;
785 // If there are no bits set that aren't demanded, nothing to do.
786 if ((~Demanded & OpC->getZExtValue()) == 0)
789 // This is producing any bits that are not needed, shrink the RHS.
790 uint64_t Val = Demanded & OpC->getZExtValue();
791 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Val));
795 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
796 // set of known zero and one bits, compute the maximum and minimum values that
797 // could have the specified known zero and known one bits, returning them in
799 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
802 int64_t &Min, int64_t &Max) {
803 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
804 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
806 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
808 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
809 // bit if it is unknown.
811 Max = KnownOne|UnknownBits;
813 if (SignBit & UnknownBits) { // Sign bit is unknown
818 // Sign extend the min/max values.
819 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
820 Min = (Min << ShAmt) >> ShAmt;
821 Max = (Max << ShAmt) >> ShAmt;
824 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
825 // a set of known zero and one bits, compute the maximum and minimum values that
826 // could have the specified known zero and known one bits, returning them in
828 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
833 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
834 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
836 // The minimum value is when the unknown bits are all zeros.
838 // The maximum value is when the unknown bits are all ones.
839 Max = KnownOne|UnknownBits;
843 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
844 /// DemandedMask bits of the result of V are ever used downstream. If we can
845 /// use this information to simplify V, do so and return true. Otherwise,
846 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
847 /// the expression (used to simplify the caller). The KnownZero/One bits may
848 /// only be accurate for those bits in the DemandedMask.
849 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
850 uint64_t &KnownZero, uint64_t &KnownOne,
852 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
853 // We know all of the bits for a constant!
854 KnownOne = CI->getZExtValue() & DemandedMask;
855 KnownZero = ~KnownOne & DemandedMask;
859 KnownZero = KnownOne = 0;
860 if (!V->hasOneUse()) { // Other users may use these bits.
861 if (Depth != 0) { // Not at the root.
862 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
863 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
866 // If this is the root being simplified, allow it to have multiple uses,
867 // just set the DemandedMask to all bits.
868 DemandedMask = cast<IntegerType>(V->getType())->getBitMask();
869 } else if (DemandedMask == 0) { // Not demanding any bits from V.
870 if (V != UndefValue::get(V->getType()))
871 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
873 } else if (Depth == 6) { // Limit search depth.
877 Instruction *I = dyn_cast<Instruction>(V);
878 if (!I) return false; // Only analyze instructions.
880 DemandedMask &= cast<IntegerType>(V->getType())->getBitMask();
882 uint64_t KnownZero2 = 0, KnownOne2 = 0;
883 switch (I->getOpcode()) {
885 case Instruction::And:
886 // If either the LHS or the RHS are Zero, the result is zero.
887 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
888 KnownZero, KnownOne, Depth+1))
890 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
892 // If something is known zero on the RHS, the bits aren't demanded on the
894 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
895 KnownZero2, KnownOne2, Depth+1))
897 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
899 // If all of the demanded bits are known 1 on one side, return the other.
900 // These bits cannot contribute to the result of the 'and'.
901 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
902 return UpdateValueUsesWith(I, I->getOperand(0));
903 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
904 return UpdateValueUsesWith(I, I->getOperand(1));
906 // If all of the demanded bits in the inputs are known zeros, return zero.
907 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
908 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
910 // If the RHS is a constant, see if we can simplify it.
911 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
912 return UpdateValueUsesWith(I, I);
914 // Output known-1 bits are only known if set in both the LHS & RHS.
915 KnownOne &= KnownOne2;
916 // Output known-0 are known to be clear if zero in either the LHS | RHS.
917 KnownZero |= KnownZero2;
919 case Instruction::Or:
920 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
921 KnownZero, KnownOne, Depth+1))
923 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
924 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
925 KnownZero2, KnownOne2, Depth+1))
927 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
929 // If all of the demanded bits are known zero on one side, return the other.
930 // These bits cannot contribute to the result of the 'or'.
931 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
932 return UpdateValueUsesWith(I, I->getOperand(0));
933 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
934 return UpdateValueUsesWith(I, I->getOperand(1));
936 // If all of the potentially set bits on one side are known to be set on
937 // the other side, just use the 'other' side.
938 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
939 (DemandedMask & (~KnownZero)))
940 return UpdateValueUsesWith(I, I->getOperand(0));
941 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
942 (DemandedMask & (~KnownZero2)))
943 return UpdateValueUsesWith(I, I->getOperand(1));
945 // If the RHS is a constant, see if we can simplify it.
946 if (ShrinkDemandedConstant(I, 1, DemandedMask))
947 return UpdateValueUsesWith(I, I);
949 // Output known-0 bits are only known if clear in both the LHS & RHS.
950 KnownZero &= KnownZero2;
951 // Output known-1 are known to be set if set in either the LHS | RHS.
952 KnownOne |= KnownOne2;
954 case Instruction::Xor: {
955 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
956 KnownZero, KnownOne, Depth+1))
958 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
959 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
960 KnownZero2, KnownOne2, Depth+1))
962 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
964 // If all of the demanded bits are known zero on one side, return the other.
965 // These bits cannot contribute to the result of the 'xor'.
966 if ((DemandedMask & KnownZero) == DemandedMask)
967 return UpdateValueUsesWith(I, I->getOperand(0));
968 if ((DemandedMask & KnownZero2) == DemandedMask)
969 return UpdateValueUsesWith(I, I->getOperand(1));
971 // Output known-0 bits are known if clear or set in both the LHS & RHS.
972 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
973 // Output known-1 are known to be set if set in only one of the LHS, RHS.
974 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
976 // If all of the demanded bits are known to be zero on one side or the
977 // other, turn this into an *inclusive* or.
978 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
979 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
981 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
983 InsertNewInstBefore(Or, *I);
984 return UpdateValueUsesWith(I, Or);
987 // If all of the demanded bits on one side are known, and all of the set
988 // bits on that side are also known to be set on the other side, turn this
989 // into an AND, as we know the bits will be cleared.
990 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
991 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
992 if ((KnownOne & KnownOne2) == KnownOne) {
993 Constant *AndC = ConstantInt::get(I->getType(),
994 ~KnownOne & DemandedMask);
996 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
997 InsertNewInstBefore(And, *I);
998 return UpdateValueUsesWith(I, And);
1002 // If the RHS is a constant, see if we can simplify it.
1003 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1004 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1005 return UpdateValueUsesWith(I, I);
1007 KnownZero = KnownZeroOut;
1008 KnownOne = KnownOneOut;
1011 case Instruction::Select:
1012 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1013 KnownZero, KnownOne, Depth+1))
1015 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1016 KnownZero2, KnownOne2, Depth+1))
1018 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1019 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1021 // If the operands are constants, see if we can simplify them.
1022 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1023 return UpdateValueUsesWith(I, I);
1024 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1025 return UpdateValueUsesWith(I, I);
1027 // Only known if known in both the LHS and RHS.
1028 KnownOne &= KnownOne2;
1029 KnownZero &= KnownZero2;
1031 case Instruction::Trunc:
1032 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1033 KnownZero, KnownOne, Depth+1))
1035 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1037 case Instruction::BitCast:
1038 if (!I->getOperand(0)->getType()->isInteger())
1041 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1042 KnownZero, KnownOne, Depth+1))
1044 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1046 case Instruction::ZExt: {
1047 // Compute the bits in the result that are not present in the input.
1048 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1049 uint64_t NotIn = ~SrcTy->getBitMask();
1050 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
1052 DemandedMask &= SrcTy->getBitMask();
1053 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1054 KnownZero, KnownOne, Depth+1))
1056 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1057 // The top bits are known to be zero.
1058 KnownZero |= NewBits;
1061 case Instruction::SExt: {
1062 // Compute the bits in the result that are not present in the input.
1063 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1064 uint64_t NotIn = ~SrcTy->getBitMask();
1065 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
1067 // Get the sign bit for the source type
1068 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1069 int64_t InputDemandedBits = DemandedMask & SrcTy->getBitMask();
1071 // If any of the sign extended bits are demanded, we know that the sign
1073 if (NewBits & DemandedMask)
1074 InputDemandedBits |= InSignBit;
1076 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1077 KnownZero, KnownOne, Depth+1))
1079 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1081 // If the sign bit of the input is known set or clear, then we know the
1082 // top bits of the result.
1084 // If the input sign bit is known zero, or if the NewBits are not demanded
1085 // convert this into a zero extension.
1086 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1087 // Convert to ZExt cast
1088 CastInst *NewCast = CastInst::create(
1089 Instruction::ZExt, I->getOperand(0), I->getType(), I->getName(), I);
1090 return UpdateValueUsesWith(I, NewCast);
1091 } else if (KnownOne & InSignBit) { // Input sign bit known set
1092 KnownOne |= NewBits;
1093 KnownZero &= ~NewBits;
1094 } else { // Input sign bit unknown
1095 KnownZero &= ~NewBits;
1096 KnownOne &= ~NewBits;
1100 case Instruction::Add:
1101 // If there is a constant on the RHS, there are a variety of xformations
1103 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1104 // If null, this should be simplified elsewhere. Some of the xforms here
1105 // won't work if the RHS is zero.
1106 if (RHS->isNullValue())
1109 // Figure out what the input bits are. If the top bits of the and result
1110 // are not demanded, then the add doesn't demand them from its input
1113 // Shift the demanded mask up so that it's at the top of the uint64_t.
1114 unsigned BitWidth = I->getType()->getPrimitiveSizeInBits();
1115 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1117 // If the top bit of the output is demanded, demand everything from the
1118 // input. Otherwise, we demand all the input bits except NLZ top bits.
1119 uint64_t InDemandedBits = ~0ULL >> (64-BitWidth+NLZ);
1121 // Find information about known zero/one bits in the input.
1122 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1123 KnownZero2, KnownOne2, Depth+1))
1126 // If the RHS of the add has bits set that can't affect the input, reduce
1128 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1129 return UpdateValueUsesWith(I, I);
1131 // Avoid excess work.
1132 if (KnownZero2 == 0 && KnownOne2 == 0)
1135 // Turn it into OR if input bits are zero.
1136 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1138 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1140 InsertNewInstBefore(Or, *I);
1141 return UpdateValueUsesWith(I, Or);
1144 // We can say something about the output known-zero and known-one bits,
1145 // depending on potential carries from the input constant and the
1146 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1147 // bits set and the RHS constant is 0x01001, then we know we have a known
1148 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1150 // To compute this, we first compute the potential carry bits. These are
1151 // the bits which may be modified. I'm not aware of a better way to do
1153 uint64_t RHSVal = RHS->getZExtValue();
1155 bool CarryIn = false;
1156 uint64_t CarryBits = 0;
1157 uint64_t CurBit = 1;
1158 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1159 // Record the current carry in.
1160 if (CarryIn) CarryBits |= CurBit;
1164 // This bit has a carry out unless it is "zero + zero" or
1165 // "zero + anything" with no carry in.
1166 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1167 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1168 } else if (!CarryIn &&
1169 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1170 CarryOut = false; // 0 + anything has no carry out if no carry in.
1172 // Otherwise, we have to assume we have a carry out.
1176 // This stage's carry out becomes the next stage's carry-in.
1180 // Now that we know which bits have carries, compute the known-1/0 sets.
1182 // Bits are known one if they are known zero in one operand and one in the
1183 // other, and there is no input carry.
1184 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1186 // Bits are known zero if they are known zero in both operands and there
1187 // is no input carry.
1188 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1191 case Instruction::Shl:
1192 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1193 uint64_t ShiftAmt = SA->getZExtValue();
1194 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1195 KnownZero, KnownOne, Depth+1))
1197 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1198 KnownZero <<= ShiftAmt;
1199 KnownOne <<= ShiftAmt;
1200 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1203 case Instruction::LShr:
1204 // For a logical shift right
1205 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1206 unsigned ShiftAmt = SA->getZExtValue();
1208 // Compute the new bits that are at the top now.
1209 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1210 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1211 uint64_t TypeMask = cast<IntegerType>(I->getType())->getBitMask();
1212 // Unsigned shift right.
1213 if (SimplifyDemandedBits(I->getOperand(0),
1214 (DemandedMask << ShiftAmt) & TypeMask,
1215 KnownZero, KnownOne, Depth+1))
1217 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1218 KnownZero &= TypeMask;
1219 KnownOne &= TypeMask;
1220 KnownZero >>= ShiftAmt;
1221 KnownOne >>= ShiftAmt;
1222 KnownZero |= HighBits; // high bits known zero.
1225 case Instruction::AShr:
1226 // If this is an arithmetic shift right and only the low-bit is set, we can
1227 // always convert this into a logical shr, even if the shift amount is
1228 // variable. The low bit of the shift cannot be an input sign bit unless
1229 // the shift amount is >= the size of the datatype, which is undefined.
1230 if (DemandedMask == 1) {
1231 // Perform the logical shift right.
1232 Value *NewVal = BinaryOperator::createLShr(
1233 I->getOperand(0), I->getOperand(1), I->getName());
1234 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1235 return UpdateValueUsesWith(I, NewVal);
1238 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1239 unsigned ShiftAmt = SA->getZExtValue();
1241 // Compute the new bits that are at the top now.
1242 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1243 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1244 uint64_t TypeMask = cast<IntegerType>(I->getType())->getBitMask();
1245 // Signed shift right.
1246 if (SimplifyDemandedBits(I->getOperand(0),
1247 (DemandedMask << ShiftAmt) & TypeMask,
1248 KnownZero, KnownOne, Depth+1))
1250 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1251 KnownZero &= TypeMask;
1252 KnownOne &= TypeMask;
1253 KnownZero >>= ShiftAmt;
1254 KnownOne >>= ShiftAmt;
1256 // Handle the sign bits.
1257 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1258 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1260 // If the input sign bit is known to be zero, or if none of the top bits
1261 // are demanded, turn this into an unsigned shift right.
1262 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1263 // Perform the logical shift right.
1264 Value *NewVal = BinaryOperator::createLShr(
1265 I->getOperand(0), SA, I->getName());
1266 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1267 return UpdateValueUsesWith(I, NewVal);
1268 } else if (KnownOne & SignBit) { // New bits are known one.
1269 KnownOne |= HighBits;
1275 // If the client is only demanding bits that we know, return the known
1277 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1278 return UpdateValueUsesWith(I, ConstantInt::get(I->getType(), KnownOne));
1283 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1284 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1285 /// actually used by the caller. This method analyzes which elements of the
1286 /// operand are undef and returns that information in UndefElts.
1288 /// If the information about demanded elements can be used to simplify the
1289 /// operation, the operation is simplified, then the resultant value is
1290 /// returned. This returns null if no change was made.
1291 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1292 uint64_t &UndefElts,
1294 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1295 assert(VWidth <= 64 && "Vector too wide to analyze!");
1296 uint64_t EltMask = ~0ULL >> (64-VWidth);
1297 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1298 "Invalid DemandedElts!");
1300 if (isa<UndefValue>(V)) {
1301 // If the entire vector is undefined, just return this info.
1302 UndefElts = EltMask;
1304 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1305 UndefElts = EltMask;
1306 return UndefValue::get(V->getType());
1310 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1311 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1312 Constant *Undef = UndefValue::get(EltTy);
1314 std::vector<Constant*> Elts;
1315 for (unsigned i = 0; i != VWidth; ++i)
1316 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1317 Elts.push_back(Undef);
1318 UndefElts |= (1ULL << i);
1319 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1320 Elts.push_back(Undef);
1321 UndefElts |= (1ULL << i);
1322 } else { // Otherwise, defined.
1323 Elts.push_back(CP->getOperand(i));
1326 // If we changed the constant, return it.
1327 Constant *NewCP = ConstantVector::get(Elts);
1328 return NewCP != CP ? NewCP : 0;
1329 } else if (isa<ConstantAggregateZero>(V)) {
1330 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1332 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1333 Constant *Zero = Constant::getNullValue(EltTy);
1334 Constant *Undef = UndefValue::get(EltTy);
1335 std::vector<Constant*> Elts;
1336 for (unsigned i = 0; i != VWidth; ++i)
1337 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1338 UndefElts = DemandedElts ^ EltMask;
1339 return ConstantVector::get(Elts);
1342 if (!V->hasOneUse()) { // Other users may use these bits.
1343 if (Depth != 0) { // Not at the root.
1344 // TODO: Just compute the UndefElts information recursively.
1348 } else if (Depth == 10) { // Limit search depth.
1352 Instruction *I = dyn_cast<Instruction>(V);
1353 if (!I) return false; // Only analyze instructions.
1355 bool MadeChange = false;
1356 uint64_t UndefElts2;
1358 switch (I->getOpcode()) {
1361 case Instruction::InsertElement: {
1362 // If this is a variable index, we don't know which element it overwrites.
1363 // demand exactly the same input as we produce.
1364 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1366 // Note that we can't propagate undef elt info, because we don't know
1367 // which elt is getting updated.
1368 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1369 UndefElts2, Depth+1);
1370 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1374 // If this is inserting an element that isn't demanded, remove this
1376 unsigned IdxNo = Idx->getZExtValue();
1377 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1378 return AddSoonDeadInstToWorklist(*I, 0);
1380 // Otherwise, the element inserted overwrites whatever was there, so the
1381 // input demanded set is simpler than the output set.
1382 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1383 DemandedElts & ~(1ULL << IdxNo),
1384 UndefElts, Depth+1);
1385 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1387 // The inserted element is defined.
1388 UndefElts |= 1ULL << IdxNo;
1392 case Instruction::And:
1393 case Instruction::Or:
1394 case Instruction::Xor:
1395 case Instruction::Add:
1396 case Instruction::Sub:
1397 case Instruction::Mul:
1398 // div/rem demand all inputs, because they don't want divide by zero.
1399 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1400 UndefElts, Depth+1);
1401 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1402 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1403 UndefElts2, Depth+1);
1404 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1406 // Output elements are undefined if both are undefined. Consider things
1407 // like undef&0. The result is known zero, not undef.
1408 UndefElts &= UndefElts2;
1411 case Instruction::Call: {
1412 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1414 switch (II->getIntrinsicID()) {
1417 // Binary vector operations that work column-wise. A dest element is a
1418 // function of the corresponding input elements from the two inputs.
1419 case Intrinsic::x86_sse_sub_ss:
1420 case Intrinsic::x86_sse_mul_ss:
1421 case Intrinsic::x86_sse_min_ss:
1422 case Intrinsic::x86_sse_max_ss:
1423 case Intrinsic::x86_sse2_sub_sd:
1424 case Intrinsic::x86_sse2_mul_sd:
1425 case Intrinsic::x86_sse2_min_sd:
1426 case Intrinsic::x86_sse2_max_sd:
1427 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1428 UndefElts, Depth+1);
1429 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1430 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1431 UndefElts2, Depth+1);
1432 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1434 // If only the low elt is demanded and this is a scalarizable intrinsic,
1435 // scalarize it now.
1436 if (DemandedElts == 1) {
1437 switch (II->getIntrinsicID()) {
1439 case Intrinsic::x86_sse_sub_ss:
1440 case Intrinsic::x86_sse_mul_ss:
1441 case Intrinsic::x86_sse2_sub_sd:
1442 case Intrinsic::x86_sse2_mul_sd:
1443 // TODO: Lower MIN/MAX/ABS/etc
1444 Value *LHS = II->getOperand(1);
1445 Value *RHS = II->getOperand(2);
1446 // Extract the element as scalars.
1447 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1448 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1450 switch (II->getIntrinsicID()) {
1451 default: assert(0 && "Case stmts out of sync!");
1452 case Intrinsic::x86_sse_sub_ss:
1453 case Intrinsic::x86_sse2_sub_sd:
1454 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1455 II->getName()), *II);
1457 case Intrinsic::x86_sse_mul_ss:
1458 case Intrinsic::x86_sse2_mul_sd:
1459 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1460 II->getName()), *II);
1465 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1467 InsertNewInstBefore(New, *II);
1468 AddSoonDeadInstToWorklist(*II, 0);
1473 // Output elements are undefined if both are undefined. Consider things
1474 // like undef&0. The result is known zero, not undef.
1475 UndefElts &= UndefElts2;
1481 return MadeChange ? I : 0;
1484 /// @returns true if the specified compare instruction is
1485 /// true when both operands are equal...
1486 /// @brief Determine if the ICmpInst returns true if both operands are equal
1487 static bool isTrueWhenEqual(ICmpInst &ICI) {
1488 ICmpInst::Predicate pred = ICI.getPredicate();
1489 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1490 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1491 pred == ICmpInst::ICMP_SLE;
1494 /// AssociativeOpt - Perform an optimization on an associative operator. This
1495 /// function is designed to check a chain of associative operators for a
1496 /// potential to apply a certain optimization. Since the optimization may be
1497 /// applicable if the expression was reassociated, this checks the chain, then
1498 /// reassociates the expression as necessary to expose the optimization
1499 /// opportunity. This makes use of a special Functor, which must define
1500 /// 'shouldApply' and 'apply' methods.
1502 template<typename Functor>
1503 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1504 unsigned Opcode = Root.getOpcode();
1505 Value *LHS = Root.getOperand(0);
1507 // Quick check, see if the immediate LHS matches...
1508 if (F.shouldApply(LHS))
1509 return F.apply(Root);
1511 // Otherwise, if the LHS is not of the same opcode as the root, return.
1512 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1513 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1514 // Should we apply this transform to the RHS?
1515 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1517 // If not to the RHS, check to see if we should apply to the LHS...
1518 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1519 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1523 // If the functor wants to apply the optimization to the RHS of LHSI,
1524 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1526 BasicBlock *BB = Root.getParent();
1528 // Now all of the instructions are in the current basic block, go ahead
1529 // and perform the reassociation.
1530 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1532 // First move the selected RHS to the LHS of the root...
1533 Root.setOperand(0, LHSI->getOperand(1));
1535 // Make what used to be the LHS of the root be the user of the root...
1536 Value *ExtraOperand = TmpLHSI->getOperand(1);
1537 if (&Root == TmpLHSI) {
1538 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1541 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1542 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1543 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1544 BasicBlock::iterator ARI = &Root; ++ARI;
1545 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1548 // Now propagate the ExtraOperand down the chain of instructions until we
1550 while (TmpLHSI != LHSI) {
1551 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1552 // Move the instruction to immediately before the chain we are
1553 // constructing to avoid breaking dominance properties.
1554 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1555 BB->getInstList().insert(ARI, NextLHSI);
1558 Value *NextOp = NextLHSI->getOperand(1);
1559 NextLHSI->setOperand(1, ExtraOperand);
1561 ExtraOperand = NextOp;
1564 // Now that the instructions are reassociated, have the functor perform
1565 // the transformation...
1566 return F.apply(Root);
1569 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1575 // AddRHS - Implements: X + X --> X << 1
1578 AddRHS(Value *rhs) : RHS(rhs) {}
1579 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1580 Instruction *apply(BinaryOperator &Add) const {
1581 return BinaryOperator::createShl(Add.getOperand(0),
1582 ConstantInt::get(Add.getType(), 1));
1586 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1588 struct AddMaskingAnd {
1590 AddMaskingAnd(Constant *c) : C2(c) {}
1591 bool shouldApply(Value *LHS) const {
1593 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1594 ConstantExpr::getAnd(C1, C2)->isNullValue();
1596 Instruction *apply(BinaryOperator &Add) const {
1597 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1601 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1603 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1604 if (Constant *SOC = dyn_cast<Constant>(SO))
1605 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1607 return IC->InsertNewInstBefore(CastInst::create(
1608 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1611 // Figure out if the constant is the left or the right argument.
1612 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1613 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1615 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1617 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1618 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1621 Value *Op0 = SO, *Op1 = ConstOperand;
1623 std::swap(Op0, Op1);
1625 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1626 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1627 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1628 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1629 SO->getName()+".cmp");
1631 assert(0 && "Unknown binary instruction type!");
1634 return IC->InsertNewInstBefore(New, I);
1637 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1638 // constant as the other operand, try to fold the binary operator into the
1639 // select arguments. This also works for Cast instructions, which obviously do
1640 // not have a second operand.
1641 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1643 // Don't modify shared select instructions
1644 if (!SI->hasOneUse()) return 0;
1645 Value *TV = SI->getOperand(1);
1646 Value *FV = SI->getOperand(2);
1648 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1649 // Bool selects with constant operands can be folded to logical ops.
1650 if (SI->getType() == Type::Int1Ty) return 0;
1652 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1653 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1655 return new SelectInst(SI->getCondition(), SelectTrueVal,
1662 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1663 /// node as operand #0, see if we can fold the instruction into the PHI (which
1664 /// is only possible if all operands to the PHI are constants).
1665 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1666 PHINode *PN = cast<PHINode>(I.getOperand(0));
1667 unsigned NumPHIValues = PN->getNumIncomingValues();
1668 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1670 // Check to see if all of the operands of the PHI are constants. If there is
1671 // one non-constant value, remember the BB it is. If there is more than one
1672 // or if *it* is a PHI, bail out.
1673 BasicBlock *NonConstBB = 0;
1674 for (unsigned i = 0; i != NumPHIValues; ++i)
1675 if (!isa<Constant>(PN->getIncomingValue(i))) {
1676 if (NonConstBB) return 0; // More than one non-const value.
1677 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1678 NonConstBB = PN->getIncomingBlock(i);
1680 // If the incoming non-constant value is in I's block, we have an infinite
1682 if (NonConstBB == I.getParent())
1686 // If there is exactly one non-constant value, we can insert a copy of the
1687 // operation in that block. However, if this is a critical edge, we would be
1688 // inserting the computation one some other paths (e.g. inside a loop). Only
1689 // do this if the pred block is unconditionally branching into the phi block.
1691 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1692 if (!BI || !BI->isUnconditional()) return 0;
1695 // Okay, we can do the transformation: create the new PHI node.
1696 PHINode *NewPN = new PHINode(I.getType(), "");
1697 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1698 InsertNewInstBefore(NewPN, *PN);
1699 NewPN->takeName(PN);
1701 // Next, add all of the operands to the PHI.
1702 if (I.getNumOperands() == 2) {
1703 Constant *C = cast<Constant>(I.getOperand(1));
1704 for (unsigned i = 0; i != NumPHIValues; ++i) {
1706 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1707 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1708 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1710 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1712 assert(PN->getIncomingBlock(i) == NonConstBB);
1713 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1714 InV = BinaryOperator::create(BO->getOpcode(),
1715 PN->getIncomingValue(i), C, "phitmp",
1716 NonConstBB->getTerminator());
1717 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1718 InV = CmpInst::create(CI->getOpcode(),
1720 PN->getIncomingValue(i), C, "phitmp",
1721 NonConstBB->getTerminator());
1723 assert(0 && "Unknown binop!");
1725 AddToWorkList(cast<Instruction>(InV));
1727 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1730 CastInst *CI = cast<CastInst>(&I);
1731 const Type *RetTy = CI->getType();
1732 for (unsigned i = 0; i != NumPHIValues; ++i) {
1734 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1735 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1737 assert(PN->getIncomingBlock(i) == NonConstBB);
1738 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1739 I.getType(), "phitmp",
1740 NonConstBB->getTerminator());
1741 AddToWorkList(cast<Instruction>(InV));
1743 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1746 return ReplaceInstUsesWith(I, NewPN);
1749 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1750 bool Changed = SimplifyCommutative(I);
1751 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1753 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1754 // X + undef -> undef
1755 if (isa<UndefValue>(RHS))
1756 return ReplaceInstUsesWith(I, RHS);
1759 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1760 if (RHSC->isNullValue())
1761 return ReplaceInstUsesWith(I, LHS);
1762 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1763 if (CFP->isExactlyValue(-0.0))
1764 return ReplaceInstUsesWith(I, LHS);
1767 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1768 // X + (signbit) --> X ^ signbit
1769 uint64_t Val = CI->getZExtValue();
1770 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1771 return BinaryOperator::createXor(LHS, RHS);
1773 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1774 // (X & 254)+1 -> (X&254)|1
1775 uint64_t KnownZero, KnownOne;
1776 if (!isa<VectorType>(I.getType()) &&
1777 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
1778 KnownZero, KnownOne))
1782 if (isa<PHINode>(LHS))
1783 if (Instruction *NV = FoldOpIntoPhi(I))
1786 ConstantInt *XorRHS = 0;
1788 if (isa<ConstantInt>(RHSC) &&
1789 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1790 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1791 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1792 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1794 uint64_t C0080Val = 1ULL << 31;
1795 int64_t CFF80Val = -C0080Val;
1798 if (TySizeBits > Size) {
1800 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1801 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1802 if (RHSSExt == CFF80Val) {
1803 if (XorRHS->getZExtValue() == C0080Val)
1805 } else if (RHSZExt == C0080Val) {
1806 if (XorRHS->getSExtValue() == CFF80Val)
1810 // This is a sign extend if the top bits are known zero.
1811 uint64_t Mask = ~0ULL;
1812 Mask <<= 64-(TySizeBits-Size);
1813 Mask &= cast<IntegerType>(XorLHS->getType())->getBitMask();
1814 if (!MaskedValueIsZero(XorLHS, Mask))
1815 Size = 0; // Not a sign ext, but can't be any others either.
1822 } while (Size >= 8);
1825 const Type *MiddleType = 0;
1828 case 32: MiddleType = Type::Int32Ty; break;
1829 case 16: MiddleType = Type::Int16Ty; break;
1830 case 8: MiddleType = Type::Int8Ty; break;
1833 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1834 InsertNewInstBefore(NewTrunc, I);
1835 return new SExtInst(NewTrunc, I.getType());
1841 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
1842 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1844 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1845 if (RHSI->getOpcode() == Instruction::Sub)
1846 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1847 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1849 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1850 if (LHSI->getOpcode() == Instruction::Sub)
1851 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1852 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1857 if (Value *V = dyn_castNegVal(LHS))
1858 return BinaryOperator::createSub(RHS, V);
1861 if (!isa<Constant>(RHS))
1862 if (Value *V = dyn_castNegVal(RHS))
1863 return BinaryOperator::createSub(LHS, V);
1867 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1868 if (X == RHS) // X*C + X --> X * (C+1)
1869 return BinaryOperator::createMul(RHS, AddOne(C2));
1871 // X*C1 + X*C2 --> X * (C1+C2)
1873 if (X == dyn_castFoldableMul(RHS, C1))
1874 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1877 // X + X*C --> X * (C+1)
1878 if (dyn_castFoldableMul(RHS, C2) == LHS)
1879 return BinaryOperator::createMul(LHS, AddOne(C2));
1881 // X + ~X --> -1 since ~X = -X-1
1882 if (dyn_castNotVal(LHS) == RHS ||
1883 dyn_castNotVal(RHS) == LHS)
1884 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
1887 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1888 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1889 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
1892 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1894 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1895 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1896 return BinaryOperator::createSub(C, X);
1899 // (X & FF00) + xx00 -> (X+xx00) & FF00
1900 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1901 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1902 if (Anded == CRHS) {
1903 // See if all bits from the first bit set in the Add RHS up are included
1904 // in the mask. First, get the rightmost bit.
1905 uint64_t AddRHSV = CRHS->getZExtValue();
1907 // Form a mask of all bits from the lowest bit added through the top.
1908 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1909 AddRHSHighBits &= C2->getType()->getBitMask();
1911 // See if the and mask includes all of these bits.
1912 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1914 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1915 // Okay, the xform is safe. Insert the new add pronto.
1916 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1917 LHS->getName()), I);
1918 return BinaryOperator::createAnd(NewAdd, C2);
1923 // Try to fold constant add into select arguments.
1924 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1925 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1929 // add (cast *A to intptrtype) B ->
1930 // cast (GEP (cast *A to sbyte*) B) ->
1933 CastInst *CI = dyn_cast<CastInst>(LHS);
1936 CI = dyn_cast<CastInst>(RHS);
1939 if (CI && CI->getType()->isSized() &&
1940 (CI->getType()->getPrimitiveSizeInBits() ==
1941 TD->getIntPtrType()->getPrimitiveSizeInBits())
1942 && isa<PointerType>(CI->getOperand(0)->getType())) {
1943 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
1944 PointerType::get(Type::Int8Ty), I);
1945 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1946 return new PtrToIntInst(I2, CI->getType());
1950 return Changed ? &I : 0;
1953 // isSignBit - Return true if the value represented by the constant only has the
1954 // highest order bit set.
1955 static bool isSignBit(ConstantInt *CI) {
1956 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1957 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1960 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1961 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1963 if (Op0 == Op1) // sub X, X -> 0
1964 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1966 // If this is a 'B = x-(-A)', change to B = x+A...
1967 if (Value *V = dyn_castNegVal(Op1))
1968 return BinaryOperator::createAdd(Op0, V);
1970 if (isa<UndefValue>(Op0))
1971 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1972 if (isa<UndefValue>(Op1))
1973 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1975 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1976 // Replace (-1 - A) with (~A)...
1977 if (C->isAllOnesValue())
1978 return BinaryOperator::createNot(Op1);
1980 // C - ~X == X + (1+C)
1982 if (match(Op1, m_Not(m_Value(X))))
1983 return BinaryOperator::createAdd(X,
1984 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1985 // -(X >>u 31) -> (X >>s 31)
1986 // -(X >>s 31) -> (X >>u 31)
1987 if (C->isNullValue()) {
1988 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
1989 if (SI->getOpcode() == Instruction::LShr) {
1990 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1991 // Check to see if we are shifting out everything but the sign bit.
1992 if (CU->getZExtValue() ==
1993 SI->getType()->getPrimitiveSizeInBits()-1) {
1994 // Ok, the transformation is safe. Insert AShr.
1995 return BinaryOperator::create(Instruction::AShr,
1996 SI->getOperand(0), CU, SI->getName());
2000 else if (SI->getOpcode() == Instruction::AShr) {
2001 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2002 // Check to see if we are shifting out everything but the sign bit.
2003 if (CU->getZExtValue() ==
2004 SI->getType()->getPrimitiveSizeInBits()-1) {
2005 // Ok, the transformation is safe. Insert LShr.
2006 return BinaryOperator::createLShr(
2007 SI->getOperand(0), CU, SI->getName());
2013 // Try to fold constant sub into select arguments.
2014 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2015 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2018 if (isa<PHINode>(Op0))
2019 if (Instruction *NV = FoldOpIntoPhi(I))
2023 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2024 if (Op1I->getOpcode() == Instruction::Add &&
2025 !Op0->getType()->isFPOrFPVector()) {
2026 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2027 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2028 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2029 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2030 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2031 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2032 // C1-(X+C2) --> (C1-C2)-X
2033 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2034 Op1I->getOperand(0));
2038 if (Op1I->hasOneUse()) {
2039 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2040 // is not used by anyone else...
2042 if (Op1I->getOpcode() == Instruction::Sub &&
2043 !Op1I->getType()->isFPOrFPVector()) {
2044 // Swap the two operands of the subexpr...
2045 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2046 Op1I->setOperand(0, IIOp1);
2047 Op1I->setOperand(1, IIOp0);
2049 // Create the new top level add instruction...
2050 return BinaryOperator::createAdd(Op0, Op1);
2053 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2055 if (Op1I->getOpcode() == Instruction::And &&
2056 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2057 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2060 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2061 return BinaryOperator::createAnd(Op0, NewNot);
2064 // 0 - (X sdiv C) -> (X sdiv -C)
2065 if (Op1I->getOpcode() == Instruction::SDiv)
2066 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2067 if (CSI->isNullValue())
2068 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2069 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2070 ConstantExpr::getNeg(DivRHS));
2072 // X - X*C --> X * (1-C)
2073 ConstantInt *C2 = 0;
2074 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2076 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2077 return BinaryOperator::createMul(Op0, CP1);
2082 if (!Op0->getType()->isFPOrFPVector())
2083 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2084 if (Op0I->getOpcode() == Instruction::Add) {
2085 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2086 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2087 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2088 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2089 } else if (Op0I->getOpcode() == Instruction::Sub) {
2090 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2091 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2095 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2096 if (X == Op1) { // X*C - X --> X * (C-1)
2097 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2098 return BinaryOperator::createMul(Op1, CP1);
2101 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2102 if (X == dyn_castFoldableMul(Op1, C2))
2103 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2108 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2109 /// really just returns true if the most significant (sign) bit is set.
2110 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2112 case ICmpInst::ICMP_SLT:
2113 // True if LHS s< RHS and RHS == 0
2114 return RHS->isNullValue();
2115 case ICmpInst::ICMP_SLE:
2116 // True if LHS s<= RHS and RHS == -1
2117 return RHS->isAllOnesValue();
2118 case ICmpInst::ICMP_UGE:
2119 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2120 return RHS->getZExtValue() == (1ULL <<
2121 (RHS->getType()->getPrimitiveSizeInBits()-1));
2122 case ICmpInst::ICMP_UGT:
2123 // True if LHS u> RHS and RHS == high-bit-mask - 1
2124 return RHS->getZExtValue() ==
2125 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2131 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2132 bool Changed = SimplifyCommutative(I);
2133 Value *Op0 = I.getOperand(0);
2135 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2136 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2138 // Simplify mul instructions with a constant RHS...
2139 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2140 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2142 // ((X << C1)*C2) == (X * (C2 << C1))
2143 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2144 if (SI->getOpcode() == Instruction::Shl)
2145 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2146 return BinaryOperator::createMul(SI->getOperand(0),
2147 ConstantExpr::getShl(CI, ShOp));
2149 if (CI->isNullValue())
2150 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2151 if (CI->equalsInt(1)) // X * 1 == X
2152 return ReplaceInstUsesWith(I, Op0);
2153 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2154 return BinaryOperator::createNeg(Op0, I.getName());
2156 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2157 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2158 uint64_t C = Log2_64(Val);
2159 return BinaryOperator::createShl(Op0,
2160 ConstantInt::get(Op0->getType(), C));
2162 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2163 if (Op1F->isNullValue())
2164 return ReplaceInstUsesWith(I, Op1);
2166 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2167 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2168 if (Op1F->getValue() == 1.0)
2169 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2172 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2173 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2174 isa<ConstantInt>(Op0I->getOperand(1))) {
2175 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2176 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2178 InsertNewInstBefore(Add, I);
2179 Value *C1C2 = ConstantExpr::getMul(Op1,
2180 cast<Constant>(Op0I->getOperand(1)));
2181 return BinaryOperator::createAdd(Add, C1C2);
2185 // Try to fold constant mul into select arguments.
2186 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2187 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2190 if (isa<PHINode>(Op0))
2191 if (Instruction *NV = FoldOpIntoPhi(I))
2195 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2196 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2197 return BinaryOperator::createMul(Op0v, Op1v);
2199 // If one of the operands of the multiply is a cast from a boolean value, then
2200 // we know the bool is either zero or one, so this is a 'masking' multiply.
2201 // See if we can simplify things based on how the boolean was originally
2203 CastInst *BoolCast = 0;
2204 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2205 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2208 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2209 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2212 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2213 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2214 const Type *SCOpTy = SCIOp0->getType();
2216 // If the icmp is true iff the sign bit of X is set, then convert this
2217 // multiply into a shift/and combination.
2218 if (isa<ConstantInt>(SCIOp1) &&
2219 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2220 // Shift the X value right to turn it into "all signbits".
2221 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2222 SCOpTy->getPrimitiveSizeInBits()-1);
2224 InsertNewInstBefore(
2225 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2226 BoolCast->getOperand(0)->getName()+
2229 // If the multiply type is not the same as the source type, sign extend
2230 // or truncate to the multiply type.
2231 if (I.getType() != V->getType()) {
2232 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2233 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2234 Instruction::CastOps opcode =
2235 (SrcBits == DstBits ? Instruction::BitCast :
2236 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2237 V = InsertCastBefore(opcode, V, I.getType(), I);
2240 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2241 return BinaryOperator::createAnd(V, OtherOp);
2246 return Changed ? &I : 0;
2249 /// This function implements the transforms on div instructions that work
2250 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2251 /// used by the visitors to those instructions.
2252 /// @brief Transforms common to all three div instructions
2253 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2254 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2257 if (isa<UndefValue>(Op0))
2258 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2260 // X / undef -> undef
2261 if (isa<UndefValue>(Op1))
2262 return ReplaceInstUsesWith(I, Op1);
2264 // Handle cases involving: div X, (select Cond, Y, Z)
2265 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2266 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2267 // same basic block, then we replace the select with Y, and the condition
2268 // of the select with false (if the cond value is in the same BB). If the
2269 // select has uses other than the div, this allows them to be simplified
2270 // also. Note that div X, Y is just as good as div X, 0 (undef)
2271 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2272 if (ST->isNullValue()) {
2273 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2274 if (CondI && CondI->getParent() == I.getParent())
2275 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2276 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2277 I.setOperand(1, SI->getOperand(2));
2279 UpdateValueUsesWith(SI, SI->getOperand(2));
2283 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2284 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2285 if (ST->isNullValue()) {
2286 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2287 if (CondI && CondI->getParent() == I.getParent())
2288 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2289 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2290 I.setOperand(1, SI->getOperand(1));
2292 UpdateValueUsesWith(SI, SI->getOperand(1));
2300 /// This function implements the transforms common to both integer division
2301 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2302 /// division instructions.
2303 /// @brief Common integer divide transforms
2304 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2305 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2307 if (Instruction *Common = commonDivTransforms(I))
2310 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2312 if (RHS->equalsInt(1))
2313 return ReplaceInstUsesWith(I, Op0);
2315 // (X / C1) / C2 -> X / (C1*C2)
2316 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2317 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2318 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2319 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2320 ConstantExpr::getMul(RHS, LHSRHS));
2323 if (!RHS->isNullValue()) { // avoid X udiv 0
2324 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2325 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2327 if (isa<PHINode>(Op0))
2328 if (Instruction *NV = FoldOpIntoPhi(I))
2333 // 0 / X == 0, we don't need to preserve faults!
2334 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2335 if (LHS->equalsInt(0))
2336 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2341 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2342 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2344 // Handle the integer div common cases
2345 if (Instruction *Common = commonIDivTransforms(I))
2348 // X udiv C^2 -> X >> C
2349 // Check to see if this is an unsigned division with an exact power of 2,
2350 // if so, convert to a right shift.
2351 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2352 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2353 if (isPowerOf2_64(Val)) {
2354 uint64_t ShiftAmt = Log2_64(Val);
2355 return BinaryOperator::createLShr(Op0,
2356 ConstantInt::get(Op0->getType(), ShiftAmt));
2360 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2361 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2362 if (RHSI->getOpcode() == Instruction::Shl &&
2363 isa<ConstantInt>(RHSI->getOperand(0))) {
2364 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2365 if (isPowerOf2_64(C1)) {
2366 Value *N = RHSI->getOperand(1);
2367 const Type *NTy = N->getType();
2368 if (uint64_t C2 = Log2_64(C1)) {
2369 Constant *C2V = ConstantInt::get(NTy, C2);
2370 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2372 return BinaryOperator::createLShr(Op0, N);
2377 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2378 // where C1&C2 are powers of two.
2379 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2380 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2381 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2382 if (!STO->isNullValue() && !STO->isNullValue()) {
2383 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2384 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2385 // Compute the shift amounts
2386 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2387 // Construct the "on true" case of the select
2388 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2389 Instruction *TSI = BinaryOperator::createLShr(
2390 Op0, TC, SI->getName()+".t");
2391 TSI = InsertNewInstBefore(TSI, I);
2393 // Construct the "on false" case of the select
2394 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2395 Instruction *FSI = BinaryOperator::createLShr(
2396 Op0, FC, SI->getName()+".f");
2397 FSI = InsertNewInstBefore(FSI, I);
2399 // construct the select instruction and return it.
2400 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2407 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2408 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2410 // Handle the integer div common cases
2411 if (Instruction *Common = commonIDivTransforms(I))
2414 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2416 if (RHS->isAllOnesValue())
2417 return BinaryOperator::createNeg(Op0);
2420 if (Value *LHSNeg = dyn_castNegVal(Op0))
2421 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2424 // If the sign bits of both operands are zero (i.e. we can prove they are
2425 // unsigned inputs), turn this into a udiv.
2426 if (I.getType()->isInteger()) {
2427 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2428 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2429 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2436 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2437 return commonDivTransforms(I);
2440 /// GetFactor - If we can prove that the specified value is at least a multiple
2441 /// of some factor, return that factor.
2442 static Constant *GetFactor(Value *V) {
2443 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2446 // Unless we can be tricky, we know this is a multiple of 1.
2447 Constant *Result = ConstantInt::get(V->getType(), 1);
2449 Instruction *I = dyn_cast<Instruction>(V);
2450 if (!I) return Result;
2452 if (I->getOpcode() == Instruction::Mul) {
2453 // Handle multiplies by a constant, etc.
2454 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2455 GetFactor(I->getOperand(1)));
2456 } else if (I->getOpcode() == Instruction::Shl) {
2457 // (X<<C) -> X * (1 << C)
2458 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2459 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2460 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2462 } else if (I->getOpcode() == Instruction::And) {
2463 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2464 // X & 0xFFF0 is known to be a multiple of 16.
2465 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2466 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2467 return ConstantExpr::getShl(Result,
2468 ConstantInt::get(Result->getType(), Zeros));
2470 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2471 // Only handle int->int casts.
2472 if (!CI->isIntegerCast())
2474 Value *Op = CI->getOperand(0);
2475 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2480 /// This function implements the transforms on rem instructions that work
2481 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2482 /// is used by the visitors to those instructions.
2483 /// @brief Transforms common to all three rem instructions
2484 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2485 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2487 // 0 % X == 0, we don't need to preserve faults!
2488 if (Constant *LHS = dyn_cast<Constant>(Op0))
2489 if (LHS->isNullValue())
2490 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2492 if (isa<UndefValue>(Op0)) // undef % X -> 0
2493 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2494 if (isa<UndefValue>(Op1))
2495 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2497 // Handle cases involving: rem X, (select Cond, Y, Z)
2498 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2499 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2500 // the same basic block, then we replace the select with Y, and the
2501 // condition of the select with false (if the cond value is in the same
2502 // BB). If the select has uses other than the div, this allows them to be
2504 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2505 if (ST->isNullValue()) {
2506 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2507 if (CondI && CondI->getParent() == I.getParent())
2508 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2509 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2510 I.setOperand(1, SI->getOperand(2));
2512 UpdateValueUsesWith(SI, SI->getOperand(2));
2515 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2516 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2517 if (ST->isNullValue()) {
2518 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2519 if (CondI && CondI->getParent() == I.getParent())
2520 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2521 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2522 I.setOperand(1, SI->getOperand(1));
2524 UpdateValueUsesWith(SI, SI->getOperand(1));
2532 /// This function implements the transforms common to both integer remainder
2533 /// instructions (urem and srem). It is called by the visitors to those integer
2534 /// remainder instructions.
2535 /// @brief Common integer remainder transforms
2536 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2537 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2539 if (Instruction *common = commonRemTransforms(I))
2542 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2543 // X % 0 == undef, we don't need to preserve faults!
2544 if (RHS->equalsInt(0))
2545 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2547 if (RHS->equalsInt(1)) // X % 1 == 0
2548 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2550 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2551 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2552 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2554 } else if (isa<PHINode>(Op0I)) {
2555 if (Instruction *NV = FoldOpIntoPhi(I))
2558 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2559 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2560 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2567 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2568 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2570 if (Instruction *common = commonIRemTransforms(I))
2573 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2574 // X urem C^2 -> X and C
2575 // Check to see if this is an unsigned remainder with an exact power of 2,
2576 // if so, convert to a bitwise and.
2577 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2578 if (isPowerOf2_64(C->getZExtValue()))
2579 return BinaryOperator::createAnd(Op0, SubOne(C));
2582 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2583 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2584 if (RHSI->getOpcode() == Instruction::Shl &&
2585 isa<ConstantInt>(RHSI->getOperand(0))) {
2586 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2587 if (isPowerOf2_64(C1)) {
2588 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2589 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2591 return BinaryOperator::createAnd(Op0, Add);
2596 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2597 // where C1&C2 are powers of two.
2598 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2599 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2600 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2601 // STO == 0 and SFO == 0 handled above.
2602 if (isPowerOf2_64(STO->getZExtValue()) &&
2603 isPowerOf2_64(SFO->getZExtValue())) {
2604 Value *TrueAnd = InsertNewInstBefore(
2605 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2606 Value *FalseAnd = InsertNewInstBefore(
2607 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2608 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2616 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2617 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2619 if (Instruction *common = commonIRemTransforms(I))
2622 if (Value *RHSNeg = dyn_castNegVal(Op1))
2623 if (!isa<ConstantInt>(RHSNeg) ||
2624 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2626 AddUsesToWorkList(I);
2627 I.setOperand(1, RHSNeg);
2631 // If the top bits of both operands are zero (i.e. we can prove they are
2632 // unsigned inputs), turn this into a urem.
2633 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2634 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2635 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2636 return BinaryOperator::createURem(Op0, Op1, I.getName());
2642 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2643 return commonRemTransforms(I);
2646 // isMaxValueMinusOne - return true if this is Max-1
2647 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2649 // Calculate 0111111111..11111
2650 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2651 int64_t Val = INT64_MAX; // All ones
2652 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2653 return C->getSExtValue() == Val-1;
2655 return C->getZExtValue() == C->getType()->getBitMask()-1;
2658 // isMinValuePlusOne - return true if this is Min+1
2659 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2661 // Calculate 1111111111000000000000
2662 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2663 int64_t Val = -1; // All ones
2664 Val <<= TypeBits-1; // Shift over to the right spot
2665 return C->getSExtValue() == Val+1;
2667 return C->getZExtValue() == 1; // unsigned
2670 // isOneBitSet - Return true if there is exactly one bit set in the specified
2672 static bool isOneBitSet(const ConstantInt *CI) {
2673 uint64_t V = CI->getZExtValue();
2674 return V && (V & (V-1)) == 0;
2677 #if 0 // Currently unused
2678 // isLowOnes - Return true if the constant is of the form 0+1+.
2679 static bool isLowOnes(const ConstantInt *CI) {
2680 uint64_t V = CI->getZExtValue();
2682 // There won't be bits set in parts that the type doesn't contain.
2683 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2685 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2686 return U && V && (U & V) == 0;
2690 // isHighOnes - Return true if the constant is of the form 1+0+.
2691 // This is the same as lowones(~X).
2692 static bool isHighOnes(const ConstantInt *CI) {
2693 uint64_t V = ~CI->getZExtValue();
2694 if (~V == 0) return false; // 0's does not match "1+"
2696 // There won't be bits set in parts that the type doesn't contain.
2697 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2699 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2700 return U && V && (U & V) == 0;
2703 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2704 /// are carefully arranged to allow folding of expressions such as:
2706 /// (A < B) | (A > B) --> (A != B)
2708 /// Note that this is only valid if the first and second predicates have the
2709 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2711 /// Three bits are used to represent the condition, as follows:
2716 /// <=> Value Definition
2717 /// 000 0 Always false
2724 /// 111 7 Always true
2726 static unsigned getICmpCode(const ICmpInst *ICI) {
2727 switch (ICI->getPredicate()) {
2729 case ICmpInst::ICMP_UGT: return 1; // 001
2730 case ICmpInst::ICMP_SGT: return 1; // 001
2731 case ICmpInst::ICMP_EQ: return 2; // 010
2732 case ICmpInst::ICMP_UGE: return 3; // 011
2733 case ICmpInst::ICMP_SGE: return 3; // 011
2734 case ICmpInst::ICMP_ULT: return 4; // 100
2735 case ICmpInst::ICMP_SLT: return 4; // 100
2736 case ICmpInst::ICMP_NE: return 5; // 101
2737 case ICmpInst::ICMP_ULE: return 6; // 110
2738 case ICmpInst::ICMP_SLE: return 6; // 110
2741 assert(0 && "Invalid ICmp predicate!");
2746 /// getICmpValue - This is the complement of getICmpCode, which turns an
2747 /// opcode and two operands into either a constant true or false, or a brand
2748 /// new /// ICmp instruction. The sign is passed in to determine which kind
2749 /// of predicate to use in new icmp instructions.
2750 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2752 default: assert(0 && "Illegal ICmp code!");
2753 case 0: return ConstantInt::getFalse();
2756 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2758 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2759 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2762 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2764 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2767 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2769 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2770 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2773 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2775 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2776 case 7: return ConstantInt::getTrue();
2780 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2781 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2782 (ICmpInst::isSignedPredicate(p1) &&
2783 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2784 (ICmpInst::isSignedPredicate(p2) &&
2785 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2789 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2790 struct FoldICmpLogical {
2793 ICmpInst::Predicate pred;
2794 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2795 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2796 pred(ICI->getPredicate()) {}
2797 bool shouldApply(Value *V) const {
2798 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2799 if (PredicatesFoldable(pred, ICI->getPredicate()))
2800 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2801 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2804 Instruction *apply(Instruction &Log) const {
2805 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2806 if (ICI->getOperand(0) != LHS) {
2807 assert(ICI->getOperand(1) == LHS);
2808 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2811 unsigned LHSCode = getICmpCode(ICI);
2812 unsigned RHSCode = getICmpCode(cast<ICmpInst>(Log.getOperand(1)));
2814 switch (Log.getOpcode()) {
2815 case Instruction::And: Code = LHSCode & RHSCode; break;
2816 case Instruction::Or: Code = LHSCode | RHSCode; break;
2817 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2818 default: assert(0 && "Illegal logical opcode!"); return 0;
2821 Value *RV = getICmpValue(ICmpInst::isSignedPredicate(pred), Code, LHS, RHS);
2822 if (Instruction *I = dyn_cast<Instruction>(RV))
2824 // Otherwise, it's a constant boolean value...
2825 return IC.ReplaceInstUsesWith(Log, RV);
2828 } // end anonymous namespace
2830 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2831 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2832 // guaranteed to be a binary operator.
2833 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2835 ConstantInt *AndRHS,
2836 BinaryOperator &TheAnd) {
2837 Value *X = Op->getOperand(0);
2838 Constant *Together = 0;
2840 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2842 switch (Op->getOpcode()) {
2843 case Instruction::Xor:
2844 if (Op->hasOneUse()) {
2845 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2846 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2847 InsertNewInstBefore(And, TheAnd);
2849 return BinaryOperator::createXor(And, Together);
2852 case Instruction::Or:
2853 if (Together == AndRHS) // (X | C) & C --> C
2854 return ReplaceInstUsesWith(TheAnd, AndRHS);
2856 if (Op->hasOneUse() && Together != OpRHS) {
2857 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2858 Instruction *Or = BinaryOperator::createOr(X, Together);
2859 InsertNewInstBefore(Or, TheAnd);
2861 return BinaryOperator::createAnd(Or, AndRHS);
2864 case Instruction::Add:
2865 if (Op->hasOneUse()) {
2866 // Adding a one to a single bit bit-field should be turned into an XOR
2867 // of the bit. First thing to check is to see if this AND is with a
2868 // single bit constant.
2869 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2871 // Clear bits that are not part of the constant.
2872 AndRHSV &= AndRHS->getType()->getBitMask();
2874 // If there is only one bit set...
2875 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2876 // Ok, at this point, we know that we are masking the result of the
2877 // ADD down to exactly one bit. If the constant we are adding has
2878 // no bits set below this bit, then we can eliminate the ADD.
2879 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2881 // Check to see if any bits below the one bit set in AndRHSV are set.
2882 if ((AddRHS & (AndRHSV-1)) == 0) {
2883 // If not, the only thing that can effect the output of the AND is
2884 // the bit specified by AndRHSV. If that bit is set, the effect of
2885 // the XOR is to toggle the bit. If it is clear, then the ADD has
2887 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2888 TheAnd.setOperand(0, X);
2891 // Pull the XOR out of the AND.
2892 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
2893 InsertNewInstBefore(NewAnd, TheAnd);
2894 NewAnd->takeName(Op);
2895 return BinaryOperator::createXor(NewAnd, AndRHS);
2902 case Instruction::Shl: {
2903 // We know that the AND will not produce any of the bits shifted in, so if
2904 // the anded constant includes them, clear them now!
2906 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2907 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2908 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2910 if (CI == ShlMask) { // Masking out bits that the shift already masks
2911 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2912 } else if (CI != AndRHS) { // Reducing bits set in and.
2913 TheAnd.setOperand(1, CI);
2918 case Instruction::LShr:
2920 // We know that the AND will not produce any of the bits shifted in, so if
2921 // the anded constant includes them, clear them now! This only applies to
2922 // unsigned shifts, because a signed shr may bring in set bits!
2924 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2925 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2926 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2928 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2929 return ReplaceInstUsesWith(TheAnd, Op);
2930 } else if (CI != AndRHS) {
2931 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2936 case Instruction::AShr:
2938 // See if this is shifting in some sign extension, then masking it out
2940 if (Op->hasOneUse()) {
2941 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2942 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2943 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
2944 if (C == AndRHS) { // Masking out bits shifted in.
2945 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2946 // Make the argument unsigned.
2947 Value *ShVal = Op->getOperand(0);
2948 ShVal = InsertNewInstBefore(
2949 BinaryOperator::createLShr(ShVal, OpRHS,
2950 Op->getName()), TheAnd);
2951 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
2960 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2961 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2962 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
2963 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
2964 /// insert new instructions.
2965 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2966 bool isSigned, bool Inside,
2968 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
2969 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
2970 "Lo is not <= Hi in range emission code!");
2973 if (Lo == Hi) // Trivially false.
2974 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
2976 // V >= Min && V < Hi --> V < Hi
2977 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2978 ICmpInst::Predicate pred = (isSigned ?
2979 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
2980 return new ICmpInst(pred, V, Hi);
2983 // Emit V-Lo <u Hi-Lo
2984 Constant *NegLo = ConstantExpr::getNeg(Lo);
2985 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2986 InsertNewInstBefore(Add, IB);
2987 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
2988 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
2991 if (Lo == Hi) // Trivially true.
2992 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
2994 // V < Min || V >= Hi ->'V > Hi-1'
2995 Hi = SubOne(cast<ConstantInt>(Hi));
2996 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2997 ICmpInst::Predicate pred = (isSigned ?
2998 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
2999 return new ICmpInst(pred, V, Hi);
3002 // Emit V-Lo > Hi-1-Lo
3003 Constant *NegLo = ConstantExpr::getNeg(Lo);
3004 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3005 InsertNewInstBefore(Add, IB);
3006 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3007 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3010 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3011 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3012 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3013 // not, since all 1s are not contiguous.
3014 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
3015 uint64_t V = Val->getZExtValue();
3016 if (!isShiftedMask_64(V)) return false;
3018 // look for the first zero bit after the run of ones
3019 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
3020 // look for the first non-zero bit
3021 ME = 64-CountLeadingZeros_64(V);
3027 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3028 /// where isSub determines whether the operator is a sub. If we can fold one of
3029 /// the following xforms:
3031 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3032 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3033 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3035 /// return (A +/- B).
3037 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3038 ConstantInt *Mask, bool isSub,
3040 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3041 if (!LHSI || LHSI->getNumOperands() != 2 ||
3042 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3044 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3046 switch (LHSI->getOpcode()) {
3048 case Instruction::And:
3049 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3050 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3051 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3054 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3055 // part, we don't need any explicit masks to take them out of A. If that
3056 // is all N is, ignore it.
3058 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3059 uint64_t Mask = cast<IntegerType>(RHS->getType())->getBitMask();
3061 if (MaskedValueIsZero(RHS, Mask))
3066 case Instruction::Or:
3067 case Instruction::Xor:
3068 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3069 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3070 ConstantExpr::getAnd(N, Mask)->isNullValue())
3077 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3079 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3080 return InsertNewInstBefore(New, I);
3083 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3084 bool Changed = SimplifyCommutative(I);
3085 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3087 if (isa<UndefValue>(Op1)) // X & undef -> 0
3088 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3092 return ReplaceInstUsesWith(I, Op1);
3094 // See if we can simplify any instructions used by the instruction whose sole
3095 // purpose is to compute bits we don't care about.
3096 uint64_t KnownZero, KnownOne;
3097 if (!isa<VectorType>(I.getType())) {
3098 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3099 KnownZero, KnownOne))
3102 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3103 if (CP->isAllOnesValue())
3104 return ReplaceInstUsesWith(I, I.getOperand(0));
3108 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3109 uint64_t AndRHSMask = AndRHS->getZExtValue();
3110 uint64_t TypeMask = cast<IntegerType>(Op0->getType())->getBitMask();
3111 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3113 // Optimize a variety of ((val OP C1) & C2) combinations...
3114 if (isa<BinaryOperator>(Op0)) {
3115 Instruction *Op0I = cast<Instruction>(Op0);
3116 Value *Op0LHS = Op0I->getOperand(0);
3117 Value *Op0RHS = Op0I->getOperand(1);
3118 switch (Op0I->getOpcode()) {
3119 case Instruction::Xor:
3120 case Instruction::Or:
3121 // If the mask is only needed on one incoming arm, push it up.
3122 if (Op0I->hasOneUse()) {
3123 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3124 // Not masking anything out for the LHS, move to RHS.
3125 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3126 Op0RHS->getName()+".masked");
3127 InsertNewInstBefore(NewRHS, I);
3128 return BinaryOperator::create(
3129 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3131 if (!isa<Constant>(Op0RHS) &&
3132 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3133 // Not masking anything out for the RHS, move to LHS.
3134 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3135 Op0LHS->getName()+".masked");
3136 InsertNewInstBefore(NewLHS, I);
3137 return BinaryOperator::create(
3138 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3143 case Instruction::Add:
3144 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3145 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3146 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3147 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3148 return BinaryOperator::createAnd(V, AndRHS);
3149 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3150 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3153 case Instruction::Sub:
3154 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3155 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3156 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3157 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3158 return BinaryOperator::createAnd(V, AndRHS);
3162 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3163 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3165 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3166 // If this is an integer truncation or change from signed-to-unsigned, and
3167 // if the source is an and/or with immediate, transform it. This
3168 // frequently occurs for bitfield accesses.
3169 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3170 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3171 CastOp->getNumOperands() == 2)
3172 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3173 if (CastOp->getOpcode() == Instruction::And) {
3174 // Change: and (cast (and X, C1) to T), C2
3175 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3176 // This will fold the two constants together, which may allow
3177 // other simplifications.
3178 Instruction *NewCast = CastInst::createTruncOrBitCast(
3179 CastOp->getOperand(0), I.getType(),
3180 CastOp->getName()+".shrunk");
3181 NewCast = InsertNewInstBefore(NewCast, I);
3182 // trunc_or_bitcast(C1)&C2
3183 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3184 C3 = ConstantExpr::getAnd(C3, AndRHS);
3185 return BinaryOperator::createAnd(NewCast, C3);
3186 } else if (CastOp->getOpcode() == Instruction::Or) {
3187 // Change: and (cast (or X, C1) to T), C2
3188 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3189 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3190 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3191 return ReplaceInstUsesWith(I, AndRHS);
3196 // Try to fold constant and into select arguments.
3197 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3198 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3200 if (isa<PHINode>(Op0))
3201 if (Instruction *NV = FoldOpIntoPhi(I))
3205 Value *Op0NotVal = dyn_castNotVal(Op0);
3206 Value *Op1NotVal = dyn_castNotVal(Op1);
3208 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3209 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3211 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3212 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3213 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3214 I.getName()+".demorgan");
3215 InsertNewInstBefore(Or, I);
3216 return BinaryOperator::createNot(Or);
3220 Value *A = 0, *B = 0;
3221 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3222 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3223 return ReplaceInstUsesWith(I, Op1);
3224 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3225 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3226 return ReplaceInstUsesWith(I, Op0);
3228 if (Op0->hasOneUse() &&
3229 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3230 if (A == Op1) { // (A^B)&A -> A&(A^B)
3231 I.swapOperands(); // Simplify below
3232 std::swap(Op0, Op1);
3233 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3234 cast<BinaryOperator>(Op0)->swapOperands();
3235 I.swapOperands(); // Simplify below
3236 std::swap(Op0, Op1);
3239 if (Op1->hasOneUse() &&
3240 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3241 if (B == Op0) { // B&(A^B) -> B&(B^A)
3242 cast<BinaryOperator>(Op1)->swapOperands();
3245 if (A == Op0) { // A&(A^B) -> A & ~B
3246 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3247 InsertNewInstBefore(NotB, I);
3248 return BinaryOperator::createAnd(A, NotB);
3253 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3254 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3255 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3258 Value *LHSVal, *RHSVal;
3259 ConstantInt *LHSCst, *RHSCst;
3260 ICmpInst::Predicate LHSCC, RHSCC;
3261 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3262 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3263 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3264 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3265 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3266 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3267 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3268 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3269 // Ensure that the larger constant is on the RHS.
3270 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3271 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3272 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3273 ICmpInst *LHS = cast<ICmpInst>(Op0);
3274 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3275 std::swap(LHS, RHS);
3276 std::swap(LHSCst, RHSCst);
3277 std::swap(LHSCC, RHSCC);
3280 // At this point, we know we have have two icmp instructions
3281 // comparing a value against two constants and and'ing the result
3282 // together. Because of the above check, we know that we only have
3283 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3284 // (from the FoldICmpLogical check above), that the two constants
3285 // are not equal and that the larger constant is on the RHS
3286 assert(LHSCst != RHSCst && "Compares not folded above?");
3289 default: assert(0 && "Unknown integer condition code!");
3290 case ICmpInst::ICMP_EQ:
3292 default: assert(0 && "Unknown integer condition code!");
3293 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3294 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3295 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3296 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3297 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3298 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3299 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3300 return ReplaceInstUsesWith(I, LHS);
3302 case ICmpInst::ICMP_NE:
3304 default: assert(0 && "Unknown integer condition code!");
3305 case ICmpInst::ICMP_ULT:
3306 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3307 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3308 break; // (X != 13 & X u< 15) -> no change
3309 case ICmpInst::ICMP_SLT:
3310 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3311 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3312 break; // (X != 13 & X s< 15) -> no change
3313 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3314 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3315 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3316 return ReplaceInstUsesWith(I, RHS);
3317 case ICmpInst::ICMP_NE:
3318 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3319 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3320 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3321 LHSVal->getName()+".off");
3322 InsertNewInstBefore(Add, I);
3323 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3324 ConstantInt::get(Add->getType(), 1));
3326 break; // (X != 13 & X != 15) -> no change
3329 case ICmpInst::ICMP_ULT:
3331 default: assert(0 && "Unknown integer condition code!");
3332 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3333 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3334 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3335 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3337 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3338 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3339 return ReplaceInstUsesWith(I, LHS);
3340 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3344 case ICmpInst::ICMP_SLT:
3346 default: assert(0 && "Unknown integer condition code!");
3347 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3348 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3349 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3350 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3352 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3353 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3354 return ReplaceInstUsesWith(I, LHS);
3355 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3359 case ICmpInst::ICMP_UGT:
3361 default: assert(0 && "Unknown integer condition code!");
3362 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3363 return ReplaceInstUsesWith(I, LHS);
3364 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3365 return ReplaceInstUsesWith(I, RHS);
3366 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3368 case ICmpInst::ICMP_NE:
3369 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3370 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3371 break; // (X u> 13 & X != 15) -> no change
3372 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3373 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3375 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3379 case ICmpInst::ICMP_SGT:
3381 default: assert(0 && "Unknown integer condition code!");
3382 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3383 return ReplaceInstUsesWith(I, LHS);
3384 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3385 return ReplaceInstUsesWith(I, RHS);
3386 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3388 case ICmpInst::ICMP_NE:
3389 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3390 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3391 break; // (X s> 13 & X != 15) -> no change
3392 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3393 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3395 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3403 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3404 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3405 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3406 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3407 const Type *SrcTy = Op0C->getOperand(0)->getType();
3408 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3409 // Only do this if the casts both really cause code to be generated.
3410 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3412 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3414 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3415 Op1C->getOperand(0),
3417 InsertNewInstBefore(NewOp, I);
3418 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3422 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3423 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3424 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3425 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3426 SI0->getOperand(1) == SI1->getOperand(1) &&
3427 (SI0->hasOneUse() || SI1->hasOneUse())) {
3428 Instruction *NewOp =
3429 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3431 SI0->getName()), I);
3432 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3433 SI1->getOperand(1));
3437 return Changed ? &I : 0;
3440 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3441 /// in the result. If it does, and if the specified byte hasn't been filled in
3442 /// yet, fill it in and return false.
3443 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3444 Instruction *I = dyn_cast<Instruction>(V);
3445 if (I == 0) return true;
3447 // If this is an or instruction, it is an inner node of the bswap.
3448 if (I->getOpcode() == Instruction::Or)
3449 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3450 CollectBSwapParts(I->getOperand(1), ByteValues);
3452 // If this is a shift by a constant int, and it is "24", then its operand
3453 // defines a byte. We only handle unsigned types here.
3454 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3455 // Not shifting the entire input by N-1 bytes?
3456 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3457 8*(ByteValues.size()-1))
3461 if (I->getOpcode() == Instruction::Shl) {
3462 // X << 24 defines the top byte with the lowest of the input bytes.
3463 DestNo = ByteValues.size()-1;
3465 // X >>u 24 defines the low byte with the highest of the input bytes.
3469 // If the destination byte value is already defined, the values are or'd
3470 // together, which isn't a bswap (unless it's an or of the same bits).
3471 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3473 ByteValues[DestNo] = I->getOperand(0);
3477 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3479 Value *Shift = 0, *ShiftLHS = 0;
3480 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3481 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3482 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3484 Instruction *SI = cast<Instruction>(Shift);
3486 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3487 if (ShiftAmt->getZExtValue() & 7 ||
3488 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3491 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3493 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3494 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3496 // Unknown mask for bswap.
3497 if (DestByte == ByteValues.size()) return true;
3499 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3501 if (SI->getOpcode() == Instruction::Shl)
3502 SrcByte = DestByte - ShiftBytes;
3504 SrcByte = DestByte + ShiftBytes;
3506 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3507 if (SrcByte != ByteValues.size()-DestByte-1)
3510 // If the destination byte value is already defined, the values are or'd
3511 // together, which isn't a bswap (unless it's an or of the same bits).
3512 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3514 ByteValues[DestByte] = SI->getOperand(0);
3518 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3519 /// If so, insert the new bswap intrinsic and return it.
3520 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3521 // We cannot bswap one byte.
3522 if (I.getType() == Type::Int8Ty)
3525 /// ByteValues - For each byte of the result, we keep track of which value
3526 /// defines each byte.
3527 SmallVector<Value*, 8> ByteValues;
3528 ByteValues.resize(TD->getTypeSize(I.getType()));
3530 // Try to find all the pieces corresponding to the bswap.
3531 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3532 CollectBSwapParts(I.getOperand(1), ByteValues))
3535 // Check to see if all of the bytes come from the same value.
3536 Value *V = ByteValues[0];
3537 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3539 // Check to make sure that all of the bytes come from the same value.
3540 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3541 if (ByteValues[i] != V)
3544 // If they do then *success* we can turn this into a bswap. Figure out what
3545 // bswap to make it into.
3546 Module *M = I.getParent()->getParent()->getParent();
3547 const char *FnName = 0;
3548 if (I.getType() == Type::Int16Ty)
3549 FnName = "llvm.bswap.i16";
3550 else if (I.getType() == Type::Int32Ty)
3551 FnName = "llvm.bswap.i32";
3552 else if (I.getType() == Type::Int64Ty)
3553 FnName = "llvm.bswap.i64";
3555 assert(0 && "Unknown integer type!");
3556 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3557 return new CallInst(F, V);
3561 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3562 bool Changed = SimplifyCommutative(I);
3563 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3565 if (isa<UndefValue>(Op1))
3566 return ReplaceInstUsesWith(I, // X | undef -> -1
3567 ConstantInt::getAllOnesValue(I.getType()));
3571 return ReplaceInstUsesWith(I, Op0);
3573 // See if we can simplify any instructions used by the instruction whose sole
3574 // purpose is to compute bits we don't care about.
3575 uint64_t KnownZero, KnownOne;
3576 if (!isa<VectorType>(I.getType()) &&
3577 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3578 KnownZero, KnownOne))
3582 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3583 ConstantInt *C1 = 0; Value *X = 0;
3584 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3585 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3586 Instruction *Or = BinaryOperator::createOr(X, RHS);
3587 InsertNewInstBefore(Or, I);
3589 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3592 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3593 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3594 Instruction *Or = BinaryOperator::createOr(X, RHS);
3595 InsertNewInstBefore(Or, I);
3597 return BinaryOperator::createXor(Or,
3598 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3601 // Try to fold constant and into select arguments.
3602 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3603 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3605 if (isa<PHINode>(Op0))
3606 if (Instruction *NV = FoldOpIntoPhi(I))
3610 Value *A = 0, *B = 0;
3611 ConstantInt *C1 = 0, *C2 = 0;
3613 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3614 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3615 return ReplaceInstUsesWith(I, Op1);
3616 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3617 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3618 return ReplaceInstUsesWith(I, Op0);
3620 // (A | B) | C and A | (B | C) -> bswap if possible.
3621 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3622 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3623 match(Op1, m_Or(m_Value(), m_Value())) ||
3624 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3625 match(Op1, m_Shift(m_Value(), m_Value())))) {
3626 if (Instruction *BSwap = MatchBSwap(I))
3630 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3631 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3632 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3633 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3634 InsertNewInstBefore(NOr, I);
3636 return BinaryOperator::createXor(NOr, C1);
3639 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3640 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3641 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3642 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3643 InsertNewInstBefore(NOr, I);
3645 return BinaryOperator::createXor(NOr, C1);
3648 // (A & C1)|(B & C2)
3649 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3650 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3652 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3653 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3656 // If we have: ((V + N) & C1) | (V & C2)
3657 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3658 // replace with V+N.
3659 if (C1 == ConstantExpr::getNot(C2)) {
3660 Value *V1 = 0, *V2 = 0;
3661 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3662 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3663 // Add commutes, try both ways.
3664 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3665 return ReplaceInstUsesWith(I, A);
3666 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3667 return ReplaceInstUsesWith(I, A);
3669 // Or commutes, try both ways.
3670 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3671 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3672 // Add commutes, try both ways.
3673 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3674 return ReplaceInstUsesWith(I, B);
3675 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3676 return ReplaceInstUsesWith(I, B);
3681 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3682 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3683 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3684 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3685 SI0->getOperand(1) == SI1->getOperand(1) &&
3686 (SI0->hasOneUse() || SI1->hasOneUse())) {
3687 Instruction *NewOp =
3688 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3690 SI0->getName()), I);
3691 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3692 SI1->getOperand(1));
3696 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3697 if (A == Op1) // ~A | A == -1
3698 return ReplaceInstUsesWith(I,
3699 ConstantInt::getAllOnesValue(I.getType()));
3703 // Note, A is still live here!
3704 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3706 return ReplaceInstUsesWith(I,
3707 ConstantInt::getAllOnesValue(I.getType()));
3709 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3710 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3711 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3712 I.getName()+".demorgan"), I);
3713 return BinaryOperator::createNot(And);
3717 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3718 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3719 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3722 Value *LHSVal, *RHSVal;
3723 ConstantInt *LHSCst, *RHSCst;
3724 ICmpInst::Predicate LHSCC, RHSCC;
3725 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3726 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3727 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3728 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3729 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3730 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3731 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3732 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3733 // Ensure that the larger constant is on the RHS.
3734 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3735 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3736 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3737 ICmpInst *LHS = cast<ICmpInst>(Op0);
3738 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3739 std::swap(LHS, RHS);
3740 std::swap(LHSCst, RHSCst);
3741 std::swap(LHSCC, RHSCC);
3744 // At this point, we know we have have two icmp instructions
3745 // comparing a value against two constants and or'ing the result
3746 // together. Because of the above check, we know that we only have
3747 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3748 // FoldICmpLogical check above), that the two constants are not
3750 assert(LHSCst != RHSCst && "Compares not folded above?");
3753 default: assert(0 && "Unknown integer condition code!");
3754 case ICmpInst::ICMP_EQ:
3756 default: assert(0 && "Unknown integer condition code!");
3757 case ICmpInst::ICMP_EQ:
3758 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3759 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3760 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3761 LHSVal->getName()+".off");
3762 InsertNewInstBefore(Add, I);
3763 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3764 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3766 break; // (X == 13 | X == 15) -> no change
3767 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3768 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3770 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3771 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3772 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3773 return ReplaceInstUsesWith(I, RHS);
3776 case ICmpInst::ICMP_NE:
3778 default: assert(0 && "Unknown integer condition code!");
3779 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3780 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3781 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3782 return ReplaceInstUsesWith(I, LHS);
3783 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3784 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3785 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3786 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3789 case ICmpInst::ICMP_ULT:
3791 default: assert(0 && "Unknown integer condition code!");
3792 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3794 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3795 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3797 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3799 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3800 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3801 return ReplaceInstUsesWith(I, RHS);
3802 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3806 case ICmpInst::ICMP_SLT:
3808 default: assert(0 && "Unknown integer condition code!");
3809 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3811 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3812 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3814 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3816 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3817 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3818 return ReplaceInstUsesWith(I, RHS);
3819 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3823 case ICmpInst::ICMP_UGT:
3825 default: assert(0 && "Unknown integer condition code!");
3826 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3827 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3828 return ReplaceInstUsesWith(I, LHS);
3829 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3831 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3832 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3833 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3834 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3838 case ICmpInst::ICMP_SGT:
3840 default: assert(0 && "Unknown integer condition code!");
3841 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3842 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3843 return ReplaceInstUsesWith(I, LHS);
3844 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3846 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3847 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3848 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3849 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3857 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3858 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3859 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3860 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3861 const Type *SrcTy = Op0C->getOperand(0)->getType();
3862 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3863 // Only do this if the casts both really cause code to be generated.
3864 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3866 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3868 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3869 Op1C->getOperand(0),
3871 InsertNewInstBefore(NewOp, I);
3872 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3877 return Changed ? &I : 0;
3880 // XorSelf - Implements: X ^ X --> 0
3883 XorSelf(Value *rhs) : RHS(rhs) {}
3884 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3885 Instruction *apply(BinaryOperator &Xor) const {
3891 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3892 bool Changed = SimplifyCommutative(I);
3893 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3895 if (isa<UndefValue>(Op1))
3896 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3898 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3899 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3900 assert(Result == &I && "AssociativeOpt didn't work?");
3901 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3904 // See if we can simplify any instructions used by the instruction whose sole
3905 // purpose is to compute bits we don't care about.
3906 uint64_t KnownZero, KnownOne;
3907 if (!isa<VectorType>(I.getType()) &&
3908 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3909 KnownZero, KnownOne))
3912 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3913 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
3914 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3915 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
3916 return new ICmpInst(ICI->getInversePredicate(),
3917 ICI->getOperand(0), ICI->getOperand(1));
3919 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3920 // ~(c-X) == X-c-1 == X+(-c-1)
3921 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3922 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3923 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3924 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3925 ConstantInt::get(I.getType(), 1));
3926 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3929 // ~(~X & Y) --> (X | ~Y)
3930 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3931 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3932 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3934 BinaryOperator::createNot(Op0I->getOperand(1),
3935 Op0I->getOperand(1)->getName()+".not");
3936 InsertNewInstBefore(NotY, I);
3937 return BinaryOperator::createOr(Op0NotVal, NotY);
3941 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3942 if (Op0I->getOpcode() == Instruction::Add) {
3943 // ~(X-c) --> (-c-1)-X
3944 if (RHS->isAllOnesValue()) {
3945 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3946 return BinaryOperator::createSub(
3947 ConstantExpr::getSub(NegOp0CI,
3948 ConstantInt::get(I.getType(), 1)),
3949 Op0I->getOperand(0));
3951 } else if (Op0I->getOpcode() == Instruction::Or) {
3952 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3953 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3954 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3955 // Anything in both C1 and C2 is known to be zero, remove it from
3957 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3958 NewRHS = ConstantExpr::getAnd(NewRHS,
3959 ConstantExpr::getNot(CommonBits));
3960 AddToWorkList(Op0I);
3961 I.setOperand(0, Op0I->getOperand(0));
3962 I.setOperand(1, NewRHS);
3968 // Try to fold constant and into select arguments.
3969 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3970 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3972 if (isa<PHINode>(Op0))
3973 if (Instruction *NV = FoldOpIntoPhi(I))
3977 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3979 return ReplaceInstUsesWith(I,
3980 ConstantInt::getAllOnesValue(I.getType()));
3982 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3984 return ReplaceInstUsesWith(I,
3985 ConstantInt::getAllOnesValue(I.getType()));
3987 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3988 if (Op1I->getOpcode() == Instruction::Or) {
3989 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3990 Op1I->swapOperands();
3992 std::swap(Op0, Op1);
3993 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3994 I.swapOperands(); // Simplified below.
3995 std::swap(Op0, Op1);
3997 } else if (Op1I->getOpcode() == Instruction::Xor) {
3998 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3999 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
4000 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
4001 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
4002 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
4003 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
4004 Op1I->swapOperands();
4005 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
4006 I.swapOperands(); // Simplified below.
4007 std::swap(Op0, Op1);
4011 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
4012 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
4013 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
4014 Op0I->swapOperands();
4015 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
4016 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
4017 InsertNewInstBefore(NotB, I);
4018 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
4020 } else if (Op0I->getOpcode() == Instruction::Xor) {
4021 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
4022 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
4023 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
4024 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
4025 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
4026 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
4027 Op0I->swapOperands();
4028 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
4029 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4030 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
4031 InsertNewInstBefore(N, I);
4032 return BinaryOperator::createAnd(N, Op1);
4036 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4037 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4038 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4041 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4042 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4043 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4044 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4045 const Type *SrcTy = Op0C->getOperand(0)->getType();
4046 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4047 // Only do this if the casts both really cause code to be generated.
4048 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4050 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4052 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4053 Op1C->getOperand(0),
4055 InsertNewInstBefore(NewOp, I);
4056 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4060 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4061 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4062 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4063 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4064 SI0->getOperand(1) == SI1->getOperand(1) &&
4065 (SI0->hasOneUse() || SI1->hasOneUse())) {
4066 Instruction *NewOp =
4067 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
4069 SI0->getName()), I);
4070 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4071 SI1->getOperand(1));
4075 return Changed ? &I : 0;
4078 static bool isPositive(ConstantInt *C) {
4079 return C->getSExtValue() >= 0;
4082 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4083 /// overflowed for this type.
4084 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4086 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4088 return cast<ConstantInt>(Result)->getZExtValue() <
4089 cast<ConstantInt>(In1)->getZExtValue();
4092 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4093 /// code necessary to compute the offset from the base pointer (without adding
4094 /// in the base pointer). Return the result as a signed integer of intptr size.
4095 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4096 TargetData &TD = IC.getTargetData();
4097 gep_type_iterator GTI = gep_type_begin(GEP);
4098 const Type *IntPtrTy = TD.getIntPtrType();
4099 Value *Result = Constant::getNullValue(IntPtrTy);
4101 // Build a mask for high order bits.
4102 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4104 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4105 Value *Op = GEP->getOperand(i);
4106 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4107 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4108 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4109 if (!OpC->isNullValue()) {
4110 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4111 Scale = ConstantExpr::getMul(OpC, Scale);
4112 if (Constant *RC = dyn_cast<Constant>(Result))
4113 Result = ConstantExpr::getAdd(RC, Scale);
4115 // Emit an add instruction.
4116 Result = IC.InsertNewInstBefore(
4117 BinaryOperator::createAdd(Result, Scale,
4118 GEP->getName()+".offs"), I);
4122 // Convert to correct type.
4123 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4124 Op->getName()+".c"), I);
4126 // We'll let instcombine(mul) convert this to a shl if possible.
4127 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4128 GEP->getName()+".idx"), I);
4130 // Emit an add instruction.
4131 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4132 GEP->getName()+".offs"), I);
4138 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4139 /// else. At this point we know that the GEP is on the LHS of the comparison.
4140 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4141 ICmpInst::Predicate Cond,
4143 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4145 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4146 if (isa<PointerType>(CI->getOperand(0)->getType()))
4147 RHS = CI->getOperand(0);
4149 Value *PtrBase = GEPLHS->getOperand(0);
4150 if (PtrBase == RHS) {
4151 // As an optimization, we don't actually have to compute the actual value of
4152 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4153 // each index is zero or not.
4154 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4155 Instruction *InVal = 0;
4156 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4157 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4159 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4160 if (isa<UndefValue>(C)) // undef index -> undef.
4161 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4162 if (C->isNullValue())
4164 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4165 EmitIt = false; // This is indexing into a zero sized array?
4166 } else if (isa<ConstantInt>(C))
4167 return ReplaceInstUsesWith(I, // No comparison is needed here.
4168 ConstantInt::get(Type::Int1Ty,
4169 Cond == ICmpInst::ICMP_NE));
4174 new ICmpInst(Cond, GEPLHS->getOperand(i),
4175 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4179 InVal = InsertNewInstBefore(InVal, I);
4180 InsertNewInstBefore(Comp, I);
4181 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4182 InVal = BinaryOperator::createOr(InVal, Comp);
4183 else // True if all are equal
4184 InVal = BinaryOperator::createAnd(InVal, Comp);
4192 // No comparison is needed here, all indexes = 0
4193 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4194 Cond == ICmpInst::ICMP_EQ));
4197 // Only lower this if the icmp is the only user of the GEP or if we expect
4198 // the result to fold to a constant!
4199 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4200 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4201 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4202 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4203 Constant::getNullValue(Offset->getType()));
4205 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4206 // If the base pointers are different, but the indices are the same, just
4207 // compare the base pointer.
4208 if (PtrBase != GEPRHS->getOperand(0)) {
4209 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4210 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4211 GEPRHS->getOperand(0)->getType();
4213 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4214 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4215 IndicesTheSame = false;
4219 // If all indices are the same, just compare the base pointers.
4221 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4222 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4224 // Otherwise, the base pointers are different and the indices are
4225 // different, bail out.
4229 // If one of the GEPs has all zero indices, recurse.
4230 bool AllZeros = true;
4231 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4232 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4233 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4238 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4239 ICmpInst::getSwappedPredicate(Cond), I);
4241 // If the other GEP has all zero indices, recurse.
4243 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4244 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4245 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4250 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4252 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4253 // If the GEPs only differ by one index, compare it.
4254 unsigned NumDifferences = 0; // Keep track of # differences.
4255 unsigned DiffOperand = 0; // The operand that differs.
4256 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4257 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4258 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4259 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4260 // Irreconcilable differences.
4264 if (NumDifferences++) break;
4269 if (NumDifferences == 0) // SAME GEP?
4270 return ReplaceInstUsesWith(I, // No comparison is needed here.
4271 ConstantInt::get(Type::Int1Ty,
4272 Cond == ICmpInst::ICMP_EQ));
4273 else if (NumDifferences == 1) {
4274 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4275 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4276 // Make sure we do a signed comparison here.
4277 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4281 // Only lower this if the icmp is the only user of the GEP or if we expect
4282 // the result to fold to a constant!
4283 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4284 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4285 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4286 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4287 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4288 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4294 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4295 bool Changed = SimplifyCompare(I);
4296 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4298 // Fold trivial predicates.
4299 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4300 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4301 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4302 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4304 // Simplify 'fcmp pred X, X'
4306 switch (I.getPredicate()) {
4307 default: assert(0 && "Unknown predicate!");
4308 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4309 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4310 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4311 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4312 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4313 case FCmpInst::FCMP_OLT: // True if ordered and less than
4314 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4315 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4317 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4318 case FCmpInst::FCMP_ULT: // True if unordered or less than
4319 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4320 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4321 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4322 I.setPredicate(FCmpInst::FCMP_UNO);
4323 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4326 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4327 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4328 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4329 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4330 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4331 I.setPredicate(FCmpInst::FCMP_ORD);
4332 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4337 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4338 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4340 // Handle fcmp with constant RHS
4341 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4342 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4343 switch (LHSI->getOpcode()) {
4344 case Instruction::PHI:
4345 if (Instruction *NV = FoldOpIntoPhi(I))
4348 case Instruction::Select:
4349 // If either operand of the select is a constant, we can fold the
4350 // comparison into the select arms, which will cause one to be
4351 // constant folded and the select turned into a bitwise or.
4352 Value *Op1 = 0, *Op2 = 0;
4353 if (LHSI->hasOneUse()) {
4354 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4355 // Fold the known value into the constant operand.
4356 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4357 // Insert a new FCmp of the other select operand.
4358 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4359 LHSI->getOperand(2), RHSC,
4361 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4362 // Fold the known value into the constant operand.
4363 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4364 // Insert a new FCmp of the other select operand.
4365 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4366 LHSI->getOperand(1), RHSC,
4372 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4377 return Changed ? &I : 0;
4380 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4381 bool Changed = SimplifyCompare(I);
4382 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4383 const Type *Ty = Op0->getType();
4387 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4388 isTrueWhenEqual(I)));
4390 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4391 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4393 // icmp of GlobalValues can never equal each other as long as they aren't
4394 // external weak linkage type.
4395 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4396 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4397 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4398 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4399 !isTrueWhenEqual(I)));
4401 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4402 // addresses never equal each other! We already know that Op0 != Op1.
4403 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4404 isa<ConstantPointerNull>(Op0)) &&
4405 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4406 isa<ConstantPointerNull>(Op1)))
4407 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4408 !isTrueWhenEqual(I)));
4410 // icmp's with boolean values can always be turned into bitwise operations
4411 if (Ty == Type::Int1Ty) {
4412 switch (I.getPredicate()) {
4413 default: assert(0 && "Invalid icmp instruction!");
4414 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4415 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4416 InsertNewInstBefore(Xor, I);
4417 return BinaryOperator::createNot(Xor);
4419 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4420 return BinaryOperator::createXor(Op0, Op1);
4422 case ICmpInst::ICMP_UGT:
4423 case ICmpInst::ICMP_SGT:
4424 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4426 case ICmpInst::ICMP_ULT:
4427 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4428 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4429 InsertNewInstBefore(Not, I);
4430 return BinaryOperator::createAnd(Not, Op1);
4432 case ICmpInst::ICMP_UGE:
4433 case ICmpInst::ICMP_SGE:
4434 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4436 case ICmpInst::ICMP_ULE:
4437 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4438 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4439 InsertNewInstBefore(Not, I);
4440 return BinaryOperator::createOr(Not, Op1);
4445 // See if we are doing a comparison between a constant and an instruction that
4446 // can be folded into the comparison.
4447 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4448 switch (I.getPredicate()) {
4450 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4451 if (CI->isMinValue(false))
4452 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4453 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4454 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4455 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4456 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4459 case ICmpInst::ICMP_SLT:
4460 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4461 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4462 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4463 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4464 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4465 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4468 case ICmpInst::ICMP_UGT:
4469 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4470 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4471 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4472 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4473 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4474 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4477 case ICmpInst::ICMP_SGT:
4478 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4479 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4480 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4481 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4482 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4483 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4486 case ICmpInst::ICMP_ULE:
4487 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4488 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4489 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4490 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4491 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4492 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4495 case ICmpInst::ICMP_SLE:
4496 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4497 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4498 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4499 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4500 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4501 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4504 case ICmpInst::ICMP_UGE:
4505 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4506 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4507 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4508 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4509 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4510 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4513 case ICmpInst::ICMP_SGE:
4514 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4515 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4516 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4517 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4518 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4519 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4523 // If we still have a icmp le or icmp ge instruction, turn it into the
4524 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4525 // already been handled above, this requires little checking.
4527 if (I.getPredicate() == ICmpInst::ICMP_ULE)
4528 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4529 if (I.getPredicate() == ICmpInst::ICMP_SLE)
4530 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4531 if (I.getPredicate() == ICmpInst::ICMP_UGE)
4532 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4533 if (I.getPredicate() == ICmpInst::ICMP_SGE)
4534 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4536 // See if we can fold the comparison based on bits known to be zero or one
4538 uint64_t KnownZero, KnownOne;
4539 if (SimplifyDemandedBits(Op0, cast<IntegerType>(Ty)->getBitMask(),
4540 KnownZero, KnownOne, 0))
4543 // Given the known and unknown bits, compute a range that the LHS could be
4545 if (KnownOne | KnownZero) {
4546 // Compute the Min, Max and RHS values based on the known bits. For the
4547 // EQ and NE we use unsigned values.
4548 uint64_t UMin = 0, UMax = 0, URHSVal = 0;
4549 int64_t SMin = 0, SMax = 0, SRHSVal = 0;
4550 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4551 SRHSVal = CI->getSExtValue();
4552 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
4555 URHSVal = CI->getZExtValue();
4556 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
4559 switch (I.getPredicate()) { // LE/GE have been folded already.
4560 default: assert(0 && "Unknown icmp opcode!");
4561 case ICmpInst::ICMP_EQ:
4562 if (UMax < URHSVal || UMin > URHSVal)
4563 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4565 case ICmpInst::ICMP_NE:
4566 if (UMax < URHSVal || UMin > URHSVal)
4567 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4569 case ICmpInst::ICMP_ULT:
4571 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4573 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4575 case ICmpInst::ICMP_UGT:
4577 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4579 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4581 case ICmpInst::ICMP_SLT:
4583 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4585 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4587 case ICmpInst::ICMP_SGT:
4589 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4591 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4596 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4597 // instruction, see if that instruction also has constants so that the
4598 // instruction can be folded into the icmp
4599 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4600 switch (LHSI->getOpcode()) {
4601 case Instruction::And:
4602 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4603 LHSI->getOperand(0)->hasOneUse()) {
4604 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4606 // If the LHS is an AND of a truncating cast, we can widen the
4607 // and/compare to be the input width without changing the value
4608 // produced, eliminating a cast.
4609 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4610 // We can do this transformation if either the AND constant does not
4611 // have its sign bit set or if it is an equality comparison.
4612 // Extending a relational comparison when we're checking the sign
4613 // bit would not work.
4614 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4616 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4617 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4618 ConstantInt *NewCST;
4620 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4621 AndCST->getZExtValue());
4622 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4623 CI->getZExtValue());
4624 Instruction *NewAnd =
4625 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4627 InsertNewInstBefore(NewAnd, I);
4628 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4632 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4633 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4634 // happens a LOT in code produced by the C front-end, for bitfield
4636 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
4637 if (Shift && !Shift->isShift())
4641 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4642 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4643 const Type *AndTy = AndCST->getType(); // Type of the and.
4645 // We can fold this as long as we can't shift unknown bits
4646 // into the mask. This can only happen with signed shift
4647 // rights, as they sign-extend.
4649 bool CanFold = Shift->isLogicalShift();
4651 // To test for the bad case of the signed shr, see if any
4652 // of the bits shifted in could be tested after the mask.
4653 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4654 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4656 Constant *OShAmt = ConstantInt::get(AndTy, ShAmtVal);
4658 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4660 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4666 if (Shift->getOpcode() == Instruction::Shl)
4667 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4669 NewCst = ConstantExpr::getShl(CI, ShAmt);
4671 // Check to see if we are shifting out any of the bits being
4673 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4674 // If we shifted bits out, the fold is not going to work out.
4675 // As a special case, check to see if this means that the
4676 // result is always true or false now.
4677 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4678 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4679 if (I.getPredicate() == ICmpInst::ICMP_NE)
4680 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4682 I.setOperand(1, NewCst);
4683 Constant *NewAndCST;
4684 if (Shift->getOpcode() == Instruction::Shl)
4685 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4687 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4688 LHSI->setOperand(1, NewAndCST);
4689 LHSI->setOperand(0, Shift->getOperand(0));
4690 AddToWorkList(Shift); // Shift is dead.
4691 AddUsesToWorkList(I);
4697 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4698 // preferable because it allows the C<<Y expression to be hoisted out
4699 // of a loop if Y is invariant and X is not.
4700 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4701 I.isEquality() && !Shift->isArithmeticShift() &&
4702 isa<Instruction>(Shift->getOperand(0))) {
4705 if (Shift->getOpcode() == Instruction::LShr) {
4706 NS = BinaryOperator::createShl(AndCST,
4707 Shift->getOperand(1), "tmp");
4709 // Insert a logical shift.
4710 NS = BinaryOperator::createLShr(AndCST,
4711 Shift->getOperand(1), "tmp");
4713 InsertNewInstBefore(cast<Instruction>(NS), I);
4715 // Compute X & (C << Y).
4716 Instruction *NewAnd = BinaryOperator::createAnd(
4717 Shift->getOperand(0), NS, LHSI->getName());
4718 InsertNewInstBefore(NewAnd, I);
4720 I.setOperand(0, NewAnd);
4726 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4727 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4728 if (I.isEquality()) {
4729 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4731 // Check that the shift amount is in range. If not, don't perform
4732 // undefined shifts. When the shift is visited it will be
4734 if (ShAmt->getZExtValue() >= TypeBits)
4737 // If we are comparing against bits always shifted out, the
4738 // comparison cannot succeed.
4740 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4741 if (Comp != CI) {// Comparing against a bit that we know is zero.
4742 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4743 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4744 return ReplaceInstUsesWith(I, Cst);
4747 if (LHSI->hasOneUse()) {
4748 // Otherwise strength reduce the shift into an and.
4749 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4750 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4751 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4754 BinaryOperator::createAnd(LHSI->getOperand(0),
4755 Mask, LHSI->getName()+".mask");
4756 Value *And = InsertNewInstBefore(AndI, I);
4757 return new ICmpInst(I.getPredicate(), And,
4758 ConstantExpr::getLShr(CI, ShAmt));
4764 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4765 case Instruction::AShr:
4766 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4767 if (I.isEquality()) {
4768 // Check that the shift amount is in range. If not, don't perform
4769 // undefined shifts. When the shift is visited it will be
4771 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4772 if (ShAmt->getZExtValue() >= TypeBits)
4775 // If we are comparing against bits always shifted out, the
4776 // comparison cannot succeed.
4778 if (LHSI->getOpcode() == Instruction::LShr)
4779 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4782 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4785 if (Comp != CI) {// Comparing against a bit that we know is zero.
4786 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4787 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4788 return ReplaceInstUsesWith(I, Cst);
4791 if (LHSI->hasOneUse() || CI->isNullValue()) {
4792 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4794 // Otherwise strength reduce the shift into an and.
4795 uint64_t Val = ~0ULL; // All ones.
4796 Val <<= ShAmtVal; // Shift over to the right spot.
4797 Val &= ~0ULL >> (64-TypeBits);
4798 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4801 BinaryOperator::createAnd(LHSI->getOperand(0),
4802 Mask, LHSI->getName()+".mask");
4803 Value *And = InsertNewInstBefore(AndI, I);
4804 return new ICmpInst(I.getPredicate(), And,
4805 ConstantExpr::getShl(CI, ShAmt));
4811 case Instruction::SDiv:
4812 case Instruction::UDiv:
4813 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4814 // Fold this div into the comparison, producing a range check.
4815 // Determine, based on the divide type, what the range is being
4816 // checked. If there is an overflow on the low or high side, remember
4817 // it, otherwise compute the range [low, hi) bounding the new value.
4818 // See: InsertRangeTest above for the kinds of replacements possible.
4819 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4820 // FIXME: If the operand types don't match the type of the divide
4821 // then don't attempt this transform. The code below doesn't have the
4822 // logic to deal with a signed divide and an unsigned compare (and
4823 // vice versa). This is because (x /s C1) <s C2 produces different
4824 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4825 // (x /u C1) <u C2. Simply casting the operands and result won't
4826 // work. :( The if statement below tests that condition and bails
4828 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4829 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4832 // Initialize the variables that will indicate the nature of the
4834 bool LoOverflow = false, HiOverflow = false;
4835 ConstantInt *LoBound = 0, *HiBound = 0;
4837 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4838 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4839 // C2 (CI). By solving for X we can turn this into a range check
4840 // instead of computing a divide.
4842 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4844 // Determine if the product overflows by seeing if the product is
4845 // not equal to the divide. Make sure we do the same kind of divide
4846 // as in the LHS instruction that we're folding.
4847 bool ProdOV = !DivRHS->isNullValue() &&
4848 (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4849 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4851 // Get the ICmp opcode
4852 ICmpInst::Predicate predicate = I.getPredicate();
4854 if (DivRHS->isNullValue()) {
4855 // Don't hack on divide by zeros!
4856 } else if (!DivIsSigned) { // udiv
4858 LoOverflow = ProdOV;
4859 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4860 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4861 if (CI->isNullValue()) { // (X / pos) op 0
4863 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4865 } else if (isPositive(CI)) { // (X / pos) op pos
4867 LoOverflow = ProdOV;
4868 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4869 } else { // (X / pos) op neg
4870 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4871 LoOverflow = AddWithOverflow(LoBound, Prod,
4872 cast<ConstantInt>(DivRHSH));
4874 HiOverflow = ProdOV;
4876 } else { // Divisor is < 0.
4877 if (CI->isNullValue()) { // (X / neg) op 0
4878 LoBound = AddOne(DivRHS);
4879 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4880 if (HiBound == DivRHS)
4881 LoBound = 0; // - INTMIN = INTMIN
4882 } else if (isPositive(CI)) { // (X / neg) op pos
4883 HiOverflow = LoOverflow = ProdOV;
4885 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4886 HiBound = AddOne(Prod);
4887 } else { // (X / neg) op neg
4889 LoOverflow = HiOverflow = ProdOV;
4890 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4893 // Dividing by a negate swaps the condition.
4894 predicate = ICmpInst::getSwappedPredicate(predicate);
4898 Value *X = LHSI->getOperand(0);
4899 switch (predicate) {
4900 default: assert(0 && "Unhandled icmp opcode!");
4901 case ICmpInst::ICMP_EQ:
4902 if (LoOverflow && HiOverflow)
4903 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4904 else if (HiOverflow)
4905 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4906 ICmpInst::ICMP_UGE, X, LoBound);
4907 else if (LoOverflow)
4908 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4909 ICmpInst::ICMP_ULT, X, HiBound);
4911 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4913 case ICmpInst::ICMP_NE:
4914 if (LoOverflow && HiOverflow)
4915 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4916 else if (HiOverflow)
4917 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4918 ICmpInst::ICMP_ULT, X, LoBound);
4919 else if (LoOverflow)
4920 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4921 ICmpInst::ICMP_UGE, X, HiBound);
4923 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4925 case ICmpInst::ICMP_ULT:
4926 case ICmpInst::ICMP_SLT:
4928 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4929 return new ICmpInst(predicate, X, LoBound);
4930 case ICmpInst::ICMP_UGT:
4931 case ICmpInst::ICMP_SGT:
4933 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4934 if (predicate == ICmpInst::ICMP_UGT)
4935 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
4937 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
4944 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
4945 if (I.isEquality()) {
4946 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4948 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4949 // the second operand is a constant, simplify a bit.
4950 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4951 switch (BO->getOpcode()) {
4952 case Instruction::SRem:
4953 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4954 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4956 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4957 if (V > 1 && isPowerOf2_64(V)) {
4958 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4959 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4960 return new ICmpInst(I.getPredicate(), NewRem,
4961 Constant::getNullValue(BO->getType()));
4965 case Instruction::Add:
4966 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4967 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4968 if (BO->hasOneUse())
4969 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4970 ConstantExpr::getSub(CI, BOp1C));
4971 } else if (CI->isNullValue()) {
4972 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4973 // efficiently invertible, or if the add has just this one use.
4974 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4976 if (Value *NegVal = dyn_castNegVal(BOp1))
4977 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
4978 else if (Value *NegVal = dyn_castNegVal(BOp0))
4979 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
4980 else if (BO->hasOneUse()) {
4981 Instruction *Neg = BinaryOperator::createNeg(BOp1);
4982 InsertNewInstBefore(Neg, I);
4984 return new ICmpInst(I.getPredicate(), BOp0, Neg);
4988 case Instruction::Xor:
4989 // For the xor case, we can xor two constants together, eliminating
4990 // the explicit xor.
4991 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4992 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4993 ConstantExpr::getXor(CI, BOC));
4996 case Instruction::Sub:
4997 // Replace (([sub|xor] A, B) != 0) with (A != B)
4998 if (CI->isNullValue())
4999 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5003 case Instruction::Or:
5004 // If bits are being or'd in that are not present in the constant we
5005 // are comparing against, then the comparison could never succeed!
5006 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5007 Constant *NotCI = ConstantExpr::getNot(CI);
5008 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5009 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5014 case Instruction::And:
5015 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5016 // If bits are being compared against that are and'd out, then the
5017 // comparison can never succeed!
5018 if (!ConstantExpr::getAnd(CI,
5019 ConstantExpr::getNot(BOC))->isNullValue())
5020 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5023 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5024 if (CI == BOC && isOneBitSet(CI))
5025 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5026 ICmpInst::ICMP_NE, Op0,
5027 Constant::getNullValue(CI->getType()));
5029 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5030 if (isSignBit(BOC)) {
5031 Value *X = BO->getOperand(0);
5032 Constant *Zero = Constant::getNullValue(X->getType());
5033 ICmpInst::Predicate pred = isICMP_NE ?
5034 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5035 return new ICmpInst(pred, X, Zero);
5038 // ((X & ~7) == 0) --> X < 8
5039 if (CI->isNullValue() && isHighOnes(BOC)) {
5040 Value *X = BO->getOperand(0);
5041 Constant *NegX = ConstantExpr::getNeg(BOC);
5042 ICmpInst::Predicate pred = isICMP_NE ?
5043 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5044 return new ICmpInst(pred, X, NegX);
5050 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5051 // Handle set{eq|ne} <intrinsic>, intcst.
5052 switch (II->getIntrinsicID()) {
5054 case Intrinsic::bswap_i16:
5055 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5056 AddToWorkList(II); // Dead?
5057 I.setOperand(0, II->getOperand(1));
5058 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5059 ByteSwap_16(CI->getZExtValue())));
5061 case Intrinsic::bswap_i32:
5062 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5063 AddToWorkList(II); // Dead?
5064 I.setOperand(0, II->getOperand(1));
5065 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5066 ByteSwap_32(CI->getZExtValue())));
5068 case Intrinsic::bswap_i64:
5069 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5070 AddToWorkList(II); // Dead?
5071 I.setOperand(0, II->getOperand(1));
5072 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5073 ByteSwap_64(CI->getZExtValue())));
5077 } else { // Not a ICMP_EQ/ICMP_NE
5078 // If the LHS is a cast from an integral value of the same size, then
5079 // since we know the RHS is a constant, try to simlify.
5080 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5081 Value *CastOp = Cast->getOperand(0);
5082 const Type *SrcTy = CastOp->getType();
5083 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5084 if (SrcTy->isInteger() &&
5085 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5086 // If this is an unsigned comparison, try to make the comparison use
5087 // smaller constant values.
5088 switch (I.getPredicate()) {
5090 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5091 ConstantInt *CUI = cast<ConstantInt>(CI);
5092 if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
5093 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5094 ConstantInt::get(SrcTy, -1ULL));
5097 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5098 ConstantInt *CUI = cast<ConstantInt>(CI);
5099 if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
5100 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5101 Constant::getNullValue(SrcTy));
5111 // Handle icmp with constant RHS
5112 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5113 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5114 switch (LHSI->getOpcode()) {
5115 case Instruction::GetElementPtr:
5116 if (RHSC->isNullValue()) {
5117 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5118 bool isAllZeros = true;
5119 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5120 if (!isa<Constant>(LHSI->getOperand(i)) ||
5121 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5126 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5127 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5131 case Instruction::PHI:
5132 if (Instruction *NV = FoldOpIntoPhi(I))
5135 case Instruction::Select:
5136 // If either operand of the select is a constant, we can fold the
5137 // comparison into the select arms, which will cause one to be
5138 // constant folded and the select turned into a bitwise or.
5139 Value *Op1 = 0, *Op2 = 0;
5140 if (LHSI->hasOneUse()) {
5141 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5142 // Fold the known value into the constant operand.
5143 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5144 // Insert a new ICmp of the other select operand.
5145 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5146 LHSI->getOperand(2), RHSC,
5148 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5149 // Fold the known value into the constant operand.
5150 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5151 // Insert a new ICmp of the other select operand.
5152 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5153 LHSI->getOperand(1), RHSC,
5159 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5164 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5165 if (User *GEP = dyn_castGetElementPtr(Op0))
5166 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5168 if (User *GEP = dyn_castGetElementPtr(Op1))
5169 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5170 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5173 // Test to see if the operands of the icmp are casted versions of other
5174 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5176 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5177 if (isa<PointerType>(Op0->getType()) &&
5178 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5179 // We keep moving the cast from the left operand over to the right
5180 // operand, where it can often be eliminated completely.
5181 Op0 = CI->getOperand(0);
5183 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5184 // so eliminate it as well.
5185 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5186 Op1 = CI2->getOperand(0);
5188 // If Op1 is a constant, we can fold the cast into the constant.
5189 if (Op0->getType() != Op1->getType())
5190 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5191 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5193 // Otherwise, cast the RHS right before the icmp
5194 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5196 return new ICmpInst(I.getPredicate(), Op0, Op1);
5200 if (isa<CastInst>(Op0)) {
5201 // Handle the special case of: icmp (cast bool to X), <cst>
5202 // This comes up when you have code like
5205 // For generality, we handle any zero-extension of any operand comparison
5206 // with a constant or another cast from the same type.
5207 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5208 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5212 if (I.isEquality()) {
5213 Value *A, *B, *C, *D;
5214 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5215 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5216 Value *OtherVal = A == Op1 ? B : A;
5217 return new ICmpInst(I.getPredicate(), OtherVal,
5218 Constant::getNullValue(A->getType()));
5221 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5222 // A^c1 == C^c2 --> A == C^(c1^c2)
5223 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5224 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5225 if (Op1->hasOneUse()) {
5226 Constant *NC = ConstantExpr::getXor(C1, C2);
5227 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5228 return new ICmpInst(I.getPredicate(), A,
5229 InsertNewInstBefore(Xor, I));
5232 // A^B == A^D -> B == D
5233 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5234 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5235 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5236 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5240 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5241 (A == Op0 || B == Op0)) {
5242 // A == (A^B) -> B == 0
5243 Value *OtherVal = A == Op0 ? B : A;
5244 return new ICmpInst(I.getPredicate(), OtherVal,
5245 Constant::getNullValue(A->getType()));
5247 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5248 // (A-B) == A -> B == 0
5249 return new ICmpInst(I.getPredicate(), B,
5250 Constant::getNullValue(B->getType()));
5252 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5253 // A == (A-B) -> B == 0
5254 return new ICmpInst(I.getPredicate(), B,
5255 Constant::getNullValue(B->getType()));
5258 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5259 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5260 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5261 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5262 Value *X = 0, *Y = 0, *Z = 0;
5265 X = B; Y = D; Z = A;
5266 } else if (A == D) {
5267 X = B; Y = C; Z = A;
5268 } else if (B == C) {
5269 X = A; Y = D; Z = B;
5270 } else if (B == D) {
5271 X = A; Y = C; Z = B;
5274 if (X) { // Build (X^Y) & Z
5275 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5276 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5277 I.setOperand(0, Op1);
5278 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5283 return Changed ? &I : 0;
5286 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5287 // We only handle extending casts so far.
5289 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5290 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5291 Value *LHSCIOp = LHSCI->getOperand(0);
5292 const Type *SrcTy = LHSCIOp->getType();
5293 const Type *DestTy = LHSCI->getType();
5296 // We only handle extension cast instructions, so far. Enforce this.
5297 if (LHSCI->getOpcode() != Instruction::ZExt &&
5298 LHSCI->getOpcode() != Instruction::SExt)
5301 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5302 bool isSignedCmp = ICI.isSignedPredicate();
5304 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5305 // Not an extension from the same type?
5306 RHSCIOp = CI->getOperand(0);
5307 if (RHSCIOp->getType() != LHSCIOp->getType())
5310 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5311 // and the other is a zext), then we can't handle this.
5312 if (CI->getOpcode() != LHSCI->getOpcode())
5315 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5316 // then we can't handle this.
5317 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5320 // Okay, just insert a compare of the reduced operands now!
5321 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5324 // If we aren't dealing with a constant on the RHS, exit early
5325 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5329 // Compute the constant that would happen if we truncated to SrcTy then
5330 // reextended to DestTy.
5331 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5332 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5334 // If the re-extended constant didn't change...
5336 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5337 // For example, we might have:
5338 // %A = sext short %X to uint
5339 // %B = icmp ugt uint %A, 1330
5340 // It is incorrect to transform this into
5341 // %B = icmp ugt short %X, 1330
5342 // because %A may have negative value.
5344 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5345 // OR operation is EQ/NE.
5346 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5347 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5352 // The re-extended constant changed so the constant cannot be represented
5353 // in the shorter type. Consequently, we cannot emit a simple comparison.
5355 // First, handle some easy cases. We know the result cannot be equal at this
5356 // point so handle the ICI.isEquality() cases
5357 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5358 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5359 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5360 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5362 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5363 // should have been folded away previously and not enter in here.
5366 // We're performing a signed comparison.
5367 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5368 Result = ConstantInt::getFalse(); // X < (small) --> false
5370 Result = ConstantInt::getTrue(); // X < (large) --> true
5372 // We're performing an unsigned comparison.
5374 // We're performing an unsigned comp with a sign extended value.
5375 // This is true if the input is >= 0. [aka >s -1]
5376 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5377 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5378 NegOne, ICI.getName()), ICI);
5380 // Unsigned extend & unsigned compare -> always true.
5381 Result = ConstantInt::getTrue();
5385 // Finally, return the value computed.
5386 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5387 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5388 return ReplaceInstUsesWith(ICI, Result);
5390 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5391 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5392 "ICmp should be folded!");
5393 if (Constant *CI = dyn_cast<Constant>(Result))
5394 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5396 return BinaryOperator::createNot(Result);
5400 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5401 return commonShiftTransforms(I);
5404 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5405 return commonShiftTransforms(I);
5408 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5409 return commonShiftTransforms(I);
5412 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5413 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5414 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5416 // shl X, 0 == X and shr X, 0 == X
5417 // shl 0, X == 0 and shr 0, X == 0
5418 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5419 Op0 == Constant::getNullValue(Op0->getType()))
5420 return ReplaceInstUsesWith(I, Op0);
5422 if (isa<UndefValue>(Op0)) {
5423 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5424 return ReplaceInstUsesWith(I, Op0);
5425 else // undef << X -> 0, undef >>u X -> 0
5426 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5428 if (isa<UndefValue>(Op1)) {
5429 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5430 return ReplaceInstUsesWith(I, Op0);
5431 else // X << undef, X >>u undef -> 0
5432 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5435 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5436 if (I.getOpcode() == Instruction::AShr)
5437 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5438 if (CSI->isAllOnesValue())
5439 return ReplaceInstUsesWith(I, CSI);
5441 // Try to fold constant and into select arguments.
5442 if (isa<Constant>(Op0))
5443 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5444 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5447 // See if we can turn a signed shr into an unsigned shr.
5448 if (I.isArithmeticShift()) {
5449 if (MaskedValueIsZero(Op0,
5450 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5451 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5455 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5456 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5461 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5462 BinaryOperator &I) {
5463 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5465 // See if we can simplify any instructions used by the instruction whose sole
5466 // purpose is to compute bits we don't care about.
5467 uint64_t KnownZero, KnownOne;
5468 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
5469 KnownZero, KnownOne))
5472 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5473 // of a signed value.
5475 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5476 if (Op1->getZExtValue() >= TypeBits) {
5477 if (I.getOpcode() != Instruction::AShr)
5478 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5480 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5485 // ((X*C1) << C2) == (X * (C1 << C2))
5486 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5487 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5488 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5489 return BinaryOperator::createMul(BO->getOperand(0),
5490 ConstantExpr::getShl(BOOp, Op1));
5492 // Try to fold constant and into select arguments.
5493 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5494 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5496 if (isa<PHINode>(Op0))
5497 if (Instruction *NV = FoldOpIntoPhi(I))
5500 if (Op0->hasOneUse()) {
5501 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5502 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5505 switch (Op0BO->getOpcode()) {
5507 case Instruction::Add:
5508 case Instruction::And:
5509 case Instruction::Or:
5510 case Instruction::Xor: {
5511 // These operators commute.
5512 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5513 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5514 match(Op0BO->getOperand(1),
5515 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5516 Instruction *YS = BinaryOperator::createShl(
5517 Op0BO->getOperand(0), Op1,
5519 InsertNewInstBefore(YS, I); // (Y << C)
5521 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5522 Op0BO->getOperand(1)->getName());
5523 InsertNewInstBefore(X, I); // (X + (Y << C))
5524 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5525 C2 = ConstantExpr::getShl(C2, Op1);
5526 return BinaryOperator::createAnd(X, C2);
5529 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5530 Value *Op0BOOp1 = Op0BO->getOperand(1);
5531 if (isLeftShift && Op0BOOp1->hasOneUse() && V2 == Op1 &&
5533 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5534 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)-> hasOneUse()) {
5535 Instruction *YS = BinaryOperator::createShl(
5536 Op0BO->getOperand(0), Op1,
5538 InsertNewInstBefore(YS, I); // (Y << C)
5540 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5541 V1->getName()+".mask");
5542 InsertNewInstBefore(XM, I); // X & (CC << C)
5544 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5549 case Instruction::Sub: {
5550 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5551 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5552 match(Op0BO->getOperand(0),
5553 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5554 Instruction *YS = BinaryOperator::createShl(
5555 Op0BO->getOperand(1), Op1,
5557 InsertNewInstBefore(YS, I); // (Y << C)
5559 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5560 Op0BO->getOperand(0)->getName());
5561 InsertNewInstBefore(X, I); // (X + (Y << C))
5562 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5563 C2 = ConstantExpr::getShl(C2, Op1);
5564 return BinaryOperator::createAnd(X, C2);
5567 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5568 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5569 match(Op0BO->getOperand(0),
5570 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5571 m_ConstantInt(CC))) && V2 == Op1 &&
5572 cast<BinaryOperator>(Op0BO->getOperand(0))
5573 ->getOperand(0)->hasOneUse()) {
5574 Instruction *YS = BinaryOperator::createShl(
5575 Op0BO->getOperand(1), Op1,
5577 InsertNewInstBefore(YS, I); // (Y << C)
5579 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5580 V1->getName()+".mask");
5581 InsertNewInstBefore(XM, I); // X & (CC << C)
5583 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5591 // If the operand is an bitwise operator with a constant RHS, and the
5592 // shift is the only use, we can pull it out of the shift.
5593 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5594 bool isValid = true; // Valid only for And, Or, Xor
5595 bool highBitSet = false; // Transform if high bit of constant set?
5597 switch (Op0BO->getOpcode()) {
5598 default: isValid = false; break; // Do not perform transform!
5599 case Instruction::Add:
5600 isValid = isLeftShift;
5602 case Instruction::Or:
5603 case Instruction::Xor:
5606 case Instruction::And:
5611 // If this is a signed shift right, and the high bit is modified
5612 // by the logical operation, do not perform the transformation.
5613 // The highBitSet boolean indicates the value of the high bit of
5614 // the constant which would cause it to be modified for this
5617 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
5618 uint64_t Val = Op0C->getZExtValue();
5619 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5623 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5625 Instruction *NewShift =
5626 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
5627 InsertNewInstBefore(NewShift, I);
5628 NewShift->takeName(Op0BO);
5630 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5637 // Find out if this is a shift of a shift by a constant.
5638 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
5639 if (ShiftOp && !ShiftOp->isShift())
5642 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5643 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5644 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5645 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5646 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
5647 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
5648 Value *X = ShiftOp->getOperand(0);
5650 unsigned AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5651 if (AmtSum > I.getType()->getPrimitiveSizeInBits())
5652 AmtSum = I.getType()->getPrimitiveSizeInBits();
5654 const IntegerType *Ty = cast<IntegerType>(I.getType());
5656 // Check for (X << c1) << c2 and (X >> c1) >> c2
5657 if (I.getOpcode() == ShiftOp->getOpcode()) {
5658 return BinaryOperator::create(I.getOpcode(), X,
5659 ConstantInt::get(Ty, AmtSum));
5660 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
5661 I.getOpcode() == Instruction::AShr) {
5662 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
5663 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
5664 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
5665 I.getOpcode() == Instruction::LShr) {
5666 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
5667 Instruction *Shift =
5668 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
5669 InsertNewInstBefore(Shift, I);
5671 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5672 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5675 // Okay, if we get here, one shift must be left, and the other shift must be
5676 // right. See if the amounts are equal.
5677 if (ShiftAmt1 == ShiftAmt2) {
5678 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
5679 if (I.getOpcode() == Instruction::Shl) {
5680 uint64_t Mask = Ty->getBitMask() << ShiftAmt1;
5681 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
5683 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
5684 if (I.getOpcode() == Instruction::LShr) {
5685 uint64_t Mask = Ty->getBitMask() >> ShiftAmt1;
5686 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
5688 // We can simplify ((X << C) >>s C) into a trunc + sext.
5689 // NOTE: we could do this for any C, but that would make 'unusual' integer
5690 // types. For now, just stick to ones well-supported by the code
5692 const Type *SExtType = 0;
5693 switch (Ty->getBitWidth() - ShiftAmt1) {
5694 case 8 : SExtType = Type::Int8Ty; break;
5695 case 16: SExtType = Type::Int16Ty; break;
5696 case 32: SExtType = Type::Int32Ty; break;
5700 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
5701 InsertNewInstBefore(NewTrunc, I);
5702 return new SExtInst(NewTrunc, Ty);
5704 // Otherwise, we can't handle it yet.
5705 } else if (ShiftAmt1 < ShiftAmt2) {
5706 unsigned ShiftDiff = ShiftAmt2-ShiftAmt1;
5708 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
5709 if (I.getOpcode() == Instruction::Shl) {
5710 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5711 ShiftOp->getOpcode() == Instruction::AShr);
5712 Instruction *Shift =
5713 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5714 InsertNewInstBefore(Shift, I);
5716 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
5717 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5720 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
5721 if (I.getOpcode() == Instruction::LShr) {
5722 assert(ShiftOp->getOpcode() == Instruction::Shl);
5723 Instruction *Shift =
5724 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
5725 InsertNewInstBefore(Shift, I);
5727 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5728 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5731 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
5733 assert(ShiftAmt2 < ShiftAmt1);
5734 unsigned ShiftDiff = ShiftAmt1-ShiftAmt2;
5736 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
5737 if (I.getOpcode() == Instruction::Shl) {
5738 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5739 ShiftOp->getOpcode() == Instruction::AShr);
5740 Instruction *Shift =
5741 BinaryOperator::create(ShiftOp->getOpcode(), X,
5742 ConstantInt::get(Ty, ShiftDiff));
5743 InsertNewInstBefore(Shift, I);
5745 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
5746 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5749 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
5750 if (I.getOpcode() == Instruction::LShr) {
5751 assert(ShiftOp->getOpcode() == Instruction::Shl);
5752 Instruction *Shift =
5753 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5754 InsertNewInstBefore(Shift, I);
5756 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5757 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5760 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
5767 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5768 /// expression. If so, decompose it, returning some value X, such that Val is
5771 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5773 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5774 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5775 Offset = CI->getZExtValue();
5777 return ConstantInt::get(Type::Int32Ty, 0);
5778 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5779 if (I->getNumOperands() == 2) {
5780 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5781 if (I->getOpcode() == Instruction::Shl) {
5782 // This is a value scaled by '1 << the shift amt'.
5783 Scale = 1U << CUI->getZExtValue();
5785 return I->getOperand(0);
5786 } else if (I->getOpcode() == Instruction::Mul) {
5787 // This value is scaled by 'CUI'.
5788 Scale = CUI->getZExtValue();
5790 return I->getOperand(0);
5791 } else if (I->getOpcode() == Instruction::Add) {
5792 // We have X+C. Check to see if we really have (X*C2)+C1,
5793 // where C1 is divisible by C2.
5796 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5797 Offset += CUI->getZExtValue();
5798 if (SubScale > 1 && (Offset % SubScale == 0)) {
5807 // Otherwise, we can't look past this.
5814 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5815 /// try to eliminate the cast by moving the type information into the alloc.
5816 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5817 AllocationInst &AI) {
5818 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5819 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5821 // Remove any uses of AI that are dead.
5822 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5824 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5825 Instruction *User = cast<Instruction>(*UI++);
5826 if (isInstructionTriviallyDead(User)) {
5827 while (UI != E && *UI == User)
5828 ++UI; // If this instruction uses AI more than once, don't break UI.
5831 DOUT << "IC: DCE: " << *User;
5832 EraseInstFromFunction(*User);
5836 // Get the type really allocated and the type casted to.
5837 const Type *AllocElTy = AI.getAllocatedType();
5838 const Type *CastElTy = PTy->getElementType();
5839 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5841 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
5842 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
5843 if (CastElTyAlign < AllocElTyAlign) return 0;
5845 // If the allocation has multiple uses, only promote it if we are strictly
5846 // increasing the alignment of the resultant allocation. If we keep it the
5847 // same, we open the door to infinite loops of various kinds.
5848 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5850 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5851 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5852 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5854 // See if we can satisfy the modulus by pulling a scale out of the array
5856 unsigned ArraySizeScale, ArrayOffset;
5857 Value *NumElements = // See if the array size is a decomposable linear expr.
5858 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5860 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5862 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5863 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5865 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5870 // If the allocation size is constant, form a constant mul expression
5871 Amt = ConstantInt::get(Type::Int32Ty, Scale);
5872 if (isa<ConstantInt>(NumElements))
5873 Amt = ConstantExpr::getMul(
5874 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5875 // otherwise multiply the amount and the number of elements
5876 else if (Scale != 1) {
5877 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5878 Amt = InsertNewInstBefore(Tmp, AI);
5882 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5883 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
5884 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5885 Amt = InsertNewInstBefore(Tmp, AI);
5888 AllocationInst *New;
5889 if (isa<MallocInst>(AI))
5890 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
5892 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
5893 InsertNewInstBefore(New, AI);
5896 // If the allocation has multiple uses, insert a cast and change all things
5897 // that used it to use the new cast. This will also hack on CI, but it will
5899 if (!AI.hasOneUse()) {
5900 AddUsesToWorkList(AI);
5901 // New is the allocation instruction, pointer typed. AI is the original
5902 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5903 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5904 InsertNewInstBefore(NewCast, AI);
5905 AI.replaceAllUsesWith(NewCast);
5907 return ReplaceInstUsesWith(CI, New);
5910 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5911 /// and return it without inserting any new casts. This is used by code that
5912 /// tries to decide whether promoting or shrinking integer operations to wider
5913 /// or smaller types will allow us to eliminate a truncate or extend.
5914 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5915 int &NumCastsRemoved) {
5916 if (isa<Constant>(V)) return true;
5918 Instruction *I = dyn_cast<Instruction>(V);
5919 if (!I || !I->hasOneUse()) return false;
5921 switch (I->getOpcode()) {
5922 case Instruction::And:
5923 case Instruction::Or:
5924 case Instruction::Xor:
5925 // These operators can all arbitrarily be extended or truncated.
5926 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5927 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5928 case Instruction::AShr:
5929 case Instruction::LShr:
5930 case Instruction::Shl:
5931 // If this is just a bitcast changing the sign of the operation, we can
5932 // convert if the operand can be converted.
5933 if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
5934 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5936 case Instruction::Trunc:
5937 case Instruction::ZExt:
5938 case Instruction::SExt:
5939 case Instruction::BitCast:
5940 // If this is a cast from the destination type, we can trivially eliminate
5941 // it, and this will remove a cast overall.
5942 if (I->getOperand(0)->getType() == Ty) {
5943 // If the first operand is itself a cast, and is eliminable, do not count
5944 // this as an eliminable cast. We would prefer to eliminate those two
5946 if (isa<CastInst>(I->getOperand(0)))
5954 // TODO: Can handle more cases here.
5961 /// EvaluateInDifferentType - Given an expression that
5962 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5963 /// evaluate the expression.
5964 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
5966 if (Constant *C = dyn_cast<Constant>(V))
5967 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
5969 // Otherwise, it must be an instruction.
5970 Instruction *I = cast<Instruction>(V);
5971 Instruction *Res = 0;
5972 switch (I->getOpcode()) {
5973 case Instruction::And:
5974 case Instruction::Or:
5975 case Instruction::Xor: {
5976 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5977 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
5978 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5979 LHS, RHS, I->getName());
5982 case Instruction::AShr:
5983 case Instruction::LShr:
5984 case Instruction::Shl: {
5985 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5986 Res = BinaryOperator::create(Instruction::BinaryOps(I->getOpcode()), LHS,
5987 I->getOperand(1), I->getName());
5990 case Instruction::Trunc:
5991 case Instruction::ZExt:
5992 case Instruction::SExt:
5993 case Instruction::BitCast:
5994 // If the source type of the cast is the type we're trying for then we can
5995 // just return the source. There's no need to insert it because its not new.
5996 if (I->getOperand(0)->getType() == Ty)
5997 return I->getOperand(0);
5999 // Some other kind of cast, which shouldn't happen, so just ..
6002 // TODO: Can handle more cases here.
6003 assert(0 && "Unreachable!");
6007 return InsertNewInstBefore(Res, *I);
6010 /// @brief Implement the transforms common to all CastInst visitors.
6011 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6012 Value *Src = CI.getOperand(0);
6014 // Casting undef to anything results in undef so might as just replace it and
6015 // get rid of the cast.
6016 if (isa<UndefValue>(Src)) // cast undef -> undef
6017 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
6019 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
6020 // eliminate it now.
6021 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6022 if (Instruction::CastOps opc =
6023 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6024 // The first cast (CSrc) is eliminable so we need to fix up or replace
6025 // the second cast (CI). CSrc will then have a good chance of being dead.
6026 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6030 // If casting the result of a getelementptr instruction with no offset, turn
6031 // this into a cast of the original pointer!
6033 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6034 bool AllZeroOperands = true;
6035 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
6036 if (!isa<Constant>(GEP->getOperand(i)) ||
6037 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
6038 AllZeroOperands = false;
6041 if (AllZeroOperands) {
6042 // Changing the cast operand is usually not a good idea but it is safe
6043 // here because the pointer operand is being replaced with another
6044 // pointer operand so the opcode doesn't need to change.
6045 CI.setOperand(0, GEP->getOperand(0));
6050 // If we are casting a malloc or alloca to a pointer to a type of the same
6051 // size, rewrite the allocation instruction to allocate the "right" type.
6052 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6053 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6056 // If we are casting a select then fold the cast into the select
6057 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6058 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6061 // If we are casting a PHI then fold the cast into the PHI
6062 if (isa<PHINode>(Src))
6063 if (Instruction *NV = FoldOpIntoPhi(CI))
6069 /// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
6070 /// integers. This function implements the common transforms for all those
6072 /// @brief Implement the transforms common to CastInst with integer operands
6073 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6074 if (Instruction *Result = commonCastTransforms(CI))
6077 Value *Src = CI.getOperand(0);
6078 const Type *SrcTy = Src->getType();
6079 const Type *DestTy = CI.getType();
6080 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6081 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6083 // See if we can simplify any instructions used by the LHS whose sole
6084 // purpose is to compute bits we don't care about.
6085 uint64_t KnownZero = 0, KnownOne = 0;
6086 if (SimplifyDemandedBits(&CI, cast<IntegerType>(DestTy)->getBitMask(),
6087 KnownZero, KnownOne))
6090 // If the source isn't an instruction or has more than one use then we
6091 // can't do anything more.
6092 Instruction *SrcI = dyn_cast<Instruction>(Src);
6093 if (!SrcI || !Src->hasOneUse())
6096 // Attempt to propagate the cast into the instruction.
6097 int NumCastsRemoved = 0;
6098 if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
6099 // If this cast is a truncate, evaluting in a different type always
6100 // eliminates the cast, so it is always a win. If this is a noop-cast
6101 // this just removes a noop cast which isn't pointful, but simplifies
6102 // the code. If this is a zero-extension, we need to do an AND to
6103 // maintain the clear top-part of the computation, so we require that
6104 // the input have eliminated at least one cast. If this is a sign
6105 // extension, we insert two new casts (to do the extension) so we
6106 // require that two casts have been eliminated.
6107 bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
6109 switch (CI.getOpcode()) {
6110 case Instruction::Trunc:
6113 case Instruction::ZExt:
6114 DoXForm = NumCastsRemoved >= 1;
6116 case Instruction::SExt:
6117 DoXForm = NumCastsRemoved >= 2;
6119 case Instruction::BitCast:
6123 // All the others use floating point so we shouldn't actually
6124 // get here because of the check above.
6125 assert(!"Unknown cast type .. unreachable");
6131 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6132 CI.getOpcode() == Instruction::SExt);
6133 assert(Res->getType() == DestTy);
6134 switch (CI.getOpcode()) {
6135 default: assert(0 && "Unknown cast type!");
6136 case Instruction::Trunc:
6137 case Instruction::BitCast:
6138 // Just replace this cast with the result.
6139 return ReplaceInstUsesWith(CI, Res);
6140 case Instruction::ZExt: {
6141 // We need to emit an AND to clear the high bits.
6142 assert(SrcBitSize < DestBitSize && "Not a zext?");
6144 ConstantInt::get(Type::Int64Ty, (1ULL << SrcBitSize)-1);
6145 if (DestBitSize < 64)
6146 C = ConstantExpr::getTrunc(C, DestTy);
6147 return BinaryOperator::createAnd(Res, C);
6149 case Instruction::SExt:
6150 // We need to emit a cast to truncate, then a cast to sext.
6151 return CastInst::create(Instruction::SExt,
6152 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6158 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6159 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6161 switch (SrcI->getOpcode()) {
6162 case Instruction::Add:
6163 case Instruction::Mul:
6164 case Instruction::And:
6165 case Instruction::Or:
6166 case Instruction::Xor:
6167 // If we are discarding information, or just changing the sign,
6169 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6170 // Don't insert two casts if they cannot be eliminated. We allow
6171 // two casts to be inserted if the sizes are the same. This could
6172 // only be converting signedness, which is a noop.
6173 if (DestBitSize == SrcBitSize ||
6174 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6175 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6176 Instruction::CastOps opcode = CI.getOpcode();
6177 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6178 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6179 return BinaryOperator::create(
6180 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6184 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6185 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6186 SrcI->getOpcode() == Instruction::Xor &&
6187 Op1 == ConstantInt::getTrue() &&
6188 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6189 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6190 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6193 case Instruction::SDiv:
6194 case Instruction::UDiv:
6195 case Instruction::SRem:
6196 case Instruction::URem:
6197 // If we are just changing the sign, rewrite.
6198 if (DestBitSize == SrcBitSize) {
6199 // Don't insert two casts if they cannot be eliminated. We allow
6200 // two casts to be inserted if the sizes are the same. This could
6201 // only be converting signedness, which is a noop.
6202 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6203 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6204 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6206 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6208 return BinaryOperator::create(
6209 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6214 case Instruction::Shl:
6215 // Allow changing the sign of the source operand. Do not allow
6216 // changing the size of the shift, UNLESS the shift amount is a
6217 // constant. We must not change variable sized shifts to a smaller
6218 // size, because it is undefined to shift more bits out than exist
6220 if (DestBitSize == SrcBitSize ||
6221 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6222 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6223 Instruction::BitCast : Instruction::Trunc);
6224 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6225 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6226 return BinaryOperator::createShl(Op0c, Op1c);
6229 case Instruction::AShr:
6230 // If this is a signed shr, and if all bits shifted in are about to be
6231 // truncated off, turn it into an unsigned shr to allow greater
6233 if (DestBitSize < SrcBitSize &&
6234 isa<ConstantInt>(Op1)) {
6235 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6236 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6237 // Insert the new logical shift right.
6238 return BinaryOperator::createLShr(Op0, Op1);
6243 case Instruction::ICmp:
6244 // If we are just checking for a icmp eq of a single bit and casting it
6245 // to an integer, then shift the bit to the appropriate place and then
6246 // cast to integer to avoid the comparison.
6247 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6248 uint64_t Op1CV = Op1C->getZExtValue();
6249 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6250 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6251 // cast (X == 1) to int --> X iff X has only the low bit set.
6252 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6253 // cast (X != 0) to int --> X iff X has only the low bit set.
6254 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6255 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6256 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6257 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
6258 // If Op1C some other power of two, convert:
6259 uint64_t KnownZero, KnownOne;
6260 uint64_t TypeMask = Op1C->getType()->getBitMask();
6261 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6263 // This only works for EQ and NE
6264 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6265 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6268 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
6269 bool isNE = pred == ICmpInst::ICMP_NE;
6270 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
6271 // (X&4) == 2 --> false
6272 // (X&4) != 2 --> true
6273 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6274 Res = ConstantExpr::getZExt(Res, CI.getType());
6275 return ReplaceInstUsesWith(CI, Res);
6278 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
6281 // Perform a logical shr by shiftamt.
6282 // Insert the shift to put the result in the low bit.
6283 In = InsertNewInstBefore(
6284 BinaryOperator::createLShr(In,
6285 ConstantInt::get(In->getType(), ShiftAmt),
6286 In->getName()+".lobit"), CI);
6289 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6290 Constant *One = ConstantInt::get(In->getType(), 1);
6291 In = BinaryOperator::createXor(In, One, "tmp");
6292 InsertNewInstBefore(cast<Instruction>(In), CI);
6295 if (CI.getType() == In->getType())
6296 return ReplaceInstUsesWith(CI, In);
6298 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6307 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6308 if (Instruction *Result = commonIntCastTransforms(CI))
6311 Value *Src = CI.getOperand(0);
6312 const Type *Ty = CI.getType();
6313 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6315 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6316 switch (SrcI->getOpcode()) {
6318 case Instruction::LShr:
6319 // We can shrink lshr to something smaller if we know the bits shifted in
6320 // are already zeros.
6321 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6322 unsigned ShAmt = ShAmtV->getZExtValue();
6324 // Get a mask for the bits shifting in.
6325 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
6326 Value* SrcIOp0 = SrcI->getOperand(0);
6327 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6328 if (ShAmt >= DestBitWidth) // All zeros.
6329 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6331 // Okay, we can shrink this. Truncate the input, then return a new
6333 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6334 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6336 return BinaryOperator::createLShr(V1, V2);
6338 } else { // This is a variable shr.
6340 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6341 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6342 // loop-invariant and CSE'd.
6343 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6344 Value *One = ConstantInt::get(SrcI->getType(), 1);
6346 Value *V = InsertNewInstBefore(
6347 BinaryOperator::createShl(One, SrcI->getOperand(1),
6349 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6350 SrcI->getOperand(0),
6352 Value *Zero = Constant::getNullValue(V->getType());
6353 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6363 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6364 // If one of the common conversion will work ..
6365 if (Instruction *Result = commonIntCastTransforms(CI))
6368 Value *Src = CI.getOperand(0);
6370 // If this is a cast of a cast
6371 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6372 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6373 // types and if the sizes are just right we can convert this into a logical
6374 // 'and' which will be much cheaper than the pair of casts.
6375 if (isa<TruncInst>(CSrc)) {
6376 // Get the sizes of the types involved
6377 Value *A = CSrc->getOperand(0);
6378 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6379 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6380 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6381 // If we're actually extending zero bits and the trunc is a no-op
6382 if (MidSize < DstSize && SrcSize == DstSize) {
6383 // Replace both of the casts with an And of the type mask.
6384 uint64_t AndValue = cast<IntegerType>(CSrc->getType())->getBitMask();
6385 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6387 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6388 // Unfortunately, if the type changed, we need to cast it back.
6389 if (And->getType() != CI.getType()) {
6390 And->setName(CSrc->getName()+".mask");
6391 InsertNewInstBefore(And, CI);
6392 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6402 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6403 return commonIntCastTransforms(CI);
6406 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6407 return commonCastTransforms(CI);
6410 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6411 return commonCastTransforms(CI);
6414 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6415 return commonCastTransforms(CI);
6418 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6419 return commonCastTransforms(CI);
6422 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6423 return commonCastTransforms(CI);
6426 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6427 return commonCastTransforms(CI);
6430 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6431 return commonCastTransforms(CI);
6434 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6435 return commonCastTransforms(CI);
6438 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6440 // If the operands are integer typed then apply the integer transforms,
6441 // otherwise just apply the common ones.
6442 Value *Src = CI.getOperand(0);
6443 const Type *SrcTy = Src->getType();
6444 const Type *DestTy = CI.getType();
6446 if (SrcTy->isInteger() && DestTy->isInteger()) {
6447 if (Instruction *Result = commonIntCastTransforms(CI))
6450 if (Instruction *Result = commonCastTransforms(CI))
6455 // Get rid of casts from one type to the same type. These are useless and can
6456 // be replaced by the operand.
6457 if (DestTy == Src->getType())
6458 return ReplaceInstUsesWith(CI, Src);
6460 // If the source and destination are pointers, and this cast is equivalent to
6461 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6462 // This can enhance SROA and other transforms that want type-safe pointers.
6463 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6464 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6465 const Type *DstElTy = DstPTy->getElementType();
6466 const Type *SrcElTy = SrcPTy->getElementType();
6468 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6469 unsigned NumZeros = 0;
6470 while (SrcElTy != DstElTy &&
6471 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6472 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6473 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6477 // If we found a path from the src to dest, create the getelementptr now.
6478 if (SrcElTy == DstElTy) {
6479 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
6480 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
6485 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6486 if (SVI->hasOneUse()) {
6487 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6488 // a bitconvert to a vector with the same # elts.
6489 if (isa<VectorType>(DestTy) &&
6490 cast<VectorType>(DestTy)->getNumElements() ==
6491 SVI->getType()->getNumElements()) {
6493 // If either of the operands is a cast from CI.getType(), then
6494 // evaluating the shuffle in the casted destination's type will allow
6495 // us to eliminate at least one cast.
6496 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6497 Tmp->getOperand(0)->getType() == DestTy) ||
6498 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6499 Tmp->getOperand(0)->getType() == DestTy)) {
6500 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6501 SVI->getOperand(0), DestTy, &CI);
6502 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6503 SVI->getOperand(1), DestTy, &CI);
6504 // Return a new shuffle vector. Use the same element ID's, as we
6505 // know the vector types match #elts.
6506 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6514 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6516 /// %D = select %cond, %C, %A
6518 /// %C = select %cond, %B, 0
6521 /// Assuming that the specified instruction is an operand to the select, return
6522 /// a bitmask indicating which operands of this instruction are foldable if they
6523 /// equal the other incoming value of the select.
6525 static unsigned GetSelectFoldableOperands(Instruction *I) {
6526 switch (I->getOpcode()) {
6527 case Instruction::Add:
6528 case Instruction::Mul:
6529 case Instruction::And:
6530 case Instruction::Or:
6531 case Instruction::Xor:
6532 return 3; // Can fold through either operand.
6533 case Instruction::Sub: // Can only fold on the amount subtracted.
6534 case Instruction::Shl: // Can only fold on the shift amount.
6535 case Instruction::LShr:
6536 case Instruction::AShr:
6539 return 0; // Cannot fold
6543 /// GetSelectFoldableConstant - For the same transformation as the previous
6544 /// function, return the identity constant that goes into the select.
6545 static Constant *GetSelectFoldableConstant(Instruction *I) {
6546 switch (I->getOpcode()) {
6547 default: assert(0 && "This cannot happen!"); abort();
6548 case Instruction::Add:
6549 case Instruction::Sub:
6550 case Instruction::Or:
6551 case Instruction::Xor:
6552 case Instruction::Shl:
6553 case Instruction::LShr:
6554 case Instruction::AShr:
6555 return Constant::getNullValue(I->getType());
6556 case Instruction::And:
6557 return ConstantInt::getAllOnesValue(I->getType());
6558 case Instruction::Mul:
6559 return ConstantInt::get(I->getType(), 1);
6563 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6564 /// have the same opcode and only one use each. Try to simplify this.
6565 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6567 if (TI->getNumOperands() == 1) {
6568 // If this is a non-volatile load or a cast from the same type,
6571 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6574 return 0; // unknown unary op.
6577 // Fold this by inserting a select from the input values.
6578 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6579 FI->getOperand(0), SI.getName()+".v");
6580 InsertNewInstBefore(NewSI, SI);
6581 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6585 // Only handle binary operators here.
6586 if (!isa<BinaryOperator>(TI))
6589 // Figure out if the operations have any operands in common.
6590 Value *MatchOp, *OtherOpT, *OtherOpF;
6592 if (TI->getOperand(0) == FI->getOperand(0)) {
6593 MatchOp = TI->getOperand(0);
6594 OtherOpT = TI->getOperand(1);
6595 OtherOpF = FI->getOperand(1);
6596 MatchIsOpZero = true;
6597 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6598 MatchOp = TI->getOperand(1);
6599 OtherOpT = TI->getOperand(0);
6600 OtherOpF = FI->getOperand(0);
6601 MatchIsOpZero = false;
6602 } else if (!TI->isCommutative()) {
6604 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6605 MatchOp = TI->getOperand(0);
6606 OtherOpT = TI->getOperand(1);
6607 OtherOpF = FI->getOperand(0);
6608 MatchIsOpZero = true;
6609 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6610 MatchOp = TI->getOperand(1);
6611 OtherOpT = TI->getOperand(0);
6612 OtherOpF = FI->getOperand(1);
6613 MatchIsOpZero = true;
6618 // If we reach here, they do have operations in common.
6619 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6620 OtherOpF, SI.getName()+".v");
6621 InsertNewInstBefore(NewSI, SI);
6623 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6625 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6627 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6629 assert(0 && "Shouldn't get here");
6633 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6634 Value *CondVal = SI.getCondition();
6635 Value *TrueVal = SI.getTrueValue();
6636 Value *FalseVal = SI.getFalseValue();
6638 // select true, X, Y -> X
6639 // select false, X, Y -> Y
6640 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6641 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
6643 // select C, X, X -> X
6644 if (TrueVal == FalseVal)
6645 return ReplaceInstUsesWith(SI, TrueVal);
6647 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6648 return ReplaceInstUsesWith(SI, FalseVal);
6649 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6650 return ReplaceInstUsesWith(SI, TrueVal);
6651 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6652 if (isa<Constant>(TrueVal))
6653 return ReplaceInstUsesWith(SI, TrueVal);
6655 return ReplaceInstUsesWith(SI, FalseVal);
6658 if (SI.getType() == Type::Int1Ty) {
6659 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
6660 if (C->getZExtValue()) {
6661 // Change: A = select B, true, C --> A = or B, C
6662 return BinaryOperator::createOr(CondVal, FalseVal);
6664 // Change: A = select B, false, C --> A = and !B, C
6666 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6667 "not."+CondVal->getName()), SI);
6668 return BinaryOperator::createAnd(NotCond, FalseVal);
6670 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
6671 if (C->getZExtValue() == false) {
6672 // Change: A = select B, C, false --> A = and B, C
6673 return BinaryOperator::createAnd(CondVal, TrueVal);
6675 // Change: A = select B, C, true --> A = or !B, C
6677 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6678 "not."+CondVal->getName()), SI);
6679 return BinaryOperator::createOr(NotCond, TrueVal);
6684 // Selecting between two integer constants?
6685 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6686 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6687 // select C, 1, 0 -> cast C to int
6688 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6689 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6690 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6691 // select C, 0, 1 -> cast !C to int
6693 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6694 "not."+CondVal->getName()), SI);
6695 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6698 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6700 // (x <s 0) ? -1 : 0 -> ashr x, 31
6701 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6702 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6703 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6704 bool CanXForm = false;
6705 if (IC->isSignedPredicate())
6706 CanXForm = CmpCst->isNullValue() &&
6707 IC->getPredicate() == ICmpInst::ICMP_SLT;
6709 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6710 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6711 IC->getPredicate() == ICmpInst::ICMP_UGT;
6715 // The comparison constant and the result are not neccessarily the
6716 // same width. Make an all-ones value by inserting a AShr.
6717 Value *X = IC->getOperand(0);
6718 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6719 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
6720 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
6722 InsertNewInstBefore(SRA, SI);
6724 // Finally, convert to the type of the select RHS. We figure out
6725 // if this requires a SExt, Trunc or BitCast based on the sizes.
6726 Instruction::CastOps opc = Instruction::BitCast;
6727 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6728 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6729 if (SRASize < SISize)
6730 opc = Instruction::SExt;
6731 else if (SRASize > SISize)
6732 opc = Instruction::Trunc;
6733 return CastInst::create(opc, SRA, SI.getType());
6738 // If one of the constants is zero (we know they can't both be) and we
6739 // have a fcmp instruction with zero, and we have an 'and' with the
6740 // non-constant value, eliminate this whole mess. This corresponds to
6741 // cases like this: ((X & 27) ? 27 : 0)
6742 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6743 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6744 cast<Constant>(IC->getOperand(1))->isNullValue())
6745 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6746 if (ICA->getOpcode() == Instruction::And &&
6747 isa<ConstantInt>(ICA->getOperand(1)) &&
6748 (ICA->getOperand(1) == TrueValC ||
6749 ICA->getOperand(1) == FalseValC) &&
6750 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6751 // Okay, now we know that everything is set up, we just don't
6752 // know whether we have a icmp_ne or icmp_eq and whether the
6753 // true or false val is the zero.
6754 bool ShouldNotVal = !TrueValC->isNullValue();
6755 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6758 V = InsertNewInstBefore(BinaryOperator::create(
6759 Instruction::Xor, V, ICA->getOperand(1)), SI);
6760 return ReplaceInstUsesWith(SI, V);
6765 // See if we are selecting two values based on a comparison of the two values.
6766 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6767 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6768 // Transform (X == Y) ? X : Y -> Y
6769 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6770 return ReplaceInstUsesWith(SI, FalseVal);
6771 // Transform (X != Y) ? X : Y -> X
6772 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6773 return ReplaceInstUsesWith(SI, TrueVal);
6774 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6776 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6777 // Transform (X == Y) ? Y : X -> X
6778 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6779 return ReplaceInstUsesWith(SI, FalseVal);
6780 // Transform (X != Y) ? Y : X -> Y
6781 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6782 return ReplaceInstUsesWith(SI, TrueVal);
6783 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6787 // See if we are selecting two values based on a comparison of the two values.
6788 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6789 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6790 // Transform (X == Y) ? X : Y -> Y
6791 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6792 return ReplaceInstUsesWith(SI, FalseVal);
6793 // Transform (X != Y) ? X : Y -> X
6794 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6795 return ReplaceInstUsesWith(SI, TrueVal);
6796 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6798 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6799 // Transform (X == Y) ? Y : X -> X
6800 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6801 return ReplaceInstUsesWith(SI, FalseVal);
6802 // Transform (X != Y) ? Y : X -> Y
6803 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6804 return ReplaceInstUsesWith(SI, TrueVal);
6805 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6809 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6810 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6811 if (TI->hasOneUse() && FI->hasOneUse()) {
6812 Instruction *AddOp = 0, *SubOp = 0;
6814 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6815 if (TI->getOpcode() == FI->getOpcode())
6816 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6819 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6820 // even legal for FP.
6821 if (TI->getOpcode() == Instruction::Sub &&
6822 FI->getOpcode() == Instruction::Add) {
6823 AddOp = FI; SubOp = TI;
6824 } else if (FI->getOpcode() == Instruction::Sub &&
6825 TI->getOpcode() == Instruction::Add) {
6826 AddOp = TI; SubOp = FI;
6830 Value *OtherAddOp = 0;
6831 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6832 OtherAddOp = AddOp->getOperand(1);
6833 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6834 OtherAddOp = AddOp->getOperand(0);
6838 // So at this point we know we have (Y -> OtherAddOp):
6839 // select C, (add X, Y), (sub X, Z)
6840 Value *NegVal; // Compute -Z
6841 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6842 NegVal = ConstantExpr::getNeg(C);
6844 NegVal = InsertNewInstBefore(
6845 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6848 Value *NewTrueOp = OtherAddOp;
6849 Value *NewFalseOp = NegVal;
6851 std::swap(NewTrueOp, NewFalseOp);
6852 Instruction *NewSel =
6853 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6855 NewSel = InsertNewInstBefore(NewSel, SI);
6856 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6861 // See if we can fold the select into one of our operands.
6862 if (SI.getType()->isInteger()) {
6863 // See the comment above GetSelectFoldableOperands for a description of the
6864 // transformation we are doing here.
6865 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6866 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6867 !isa<Constant>(FalseVal))
6868 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6869 unsigned OpToFold = 0;
6870 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6872 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6877 Constant *C = GetSelectFoldableConstant(TVI);
6878 Instruction *NewSel =
6879 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
6880 InsertNewInstBefore(NewSel, SI);
6881 NewSel->takeName(TVI);
6882 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6883 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6885 assert(0 && "Unknown instruction!!");
6890 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6891 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6892 !isa<Constant>(TrueVal))
6893 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6894 unsigned OpToFold = 0;
6895 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6897 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6902 Constant *C = GetSelectFoldableConstant(FVI);
6903 Instruction *NewSel =
6904 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
6905 InsertNewInstBefore(NewSel, SI);
6906 NewSel->takeName(FVI);
6907 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6908 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6910 assert(0 && "Unknown instruction!!");
6915 if (BinaryOperator::isNot(CondVal)) {
6916 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6917 SI.setOperand(1, FalseVal);
6918 SI.setOperand(2, TrueVal);
6925 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6926 /// determine, return it, otherwise return 0.
6927 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6928 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6929 unsigned Align = GV->getAlignment();
6930 if (Align == 0 && TD)
6931 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
6933 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6934 unsigned Align = AI->getAlignment();
6935 if (Align == 0 && TD) {
6936 if (isa<AllocaInst>(AI))
6937 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
6938 else if (isa<MallocInst>(AI)) {
6939 // Malloc returns maximally aligned memory.
6940 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
6943 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
6946 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
6950 } else if (isa<BitCastInst>(V) ||
6951 (isa<ConstantExpr>(V) &&
6952 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
6953 User *CI = cast<User>(V);
6954 if (isa<PointerType>(CI->getOperand(0)->getType()))
6955 return GetKnownAlignment(CI->getOperand(0), TD);
6957 } else if (isa<GetElementPtrInst>(V) ||
6958 (isa<ConstantExpr>(V) &&
6959 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
6960 User *GEPI = cast<User>(V);
6961 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
6962 if (BaseAlignment == 0) return 0;
6964 // If all indexes are zero, it is just the alignment of the base pointer.
6965 bool AllZeroOperands = true;
6966 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
6967 if (!isa<Constant>(GEPI->getOperand(i)) ||
6968 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
6969 AllZeroOperands = false;
6972 if (AllZeroOperands)
6973 return BaseAlignment;
6975 // Otherwise, if the base alignment is >= the alignment we expect for the
6976 // base pointer type, then we know that the resultant pointer is aligned at
6977 // least as much as its type requires.
6980 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
6981 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
6982 if (TD->getABITypeAlignment(PtrTy->getElementType())
6984 const Type *GEPTy = GEPI->getType();
6985 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
6986 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
6994 /// visitCallInst - CallInst simplification. This mostly only handles folding
6995 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
6996 /// the heavy lifting.
6998 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
6999 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7000 if (!II) return visitCallSite(&CI);
7002 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7004 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7005 bool Changed = false;
7007 // memmove/cpy/set of zero bytes is a noop.
7008 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7009 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7011 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7012 if (CI->getZExtValue() == 1) {
7013 // Replace the instruction with just byte operations. We would
7014 // transform other cases to loads/stores, but we don't know if
7015 // alignment is sufficient.
7019 // If we have a memmove and the source operation is a constant global,
7020 // then the source and dest pointers can't alias, so we can change this
7021 // into a call to memcpy.
7022 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7023 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7024 if (GVSrc->isConstant()) {
7025 Module *M = CI.getParent()->getParent()->getParent();
7027 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7029 Name = "llvm.memcpy.i32";
7031 Name = "llvm.memcpy.i64";
7032 Constant *MemCpy = M->getOrInsertFunction(Name,
7033 CI.getCalledFunction()->getFunctionType());
7034 CI.setOperand(0, MemCpy);
7039 // If we can determine a pointer alignment that is bigger than currently
7040 // set, update the alignment.
7041 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7042 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7043 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7044 unsigned Align = std::min(Alignment1, Alignment2);
7045 if (MI->getAlignment()->getZExtValue() < Align) {
7046 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7049 } else if (isa<MemSetInst>(MI)) {
7050 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7051 if (MI->getAlignment()->getZExtValue() < Alignment) {
7052 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7057 if (Changed) return II;
7059 switch (II->getIntrinsicID()) {
7061 case Intrinsic::ppc_altivec_lvx:
7062 case Intrinsic::ppc_altivec_lvxl:
7063 case Intrinsic::x86_sse_loadu_ps:
7064 case Intrinsic::x86_sse2_loadu_pd:
7065 case Intrinsic::x86_sse2_loadu_dq:
7066 // Turn PPC lvx -> load if the pointer is known aligned.
7067 // Turn X86 loadups -> load if the pointer is known aligned.
7068 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7069 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7070 PointerType::get(II->getType()), CI);
7071 return new LoadInst(Ptr);
7074 case Intrinsic::ppc_altivec_stvx:
7075 case Intrinsic::ppc_altivec_stvxl:
7076 // Turn stvx -> store if the pointer is known aligned.
7077 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7078 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7079 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7081 return new StoreInst(II->getOperand(1), Ptr);
7084 case Intrinsic::x86_sse_storeu_ps:
7085 case Intrinsic::x86_sse2_storeu_pd:
7086 case Intrinsic::x86_sse2_storeu_dq:
7087 case Intrinsic::x86_sse2_storel_dq:
7088 // Turn X86 storeu -> store if the pointer is known aligned.
7089 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7090 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7091 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7093 return new StoreInst(II->getOperand(2), Ptr);
7097 case Intrinsic::x86_sse_cvttss2si: {
7098 // These intrinsics only demands the 0th element of its input vector. If
7099 // we can simplify the input based on that, do so now.
7101 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7103 II->setOperand(1, V);
7109 case Intrinsic::ppc_altivec_vperm:
7110 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7111 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7112 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7114 // Check that all of the elements are integer constants or undefs.
7115 bool AllEltsOk = true;
7116 for (unsigned i = 0; i != 16; ++i) {
7117 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7118 !isa<UndefValue>(Mask->getOperand(i))) {
7125 // Cast the input vectors to byte vectors.
7126 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7127 II->getOperand(1), Mask->getType(), CI);
7128 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7129 II->getOperand(2), Mask->getType(), CI);
7130 Value *Result = UndefValue::get(Op0->getType());
7132 // Only extract each element once.
7133 Value *ExtractedElts[32];
7134 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7136 for (unsigned i = 0; i != 16; ++i) {
7137 if (isa<UndefValue>(Mask->getOperand(i)))
7139 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7140 Idx &= 31; // Match the hardware behavior.
7142 if (ExtractedElts[Idx] == 0) {
7144 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7145 InsertNewInstBefore(Elt, CI);
7146 ExtractedElts[Idx] = Elt;
7149 // Insert this value into the result vector.
7150 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7151 InsertNewInstBefore(cast<Instruction>(Result), CI);
7153 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7158 case Intrinsic::stackrestore: {
7159 // If the save is right next to the restore, remove the restore. This can
7160 // happen when variable allocas are DCE'd.
7161 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7162 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7163 BasicBlock::iterator BI = SS;
7165 return EraseInstFromFunction(CI);
7169 // If the stack restore is in a return/unwind block and if there are no
7170 // allocas or calls between the restore and the return, nuke the restore.
7171 TerminatorInst *TI = II->getParent()->getTerminator();
7172 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7173 BasicBlock::iterator BI = II;
7174 bool CannotRemove = false;
7175 for (++BI; &*BI != TI; ++BI) {
7176 if (isa<AllocaInst>(BI) ||
7177 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7178 CannotRemove = true;
7183 return EraseInstFromFunction(CI);
7190 return visitCallSite(II);
7193 // InvokeInst simplification
7195 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7196 return visitCallSite(&II);
7199 // visitCallSite - Improvements for call and invoke instructions.
7201 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7202 bool Changed = false;
7204 // If the callee is a constexpr cast of a function, attempt to move the cast
7205 // to the arguments of the call/invoke.
7206 if (transformConstExprCastCall(CS)) return 0;
7208 Value *Callee = CS.getCalledValue();
7210 if (Function *CalleeF = dyn_cast<Function>(Callee))
7211 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7212 Instruction *OldCall = CS.getInstruction();
7213 // If the call and callee calling conventions don't match, this call must
7214 // be unreachable, as the call is undefined.
7215 new StoreInst(ConstantInt::getTrue(),
7216 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7217 if (!OldCall->use_empty())
7218 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7219 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7220 return EraseInstFromFunction(*OldCall);
7224 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7225 // This instruction is not reachable, just remove it. We insert a store to
7226 // undef so that we know that this code is not reachable, despite the fact
7227 // that we can't modify the CFG here.
7228 new StoreInst(ConstantInt::getTrue(),
7229 UndefValue::get(PointerType::get(Type::Int1Ty)),
7230 CS.getInstruction());
7232 if (!CS.getInstruction()->use_empty())
7233 CS.getInstruction()->
7234 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7236 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7237 // Don't break the CFG, insert a dummy cond branch.
7238 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7239 ConstantInt::getTrue(), II);
7241 return EraseInstFromFunction(*CS.getInstruction());
7244 const PointerType *PTy = cast<PointerType>(Callee->getType());
7245 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7246 if (FTy->isVarArg()) {
7247 // See if we can optimize any arguments passed through the varargs area of
7249 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7250 E = CS.arg_end(); I != E; ++I)
7251 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7252 // If this cast does not effect the value passed through the varargs
7253 // area, we can eliminate the use of the cast.
7254 Value *Op = CI->getOperand(0);
7255 if (CI->isLosslessCast()) {
7262 return Changed ? CS.getInstruction() : 0;
7265 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7266 // attempt to move the cast to the arguments of the call/invoke.
7268 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7269 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7270 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7271 if (CE->getOpcode() != Instruction::BitCast ||
7272 !isa<Function>(CE->getOperand(0)))
7274 Function *Callee = cast<Function>(CE->getOperand(0));
7275 Instruction *Caller = CS.getInstruction();
7277 // Okay, this is a cast from a function to a different type. Unless doing so
7278 // would cause a type conversion of one of our arguments, change this call to
7279 // be a direct call with arguments casted to the appropriate types.
7281 const FunctionType *FT = Callee->getFunctionType();
7282 const Type *OldRetTy = Caller->getType();
7284 // Check to see if we are changing the return type...
7285 if (OldRetTy != FT->getReturnType()) {
7286 if (Callee->isDeclaration() && !Caller->use_empty() &&
7287 OldRetTy != FT->getReturnType() &&
7288 // Conversion is ok if changing from pointer to int of same size.
7289 !(isa<PointerType>(FT->getReturnType()) &&
7290 TD->getIntPtrType() == OldRetTy))
7291 return false; // Cannot transform this return value.
7293 // If the callsite is an invoke instruction, and the return value is used by
7294 // a PHI node in a successor, we cannot change the return type of the call
7295 // because there is no place to put the cast instruction (without breaking
7296 // the critical edge). Bail out in this case.
7297 if (!Caller->use_empty())
7298 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7299 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7301 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7302 if (PN->getParent() == II->getNormalDest() ||
7303 PN->getParent() == II->getUnwindDest())
7307 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7308 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7310 CallSite::arg_iterator AI = CS.arg_begin();
7311 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7312 const Type *ParamTy = FT->getParamType(i);
7313 const Type *ActTy = (*AI)->getType();
7314 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7315 //Either we can cast directly, or we can upconvert the argument
7316 bool isConvertible = ActTy == ParamTy ||
7317 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7318 (ParamTy->isInteger() && ActTy->isInteger() &&
7319 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7320 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7321 && c->getSExtValue() > 0);
7322 if (Callee->isDeclaration() && !isConvertible) return false;
7325 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7326 Callee->isDeclaration())
7327 return false; // Do not delete arguments unless we have a function body...
7329 // Okay, we decided that this is a safe thing to do: go ahead and start
7330 // inserting cast instructions as necessary...
7331 std::vector<Value*> Args;
7332 Args.reserve(NumActualArgs);
7334 AI = CS.arg_begin();
7335 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7336 const Type *ParamTy = FT->getParamType(i);
7337 if ((*AI)->getType() == ParamTy) {
7338 Args.push_back(*AI);
7340 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7341 false, ParamTy, false);
7342 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7343 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7347 // If the function takes more arguments than the call was taking, add them
7349 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7350 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7352 // If we are removing arguments to the function, emit an obnoxious warning...
7353 if (FT->getNumParams() < NumActualArgs)
7354 if (!FT->isVarArg()) {
7355 cerr << "WARNING: While resolving call to function '"
7356 << Callee->getName() << "' arguments were dropped!\n";
7358 // Add all of the arguments in their promoted form to the arg list...
7359 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7360 const Type *PTy = getPromotedType((*AI)->getType());
7361 if (PTy != (*AI)->getType()) {
7362 // Must promote to pass through va_arg area!
7363 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7365 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7366 InsertNewInstBefore(Cast, *Caller);
7367 Args.push_back(Cast);
7369 Args.push_back(*AI);
7374 if (FT->getReturnType() == Type::VoidTy)
7375 Caller->setName(""); // Void type should not have a name.
7378 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7379 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7380 &Args[0], Args.size(), Caller->getName(), Caller);
7381 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7383 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
7384 if (cast<CallInst>(Caller)->isTailCall())
7385 cast<CallInst>(NC)->setTailCall();
7386 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7389 // Insert a cast of the return type as necessary.
7391 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7392 if (NV->getType() != Type::VoidTy) {
7393 const Type *CallerTy = Caller->getType();
7394 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7396 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7398 // If this is an invoke instruction, we should insert it after the first
7399 // non-phi, instruction in the normal successor block.
7400 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7401 BasicBlock::iterator I = II->getNormalDest()->begin();
7402 while (isa<PHINode>(I)) ++I;
7403 InsertNewInstBefore(NC, *I);
7405 // Otherwise, it's a call, just insert cast right after the call instr
7406 InsertNewInstBefore(NC, *Caller);
7408 AddUsersToWorkList(*Caller);
7410 NV = UndefValue::get(Caller->getType());
7414 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7415 Caller->replaceAllUsesWith(NV);
7416 Caller->eraseFromParent();
7417 RemoveFromWorkList(Caller);
7421 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7422 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7423 /// and a single binop.
7424 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7425 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7426 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
7427 isa<CmpInst>(FirstInst));
7428 unsigned Opc = FirstInst->getOpcode();
7429 Value *LHSVal = FirstInst->getOperand(0);
7430 Value *RHSVal = FirstInst->getOperand(1);
7432 const Type *LHSType = LHSVal->getType();
7433 const Type *RHSType = RHSVal->getType();
7435 // Scan to see if all operands are the same opcode, all have one use, and all
7436 // kill their operands (i.e. the operands have one use).
7437 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7438 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7439 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7440 // Verify type of the LHS matches so we don't fold cmp's of different
7441 // types or GEP's with different index types.
7442 I->getOperand(0)->getType() != LHSType ||
7443 I->getOperand(1)->getType() != RHSType)
7446 // If they are CmpInst instructions, check their predicates
7447 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7448 if (cast<CmpInst>(I)->getPredicate() !=
7449 cast<CmpInst>(FirstInst)->getPredicate())
7452 // Keep track of which operand needs a phi node.
7453 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7454 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7457 // Otherwise, this is safe to transform, determine if it is profitable.
7459 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7460 // Indexes are often folded into load/store instructions, so we don't want to
7461 // hide them behind a phi.
7462 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7465 Value *InLHS = FirstInst->getOperand(0);
7466 Value *InRHS = FirstInst->getOperand(1);
7467 PHINode *NewLHS = 0, *NewRHS = 0;
7469 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7470 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7471 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7472 InsertNewInstBefore(NewLHS, PN);
7477 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7478 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7479 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7480 InsertNewInstBefore(NewRHS, PN);
7484 // Add all operands to the new PHIs.
7485 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7487 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7488 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7491 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7492 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7496 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7497 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7498 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7499 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7502 assert(isa<GetElementPtrInst>(FirstInst));
7503 return new GetElementPtrInst(LHSVal, RHSVal);
7507 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7508 /// of the block that defines it. This means that it must be obvious the value
7509 /// of the load is not changed from the point of the load to the end of the
7512 /// Finally, it is safe, but not profitable, to sink a load targetting a
7513 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
7515 static bool isSafeToSinkLoad(LoadInst *L) {
7516 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7518 for (++BBI; BBI != E; ++BBI)
7519 if (BBI->mayWriteToMemory())
7522 // Check for non-address taken alloca. If not address-taken already, it isn't
7523 // profitable to do this xform.
7524 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
7525 bool isAddressTaken = false;
7526 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
7528 if (isa<LoadInst>(UI)) continue;
7529 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
7530 // If storing TO the alloca, then the address isn't taken.
7531 if (SI->getOperand(1) == AI) continue;
7533 isAddressTaken = true;
7537 if (!isAddressTaken)
7545 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7546 // operator and they all are only used by the PHI, PHI together their
7547 // inputs, and do the operation once, to the result of the PHI.
7548 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7549 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7551 // Scan the instruction, looking for input operations that can be folded away.
7552 // If all input operands to the phi are the same instruction (e.g. a cast from
7553 // the same type or "+42") we can pull the operation through the PHI, reducing
7554 // code size and simplifying code.
7555 Constant *ConstantOp = 0;
7556 const Type *CastSrcTy = 0;
7557 bool isVolatile = false;
7558 if (isa<CastInst>(FirstInst)) {
7559 CastSrcTy = FirstInst->getOperand(0)->getType();
7560 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
7561 // Can fold binop, compare or shift here if the RHS is a constant,
7562 // otherwise call FoldPHIArgBinOpIntoPHI.
7563 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7564 if (ConstantOp == 0)
7565 return FoldPHIArgBinOpIntoPHI(PN);
7566 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7567 isVolatile = LI->isVolatile();
7568 // We can't sink the load if the loaded value could be modified between the
7569 // load and the PHI.
7570 if (LI->getParent() != PN.getIncomingBlock(0) ||
7571 !isSafeToSinkLoad(LI))
7573 } else if (isa<GetElementPtrInst>(FirstInst)) {
7574 if (FirstInst->getNumOperands() == 2)
7575 return FoldPHIArgBinOpIntoPHI(PN);
7576 // Can't handle general GEPs yet.
7579 return 0; // Cannot fold this operation.
7582 // Check to see if all arguments are the same operation.
7583 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7584 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7585 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7586 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7589 if (I->getOperand(0)->getType() != CastSrcTy)
7590 return 0; // Cast operation must match.
7591 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7592 // We can't sink the load if the loaded value could be modified between
7593 // the load and the PHI.
7594 if (LI->isVolatile() != isVolatile ||
7595 LI->getParent() != PN.getIncomingBlock(i) ||
7596 !isSafeToSinkLoad(LI))
7598 } else if (I->getOperand(1) != ConstantOp) {
7603 // Okay, they are all the same operation. Create a new PHI node of the
7604 // correct type, and PHI together all of the LHS's of the instructions.
7605 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7606 PN.getName()+".in");
7607 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7609 Value *InVal = FirstInst->getOperand(0);
7610 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7612 // Add all operands to the new PHI.
7613 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7614 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7615 if (NewInVal != InVal)
7617 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7622 // The new PHI unions all of the same values together. This is really
7623 // common, so we handle it intelligently here for compile-time speed.
7627 InsertNewInstBefore(NewPN, PN);
7631 // Insert and return the new operation.
7632 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7633 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7634 else if (isa<LoadInst>(FirstInst))
7635 return new LoadInst(PhiVal, "", isVolatile);
7636 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7637 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7638 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7639 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7640 PhiVal, ConstantOp);
7642 assert(0 && "Unknown operation");
7645 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7647 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7648 if (PN->use_empty()) return true;
7649 if (!PN->hasOneUse()) return false;
7651 // Remember this node, and if we find the cycle, return.
7652 if (!PotentiallyDeadPHIs.insert(PN).second)
7655 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7656 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7661 // PHINode simplification
7663 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7664 // If LCSSA is around, don't mess with Phi nodes
7665 if (mustPreserveAnalysisID(LCSSAID)) return 0;
7667 if (Value *V = PN.hasConstantValue())
7668 return ReplaceInstUsesWith(PN, V);
7670 // If all PHI operands are the same operation, pull them through the PHI,
7671 // reducing code size.
7672 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7673 PN.getIncomingValue(0)->hasOneUse())
7674 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7677 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7678 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7679 // PHI)... break the cycle.
7680 if (PN.hasOneUse()) {
7681 Instruction *PHIUser = cast<Instruction>(PN.use_back());
7682 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
7683 std::set<PHINode*> PotentiallyDeadPHIs;
7684 PotentiallyDeadPHIs.insert(&PN);
7685 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7686 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7689 // If this phi has a single use, and if that use just computes a value for
7690 // the next iteration of a loop, delete the phi. This occurs with unused
7691 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
7692 // common case here is good because the only other things that catch this
7693 // are induction variable analysis (sometimes) and ADCE, which is only run
7695 if (PHIUser->hasOneUse() &&
7696 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
7697 PHIUser->use_back() == &PN) {
7698 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7705 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7706 Instruction *InsertPoint,
7708 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7709 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7710 // We must cast correctly to the pointer type. Ensure that we
7711 // sign extend the integer value if it is smaller as this is
7712 // used for address computation.
7713 Instruction::CastOps opcode =
7714 (VTySize < PtrSize ? Instruction::SExt :
7715 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7716 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7720 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7721 Value *PtrOp = GEP.getOperand(0);
7722 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7723 // If so, eliminate the noop.
7724 if (GEP.getNumOperands() == 1)
7725 return ReplaceInstUsesWith(GEP, PtrOp);
7727 if (isa<UndefValue>(GEP.getOperand(0)))
7728 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7730 bool HasZeroPointerIndex = false;
7731 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7732 HasZeroPointerIndex = C->isNullValue();
7734 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7735 return ReplaceInstUsesWith(GEP, PtrOp);
7737 // Eliminate unneeded casts for indices.
7738 bool MadeChange = false;
7739 gep_type_iterator GTI = gep_type_begin(GEP);
7740 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7741 if (isa<SequentialType>(*GTI)) {
7742 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7743 if (CI->getOpcode() == Instruction::ZExt ||
7744 CI->getOpcode() == Instruction::SExt) {
7745 const Type *SrcTy = CI->getOperand(0)->getType();
7746 // We can eliminate a cast from i32 to i64 iff the target
7747 // is a 32-bit pointer target.
7748 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7750 GEP.setOperand(i, CI->getOperand(0));
7754 // If we are using a wider index than needed for this platform, shrink it
7755 // to what we need. If the incoming value needs a cast instruction,
7756 // insert it. This explicit cast can make subsequent optimizations more
7758 Value *Op = GEP.getOperand(i);
7759 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
7760 if (Constant *C = dyn_cast<Constant>(Op)) {
7761 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7764 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7766 GEP.setOperand(i, Op);
7770 if (MadeChange) return &GEP;
7772 // Combine Indices - If the source pointer to this getelementptr instruction
7773 // is a getelementptr instruction, combine the indices of the two
7774 // getelementptr instructions into a single instruction.
7776 SmallVector<Value*, 8> SrcGEPOperands;
7777 if (User *Src = dyn_castGetElementPtr(PtrOp))
7778 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
7780 if (!SrcGEPOperands.empty()) {
7781 // Note that if our source is a gep chain itself that we wait for that
7782 // chain to be resolved before we perform this transformation. This
7783 // avoids us creating a TON of code in some cases.
7785 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7786 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7787 return 0; // Wait until our source is folded to completion.
7789 SmallVector<Value*, 8> Indices;
7791 // Find out whether the last index in the source GEP is a sequential idx.
7792 bool EndsWithSequential = false;
7793 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7794 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7795 EndsWithSequential = !isa<StructType>(*I);
7797 // Can we combine the two pointer arithmetics offsets?
7798 if (EndsWithSequential) {
7799 // Replace: gep (gep %P, long B), long A, ...
7800 // With: T = long A+B; gep %P, T, ...
7802 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7803 if (SO1 == Constant::getNullValue(SO1->getType())) {
7805 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7808 // If they aren't the same type, convert both to an integer of the
7809 // target's pointer size.
7810 if (SO1->getType() != GO1->getType()) {
7811 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7812 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7813 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7814 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7816 unsigned PS = TD->getPointerSize();
7817 if (TD->getTypeSize(SO1->getType()) == PS) {
7818 // Convert GO1 to SO1's type.
7819 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
7821 } else if (TD->getTypeSize(GO1->getType()) == PS) {
7822 // Convert SO1 to GO1's type.
7823 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
7825 const Type *PT = TD->getIntPtrType();
7826 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
7827 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
7831 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7832 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7834 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7835 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7839 // Recycle the GEP we already have if possible.
7840 if (SrcGEPOperands.size() == 2) {
7841 GEP.setOperand(0, SrcGEPOperands[0]);
7842 GEP.setOperand(1, Sum);
7845 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7846 SrcGEPOperands.end()-1);
7847 Indices.push_back(Sum);
7848 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7850 } else if (isa<Constant>(*GEP.idx_begin()) &&
7851 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7852 SrcGEPOperands.size() != 1) {
7853 // Otherwise we can do the fold if the first index of the GEP is a zero
7854 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7855 SrcGEPOperands.end());
7856 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7859 if (!Indices.empty())
7860 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
7861 Indices.size(), GEP.getName());
7863 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7864 // GEP of global variable. If all of the indices for this GEP are
7865 // constants, we can promote this to a constexpr instead of an instruction.
7867 // Scan for nonconstants...
7868 SmallVector<Constant*, 8> Indices;
7869 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7870 for (; I != E && isa<Constant>(*I); ++I)
7871 Indices.push_back(cast<Constant>(*I));
7873 if (I == E) { // If they are all constants...
7874 Constant *CE = ConstantExpr::getGetElementPtr(GV,
7875 &Indices[0],Indices.size());
7877 // Replace all uses of the GEP with the new constexpr...
7878 return ReplaceInstUsesWith(GEP, CE);
7880 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7881 if (!isa<PointerType>(X->getType())) {
7882 // Not interesting. Source pointer must be a cast from pointer.
7883 } else if (HasZeroPointerIndex) {
7884 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7885 // into : GEP [10 x ubyte]* X, long 0, ...
7887 // This occurs when the program declares an array extern like "int X[];"
7889 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7890 const PointerType *XTy = cast<PointerType>(X->getType());
7891 if (const ArrayType *XATy =
7892 dyn_cast<ArrayType>(XTy->getElementType()))
7893 if (const ArrayType *CATy =
7894 dyn_cast<ArrayType>(CPTy->getElementType()))
7895 if (CATy->getElementType() == XATy->getElementType()) {
7896 // At this point, we know that the cast source type is a pointer
7897 // to an array of the same type as the destination pointer
7898 // array. Because the array type is never stepped over (there
7899 // is a leading zero) we can fold the cast into this GEP.
7900 GEP.setOperand(0, X);
7903 } else if (GEP.getNumOperands() == 2) {
7904 // Transform things like:
7905 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7906 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7907 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7908 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7909 if (isa<ArrayType>(SrcElTy) &&
7910 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7911 TD->getTypeSize(ResElTy)) {
7912 Value *V = InsertNewInstBefore(
7913 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7914 GEP.getOperand(1), GEP.getName()), GEP);
7915 // V and GEP are both pointer types --> BitCast
7916 return new BitCastInst(V, GEP.getType());
7919 // Transform things like:
7920 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7921 // (where tmp = 8*tmp2) into:
7922 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7924 if (isa<ArrayType>(SrcElTy) &&
7925 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
7926 uint64_t ArrayEltSize =
7927 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7929 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7930 // allow either a mul, shift, or constant here.
7932 ConstantInt *Scale = 0;
7933 if (ArrayEltSize == 1) {
7934 NewIdx = GEP.getOperand(1);
7935 Scale = ConstantInt::get(NewIdx->getType(), 1);
7936 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7937 NewIdx = ConstantInt::get(CI->getType(), 1);
7939 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7940 if (Inst->getOpcode() == Instruction::Shl &&
7941 isa<ConstantInt>(Inst->getOperand(1))) {
7943 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7944 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7945 NewIdx = Inst->getOperand(0);
7946 } else if (Inst->getOpcode() == Instruction::Mul &&
7947 isa<ConstantInt>(Inst->getOperand(1))) {
7948 Scale = cast<ConstantInt>(Inst->getOperand(1));
7949 NewIdx = Inst->getOperand(0);
7953 // If the index will be to exactly the right offset with the scale taken
7954 // out, perform the transformation.
7955 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7956 if (isa<ConstantInt>(Scale))
7957 Scale = ConstantInt::get(Scale->getType(),
7958 Scale->getZExtValue() / ArrayEltSize);
7959 if (Scale->getZExtValue() != 1) {
7960 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
7962 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7963 NewIdx = InsertNewInstBefore(Sc, GEP);
7966 // Insert the new GEP instruction.
7967 Instruction *NewGEP =
7968 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7969 NewIdx, GEP.getName());
7970 NewGEP = InsertNewInstBefore(NewGEP, GEP);
7971 // The NewGEP must be pointer typed, so must the old one -> BitCast
7972 return new BitCastInst(NewGEP, GEP.getType());
7981 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7982 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7983 if (AI.isArrayAllocation()) // Check C != 1
7984 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7986 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7987 AllocationInst *New = 0;
7989 // Create and insert the replacement instruction...
7990 if (isa<MallocInst>(AI))
7991 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7993 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7994 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7997 InsertNewInstBefore(New, AI);
7999 // Scan to the end of the allocation instructions, to skip over a block of
8000 // allocas if possible...
8002 BasicBlock::iterator It = New;
8003 while (isa<AllocationInst>(*It)) ++It;
8005 // Now that I is pointing to the first non-allocation-inst in the block,
8006 // insert our getelementptr instruction...
8008 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8009 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8010 New->getName()+".sub", It);
8012 // Now make everything use the getelementptr instead of the original
8014 return ReplaceInstUsesWith(AI, V);
8015 } else if (isa<UndefValue>(AI.getArraySize())) {
8016 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8019 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8020 // Note that we only do this for alloca's, because malloc should allocate and
8021 // return a unique pointer, even for a zero byte allocation.
8022 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8023 TD->getTypeSize(AI.getAllocatedType()) == 0)
8024 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8029 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8030 Value *Op = FI.getOperand(0);
8032 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8033 if (CastInst *CI = dyn_cast<CastInst>(Op))
8034 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8035 FI.setOperand(0, CI->getOperand(0));
8039 // free undef -> unreachable.
8040 if (isa<UndefValue>(Op)) {
8041 // Insert a new store to null because we cannot modify the CFG here.
8042 new StoreInst(ConstantInt::getTrue(),
8043 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8044 return EraseInstFromFunction(FI);
8047 // If we have 'free null' delete the instruction. This can happen in stl code
8048 // when lots of inlining happens.
8049 if (isa<ConstantPointerNull>(Op))
8050 return EraseInstFromFunction(FI);
8056 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8057 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8058 User *CI = cast<User>(LI.getOperand(0));
8059 Value *CastOp = CI->getOperand(0);
8061 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8062 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8063 const Type *SrcPTy = SrcTy->getElementType();
8065 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8066 isa<VectorType>(DestPTy)) {
8067 // If the source is an array, the code below will not succeed. Check to
8068 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8070 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8071 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8072 if (ASrcTy->getNumElements() != 0) {
8074 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8075 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8076 SrcTy = cast<PointerType>(CastOp->getType());
8077 SrcPTy = SrcTy->getElementType();
8080 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8081 isa<VectorType>(SrcPTy)) &&
8082 // Do not allow turning this into a load of an integer, which is then
8083 // casted to a pointer, this pessimizes pointer analysis a lot.
8084 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8085 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8086 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8088 // Okay, we are casting from one integer or pointer type to another of
8089 // the same size. Instead of casting the pointer before the load, cast
8090 // the result of the loaded value.
8091 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8093 LI.isVolatile()),LI);
8094 // Now cast the result of the load.
8095 return new BitCastInst(NewLoad, LI.getType());
8102 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8103 /// from this value cannot trap. If it is not obviously safe to load from the
8104 /// specified pointer, we do a quick local scan of the basic block containing
8105 /// ScanFrom, to determine if the address is already accessed.
8106 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8107 // If it is an alloca or global variable, it is always safe to load from.
8108 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8110 // Otherwise, be a little bit agressive by scanning the local block where we
8111 // want to check to see if the pointer is already being loaded or stored
8112 // from/to. If so, the previous load or store would have already trapped,
8113 // so there is no harm doing an extra load (also, CSE will later eliminate
8114 // the load entirely).
8115 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8120 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8121 if (LI->getOperand(0) == V) return true;
8122 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8123 if (SI->getOperand(1) == V) return true;
8129 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8130 Value *Op = LI.getOperand(0);
8132 // load (cast X) --> cast (load X) iff safe
8133 if (isa<CastInst>(Op))
8134 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8137 // None of the following transforms are legal for volatile loads.
8138 if (LI.isVolatile()) return 0;
8140 if (&LI.getParent()->front() != &LI) {
8141 BasicBlock::iterator BBI = &LI; --BBI;
8142 // If the instruction immediately before this is a store to the same
8143 // address, do a simple form of store->load forwarding.
8144 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8145 if (SI->getOperand(1) == LI.getOperand(0))
8146 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8147 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8148 if (LIB->getOperand(0) == LI.getOperand(0))
8149 return ReplaceInstUsesWith(LI, LIB);
8152 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8153 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8154 isa<UndefValue>(GEPI->getOperand(0))) {
8155 // Insert a new store to null instruction before the load to indicate
8156 // that this code is not reachable. We do this instead of inserting
8157 // an unreachable instruction directly because we cannot modify the
8159 new StoreInst(UndefValue::get(LI.getType()),
8160 Constant::getNullValue(Op->getType()), &LI);
8161 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8164 if (Constant *C = dyn_cast<Constant>(Op)) {
8165 // load null/undef -> undef
8166 if ((C->isNullValue() || isa<UndefValue>(C))) {
8167 // Insert a new store to null instruction before the load to indicate that
8168 // this code is not reachable. We do this instead of inserting an
8169 // unreachable instruction directly because we cannot modify the CFG.
8170 new StoreInst(UndefValue::get(LI.getType()),
8171 Constant::getNullValue(Op->getType()), &LI);
8172 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8175 // Instcombine load (constant global) into the value loaded.
8176 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8177 if (GV->isConstant() && !GV->isDeclaration())
8178 return ReplaceInstUsesWith(LI, GV->getInitializer());
8180 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8181 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8182 if (CE->getOpcode() == Instruction::GetElementPtr) {
8183 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8184 if (GV->isConstant() && !GV->isDeclaration())
8186 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8187 return ReplaceInstUsesWith(LI, V);
8188 if (CE->getOperand(0)->isNullValue()) {
8189 // Insert a new store to null instruction before the load to indicate
8190 // that this code is not reachable. We do this instead of inserting
8191 // an unreachable instruction directly because we cannot modify the
8193 new StoreInst(UndefValue::get(LI.getType()),
8194 Constant::getNullValue(Op->getType()), &LI);
8195 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8198 } else if (CE->isCast()) {
8199 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8204 if (Op->hasOneUse()) {
8205 // Change select and PHI nodes to select values instead of addresses: this
8206 // helps alias analysis out a lot, allows many others simplifications, and
8207 // exposes redundancy in the code.
8209 // Note that we cannot do the transformation unless we know that the
8210 // introduced loads cannot trap! Something like this is valid as long as
8211 // the condition is always false: load (select bool %C, int* null, int* %G),
8212 // but it would not be valid if we transformed it to load from null
8215 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8216 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8217 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8218 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8219 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8220 SI->getOperand(1)->getName()+".val"), LI);
8221 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8222 SI->getOperand(2)->getName()+".val"), LI);
8223 return new SelectInst(SI->getCondition(), V1, V2);
8226 // load (select (cond, null, P)) -> load P
8227 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8228 if (C->isNullValue()) {
8229 LI.setOperand(0, SI->getOperand(2));
8233 // load (select (cond, P, null)) -> load P
8234 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8235 if (C->isNullValue()) {
8236 LI.setOperand(0, SI->getOperand(1));
8244 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8246 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8247 User *CI = cast<User>(SI.getOperand(1));
8248 Value *CastOp = CI->getOperand(0);
8250 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8251 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8252 const Type *SrcPTy = SrcTy->getElementType();
8254 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8255 // If the source is an array, the code below will not succeed. Check to
8256 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8258 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8259 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8260 if (ASrcTy->getNumElements() != 0) {
8262 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8263 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8264 SrcTy = cast<PointerType>(CastOp->getType());
8265 SrcPTy = SrcTy->getElementType();
8268 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8269 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8270 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8272 // Okay, we are casting from one integer or pointer type to another of
8273 // the same size. Instead of casting the pointer before
8274 // the store, cast the value to be stored.
8276 Value *SIOp0 = SI.getOperand(0);
8277 Instruction::CastOps opcode = Instruction::BitCast;
8278 const Type* CastSrcTy = SIOp0->getType();
8279 const Type* CastDstTy = SrcPTy;
8280 if (isa<PointerType>(CastDstTy)) {
8281 if (CastSrcTy->isInteger())
8282 opcode = Instruction::IntToPtr;
8283 } else if (isa<IntegerType>(CastDstTy)) {
8284 if (isa<PointerType>(SIOp0->getType()))
8285 opcode = Instruction::PtrToInt;
8287 if (Constant *C = dyn_cast<Constant>(SIOp0))
8288 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8290 NewCast = IC.InsertNewInstBefore(
8291 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8293 return new StoreInst(NewCast, CastOp);
8300 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8301 Value *Val = SI.getOperand(0);
8302 Value *Ptr = SI.getOperand(1);
8304 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8305 EraseInstFromFunction(SI);
8310 // If the RHS is an alloca with a single use, zapify the store, making the
8312 if (Ptr->hasOneUse()) {
8313 if (isa<AllocaInst>(Ptr)) {
8314 EraseInstFromFunction(SI);
8319 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8320 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8321 GEP->getOperand(0)->hasOneUse()) {
8322 EraseInstFromFunction(SI);
8328 // Do really simple DSE, to catch cases where there are several consequtive
8329 // stores to the same location, separated by a few arithmetic operations. This
8330 // situation often occurs with bitfield accesses.
8331 BasicBlock::iterator BBI = &SI;
8332 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8336 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8337 // Prev store isn't volatile, and stores to the same location?
8338 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8341 EraseInstFromFunction(*PrevSI);
8347 // If this is a load, we have to stop. However, if the loaded value is from
8348 // the pointer we're loading and is producing the pointer we're storing,
8349 // then *this* store is dead (X = load P; store X -> P).
8350 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8351 if (LI == Val && LI->getOperand(0) == Ptr) {
8352 EraseInstFromFunction(SI);
8356 // Otherwise, this is a load from some other location. Stores before it
8361 // Don't skip over loads or things that can modify memory.
8362 if (BBI->mayWriteToMemory())
8367 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8369 // store X, null -> turns into 'unreachable' in SimplifyCFG
8370 if (isa<ConstantPointerNull>(Ptr)) {
8371 if (!isa<UndefValue>(Val)) {
8372 SI.setOperand(0, UndefValue::get(Val->getType()));
8373 if (Instruction *U = dyn_cast<Instruction>(Val))
8374 AddToWorkList(U); // Dropped a use.
8377 return 0; // Do not modify these!
8380 // store undef, Ptr -> noop
8381 if (isa<UndefValue>(Val)) {
8382 EraseInstFromFunction(SI);
8387 // If the pointer destination is a cast, see if we can fold the cast into the
8389 if (isa<CastInst>(Ptr))
8390 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8392 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8394 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8398 // If this store is the last instruction in the basic block, and if the block
8399 // ends with an unconditional branch, try to move it to the successor block.
8401 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8402 if (BI->isUnconditional()) {
8403 // Check to see if the successor block has exactly two incoming edges. If
8404 // so, see if the other predecessor contains a store to the same location.
8405 // if so, insert a PHI node (if needed) and move the stores down.
8406 BasicBlock *Dest = BI->getSuccessor(0);
8408 pred_iterator PI = pred_begin(Dest);
8409 BasicBlock *Other = 0;
8410 if (*PI != BI->getParent())
8413 if (PI != pred_end(Dest)) {
8414 if (*PI != BI->getParent())
8419 if (++PI != pred_end(Dest))
8422 if (Other) { // If only one other pred...
8423 BBI = Other->getTerminator();
8424 // Make sure this other block ends in an unconditional branch and that
8425 // there is an instruction before the branch.
8426 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8427 BBI != Other->begin()) {
8429 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8431 // If this instruction is a store to the same location.
8432 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8433 // Okay, we know we can perform this transformation. Insert a PHI
8434 // node now if we need it.
8435 Value *MergedVal = OtherStore->getOperand(0);
8436 if (MergedVal != SI.getOperand(0)) {
8437 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8438 PN->reserveOperandSpace(2);
8439 PN->addIncoming(SI.getOperand(0), SI.getParent());
8440 PN->addIncoming(OtherStore->getOperand(0), Other);
8441 MergedVal = InsertNewInstBefore(PN, Dest->front());
8444 // Advance to a place where it is safe to insert the new store and
8446 BBI = Dest->begin();
8447 while (isa<PHINode>(BBI)) ++BBI;
8448 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8449 OtherStore->isVolatile()), *BBI);
8451 // Nuke the old stores.
8452 EraseInstFromFunction(SI);
8453 EraseInstFromFunction(*OtherStore);
8465 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8466 // Change br (not X), label True, label False to: br X, label False, True
8468 BasicBlock *TrueDest;
8469 BasicBlock *FalseDest;
8470 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8471 !isa<Constant>(X)) {
8472 // Swap Destinations and condition...
8474 BI.setSuccessor(0, FalseDest);
8475 BI.setSuccessor(1, TrueDest);
8479 // Cannonicalize fcmp_one -> fcmp_oeq
8480 FCmpInst::Predicate FPred; Value *Y;
8481 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8482 TrueDest, FalseDest)))
8483 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8484 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8485 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8486 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8487 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
8488 NewSCC->takeName(I);
8489 // Swap Destinations and condition...
8490 BI.setCondition(NewSCC);
8491 BI.setSuccessor(0, FalseDest);
8492 BI.setSuccessor(1, TrueDest);
8493 RemoveFromWorkList(I);
8494 I->eraseFromParent();
8495 AddToWorkList(NewSCC);
8499 // Cannonicalize icmp_ne -> icmp_eq
8500 ICmpInst::Predicate IPred;
8501 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8502 TrueDest, FalseDest)))
8503 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8504 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8505 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8506 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8507 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8508 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
8509 NewSCC->takeName(I);
8510 // Swap Destinations and condition...
8511 BI.setCondition(NewSCC);
8512 BI.setSuccessor(0, FalseDest);
8513 BI.setSuccessor(1, TrueDest);
8514 RemoveFromWorkList(I);
8515 I->eraseFromParent();;
8516 AddToWorkList(NewSCC);
8523 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8524 Value *Cond = SI.getCondition();
8525 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8526 if (I->getOpcode() == Instruction::Add)
8527 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8528 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8529 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8530 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8532 SI.setOperand(0, I->getOperand(0));
8540 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8541 /// is to leave as a vector operation.
8542 static bool CheapToScalarize(Value *V, bool isConstant) {
8543 if (isa<ConstantAggregateZero>(V))
8545 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
8546 if (isConstant) return true;
8547 // If all elts are the same, we can extract.
8548 Constant *Op0 = C->getOperand(0);
8549 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8550 if (C->getOperand(i) != Op0)
8554 Instruction *I = dyn_cast<Instruction>(V);
8555 if (!I) return false;
8557 // Insert element gets simplified to the inserted element or is deleted if
8558 // this is constant idx extract element and its a constant idx insertelt.
8559 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8560 isa<ConstantInt>(I->getOperand(2)))
8562 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8564 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8565 if (BO->hasOneUse() &&
8566 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8567 CheapToScalarize(BO->getOperand(1), isConstant)))
8569 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8570 if (CI->hasOneUse() &&
8571 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8572 CheapToScalarize(CI->getOperand(1), isConstant)))
8578 /// Read and decode a shufflevector mask.
8580 /// It turns undef elements into values that are larger than the number of
8581 /// elements in the input.
8582 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8583 unsigned NElts = SVI->getType()->getNumElements();
8584 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8585 return std::vector<unsigned>(NElts, 0);
8586 if (isa<UndefValue>(SVI->getOperand(2)))
8587 return std::vector<unsigned>(NElts, 2*NElts);
8589 std::vector<unsigned> Result;
8590 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
8591 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8592 if (isa<UndefValue>(CP->getOperand(i)))
8593 Result.push_back(NElts*2); // undef -> 8
8595 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8599 /// FindScalarElement - Given a vector and an element number, see if the scalar
8600 /// value is already around as a register, for example if it were inserted then
8601 /// extracted from the vector.
8602 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8603 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
8604 const VectorType *PTy = cast<VectorType>(V->getType());
8605 unsigned Width = PTy->getNumElements();
8606 if (EltNo >= Width) // Out of range access.
8607 return UndefValue::get(PTy->getElementType());
8609 if (isa<UndefValue>(V))
8610 return UndefValue::get(PTy->getElementType());
8611 else if (isa<ConstantAggregateZero>(V))
8612 return Constant::getNullValue(PTy->getElementType());
8613 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
8614 return CP->getOperand(EltNo);
8615 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8616 // If this is an insert to a variable element, we don't know what it is.
8617 if (!isa<ConstantInt>(III->getOperand(2)))
8619 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8621 // If this is an insert to the element we are looking for, return the
8624 return III->getOperand(1);
8626 // Otherwise, the insertelement doesn't modify the value, recurse on its
8628 return FindScalarElement(III->getOperand(0), EltNo);
8629 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8630 unsigned InEl = getShuffleMask(SVI)[EltNo];
8632 return FindScalarElement(SVI->getOperand(0), InEl);
8633 else if (InEl < Width*2)
8634 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8636 return UndefValue::get(PTy->getElementType());
8639 // Otherwise, we don't know.
8643 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8645 // If packed val is undef, replace extract with scalar undef.
8646 if (isa<UndefValue>(EI.getOperand(0)))
8647 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8649 // If packed val is constant 0, replace extract with scalar 0.
8650 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8651 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8653 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
8654 // If packed val is constant with uniform operands, replace EI
8655 // with that operand
8656 Constant *op0 = C->getOperand(0);
8657 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8658 if (C->getOperand(i) != op0) {
8663 return ReplaceInstUsesWith(EI, op0);
8666 // If extracting a specified index from the vector, see if we can recursively
8667 // find a previously computed scalar that was inserted into the vector.
8668 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8669 // This instruction only demands the single element from the input vector.
8670 // If the input vector has a single use, simplify it based on this use
8672 uint64_t IndexVal = IdxC->getZExtValue();
8673 if (EI.getOperand(0)->hasOneUse()) {
8675 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8678 EI.setOperand(0, V);
8683 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8684 return ReplaceInstUsesWith(EI, Elt);
8687 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8688 if (I->hasOneUse()) {
8689 // Push extractelement into predecessor operation if legal and
8690 // profitable to do so
8691 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8692 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8693 if (CheapToScalarize(BO, isConstantElt)) {
8694 ExtractElementInst *newEI0 =
8695 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8696 EI.getName()+".lhs");
8697 ExtractElementInst *newEI1 =
8698 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8699 EI.getName()+".rhs");
8700 InsertNewInstBefore(newEI0, EI);
8701 InsertNewInstBefore(newEI1, EI);
8702 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8704 } else if (isa<LoadInst>(I)) {
8705 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8706 PointerType::get(EI.getType()), EI);
8707 GetElementPtrInst *GEP =
8708 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8709 InsertNewInstBefore(GEP, EI);
8710 return new LoadInst(GEP);
8713 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8714 // Extracting the inserted element?
8715 if (IE->getOperand(2) == EI.getOperand(1))
8716 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8717 // If the inserted and extracted elements are constants, they must not
8718 // be the same value, extract from the pre-inserted value instead.
8719 if (isa<Constant>(IE->getOperand(2)) &&
8720 isa<Constant>(EI.getOperand(1))) {
8721 AddUsesToWorkList(EI);
8722 EI.setOperand(0, IE->getOperand(0));
8725 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8726 // If this is extracting an element from a shufflevector, figure out where
8727 // it came from and extract from the appropriate input element instead.
8728 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8729 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8731 if (SrcIdx < SVI->getType()->getNumElements())
8732 Src = SVI->getOperand(0);
8733 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8734 SrcIdx -= SVI->getType()->getNumElements();
8735 Src = SVI->getOperand(1);
8737 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8739 return new ExtractElementInst(Src, SrcIdx);
8746 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8747 /// elements from either LHS or RHS, return the shuffle mask and true.
8748 /// Otherwise, return false.
8749 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8750 std::vector<Constant*> &Mask) {
8751 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8752 "Invalid CollectSingleShuffleElements");
8753 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8755 if (isa<UndefValue>(V)) {
8756 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8758 } else if (V == LHS) {
8759 for (unsigned i = 0; i != NumElts; ++i)
8760 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8762 } else if (V == RHS) {
8763 for (unsigned i = 0; i != NumElts; ++i)
8764 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8766 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8767 // If this is an insert of an extract from some other vector, include it.
8768 Value *VecOp = IEI->getOperand(0);
8769 Value *ScalarOp = IEI->getOperand(1);
8770 Value *IdxOp = IEI->getOperand(2);
8772 if (!isa<ConstantInt>(IdxOp))
8774 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8776 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8777 // Okay, we can handle this if the vector we are insertinting into is
8779 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8780 // If so, update the mask to reflect the inserted undef.
8781 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
8784 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8785 if (isa<ConstantInt>(EI->getOperand(1)) &&
8786 EI->getOperand(0)->getType() == V->getType()) {
8787 unsigned ExtractedIdx =
8788 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8790 // This must be extracting from either LHS or RHS.
8791 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8792 // Okay, we can handle this if the vector we are insertinting into is
8794 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8795 // If so, update the mask to reflect the inserted value.
8796 if (EI->getOperand(0) == LHS) {
8797 Mask[InsertedIdx & (NumElts-1)] =
8798 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8800 assert(EI->getOperand(0) == RHS);
8801 Mask[InsertedIdx & (NumElts-1)] =
8802 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
8811 // TODO: Handle shufflevector here!
8816 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8817 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8818 /// that computes V and the LHS value of the shuffle.
8819 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8821 assert(isa<VectorType>(V->getType()) &&
8822 (RHS == 0 || V->getType() == RHS->getType()) &&
8823 "Invalid shuffle!");
8824 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8826 if (isa<UndefValue>(V)) {
8827 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8829 } else if (isa<ConstantAggregateZero>(V)) {
8830 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
8832 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8833 // If this is an insert of an extract from some other vector, include it.
8834 Value *VecOp = IEI->getOperand(0);
8835 Value *ScalarOp = IEI->getOperand(1);
8836 Value *IdxOp = IEI->getOperand(2);
8838 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8839 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8840 EI->getOperand(0)->getType() == V->getType()) {
8841 unsigned ExtractedIdx =
8842 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8843 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8845 // Either the extracted from or inserted into vector must be RHSVec,
8846 // otherwise we'd end up with a shuffle of three inputs.
8847 if (EI->getOperand(0) == RHS || RHS == 0) {
8848 RHS = EI->getOperand(0);
8849 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8850 Mask[InsertedIdx & (NumElts-1)] =
8851 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
8856 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8857 // Everything but the extracted element is replaced with the RHS.
8858 for (unsigned i = 0; i != NumElts; ++i) {
8859 if (i != InsertedIdx)
8860 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
8865 // If this insertelement is a chain that comes from exactly these two
8866 // vectors, return the vector and the effective shuffle.
8867 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8868 return EI->getOperand(0);
8873 // TODO: Handle shufflevector here!
8875 // Otherwise, can't do anything fancy. Return an identity vector.
8876 for (unsigned i = 0; i != NumElts; ++i)
8877 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8881 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8882 Value *VecOp = IE.getOperand(0);
8883 Value *ScalarOp = IE.getOperand(1);
8884 Value *IdxOp = IE.getOperand(2);
8886 // If the inserted element was extracted from some other vector, and if the
8887 // indexes are constant, try to turn this into a shufflevector operation.
8888 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8889 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8890 EI->getOperand(0)->getType() == IE.getType()) {
8891 unsigned NumVectorElts = IE.getType()->getNumElements();
8892 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8893 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8895 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8896 return ReplaceInstUsesWith(IE, VecOp);
8898 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8899 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8901 // If we are extracting a value from a vector, then inserting it right
8902 // back into the same place, just use the input vector.
8903 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8904 return ReplaceInstUsesWith(IE, VecOp);
8906 // We could theoretically do this for ANY input. However, doing so could
8907 // turn chains of insertelement instructions into a chain of shufflevector
8908 // instructions, and right now we do not merge shufflevectors. As such,
8909 // only do this in a situation where it is clear that there is benefit.
8910 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8911 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8912 // the values of VecOp, except then one read from EIOp0.
8913 // Build a new shuffle mask.
8914 std::vector<Constant*> Mask;
8915 if (isa<UndefValue>(VecOp))
8916 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
8918 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8919 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
8922 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8923 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8924 ConstantVector::get(Mask));
8927 // If this insertelement isn't used by some other insertelement, turn it
8928 // (and any insertelements it points to), into one big shuffle.
8929 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8930 std::vector<Constant*> Mask;
8932 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8933 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8934 // We now have a shuffle of LHS, RHS, Mask.
8935 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
8944 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8945 Value *LHS = SVI.getOperand(0);
8946 Value *RHS = SVI.getOperand(1);
8947 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8949 bool MadeChange = false;
8951 // Undefined shuffle mask -> undefined value.
8952 if (isa<UndefValue>(SVI.getOperand(2)))
8953 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8955 // If we have shuffle(x, undef, mask) and any elements of mask refer to
8956 // the undef, change them to undefs.
8957 if (isa<UndefValue>(SVI.getOperand(1))) {
8958 // Scan to see if there are any references to the RHS. If so, replace them
8959 // with undef element refs and set MadeChange to true.
8960 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8961 if (Mask[i] >= e && Mask[i] != 2*e) {
8968 // Remap any references to RHS to use LHS.
8969 std::vector<Constant*> Elts;
8970 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8972 Elts.push_back(UndefValue::get(Type::Int32Ty));
8974 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8976 SVI.setOperand(2, ConstantVector::get(Elts));
8980 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8981 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8982 if (LHS == RHS || isa<UndefValue>(LHS)) {
8983 if (isa<UndefValue>(LHS) && LHS == RHS) {
8984 // shuffle(undef,undef,mask) -> undef.
8985 return ReplaceInstUsesWith(SVI, LHS);
8988 // Remap any references to RHS to use LHS.
8989 std::vector<Constant*> Elts;
8990 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8992 Elts.push_back(UndefValue::get(Type::Int32Ty));
8994 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8995 (Mask[i] < e && isa<UndefValue>(LHS)))
8996 Mask[i] = 2*e; // Turn into undef.
8998 Mask[i] &= (e-1); // Force to LHS.
8999 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9002 SVI.setOperand(0, SVI.getOperand(1));
9003 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9004 SVI.setOperand(2, ConstantVector::get(Elts));
9005 LHS = SVI.getOperand(0);
9006 RHS = SVI.getOperand(1);
9010 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9011 bool isLHSID = true, isRHSID = true;
9013 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9014 if (Mask[i] >= e*2) continue; // Ignore undef values.
9015 // Is this an identity shuffle of the LHS value?
9016 isLHSID &= (Mask[i] == i);
9018 // Is this an identity shuffle of the RHS value?
9019 isRHSID &= (Mask[i]-e == i);
9022 // Eliminate identity shuffles.
9023 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9024 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9026 // If the LHS is a shufflevector itself, see if we can combine it with this
9027 // one without producing an unusual shuffle. Here we are really conservative:
9028 // we are absolutely afraid of producing a shuffle mask not in the input
9029 // program, because the code gen may not be smart enough to turn a merged
9030 // shuffle into two specific shuffles: it may produce worse code. As such,
9031 // we only merge two shuffles if the result is one of the two input shuffle
9032 // masks. In this case, merging the shuffles just removes one instruction,
9033 // which we know is safe. This is good for things like turning:
9034 // (splat(splat)) -> splat.
9035 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9036 if (isa<UndefValue>(RHS)) {
9037 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9039 std::vector<unsigned> NewMask;
9040 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9042 NewMask.push_back(2*e);
9044 NewMask.push_back(LHSMask[Mask[i]]);
9046 // If the result mask is equal to the src shuffle or this shuffle mask, do
9048 if (NewMask == LHSMask || NewMask == Mask) {
9049 std::vector<Constant*> Elts;
9050 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9051 if (NewMask[i] >= e*2) {
9052 Elts.push_back(UndefValue::get(Type::Int32Ty));
9054 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9057 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9058 LHSSVI->getOperand(1),
9059 ConstantVector::get(Elts));
9064 return MadeChange ? &SVI : 0;
9070 /// TryToSinkInstruction - Try to move the specified instruction from its
9071 /// current block into the beginning of DestBlock, which can only happen if it's
9072 /// safe to move the instruction past all of the instructions between it and the
9073 /// end of its block.
9074 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9075 assert(I->hasOneUse() && "Invariants didn't hold!");
9077 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9078 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9080 // Do not sink alloca instructions out of the entry block.
9081 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
9084 // We can only sink load instructions if there is nothing between the load and
9085 // the end of block that could change the value.
9086 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9087 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9089 if (Scan->mayWriteToMemory())
9093 BasicBlock::iterator InsertPos = DestBlock->begin();
9094 while (isa<PHINode>(InsertPos)) ++InsertPos;
9096 I->moveBefore(InsertPos);
9102 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9103 /// all reachable code to the worklist.
9105 /// This has a couple of tricks to make the code faster and more powerful. In
9106 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9107 /// them to the worklist (this significantly speeds up instcombine on code where
9108 /// many instructions are dead or constant). Additionally, if we find a branch
9109 /// whose condition is a known constant, we only visit the reachable successors.
9111 static void AddReachableCodeToWorklist(BasicBlock *BB,
9112 SmallPtrSet<BasicBlock*, 64> &Visited,
9114 const TargetData *TD) {
9115 // We have now visited this block! If we've already been here, bail out.
9116 if (!Visited.insert(BB)) return;
9118 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9119 Instruction *Inst = BBI++;
9121 // DCE instruction if trivially dead.
9122 if (isInstructionTriviallyDead(Inst)) {
9124 DOUT << "IC: DCE: " << *Inst;
9125 Inst->eraseFromParent();
9129 // ConstantProp instruction if trivially constant.
9130 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9131 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9132 Inst->replaceAllUsesWith(C);
9134 Inst->eraseFromParent();
9138 IC.AddToWorkList(Inst);
9141 // Recursively visit successors. If this is a branch or switch on a constant,
9142 // only visit the reachable successor.
9143 TerminatorInst *TI = BB->getTerminator();
9144 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9145 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9146 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9147 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, IC, TD);
9150 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9151 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9152 // See if this is an explicit destination.
9153 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9154 if (SI->getCaseValue(i) == Cond) {
9155 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, IC, TD);
9159 // Otherwise it is the default destination.
9160 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, IC, TD);
9165 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9166 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, IC, TD);
9169 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
9170 bool Changed = false;
9171 TD = &getAnalysis<TargetData>();
9173 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
9174 << F.getNameStr() << "\n");
9177 // Do a depth-first traversal of the function, populate the worklist with
9178 // the reachable instructions. Ignore blocks that are not reachable. Keep
9179 // track of which blocks we visit.
9180 SmallPtrSet<BasicBlock*, 64> Visited;
9181 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
9183 // Do a quick scan over the function. If we find any blocks that are
9184 // unreachable, remove any instructions inside of them. This prevents
9185 // the instcombine code from having to deal with some bad special cases.
9186 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9187 if (!Visited.count(BB)) {
9188 Instruction *Term = BB->getTerminator();
9189 while (Term != BB->begin()) { // Remove instrs bottom-up
9190 BasicBlock::iterator I = Term; --I;
9192 DOUT << "IC: DCE: " << *I;
9195 if (!I->use_empty())
9196 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9197 I->eraseFromParent();
9202 while (!Worklist.empty()) {
9203 Instruction *I = RemoveOneFromWorkList();
9204 if (I == 0) continue; // skip null values.
9206 // Check to see if we can DCE the instruction.
9207 if (isInstructionTriviallyDead(I)) {
9208 // Add operands to the worklist.
9209 if (I->getNumOperands() < 4)
9210 AddUsesToWorkList(*I);
9213 DOUT << "IC: DCE: " << *I;
9215 I->eraseFromParent();
9216 RemoveFromWorkList(I);
9220 // Instruction isn't dead, see if we can constant propagate it.
9221 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9222 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9224 // Add operands to the worklist.
9225 AddUsesToWorkList(*I);
9226 ReplaceInstUsesWith(*I, C);
9229 I->eraseFromParent();
9230 RemoveFromWorkList(I);
9234 // See if we can trivially sink this instruction to a successor basic block.
9235 if (I->hasOneUse()) {
9236 BasicBlock *BB = I->getParent();
9237 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9238 if (UserParent != BB) {
9239 bool UserIsSuccessor = false;
9240 // See if the user is one of our successors.
9241 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9242 if (*SI == UserParent) {
9243 UserIsSuccessor = true;
9247 // If the user is one of our immediate successors, and if that successor
9248 // only has us as a predecessors (we'd have to split the critical edge
9249 // otherwise), we can keep going.
9250 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9251 next(pred_begin(UserParent)) == pred_end(UserParent))
9252 // Okay, the CFG is simple enough, try to sink this instruction.
9253 Changed |= TryToSinkInstruction(I, UserParent);
9257 // Now that we have an instruction, try combining it to simplify it...
9258 if (Instruction *Result = visit(*I)) {
9260 // Should we replace the old instruction with a new one?
9262 DOUT << "IC: Old = " << *I
9263 << " New = " << *Result;
9265 // Everything uses the new instruction now.
9266 I->replaceAllUsesWith(Result);
9268 // Push the new instruction and any users onto the worklist.
9269 AddToWorkList(Result);
9270 AddUsersToWorkList(*Result);
9272 // Move the name to the new instruction first.
9273 Result->takeName(I);
9275 // Insert the new instruction into the basic block...
9276 BasicBlock *InstParent = I->getParent();
9277 BasicBlock::iterator InsertPos = I;
9279 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9280 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9283 InstParent->getInstList().insert(InsertPos, Result);
9285 // Make sure that we reprocess all operands now that we reduced their
9287 AddUsesToWorkList(*I);
9289 // Instructions can end up on the worklist more than once. Make sure
9290 // we do not process an instruction that has been deleted.
9291 RemoveFromWorkList(I);
9293 // Erase the old instruction.
9294 InstParent->getInstList().erase(I);
9296 DOUT << "IC: MOD = " << *I;
9298 // If the instruction was modified, it's possible that it is now dead.
9299 // if so, remove it.
9300 if (isInstructionTriviallyDead(I)) {
9301 // Make sure we process all operands now that we are reducing their
9303 AddUsesToWorkList(*I);
9305 // Instructions may end up in the worklist more than once. Erase all
9306 // occurrences of this instruction.
9307 RemoveFromWorkList(I);
9308 I->eraseFromParent();
9311 AddUsersToWorkList(*I);
9318 assert(WorklistMap.empty() && "Worklist empty, but map not?");
9323 bool InstCombiner::runOnFunction(Function &F) {
9324 bool EverMadeChange = false;
9326 // Iterate while there is work to do.
9327 unsigned Iteration = 0;
9328 while (DoOneIteration(F, Iteration++))
9329 EverMadeChange = true;
9330 return EverMadeChange;
9333 FunctionPass *llvm::createInstructionCombiningPass() {
9334 return new InstCombiner();