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/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/STLExtras.h"
59 using namespace llvm::PatternMatch;
61 STATISTIC(NumCombined , "Number of insts combined");
62 STATISTIC(NumConstProp, "Number of constant folds");
63 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
64 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
65 STATISTIC(NumSunkInst , "Number of instructions sunk");
68 class VISIBILITY_HIDDEN InstCombiner
69 : public FunctionPass,
70 public InstVisitor<InstCombiner, Instruction*> {
71 // Worklist of all of the instructions that need to be simplified.
72 std::vector<Instruction*> WorkList;
75 /// AddUsersToWorkList - When an instruction is simplified, add all users of
76 /// the instruction to the work lists because they might get more simplified
79 void AddUsersToWorkList(Value &I) {
80 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
82 WorkList.push_back(cast<Instruction>(*UI));
85 /// AddUsesToWorkList - When an instruction is simplified, add operands to
86 /// the work lists because they might get more simplified now.
88 void AddUsesToWorkList(Instruction &I) {
89 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
90 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
91 WorkList.push_back(Op);
94 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
95 /// dead. Add all of its operands to the worklist, turning them into
96 /// undef's to reduce the number of uses of those instructions.
98 /// Return the specified operand before it is turned into an undef.
100 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
101 Value *R = I.getOperand(op);
103 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
104 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
105 WorkList.push_back(Op);
106 // Set the operand to undef to drop the use.
107 I.setOperand(i, UndefValue::get(Op->getType()));
113 // removeFromWorkList - remove all instances of I from the worklist.
114 void removeFromWorkList(Instruction *I);
116 virtual bool runOnFunction(Function &F);
118 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
119 AU.addRequired<TargetData>();
120 AU.addPreservedID(LCSSAID);
121 AU.setPreservesCFG();
124 TargetData &getTargetData() const { return *TD; }
126 // Visitation implementation - Implement instruction combining for different
127 // instruction types. The semantics are as follows:
129 // null - No change was made
130 // I - Change was made, I is still valid, I may be dead though
131 // otherwise - Change was made, replace I with returned instruction
133 Instruction *visitAdd(BinaryOperator &I);
134 Instruction *visitSub(BinaryOperator &I);
135 Instruction *visitMul(BinaryOperator &I);
136 Instruction *visitURem(BinaryOperator &I);
137 Instruction *visitSRem(BinaryOperator &I);
138 Instruction *visitFRem(BinaryOperator &I);
139 Instruction *commonRemTransforms(BinaryOperator &I);
140 Instruction *commonIRemTransforms(BinaryOperator &I);
141 Instruction *commonDivTransforms(BinaryOperator &I);
142 Instruction *commonIDivTransforms(BinaryOperator &I);
143 Instruction *visitUDiv(BinaryOperator &I);
144 Instruction *visitSDiv(BinaryOperator &I);
145 Instruction *visitFDiv(BinaryOperator &I);
146 Instruction *visitAnd(BinaryOperator &I);
147 Instruction *visitOr (BinaryOperator &I);
148 Instruction *visitXor(BinaryOperator &I);
149 Instruction *visitShl(BinaryOperator &I);
150 Instruction *visitAShr(BinaryOperator &I);
151 Instruction *visitLShr(BinaryOperator &I);
152 Instruction *commonShiftTransforms(BinaryOperator &I);
153 Instruction *visitFCmpInst(FCmpInst &I);
154 Instruction *visitICmpInst(ICmpInst &I);
155 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
157 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
158 ICmpInst::Predicate Cond, Instruction &I);
159 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
161 Instruction *commonCastTransforms(CastInst &CI);
162 Instruction *commonIntCastTransforms(CastInst &CI);
163 Instruction *visitTrunc(CastInst &CI);
164 Instruction *visitZExt(CastInst &CI);
165 Instruction *visitSExt(CastInst &CI);
166 Instruction *visitFPTrunc(CastInst &CI);
167 Instruction *visitFPExt(CastInst &CI);
168 Instruction *visitFPToUI(CastInst &CI);
169 Instruction *visitFPToSI(CastInst &CI);
170 Instruction *visitUIToFP(CastInst &CI);
171 Instruction *visitSIToFP(CastInst &CI);
172 Instruction *visitPtrToInt(CastInst &CI);
173 Instruction *visitIntToPtr(CastInst &CI);
174 Instruction *visitBitCast(CastInst &CI);
175 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
177 Instruction *visitSelectInst(SelectInst &CI);
178 Instruction *visitCallInst(CallInst &CI);
179 Instruction *visitInvokeInst(InvokeInst &II);
180 Instruction *visitPHINode(PHINode &PN);
181 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
182 Instruction *visitAllocationInst(AllocationInst &AI);
183 Instruction *visitFreeInst(FreeInst &FI);
184 Instruction *visitLoadInst(LoadInst &LI);
185 Instruction *visitStoreInst(StoreInst &SI);
186 Instruction *visitBranchInst(BranchInst &BI);
187 Instruction *visitSwitchInst(SwitchInst &SI);
188 Instruction *visitInsertElementInst(InsertElementInst &IE);
189 Instruction *visitExtractElementInst(ExtractElementInst &EI);
190 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
192 // visitInstruction - Specify what to return for unhandled instructions...
193 Instruction *visitInstruction(Instruction &I) { return 0; }
196 Instruction *visitCallSite(CallSite CS);
197 bool transformConstExprCastCall(CallSite CS);
200 // InsertNewInstBefore - insert an instruction New before instruction Old
201 // in the program. Add the new instruction to the worklist.
203 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
204 assert(New && New->getParent() == 0 &&
205 "New instruction already inserted into a basic block!");
206 BasicBlock *BB = Old.getParent();
207 BB->getInstList().insert(&Old, New); // Insert inst
208 WorkList.push_back(New); // Add to worklist
212 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
213 /// This also adds the cast to the worklist. Finally, this returns the
215 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
217 if (V->getType() == Ty) return V;
219 if (Constant *CV = dyn_cast<Constant>(V))
220 return ConstantExpr::getCast(opc, CV, Ty);
222 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
223 WorkList.push_back(C);
227 // ReplaceInstUsesWith - This method is to be used when an instruction is
228 // found to be dead, replacable with another preexisting expression. Here
229 // we add all uses of I to the worklist, replace all uses of I with the new
230 // value, then return I, so that the inst combiner will know that I was
233 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
234 AddUsersToWorkList(I); // Add all modified instrs to worklist
236 I.replaceAllUsesWith(V);
239 // If we are replacing the instruction with itself, this must be in a
240 // segment of unreachable code, so just clobber the instruction.
241 I.replaceAllUsesWith(UndefValue::get(I.getType()));
246 // UpdateValueUsesWith - This method is to be used when an value is
247 // found to be replacable with another preexisting expression or was
248 // updated. Here we add all uses of I to the worklist, replace all uses of
249 // I with the new value (unless the instruction was just updated), then
250 // return true, so that the inst combiner will know that I was modified.
252 bool UpdateValueUsesWith(Value *Old, Value *New) {
253 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
255 Old->replaceAllUsesWith(New);
256 if (Instruction *I = dyn_cast<Instruction>(Old))
257 WorkList.push_back(I);
258 if (Instruction *I = dyn_cast<Instruction>(New))
259 WorkList.push_back(I);
263 // EraseInstFromFunction - When dealing with an instruction that has side
264 // effects or produces a void value, we can't rely on DCE to delete the
265 // instruction. Instead, visit methods should return the value returned by
267 Instruction *EraseInstFromFunction(Instruction &I) {
268 assert(I.use_empty() && "Cannot erase instruction that is used!");
269 AddUsesToWorkList(I);
270 removeFromWorkList(&I);
272 return 0; // Don't do anything with FI
276 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
277 /// InsertBefore instruction. This is specialized a bit to avoid inserting
278 /// casts that are known to not do anything...
280 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
281 Value *V, const Type *DestTy,
282 Instruction *InsertBefore);
284 /// SimplifyCommutative - This performs a few simplifications for
285 /// commutative operators.
286 bool SimplifyCommutative(BinaryOperator &I);
288 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
289 /// most-complex to least-complex order.
290 bool SimplifyCompare(CmpInst &I);
292 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
293 uint64_t &KnownZero, uint64_t &KnownOne,
296 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
297 uint64_t &UndefElts, unsigned Depth = 0);
299 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
300 // PHI node as operand #0, see if we can fold the instruction into the PHI
301 // (which is only possible if all operands to the PHI are constants).
302 Instruction *FoldOpIntoPhi(Instruction &I);
304 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
305 // operator and they all are only used by the PHI, PHI together their
306 // inputs, and do the operation once, to the result of the PHI.
307 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
308 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
311 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
312 ConstantInt *AndRHS, BinaryOperator &TheAnd);
314 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
315 bool isSub, Instruction &I);
316 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
317 bool isSigned, bool Inside, Instruction &IB);
318 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
319 Instruction *MatchBSwap(BinaryOperator &I);
321 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
324 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
327 // getComplexity: Assign a complexity or rank value to LLVM Values...
328 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
329 static unsigned getComplexity(Value *V) {
330 if (isa<Instruction>(V)) {
331 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
335 if (isa<Argument>(V)) return 3;
336 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
339 // isOnlyUse - Return true if this instruction will be deleted if we stop using
341 static bool isOnlyUse(Value *V) {
342 return V->hasOneUse() || isa<Constant>(V);
345 // getPromotedType - Return the specified type promoted as it would be to pass
346 // though a va_arg area...
347 static const Type *getPromotedType(const Type *Ty) {
348 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
349 if (ITy->getBitWidth() < 32)
350 return Type::Int32Ty;
351 } else if (Ty == Type::FloatTy)
352 return Type::DoubleTy;
356 /// getBitCastOperand - If the specified operand is a CastInst or a constant
357 /// expression bitcast, return the operand value, otherwise return null.
358 static Value *getBitCastOperand(Value *V) {
359 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
360 return I->getOperand(0);
361 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
362 if (CE->getOpcode() == Instruction::BitCast)
363 return CE->getOperand(0);
367 /// This function is a wrapper around CastInst::isEliminableCastPair. It
368 /// simply extracts arguments and returns what that function returns.
369 static Instruction::CastOps
370 isEliminableCastPair(
371 const CastInst *CI, ///< The first cast instruction
372 unsigned opcode, ///< The opcode of the second cast instruction
373 const Type *DstTy, ///< The target type for the second cast instruction
374 TargetData *TD ///< The target data for pointer size
377 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
378 const Type *MidTy = CI->getType(); // B from above
380 // Get the opcodes of the two Cast instructions
381 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
382 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
384 return Instruction::CastOps(
385 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
386 DstTy, TD->getIntPtrType()));
389 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
390 /// in any code being generated. It does not require codegen if V is simple
391 /// enough or if the cast can be folded into other casts.
392 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
393 const Type *Ty, TargetData *TD) {
394 if (V->getType() == Ty || isa<Constant>(V)) return false;
396 // If this is another cast that can be eliminated, it isn't codegen either.
397 if (const CastInst *CI = dyn_cast<CastInst>(V))
398 if (isEliminableCastPair(CI, opcode, Ty, TD))
403 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
404 /// InsertBefore instruction. This is specialized a bit to avoid inserting
405 /// casts that are known to not do anything...
407 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
408 Value *V, const Type *DestTy,
409 Instruction *InsertBefore) {
410 if (V->getType() == DestTy) return V;
411 if (Constant *C = dyn_cast<Constant>(V))
412 return ConstantExpr::getCast(opcode, C, DestTy);
414 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
417 // SimplifyCommutative - This performs a few simplifications for commutative
420 // 1. Order operands such that they are listed from right (least complex) to
421 // left (most complex). This puts constants before unary operators before
424 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
425 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
427 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
428 bool Changed = false;
429 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
430 Changed = !I.swapOperands();
432 if (!I.isAssociative()) return Changed;
433 Instruction::BinaryOps Opcode = I.getOpcode();
434 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
435 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
436 if (isa<Constant>(I.getOperand(1))) {
437 Constant *Folded = ConstantExpr::get(I.getOpcode(),
438 cast<Constant>(I.getOperand(1)),
439 cast<Constant>(Op->getOperand(1)));
440 I.setOperand(0, Op->getOperand(0));
441 I.setOperand(1, Folded);
443 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
444 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
445 isOnlyUse(Op) && isOnlyUse(Op1)) {
446 Constant *C1 = cast<Constant>(Op->getOperand(1));
447 Constant *C2 = cast<Constant>(Op1->getOperand(1));
449 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
450 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
451 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
454 WorkList.push_back(New);
455 I.setOperand(0, New);
456 I.setOperand(1, Folded);
463 /// SimplifyCompare - For a CmpInst this function just orders the operands
464 /// so that theyare listed from right (least complex) to left (most complex).
465 /// This puts constants before unary operators before binary operators.
466 bool InstCombiner::SimplifyCompare(CmpInst &I) {
467 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
470 // Compare instructions are not associative so there's nothing else we can do.
474 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
475 // if the LHS is a constant zero (which is the 'negate' form).
477 static inline Value *dyn_castNegVal(Value *V) {
478 if (BinaryOperator::isNeg(V))
479 return BinaryOperator::getNegArgument(V);
481 // Constants can be considered to be negated values if they can be folded.
482 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
483 return ConstantExpr::getNeg(C);
487 static inline Value *dyn_castNotVal(Value *V) {
488 if (BinaryOperator::isNot(V))
489 return BinaryOperator::getNotArgument(V);
491 // Constants can be considered to be not'ed values...
492 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
493 return ConstantExpr::getNot(C);
497 // dyn_castFoldableMul - If this value is a multiply that can be folded into
498 // other computations (because it has a constant operand), return the
499 // non-constant operand of the multiply, and set CST to point to the multiplier.
500 // Otherwise, return null.
502 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
503 if (V->hasOneUse() && V->getType()->isInteger())
504 if (Instruction *I = dyn_cast<Instruction>(V)) {
505 if (I->getOpcode() == Instruction::Mul)
506 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
507 return I->getOperand(0);
508 if (I->getOpcode() == Instruction::Shl)
509 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
510 // The multiplier is really 1 << CST.
511 Constant *One = ConstantInt::get(V->getType(), 1);
512 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
513 return I->getOperand(0);
519 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
520 /// expression, return it.
521 static User *dyn_castGetElementPtr(Value *V) {
522 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
523 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
524 if (CE->getOpcode() == Instruction::GetElementPtr)
525 return cast<User>(V);
529 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
530 static ConstantInt *AddOne(ConstantInt *C) {
531 return cast<ConstantInt>(ConstantExpr::getAdd(C,
532 ConstantInt::get(C->getType(), 1)));
534 static ConstantInt *SubOne(ConstantInt *C) {
535 return cast<ConstantInt>(ConstantExpr::getSub(C,
536 ConstantInt::get(C->getType(), 1)));
539 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
540 /// known to be either zero or one and return them in the KnownZero/KnownOne
541 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
543 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
544 uint64_t &KnownOne, unsigned Depth = 0) {
545 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
546 // we cannot optimize based on the assumption that it is zero without changing
547 // it to be an explicit zero. If we don't change it to zero, other code could
548 // optimized based on the contradictory assumption that it is non-zero.
549 // Because instcombine aggressively folds operations with undef args anyway,
550 // this won't lose us code quality.
551 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
552 // We know all of the bits for a constant!
553 KnownOne = CI->getZExtValue() & Mask;
554 KnownZero = ~KnownOne & Mask;
558 KnownZero = KnownOne = 0; // Don't know anything.
559 if (Depth == 6 || Mask == 0)
560 return; // Limit search depth.
562 uint64_t KnownZero2, KnownOne2;
563 Instruction *I = dyn_cast<Instruction>(V);
566 Mask &= cast<IntegerType>(V->getType())->getBitMask();
568 switch (I->getOpcode()) {
569 case Instruction::And:
570 // If either the LHS or the RHS are Zero, the result is zero.
571 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
573 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
574 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
575 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
577 // Output known-1 bits are only known if set in both the LHS & RHS.
578 KnownOne &= KnownOne2;
579 // Output known-0 are known to be clear if zero in either the LHS | RHS.
580 KnownZero |= KnownZero2;
582 case Instruction::Or:
583 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
585 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
586 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
587 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
589 // Output known-0 bits are only known if clear in both the LHS & RHS.
590 KnownZero &= KnownZero2;
591 // Output known-1 are known to be set if set in either the LHS | RHS.
592 KnownOne |= KnownOne2;
594 case Instruction::Xor: {
595 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
596 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
597 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
598 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
600 // Output known-0 bits are known if clear or set in both the LHS & RHS.
601 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
602 // Output known-1 are known to be set if set in only one of the LHS, RHS.
603 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
604 KnownZero = KnownZeroOut;
607 case Instruction::Select:
608 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
609 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
610 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
611 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
613 // Only known if known in both the LHS and RHS.
614 KnownOne &= KnownOne2;
615 KnownZero &= KnownZero2;
617 case Instruction::FPTrunc:
618 case Instruction::FPExt:
619 case Instruction::FPToUI:
620 case Instruction::FPToSI:
621 case Instruction::SIToFP:
622 case Instruction::PtrToInt:
623 case Instruction::UIToFP:
624 case Instruction::IntToPtr:
625 return; // Can't work with floating point or pointers
626 case Instruction::Trunc:
627 // All these have integer operands
628 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
630 case Instruction::BitCast: {
631 const Type *SrcTy = I->getOperand(0)->getType();
632 if (SrcTy->isInteger()) {
633 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
638 case Instruction::ZExt: {
639 // Compute the bits in the result that are not present in the input.
640 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
641 uint64_t NotIn = ~SrcTy->getBitMask();
642 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
644 Mask &= SrcTy->getBitMask();
645 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
646 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
647 // The top bits are known to be zero.
648 KnownZero |= NewBits;
651 case Instruction::SExt: {
652 // Compute the bits in the result that are not present in the input.
653 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
654 uint64_t NotIn = ~SrcTy->getBitMask();
655 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
657 Mask &= SrcTy->getBitMask();
658 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
659 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
661 // If the sign bit of the input is known set or clear, then we know the
662 // top bits of the result.
663 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
664 if (KnownZero & InSignBit) { // Input sign bit known zero
665 KnownZero |= NewBits;
666 KnownOne &= ~NewBits;
667 } else if (KnownOne & InSignBit) { // Input sign bit known set
669 KnownZero &= ~NewBits;
670 } else { // Input sign bit unknown
671 KnownZero &= ~NewBits;
672 KnownOne &= ~NewBits;
676 case Instruction::Shl:
677 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
678 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
679 uint64_t ShiftAmt = SA->getZExtValue();
681 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
682 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
683 KnownZero <<= ShiftAmt;
684 KnownOne <<= ShiftAmt;
685 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
689 case Instruction::LShr:
690 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
691 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
692 // Compute the new bits that are at the top now.
693 uint64_t ShiftAmt = SA->getZExtValue();
694 uint64_t HighBits = (1ULL << ShiftAmt)-1;
695 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
697 // Unsigned shift right.
699 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
700 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
701 KnownZero >>= ShiftAmt;
702 KnownOne >>= ShiftAmt;
703 KnownZero |= HighBits; // high bits known zero.
707 case Instruction::AShr:
708 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
709 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
710 // Compute the new bits that are at the top now.
711 uint64_t ShiftAmt = SA->getZExtValue();
712 uint64_t HighBits = (1ULL << ShiftAmt)-1;
713 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
715 // Signed shift right.
717 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
718 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
719 KnownZero >>= ShiftAmt;
720 KnownOne >>= ShiftAmt;
722 // Handle the sign bits.
723 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
724 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
726 if (KnownZero & SignBit) { // New bits are known zero.
727 KnownZero |= HighBits;
728 } else if (KnownOne & SignBit) { // New bits are known one.
729 KnownOne |= HighBits;
737 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
738 /// this predicate to simplify operations downstream. Mask is known to be zero
739 /// for bits that V cannot have.
740 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
741 uint64_t KnownZero, KnownOne;
742 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
743 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
744 return (KnownZero & Mask) == Mask;
747 /// ShrinkDemandedConstant - Check to see if the specified operand of the
748 /// specified instruction is a constant integer. If so, check to see if there
749 /// are any bits set in the constant that are not demanded. If so, shrink the
750 /// constant and return true.
751 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
753 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
754 if (!OpC) return false;
756 // If there are no bits set that aren't demanded, nothing to do.
757 if ((~Demanded & OpC->getZExtValue()) == 0)
760 // This is producing any bits that are not needed, shrink the RHS.
761 uint64_t Val = Demanded & OpC->getZExtValue();
762 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Val));
766 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
767 // set of known zero and one bits, compute the maximum and minimum values that
768 // could have the specified known zero and known one bits, returning them in
770 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
773 int64_t &Min, int64_t &Max) {
774 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
775 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
777 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
779 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
780 // bit if it is unknown.
782 Max = KnownOne|UnknownBits;
784 if (SignBit & UnknownBits) { // Sign bit is unknown
789 // Sign extend the min/max values.
790 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
791 Min = (Min << ShAmt) >> ShAmt;
792 Max = (Max << ShAmt) >> ShAmt;
795 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
796 // a 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 ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
804 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
805 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
807 // The minimum value is when the unknown bits are all zeros.
809 // The maximum value is when the unknown bits are all ones.
810 Max = KnownOne|UnknownBits;
814 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
815 /// DemandedMask bits of the result of V are ever used downstream. If we can
816 /// use this information to simplify V, do so and return true. Otherwise,
817 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
818 /// the expression (used to simplify the caller). The KnownZero/One bits may
819 /// only be accurate for those bits in the DemandedMask.
820 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
821 uint64_t &KnownZero, uint64_t &KnownOne,
823 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
824 // We know all of the bits for a constant!
825 KnownOne = CI->getZExtValue() & DemandedMask;
826 KnownZero = ~KnownOne & DemandedMask;
830 KnownZero = KnownOne = 0;
831 if (!V->hasOneUse()) { // Other users may use these bits.
832 if (Depth != 0) { // Not at the root.
833 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
834 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
837 // If this is the root being simplified, allow it to have multiple uses,
838 // just set the DemandedMask to all bits.
839 DemandedMask = cast<IntegerType>(V->getType())->getBitMask();
840 } else if (DemandedMask == 0) { // Not demanding any bits from V.
841 if (V != UndefValue::get(V->getType()))
842 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
844 } else if (Depth == 6) { // Limit search depth.
848 Instruction *I = dyn_cast<Instruction>(V);
849 if (!I) return false; // Only analyze instructions.
851 DemandedMask &= cast<IntegerType>(V->getType())->getBitMask();
853 uint64_t KnownZero2 = 0, KnownOne2 = 0;
854 switch (I->getOpcode()) {
856 case Instruction::And:
857 // If either the LHS or the RHS are Zero, the result is zero.
858 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
859 KnownZero, KnownOne, Depth+1))
861 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
863 // If something is known zero on the RHS, the bits aren't demanded on the
865 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
866 KnownZero2, KnownOne2, Depth+1))
868 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
870 // If all of the demanded bits are known 1 on one side, return the other.
871 // These bits cannot contribute to the result of the 'and'.
872 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
873 return UpdateValueUsesWith(I, I->getOperand(0));
874 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
875 return UpdateValueUsesWith(I, I->getOperand(1));
877 // If all of the demanded bits in the inputs are known zeros, return zero.
878 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
879 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
881 // If the RHS is a constant, see if we can simplify it.
882 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
883 return UpdateValueUsesWith(I, I);
885 // Output known-1 bits are only known if set in both the LHS & RHS.
886 KnownOne &= KnownOne2;
887 // Output known-0 are known to be clear if zero in either the LHS | RHS.
888 KnownZero |= KnownZero2;
890 case Instruction::Or:
891 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
892 KnownZero, KnownOne, Depth+1))
894 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
895 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
896 KnownZero2, KnownOne2, Depth+1))
898 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
900 // If all of the demanded bits are known zero on one side, return the other.
901 // These bits cannot contribute to the result of the 'or'.
902 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
903 return UpdateValueUsesWith(I, I->getOperand(0));
904 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
905 return UpdateValueUsesWith(I, I->getOperand(1));
907 // If all of the potentially set bits on one side are known to be set on
908 // the other side, just use the 'other' side.
909 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
910 (DemandedMask & (~KnownZero)))
911 return UpdateValueUsesWith(I, I->getOperand(0));
912 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
913 (DemandedMask & (~KnownZero2)))
914 return UpdateValueUsesWith(I, I->getOperand(1));
916 // If the RHS is a constant, see if we can simplify it.
917 if (ShrinkDemandedConstant(I, 1, DemandedMask))
918 return UpdateValueUsesWith(I, I);
920 // Output known-0 bits are only known if clear in both the LHS & RHS.
921 KnownZero &= KnownZero2;
922 // Output known-1 are known to be set if set in either the LHS | RHS.
923 KnownOne |= KnownOne2;
925 case Instruction::Xor: {
926 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
927 KnownZero, KnownOne, Depth+1))
929 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
930 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
931 KnownZero2, KnownOne2, Depth+1))
933 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
935 // If all of the demanded bits are known zero on one side, return the other.
936 // These bits cannot contribute to the result of the 'xor'.
937 if ((DemandedMask & KnownZero) == DemandedMask)
938 return UpdateValueUsesWith(I, I->getOperand(0));
939 if ((DemandedMask & KnownZero2) == DemandedMask)
940 return UpdateValueUsesWith(I, I->getOperand(1));
942 // Output known-0 bits are known if clear or set in both the LHS & RHS.
943 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
944 // Output known-1 are known to be set if set in only one of the LHS, RHS.
945 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
947 // If all of the demanded bits are known to be zero on one side or the
948 // other, turn this into an *inclusive* or.
949 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
950 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
952 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
954 InsertNewInstBefore(Or, *I);
955 return UpdateValueUsesWith(I, Or);
958 // If all of the demanded bits on one side are known, and all of the set
959 // bits on that side are also known to be set on the other side, turn this
960 // into an AND, as we know the bits will be cleared.
961 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
962 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
963 if ((KnownOne & KnownOne2) == KnownOne) {
964 Constant *AndC = ConstantInt::get(I->getType(),
965 ~KnownOne & DemandedMask);
967 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
968 InsertNewInstBefore(And, *I);
969 return UpdateValueUsesWith(I, And);
973 // If the RHS is a constant, see if we can simplify it.
974 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
975 if (ShrinkDemandedConstant(I, 1, DemandedMask))
976 return UpdateValueUsesWith(I, I);
978 KnownZero = KnownZeroOut;
979 KnownOne = KnownOneOut;
982 case Instruction::Select:
983 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
984 KnownZero, KnownOne, Depth+1))
986 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
987 KnownZero2, KnownOne2, Depth+1))
989 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
990 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
992 // If the operands are constants, see if we can simplify them.
993 if (ShrinkDemandedConstant(I, 1, DemandedMask))
994 return UpdateValueUsesWith(I, I);
995 if (ShrinkDemandedConstant(I, 2, DemandedMask))
996 return UpdateValueUsesWith(I, I);
998 // Only known if known in both the LHS and RHS.
999 KnownOne &= KnownOne2;
1000 KnownZero &= KnownZero2;
1002 case Instruction::Trunc:
1003 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1004 KnownZero, KnownOne, Depth+1))
1006 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1008 case Instruction::BitCast:
1009 if (!I->getOperand(0)->getType()->isInteger())
1012 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1013 KnownZero, KnownOne, Depth+1))
1015 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1017 case Instruction::ZExt: {
1018 // Compute the bits in the result that are not present in the input.
1019 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1020 uint64_t NotIn = ~SrcTy->getBitMask();
1021 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
1023 DemandedMask &= SrcTy->getBitMask();
1024 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1025 KnownZero, KnownOne, Depth+1))
1027 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1028 // The top bits are known to be zero.
1029 KnownZero |= NewBits;
1032 case Instruction::SExt: {
1033 // Compute the bits in the result that are not present in the input.
1034 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1035 uint64_t NotIn = ~SrcTy->getBitMask();
1036 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
1038 // Get the sign bit for the source type
1039 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1040 int64_t InputDemandedBits = DemandedMask & SrcTy->getBitMask();
1042 // If any of the sign extended bits are demanded, we know that the sign
1044 if (NewBits & DemandedMask)
1045 InputDemandedBits |= InSignBit;
1047 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1048 KnownZero, KnownOne, Depth+1))
1050 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1052 // If the sign bit of the input is known set or clear, then we know the
1053 // top bits of the result.
1055 // If the input sign bit is known zero, or if the NewBits are not demanded
1056 // convert this into a zero extension.
1057 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1058 // Convert to ZExt cast
1059 CastInst *NewCast = CastInst::create(
1060 Instruction::ZExt, I->getOperand(0), I->getType(), I->getName(), I);
1061 return UpdateValueUsesWith(I, NewCast);
1062 } else if (KnownOne & InSignBit) { // Input sign bit known set
1063 KnownOne |= NewBits;
1064 KnownZero &= ~NewBits;
1065 } else { // Input sign bit unknown
1066 KnownZero &= ~NewBits;
1067 KnownOne &= ~NewBits;
1071 case Instruction::Add:
1072 // If there is a constant on the RHS, there are a variety of xformations
1074 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1075 // If null, this should be simplified elsewhere. Some of the xforms here
1076 // won't work if the RHS is zero.
1077 if (RHS->isNullValue())
1080 // Figure out what the input bits are. If the top bits of the and result
1081 // are not demanded, then the add doesn't demand them from its input
1084 // Shift the demanded mask up so that it's at the top of the uint64_t.
1085 unsigned BitWidth = I->getType()->getPrimitiveSizeInBits();
1086 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1088 // If the top bit of the output is demanded, demand everything from the
1089 // input. Otherwise, we demand all the input bits except NLZ top bits.
1090 uint64_t InDemandedBits = ~0ULL >> (64-BitWidth+NLZ);
1092 // Find information about known zero/one bits in the input.
1093 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1094 KnownZero2, KnownOne2, Depth+1))
1097 // If the RHS of the add has bits set that can't affect the input, reduce
1099 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1100 return UpdateValueUsesWith(I, I);
1102 // Avoid excess work.
1103 if (KnownZero2 == 0 && KnownOne2 == 0)
1106 // Turn it into OR if input bits are zero.
1107 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1109 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1111 InsertNewInstBefore(Or, *I);
1112 return UpdateValueUsesWith(I, Or);
1115 // We can say something about the output known-zero and known-one bits,
1116 // depending on potential carries from the input constant and the
1117 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1118 // bits set and the RHS constant is 0x01001, then we know we have a known
1119 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1121 // To compute this, we first compute the potential carry bits. These are
1122 // the bits which may be modified. I'm not aware of a better way to do
1124 uint64_t RHSVal = RHS->getZExtValue();
1126 bool CarryIn = false;
1127 uint64_t CarryBits = 0;
1128 uint64_t CurBit = 1;
1129 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1130 // Record the current carry in.
1131 if (CarryIn) CarryBits |= CurBit;
1135 // This bit has a carry out unless it is "zero + zero" or
1136 // "zero + anything" with no carry in.
1137 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1138 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1139 } else if (!CarryIn &&
1140 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1141 CarryOut = false; // 0 + anything has no carry out if no carry in.
1143 // Otherwise, we have to assume we have a carry out.
1147 // This stage's carry out becomes the next stage's carry-in.
1151 // Now that we know which bits have carries, compute the known-1/0 sets.
1153 // Bits are known one if they are known zero in one operand and one in the
1154 // other, and there is no input carry.
1155 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1157 // Bits are known zero if they are known zero in both operands and there
1158 // is no input carry.
1159 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1162 case Instruction::Shl:
1163 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1164 uint64_t ShiftAmt = SA->getZExtValue();
1165 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1166 KnownZero, KnownOne, Depth+1))
1168 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1169 KnownZero <<= ShiftAmt;
1170 KnownOne <<= ShiftAmt;
1171 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1174 case Instruction::LShr:
1175 // For a logical shift right
1176 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1177 unsigned ShiftAmt = SA->getZExtValue();
1179 // Compute the new bits that are at the top now.
1180 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1181 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1182 uint64_t TypeMask = cast<IntegerType>(I->getType())->getBitMask();
1183 // Unsigned shift right.
1184 if (SimplifyDemandedBits(I->getOperand(0),
1185 (DemandedMask << ShiftAmt) & TypeMask,
1186 KnownZero, KnownOne, Depth+1))
1188 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1189 KnownZero &= TypeMask;
1190 KnownOne &= TypeMask;
1191 KnownZero >>= ShiftAmt;
1192 KnownOne >>= ShiftAmt;
1193 KnownZero |= HighBits; // high bits known zero.
1196 case Instruction::AShr:
1197 // If this is an arithmetic shift right and only the low-bit is set, we can
1198 // always convert this into a logical shr, even if the shift amount is
1199 // variable. The low bit of the shift cannot be an input sign bit unless
1200 // the shift amount is >= the size of the datatype, which is undefined.
1201 if (DemandedMask == 1) {
1202 // Perform the logical shift right.
1203 Value *NewVal = BinaryOperator::createLShr(
1204 I->getOperand(0), I->getOperand(1), I->getName());
1205 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1206 return UpdateValueUsesWith(I, NewVal);
1209 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1210 unsigned ShiftAmt = SA->getZExtValue();
1212 // Compute the new bits that are at the top now.
1213 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1214 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1215 uint64_t TypeMask = cast<IntegerType>(I->getType())->getBitMask();
1216 // Signed shift right.
1217 if (SimplifyDemandedBits(I->getOperand(0),
1218 (DemandedMask << ShiftAmt) & TypeMask,
1219 KnownZero, KnownOne, Depth+1))
1221 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1222 KnownZero &= TypeMask;
1223 KnownOne &= TypeMask;
1224 KnownZero >>= ShiftAmt;
1225 KnownOne >>= ShiftAmt;
1227 // Handle the sign bits.
1228 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1229 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1231 // If the input sign bit is known to be zero, or if none of the top bits
1232 // are demanded, turn this into an unsigned shift right.
1233 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1234 // Perform the logical shift right.
1235 Value *NewVal = BinaryOperator::createLShr(
1236 I->getOperand(0), SA, I->getName());
1237 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1238 return UpdateValueUsesWith(I, NewVal);
1239 } else if (KnownOne & SignBit) { // New bits are known one.
1240 KnownOne |= HighBits;
1246 // If the client is only demanding bits that we know, return the known
1248 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1249 return UpdateValueUsesWith(I, ConstantInt::get(I->getType(), KnownOne));
1254 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1255 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1256 /// actually used by the caller. This method analyzes which elements of the
1257 /// operand are undef and returns that information in UndefElts.
1259 /// If the information about demanded elements can be used to simplify the
1260 /// operation, the operation is simplified, then the resultant value is
1261 /// returned. This returns null if no change was made.
1262 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1263 uint64_t &UndefElts,
1265 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1266 assert(VWidth <= 64 && "Vector too wide to analyze!");
1267 uint64_t EltMask = ~0ULL >> (64-VWidth);
1268 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1269 "Invalid DemandedElts!");
1271 if (isa<UndefValue>(V)) {
1272 // If the entire vector is undefined, just return this info.
1273 UndefElts = EltMask;
1275 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1276 UndefElts = EltMask;
1277 return UndefValue::get(V->getType());
1281 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1282 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1283 Constant *Undef = UndefValue::get(EltTy);
1285 std::vector<Constant*> Elts;
1286 for (unsigned i = 0; i != VWidth; ++i)
1287 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1288 Elts.push_back(Undef);
1289 UndefElts |= (1ULL << i);
1290 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1291 Elts.push_back(Undef);
1292 UndefElts |= (1ULL << i);
1293 } else { // Otherwise, defined.
1294 Elts.push_back(CP->getOperand(i));
1297 // If we changed the constant, return it.
1298 Constant *NewCP = ConstantVector::get(Elts);
1299 return NewCP != CP ? NewCP : 0;
1300 } else if (isa<ConstantAggregateZero>(V)) {
1301 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1303 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1304 Constant *Zero = Constant::getNullValue(EltTy);
1305 Constant *Undef = UndefValue::get(EltTy);
1306 std::vector<Constant*> Elts;
1307 for (unsigned i = 0; i != VWidth; ++i)
1308 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1309 UndefElts = DemandedElts ^ EltMask;
1310 return ConstantVector::get(Elts);
1313 if (!V->hasOneUse()) { // Other users may use these bits.
1314 if (Depth != 0) { // Not at the root.
1315 // TODO: Just compute the UndefElts information recursively.
1319 } else if (Depth == 10) { // Limit search depth.
1323 Instruction *I = dyn_cast<Instruction>(V);
1324 if (!I) return false; // Only analyze instructions.
1326 bool MadeChange = false;
1327 uint64_t UndefElts2;
1329 switch (I->getOpcode()) {
1332 case Instruction::InsertElement: {
1333 // If this is a variable index, we don't know which element it overwrites.
1334 // demand exactly the same input as we produce.
1335 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1337 // Note that we can't propagate undef elt info, because we don't know
1338 // which elt is getting updated.
1339 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1340 UndefElts2, Depth+1);
1341 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1345 // If this is inserting an element that isn't demanded, remove this
1347 unsigned IdxNo = Idx->getZExtValue();
1348 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1349 return AddSoonDeadInstToWorklist(*I, 0);
1351 // Otherwise, the element inserted overwrites whatever was there, so the
1352 // input demanded set is simpler than the output set.
1353 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1354 DemandedElts & ~(1ULL << IdxNo),
1355 UndefElts, Depth+1);
1356 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1358 // The inserted element is defined.
1359 UndefElts |= 1ULL << IdxNo;
1363 case Instruction::And:
1364 case Instruction::Or:
1365 case Instruction::Xor:
1366 case Instruction::Add:
1367 case Instruction::Sub:
1368 case Instruction::Mul:
1369 // div/rem demand all inputs, because they don't want divide by zero.
1370 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1371 UndefElts, Depth+1);
1372 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1373 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1374 UndefElts2, Depth+1);
1375 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1377 // Output elements are undefined if both are undefined. Consider things
1378 // like undef&0. The result is known zero, not undef.
1379 UndefElts &= UndefElts2;
1382 case Instruction::Call: {
1383 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1385 switch (II->getIntrinsicID()) {
1388 // Binary vector operations that work column-wise. A dest element is a
1389 // function of the corresponding input elements from the two inputs.
1390 case Intrinsic::x86_sse_sub_ss:
1391 case Intrinsic::x86_sse_mul_ss:
1392 case Intrinsic::x86_sse_min_ss:
1393 case Intrinsic::x86_sse_max_ss:
1394 case Intrinsic::x86_sse2_sub_sd:
1395 case Intrinsic::x86_sse2_mul_sd:
1396 case Intrinsic::x86_sse2_min_sd:
1397 case Intrinsic::x86_sse2_max_sd:
1398 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1399 UndefElts, Depth+1);
1400 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1401 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1402 UndefElts2, Depth+1);
1403 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1405 // If only the low elt is demanded and this is a scalarizable intrinsic,
1406 // scalarize it now.
1407 if (DemandedElts == 1) {
1408 switch (II->getIntrinsicID()) {
1410 case Intrinsic::x86_sse_sub_ss:
1411 case Intrinsic::x86_sse_mul_ss:
1412 case Intrinsic::x86_sse2_sub_sd:
1413 case Intrinsic::x86_sse2_mul_sd:
1414 // TODO: Lower MIN/MAX/ABS/etc
1415 Value *LHS = II->getOperand(1);
1416 Value *RHS = II->getOperand(2);
1417 // Extract the element as scalars.
1418 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1419 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1421 switch (II->getIntrinsicID()) {
1422 default: assert(0 && "Case stmts out of sync!");
1423 case Intrinsic::x86_sse_sub_ss:
1424 case Intrinsic::x86_sse2_sub_sd:
1425 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1426 II->getName()), *II);
1428 case Intrinsic::x86_sse_mul_ss:
1429 case Intrinsic::x86_sse2_mul_sd:
1430 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1431 II->getName()), *II);
1436 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1438 InsertNewInstBefore(New, *II);
1439 AddSoonDeadInstToWorklist(*II, 0);
1444 // Output elements are undefined if both are undefined. Consider things
1445 // like undef&0. The result is known zero, not undef.
1446 UndefElts &= UndefElts2;
1452 return MadeChange ? I : 0;
1455 /// @returns true if the specified compare instruction is
1456 /// true when both operands are equal...
1457 /// @brief Determine if the ICmpInst returns true if both operands are equal
1458 static bool isTrueWhenEqual(ICmpInst &ICI) {
1459 ICmpInst::Predicate pred = ICI.getPredicate();
1460 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1461 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1462 pred == ICmpInst::ICMP_SLE;
1465 /// AssociativeOpt - Perform an optimization on an associative operator. This
1466 /// function is designed to check a chain of associative operators for a
1467 /// potential to apply a certain optimization. Since the optimization may be
1468 /// applicable if the expression was reassociated, this checks the chain, then
1469 /// reassociates the expression as necessary to expose the optimization
1470 /// opportunity. This makes use of a special Functor, which must define
1471 /// 'shouldApply' and 'apply' methods.
1473 template<typename Functor>
1474 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1475 unsigned Opcode = Root.getOpcode();
1476 Value *LHS = Root.getOperand(0);
1478 // Quick check, see if the immediate LHS matches...
1479 if (F.shouldApply(LHS))
1480 return F.apply(Root);
1482 // Otherwise, if the LHS is not of the same opcode as the root, return.
1483 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1484 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1485 // Should we apply this transform to the RHS?
1486 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1488 // If not to the RHS, check to see if we should apply to the LHS...
1489 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1490 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1494 // If the functor wants to apply the optimization to the RHS of LHSI,
1495 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1497 BasicBlock *BB = Root.getParent();
1499 // Now all of the instructions are in the current basic block, go ahead
1500 // and perform the reassociation.
1501 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1503 // First move the selected RHS to the LHS of the root...
1504 Root.setOperand(0, LHSI->getOperand(1));
1506 // Make what used to be the LHS of the root be the user of the root...
1507 Value *ExtraOperand = TmpLHSI->getOperand(1);
1508 if (&Root == TmpLHSI) {
1509 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1512 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1513 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1514 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1515 BasicBlock::iterator ARI = &Root; ++ARI;
1516 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1519 // Now propagate the ExtraOperand down the chain of instructions until we
1521 while (TmpLHSI != LHSI) {
1522 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1523 // Move the instruction to immediately before the chain we are
1524 // constructing to avoid breaking dominance properties.
1525 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1526 BB->getInstList().insert(ARI, NextLHSI);
1529 Value *NextOp = NextLHSI->getOperand(1);
1530 NextLHSI->setOperand(1, ExtraOperand);
1532 ExtraOperand = NextOp;
1535 // Now that the instructions are reassociated, have the functor perform
1536 // the transformation...
1537 return F.apply(Root);
1540 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1546 // AddRHS - Implements: X + X --> X << 1
1549 AddRHS(Value *rhs) : RHS(rhs) {}
1550 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1551 Instruction *apply(BinaryOperator &Add) const {
1552 return BinaryOperator::createShl(Add.getOperand(0),
1553 ConstantInt::get(Add.getType(), 1));
1557 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1559 struct AddMaskingAnd {
1561 AddMaskingAnd(Constant *c) : C2(c) {}
1562 bool shouldApply(Value *LHS) const {
1564 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1565 ConstantExpr::getAnd(C1, C2)->isNullValue();
1567 Instruction *apply(BinaryOperator &Add) const {
1568 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1572 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1574 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1575 if (Constant *SOC = dyn_cast<Constant>(SO))
1576 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1578 return IC->InsertNewInstBefore(CastInst::create(
1579 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1582 // Figure out if the constant is the left or the right argument.
1583 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1584 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1586 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1588 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1589 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1592 Value *Op0 = SO, *Op1 = ConstOperand;
1594 std::swap(Op0, Op1);
1596 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1597 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1598 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1599 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1600 SO->getName()+".cmp");
1602 assert(0 && "Unknown binary instruction type!");
1605 return IC->InsertNewInstBefore(New, I);
1608 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1609 // constant as the other operand, try to fold the binary operator into the
1610 // select arguments. This also works for Cast instructions, which obviously do
1611 // not have a second operand.
1612 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1614 // Don't modify shared select instructions
1615 if (!SI->hasOneUse()) return 0;
1616 Value *TV = SI->getOperand(1);
1617 Value *FV = SI->getOperand(2);
1619 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1620 // Bool selects with constant operands can be folded to logical ops.
1621 if (SI->getType() == Type::Int1Ty) return 0;
1623 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1624 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1626 return new SelectInst(SI->getCondition(), SelectTrueVal,
1633 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1634 /// node as operand #0, see if we can fold the instruction into the PHI (which
1635 /// is only possible if all operands to the PHI are constants).
1636 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1637 PHINode *PN = cast<PHINode>(I.getOperand(0));
1638 unsigned NumPHIValues = PN->getNumIncomingValues();
1639 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1641 // Check to see if all of the operands of the PHI are constants. If there is
1642 // one non-constant value, remember the BB it is. If there is more than one
1644 BasicBlock *NonConstBB = 0;
1645 for (unsigned i = 0; i != NumPHIValues; ++i)
1646 if (!isa<Constant>(PN->getIncomingValue(i))) {
1647 if (NonConstBB) return 0; // More than one non-const value.
1648 NonConstBB = PN->getIncomingBlock(i);
1650 // If the incoming non-constant value is in I's block, we have an infinite
1652 if (NonConstBB == I.getParent())
1656 // If there is exactly one non-constant value, we can insert a copy of the
1657 // operation in that block. However, if this is a critical edge, we would be
1658 // inserting the computation one some other paths (e.g. inside a loop). Only
1659 // do this if the pred block is unconditionally branching into the phi block.
1661 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1662 if (!BI || !BI->isUnconditional()) return 0;
1665 // Okay, we can do the transformation: create the new PHI node.
1666 PHINode *NewPN = new PHINode(I.getType(), "");
1667 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1668 InsertNewInstBefore(NewPN, *PN);
1669 NewPN->takeName(PN);
1671 // Next, add all of the operands to the PHI.
1672 if (I.getNumOperands() == 2) {
1673 Constant *C = cast<Constant>(I.getOperand(1));
1674 for (unsigned i = 0; i != NumPHIValues; ++i) {
1676 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1677 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1678 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1680 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1682 assert(PN->getIncomingBlock(i) == NonConstBB);
1683 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1684 InV = BinaryOperator::create(BO->getOpcode(),
1685 PN->getIncomingValue(i), C, "phitmp",
1686 NonConstBB->getTerminator());
1687 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1688 InV = CmpInst::create(CI->getOpcode(),
1690 PN->getIncomingValue(i), C, "phitmp",
1691 NonConstBB->getTerminator());
1693 assert(0 && "Unknown binop!");
1695 WorkList.push_back(cast<Instruction>(InV));
1697 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1700 CastInst *CI = cast<CastInst>(&I);
1701 const Type *RetTy = CI->getType();
1702 for (unsigned i = 0; i != NumPHIValues; ++i) {
1704 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1705 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1707 assert(PN->getIncomingBlock(i) == NonConstBB);
1708 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1709 I.getType(), "phitmp",
1710 NonConstBB->getTerminator());
1711 WorkList.push_back(cast<Instruction>(InV));
1713 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1716 return ReplaceInstUsesWith(I, NewPN);
1719 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1720 bool Changed = SimplifyCommutative(I);
1721 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1723 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1724 // X + undef -> undef
1725 if (isa<UndefValue>(RHS))
1726 return ReplaceInstUsesWith(I, RHS);
1729 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1730 if (RHSC->isNullValue())
1731 return ReplaceInstUsesWith(I, LHS);
1732 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1733 if (CFP->isExactlyValue(-0.0))
1734 return ReplaceInstUsesWith(I, LHS);
1737 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1738 // X + (signbit) --> X ^ signbit
1739 uint64_t Val = CI->getZExtValue();
1740 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1741 return BinaryOperator::createXor(LHS, RHS);
1743 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1744 // (X & 254)+1 -> (X&254)|1
1745 uint64_t KnownZero, KnownOne;
1746 if (!isa<VectorType>(I.getType()) &&
1747 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
1748 KnownZero, KnownOne))
1752 if (isa<PHINode>(LHS))
1753 if (Instruction *NV = FoldOpIntoPhi(I))
1756 ConstantInt *XorRHS = 0;
1758 if (isa<ConstantInt>(RHSC) &&
1759 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1760 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1761 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1762 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1764 uint64_t C0080Val = 1ULL << 31;
1765 int64_t CFF80Val = -C0080Val;
1768 if (TySizeBits > Size) {
1770 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1771 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1772 if (RHSSExt == CFF80Val) {
1773 if (XorRHS->getZExtValue() == C0080Val)
1775 } else if (RHSZExt == C0080Val) {
1776 if (XorRHS->getSExtValue() == CFF80Val)
1780 // This is a sign extend if the top bits are known zero.
1781 uint64_t Mask = ~0ULL;
1782 Mask <<= 64-(TySizeBits-Size);
1783 Mask &= cast<IntegerType>(XorLHS->getType())->getBitMask();
1784 if (!MaskedValueIsZero(XorLHS, Mask))
1785 Size = 0; // Not a sign ext, but can't be any others either.
1792 } while (Size >= 8);
1795 const Type *MiddleType = 0;
1798 case 32: MiddleType = Type::Int32Ty; break;
1799 case 16: MiddleType = Type::Int16Ty; break;
1800 case 8: MiddleType = Type::Int8Ty; break;
1803 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1804 InsertNewInstBefore(NewTrunc, I);
1805 return new SExtInst(NewTrunc, I.getType());
1811 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
1812 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1814 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1815 if (RHSI->getOpcode() == Instruction::Sub)
1816 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1817 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1819 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1820 if (LHSI->getOpcode() == Instruction::Sub)
1821 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1822 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1827 if (Value *V = dyn_castNegVal(LHS))
1828 return BinaryOperator::createSub(RHS, V);
1831 if (!isa<Constant>(RHS))
1832 if (Value *V = dyn_castNegVal(RHS))
1833 return BinaryOperator::createSub(LHS, V);
1837 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1838 if (X == RHS) // X*C + X --> X * (C+1)
1839 return BinaryOperator::createMul(RHS, AddOne(C2));
1841 // X*C1 + X*C2 --> X * (C1+C2)
1843 if (X == dyn_castFoldableMul(RHS, C1))
1844 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1847 // X + X*C --> X * (C+1)
1848 if (dyn_castFoldableMul(RHS, C2) == LHS)
1849 return BinaryOperator::createMul(LHS, AddOne(C2));
1851 // X + ~X --> -1 since ~X = -X-1
1852 if (dyn_castNotVal(LHS) == RHS ||
1853 dyn_castNotVal(RHS) == LHS)
1854 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
1857 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1858 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1859 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
1862 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1864 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1865 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1866 return BinaryOperator::createSub(C, X);
1869 // (X & FF00) + xx00 -> (X+xx00) & FF00
1870 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1871 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1872 if (Anded == CRHS) {
1873 // See if all bits from the first bit set in the Add RHS up are included
1874 // in the mask. First, get the rightmost bit.
1875 uint64_t AddRHSV = CRHS->getZExtValue();
1877 // Form a mask of all bits from the lowest bit added through the top.
1878 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1879 AddRHSHighBits &= C2->getType()->getBitMask();
1881 // See if the and mask includes all of these bits.
1882 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1884 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1885 // Okay, the xform is safe. Insert the new add pronto.
1886 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1887 LHS->getName()), I);
1888 return BinaryOperator::createAnd(NewAdd, C2);
1893 // Try to fold constant add into select arguments.
1894 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1895 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1899 // add (cast *A to intptrtype) B ->
1900 // cast (GEP (cast *A to sbyte*) B) ->
1903 CastInst *CI = dyn_cast<CastInst>(LHS);
1906 CI = dyn_cast<CastInst>(RHS);
1909 if (CI && CI->getType()->isSized() &&
1910 (CI->getType()->getPrimitiveSizeInBits() ==
1911 TD->getIntPtrType()->getPrimitiveSizeInBits())
1912 && isa<PointerType>(CI->getOperand(0)->getType())) {
1913 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
1914 PointerType::get(Type::Int8Ty), I);
1915 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1916 return new PtrToIntInst(I2, CI->getType());
1920 return Changed ? &I : 0;
1923 // isSignBit - Return true if the value represented by the constant only has the
1924 // highest order bit set.
1925 static bool isSignBit(ConstantInt *CI) {
1926 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1927 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1930 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1931 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1933 if (Op0 == Op1) // sub X, X -> 0
1934 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1936 // If this is a 'B = x-(-A)', change to B = x+A...
1937 if (Value *V = dyn_castNegVal(Op1))
1938 return BinaryOperator::createAdd(Op0, V);
1940 if (isa<UndefValue>(Op0))
1941 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1942 if (isa<UndefValue>(Op1))
1943 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1945 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1946 // Replace (-1 - A) with (~A)...
1947 if (C->isAllOnesValue())
1948 return BinaryOperator::createNot(Op1);
1950 // C - ~X == X + (1+C)
1952 if (match(Op1, m_Not(m_Value(X))))
1953 return BinaryOperator::createAdd(X,
1954 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1955 // -(X >>u 31) -> (X >>s 31)
1956 // -(X >>s 31) -> (X >>u 31)
1957 if (C->isNullValue()) {
1958 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
1959 if (SI->getOpcode() == Instruction::LShr) {
1960 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1961 // Check to see if we are shifting out everything but the sign bit.
1962 if (CU->getZExtValue() ==
1963 SI->getType()->getPrimitiveSizeInBits()-1) {
1964 // Ok, the transformation is safe. Insert AShr.
1965 return BinaryOperator::create(Instruction::AShr,
1966 SI->getOperand(0), CU, SI->getName());
1970 else if (SI->getOpcode() == Instruction::AShr) {
1971 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1972 // Check to see if we are shifting out everything but the sign bit.
1973 if (CU->getZExtValue() ==
1974 SI->getType()->getPrimitiveSizeInBits()-1) {
1975 // Ok, the transformation is safe. Insert LShr.
1976 return BinaryOperator::createLShr(
1977 SI->getOperand(0), CU, SI->getName());
1983 // Try to fold constant sub into select arguments.
1984 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1985 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1988 if (isa<PHINode>(Op0))
1989 if (Instruction *NV = FoldOpIntoPhi(I))
1993 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1994 if (Op1I->getOpcode() == Instruction::Add &&
1995 !Op0->getType()->isFPOrFPVector()) {
1996 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1997 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
1998 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
1999 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2000 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2001 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2002 // C1-(X+C2) --> (C1-C2)-X
2003 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2004 Op1I->getOperand(0));
2008 if (Op1I->hasOneUse()) {
2009 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2010 // is not used by anyone else...
2012 if (Op1I->getOpcode() == Instruction::Sub &&
2013 !Op1I->getType()->isFPOrFPVector()) {
2014 // Swap the two operands of the subexpr...
2015 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2016 Op1I->setOperand(0, IIOp1);
2017 Op1I->setOperand(1, IIOp0);
2019 // Create the new top level add instruction...
2020 return BinaryOperator::createAdd(Op0, Op1);
2023 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2025 if (Op1I->getOpcode() == Instruction::And &&
2026 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2027 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2030 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2031 return BinaryOperator::createAnd(Op0, NewNot);
2034 // 0 - (X sdiv C) -> (X sdiv -C)
2035 if (Op1I->getOpcode() == Instruction::SDiv)
2036 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2037 if (CSI->isNullValue())
2038 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2039 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2040 ConstantExpr::getNeg(DivRHS));
2042 // X - X*C --> X * (1-C)
2043 ConstantInt *C2 = 0;
2044 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2046 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2047 return BinaryOperator::createMul(Op0, CP1);
2052 if (!Op0->getType()->isFPOrFPVector())
2053 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2054 if (Op0I->getOpcode() == Instruction::Add) {
2055 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2056 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2057 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2058 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2059 } else if (Op0I->getOpcode() == Instruction::Sub) {
2060 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2061 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2065 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2066 if (X == Op1) { // X*C - X --> X * (C-1)
2067 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2068 return BinaryOperator::createMul(Op1, CP1);
2071 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2072 if (X == dyn_castFoldableMul(Op1, C2))
2073 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2078 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2079 /// really just returns true if the most significant (sign) bit is set.
2080 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2082 case ICmpInst::ICMP_SLT:
2083 // True if LHS s< RHS and RHS == 0
2084 return RHS->isNullValue();
2085 case ICmpInst::ICMP_SLE:
2086 // True if LHS s<= RHS and RHS == -1
2087 return RHS->isAllOnesValue();
2088 case ICmpInst::ICMP_UGE:
2089 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2090 return RHS->getZExtValue() == (1ULL <<
2091 (RHS->getType()->getPrimitiveSizeInBits()-1));
2092 case ICmpInst::ICMP_UGT:
2093 // True if LHS u> RHS and RHS == high-bit-mask - 1
2094 return RHS->getZExtValue() ==
2095 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2101 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2102 bool Changed = SimplifyCommutative(I);
2103 Value *Op0 = I.getOperand(0);
2105 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2106 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2108 // Simplify mul instructions with a constant RHS...
2109 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2110 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2112 // ((X << C1)*C2) == (X * (C2 << C1))
2113 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2114 if (SI->getOpcode() == Instruction::Shl)
2115 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2116 return BinaryOperator::createMul(SI->getOperand(0),
2117 ConstantExpr::getShl(CI, ShOp));
2119 if (CI->isNullValue())
2120 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2121 if (CI->equalsInt(1)) // X * 1 == X
2122 return ReplaceInstUsesWith(I, Op0);
2123 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2124 return BinaryOperator::createNeg(Op0, I.getName());
2126 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2127 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2128 uint64_t C = Log2_64(Val);
2129 return BinaryOperator::createShl(Op0,
2130 ConstantInt::get(Op0->getType(), C));
2132 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2133 if (Op1F->isNullValue())
2134 return ReplaceInstUsesWith(I, Op1);
2136 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2137 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2138 if (Op1F->getValue() == 1.0)
2139 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2142 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2143 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2144 isa<ConstantInt>(Op0I->getOperand(1))) {
2145 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2146 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2148 InsertNewInstBefore(Add, I);
2149 Value *C1C2 = ConstantExpr::getMul(Op1,
2150 cast<Constant>(Op0I->getOperand(1)));
2151 return BinaryOperator::createAdd(Add, C1C2);
2155 // Try to fold constant mul into select arguments.
2156 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2157 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2160 if (isa<PHINode>(Op0))
2161 if (Instruction *NV = FoldOpIntoPhi(I))
2165 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2166 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2167 return BinaryOperator::createMul(Op0v, Op1v);
2169 // If one of the operands of the multiply is a cast from a boolean value, then
2170 // we know the bool is either zero or one, so this is a 'masking' multiply.
2171 // See if we can simplify things based on how the boolean was originally
2173 CastInst *BoolCast = 0;
2174 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2175 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2178 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2179 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2182 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2183 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2184 const Type *SCOpTy = SCIOp0->getType();
2186 // If the icmp is true iff the sign bit of X is set, then convert this
2187 // multiply into a shift/and combination.
2188 if (isa<ConstantInt>(SCIOp1) &&
2189 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2190 // Shift the X value right to turn it into "all signbits".
2191 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2192 SCOpTy->getPrimitiveSizeInBits()-1);
2194 InsertNewInstBefore(
2195 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2196 BoolCast->getOperand(0)->getName()+
2199 // If the multiply type is not the same as the source type, sign extend
2200 // or truncate to the multiply type.
2201 if (I.getType() != V->getType()) {
2202 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2203 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2204 Instruction::CastOps opcode =
2205 (SrcBits == DstBits ? Instruction::BitCast :
2206 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2207 V = InsertCastBefore(opcode, V, I.getType(), I);
2210 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2211 return BinaryOperator::createAnd(V, OtherOp);
2216 return Changed ? &I : 0;
2219 /// This function implements the transforms on div instructions that work
2220 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2221 /// used by the visitors to those instructions.
2222 /// @brief Transforms common to all three div instructions
2223 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2224 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2227 if (isa<UndefValue>(Op0))
2228 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2230 // X / undef -> undef
2231 if (isa<UndefValue>(Op1))
2232 return ReplaceInstUsesWith(I, Op1);
2234 // Handle cases involving: div X, (select Cond, Y, Z)
2235 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2236 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2237 // same basic block, then we replace the select with Y, and the condition
2238 // of the select with false (if the cond value is in the same BB). If the
2239 // select has uses other than the div, this allows them to be simplified
2240 // also. Note that div X, Y is just as good as div X, 0 (undef)
2241 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2242 if (ST->isNullValue()) {
2243 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2244 if (CondI && CondI->getParent() == I.getParent())
2245 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2246 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2247 I.setOperand(1, SI->getOperand(2));
2249 UpdateValueUsesWith(SI, SI->getOperand(2));
2253 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2254 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2255 if (ST->isNullValue()) {
2256 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2257 if (CondI && CondI->getParent() == I.getParent())
2258 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2259 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2260 I.setOperand(1, SI->getOperand(1));
2262 UpdateValueUsesWith(SI, SI->getOperand(1));
2270 /// This function implements the transforms common to both integer division
2271 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2272 /// division instructions.
2273 /// @brief Common integer divide transforms
2274 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2275 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2277 if (Instruction *Common = commonDivTransforms(I))
2280 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2282 if (RHS->equalsInt(1))
2283 return ReplaceInstUsesWith(I, Op0);
2285 // (X / C1) / C2 -> X / (C1*C2)
2286 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2287 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2288 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2289 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2290 ConstantExpr::getMul(RHS, LHSRHS));
2293 if (!RHS->isNullValue()) { // avoid X udiv 0
2294 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2295 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2297 if (isa<PHINode>(Op0))
2298 if (Instruction *NV = FoldOpIntoPhi(I))
2303 // 0 / X == 0, we don't need to preserve faults!
2304 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2305 if (LHS->equalsInt(0))
2306 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2311 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2312 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2314 // Handle the integer div common cases
2315 if (Instruction *Common = commonIDivTransforms(I))
2318 // X udiv C^2 -> X >> C
2319 // Check to see if this is an unsigned division with an exact power of 2,
2320 // if so, convert to a right shift.
2321 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2322 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2323 if (isPowerOf2_64(Val)) {
2324 uint64_t ShiftAmt = Log2_64(Val);
2325 return BinaryOperator::createLShr(Op0,
2326 ConstantInt::get(Op0->getType(), ShiftAmt));
2330 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2331 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2332 if (RHSI->getOpcode() == Instruction::Shl &&
2333 isa<ConstantInt>(RHSI->getOperand(0))) {
2334 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2335 if (isPowerOf2_64(C1)) {
2336 Value *N = RHSI->getOperand(1);
2337 const Type *NTy = N->getType();
2338 if (uint64_t C2 = Log2_64(C1)) {
2339 Constant *C2V = ConstantInt::get(NTy, C2);
2340 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2342 return BinaryOperator::createLShr(Op0, N);
2347 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2348 // where C1&C2 are powers of two.
2349 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2350 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2351 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2352 if (!STO->isNullValue() && !STO->isNullValue()) {
2353 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2354 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2355 // Compute the shift amounts
2356 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2357 // Construct the "on true" case of the select
2358 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2359 Instruction *TSI = BinaryOperator::createLShr(
2360 Op0, TC, SI->getName()+".t");
2361 TSI = InsertNewInstBefore(TSI, I);
2363 // Construct the "on false" case of the select
2364 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2365 Instruction *FSI = BinaryOperator::createLShr(
2366 Op0, FC, SI->getName()+".f");
2367 FSI = InsertNewInstBefore(FSI, I);
2369 // construct the select instruction and return it.
2370 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2377 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2378 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2380 // Handle the integer div common cases
2381 if (Instruction *Common = commonIDivTransforms(I))
2384 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2386 if (RHS->isAllOnesValue())
2387 return BinaryOperator::createNeg(Op0);
2390 if (Value *LHSNeg = dyn_castNegVal(Op0))
2391 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2394 // If the sign bits of both operands are zero (i.e. we can prove they are
2395 // unsigned inputs), turn this into a udiv.
2396 if (I.getType()->isInteger()) {
2397 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2398 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2399 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2406 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2407 return commonDivTransforms(I);
2410 /// GetFactor - If we can prove that the specified value is at least a multiple
2411 /// of some factor, return that factor.
2412 static Constant *GetFactor(Value *V) {
2413 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2416 // Unless we can be tricky, we know this is a multiple of 1.
2417 Constant *Result = ConstantInt::get(V->getType(), 1);
2419 Instruction *I = dyn_cast<Instruction>(V);
2420 if (!I) return Result;
2422 if (I->getOpcode() == Instruction::Mul) {
2423 // Handle multiplies by a constant, etc.
2424 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2425 GetFactor(I->getOperand(1)));
2426 } else if (I->getOpcode() == Instruction::Shl) {
2427 // (X<<C) -> X * (1 << C)
2428 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2429 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2430 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2432 } else if (I->getOpcode() == Instruction::And) {
2433 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2434 // X & 0xFFF0 is known to be a multiple of 16.
2435 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2436 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2437 return ConstantExpr::getShl(Result,
2438 ConstantInt::get(Result->getType(), Zeros));
2440 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2441 // Only handle int->int casts.
2442 if (!CI->isIntegerCast())
2444 Value *Op = CI->getOperand(0);
2445 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2450 /// This function implements the transforms on rem instructions that work
2451 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2452 /// is used by the visitors to those instructions.
2453 /// @brief Transforms common to all three rem instructions
2454 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2455 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2457 // 0 % X == 0, we don't need to preserve faults!
2458 if (Constant *LHS = dyn_cast<Constant>(Op0))
2459 if (LHS->isNullValue())
2460 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2462 if (isa<UndefValue>(Op0)) // undef % X -> 0
2463 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2464 if (isa<UndefValue>(Op1))
2465 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2467 // Handle cases involving: rem X, (select Cond, Y, Z)
2468 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2469 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2470 // the same basic block, then we replace the select with Y, and the
2471 // condition of the select with false (if the cond value is in the same
2472 // BB). If the select has uses other than the div, this allows them to be
2474 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2475 if (ST->isNullValue()) {
2476 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2477 if (CondI && CondI->getParent() == I.getParent())
2478 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2479 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2480 I.setOperand(1, SI->getOperand(2));
2482 UpdateValueUsesWith(SI, SI->getOperand(2));
2485 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2486 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2487 if (ST->isNullValue()) {
2488 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2489 if (CondI && CondI->getParent() == I.getParent())
2490 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2491 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2492 I.setOperand(1, SI->getOperand(1));
2494 UpdateValueUsesWith(SI, SI->getOperand(1));
2502 /// This function implements the transforms common to both integer remainder
2503 /// instructions (urem and srem). It is called by the visitors to those integer
2504 /// remainder instructions.
2505 /// @brief Common integer remainder transforms
2506 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2507 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2509 if (Instruction *common = commonRemTransforms(I))
2512 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2513 // X % 0 == undef, we don't need to preserve faults!
2514 if (RHS->equalsInt(0))
2515 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2517 if (RHS->equalsInt(1)) // X % 1 == 0
2518 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2520 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2521 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2522 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2524 } else if (isa<PHINode>(Op0I)) {
2525 if (Instruction *NV = FoldOpIntoPhi(I))
2528 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2529 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2530 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2537 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2538 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2540 if (Instruction *common = commonIRemTransforms(I))
2543 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2544 // X urem C^2 -> X and C
2545 // Check to see if this is an unsigned remainder with an exact power of 2,
2546 // if so, convert to a bitwise and.
2547 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2548 if (isPowerOf2_64(C->getZExtValue()))
2549 return BinaryOperator::createAnd(Op0, SubOne(C));
2552 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2553 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2554 if (RHSI->getOpcode() == Instruction::Shl &&
2555 isa<ConstantInt>(RHSI->getOperand(0))) {
2556 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2557 if (isPowerOf2_64(C1)) {
2558 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2559 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2561 return BinaryOperator::createAnd(Op0, Add);
2566 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2567 // where C1&C2 are powers of two.
2568 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2569 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2570 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2571 // STO == 0 and SFO == 0 handled above.
2572 if (isPowerOf2_64(STO->getZExtValue()) &&
2573 isPowerOf2_64(SFO->getZExtValue())) {
2574 Value *TrueAnd = InsertNewInstBefore(
2575 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2576 Value *FalseAnd = InsertNewInstBefore(
2577 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2578 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2586 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2587 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2589 if (Instruction *common = commonIRemTransforms(I))
2592 if (Value *RHSNeg = dyn_castNegVal(Op1))
2593 if (!isa<ConstantInt>(RHSNeg) ||
2594 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2596 AddUsesToWorkList(I);
2597 I.setOperand(1, RHSNeg);
2601 // If the top bits of both operands are zero (i.e. we can prove they are
2602 // unsigned inputs), turn this into a urem.
2603 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2604 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2605 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2606 return BinaryOperator::createURem(Op0, Op1, I.getName());
2612 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2613 return commonRemTransforms(I);
2616 // isMaxValueMinusOne - return true if this is Max-1
2617 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2619 // Calculate 0111111111..11111
2620 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2621 int64_t Val = INT64_MAX; // All ones
2622 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2623 return C->getSExtValue() == Val-1;
2625 return C->getZExtValue() == C->getType()->getBitMask()-1;
2628 // isMinValuePlusOne - return true if this is Min+1
2629 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2631 // Calculate 1111111111000000000000
2632 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2633 int64_t Val = -1; // All ones
2634 Val <<= TypeBits-1; // Shift over to the right spot
2635 return C->getSExtValue() == Val+1;
2637 return C->getZExtValue() == 1; // unsigned
2640 // isOneBitSet - Return true if there is exactly one bit set in the specified
2642 static bool isOneBitSet(const ConstantInt *CI) {
2643 uint64_t V = CI->getZExtValue();
2644 return V && (V & (V-1)) == 0;
2647 #if 0 // Currently unused
2648 // isLowOnes - Return true if the constant is of the form 0+1+.
2649 static bool isLowOnes(const ConstantInt *CI) {
2650 uint64_t V = CI->getZExtValue();
2652 // There won't be bits set in parts that the type doesn't contain.
2653 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2655 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2656 return U && V && (U & V) == 0;
2660 // isHighOnes - Return true if the constant is of the form 1+0+.
2661 // This is the same as lowones(~X).
2662 static bool isHighOnes(const ConstantInt *CI) {
2663 uint64_t V = ~CI->getZExtValue();
2664 if (~V == 0) return false; // 0's does not match "1+"
2666 // There won't be bits set in parts that the type doesn't contain.
2667 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2669 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2670 return U && V && (U & V) == 0;
2673 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2674 /// are carefully arranged to allow folding of expressions such as:
2676 /// (A < B) | (A > B) --> (A != B)
2678 /// Note that this is only valid if the first and second predicates have the
2679 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2681 /// Three bits are used to represent the condition, as follows:
2686 /// <=> Value Definition
2687 /// 000 0 Always false
2694 /// 111 7 Always true
2696 static unsigned getICmpCode(const ICmpInst *ICI) {
2697 switch (ICI->getPredicate()) {
2699 case ICmpInst::ICMP_UGT: return 1; // 001
2700 case ICmpInst::ICMP_SGT: return 1; // 001
2701 case ICmpInst::ICMP_EQ: return 2; // 010
2702 case ICmpInst::ICMP_UGE: return 3; // 011
2703 case ICmpInst::ICMP_SGE: return 3; // 011
2704 case ICmpInst::ICMP_ULT: return 4; // 100
2705 case ICmpInst::ICMP_SLT: return 4; // 100
2706 case ICmpInst::ICMP_NE: return 5; // 101
2707 case ICmpInst::ICMP_ULE: return 6; // 110
2708 case ICmpInst::ICMP_SLE: return 6; // 110
2711 assert(0 && "Invalid ICmp predicate!");
2716 /// getICmpValue - This is the complement of getICmpCode, which turns an
2717 /// opcode and two operands into either a constant true or false, or a brand
2718 /// new /// ICmp instruction. The sign is passed in to determine which kind
2719 /// of predicate to use in new icmp instructions.
2720 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2722 default: assert(0 && "Illegal ICmp code!");
2723 case 0: return ConstantInt::getFalse();
2726 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2728 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2729 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2732 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2734 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2737 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2739 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2740 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2743 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2745 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2746 case 7: return ConstantInt::getTrue();
2750 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2751 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2752 (ICmpInst::isSignedPredicate(p1) &&
2753 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2754 (ICmpInst::isSignedPredicate(p2) &&
2755 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2759 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2760 struct FoldICmpLogical {
2763 ICmpInst::Predicate pred;
2764 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2765 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2766 pred(ICI->getPredicate()) {}
2767 bool shouldApply(Value *V) const {
2768 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2769 if (PredicatesFoldable(pred, ICI->getPredicate()))
2770 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2771 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2774 Instruction *apply(Instruction &Log) const {
2775 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2776 if (ICI->getOperand(0) != LHS) {
2777 assert(ICI->getOperand(1) == LHS);
2778 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2781 unsigned LHSCode = getICmpCode(ICI);
2782 unsigned RHSCode = getICmpCode(cast<ICmpInst>(Log.getOperand(1)));
2784 switch (Log.getOpcode()) {
2785 case Instruction::And: Code = LHSCode & RHSCode; break;
2786 case Instruction::Or: Code = LHSCode | RHSCode; break;
2787 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2788 default: assert(0 && "Illegal logical opcode!"); return 0;
2791 Value *RV = getICmpValue(ICmpInst::isSignedPredicate(pred), Code, LHS, RHS);
2792 if (Instruction *I = dyn_cast<Instruction>(RV))
2794 // Otherwise, it's a constant boolean value...
2795 return IC.ReplaceInstUsesWith(Log, RV);
2798 } // end anonymous namespace
2800 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2801 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2802 // guaranteed to be a binary operator.
2803 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2805 ConstantInt *AndRHS,
2806 BinaryOperator &TheAnd) {
2807 Value *X = Op->getOperand(0);
2808 Constant *Together = 0;
2810 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2812 switch (Op->getOpcode()) {
2813 case Instruction::Xor:
2814 if (Op->hasOneUse()) {
2815 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2816 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2817 InsertNewInstBefore(And, TheAnd);
2819 return BinaryOperator::createXor(And, Together);
2822 case Instruction::Or:
2823 if (Together == AndRHS) // (X | C) & C --> C
2824 return ReplaceInstUsesWith(TheAnd, AndRHS);
2826 if (Op->hasOneUse() && Together != OpRHS) {
2827 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2828 Instruction *Or = BinaryOperator::createOr(X, Together);
2829 InsertNewInstBefore(Or, TheAnd);
2831 return BinaryOperator::createAnd(Or, AndRHS);
2834 case Instruction::Add:
2835 if (Op->hasOneUse()) {
2836 // Adding a one to a single bit bit-field should be turned into an XOR
2837 // of the bit. First thing to check is to see if this AND is with a
2838 // single bit constant.
2839 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2841 // Clear bits that are not part of the constant.
2842 AndRHSV &= AndRHS->getType()->getBitMask();
2844 // If there is only one bit set...
2845 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2846 // Ok, at this point, we know that we are masking the result of the
2847 // ADD down to exactly one bit. If the constant we are adding has
2848 // no bits set below this bit, then we can eliminate the ADD.
2849 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2851 // Check to see if any bits below the one bit set in AndRHSV are set.
2852 if ((AddRHS & (AndRHSV-1)) == 0) {
2853 // If not, the only thing that can effect the output of the AND is
2854 // the bit specified by AndRHSV. If that bit is set, the effect of
2855 // the XOR is to toggle the bit. If it is clear, then the ADD has
2857 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2858 TheAnd.setOperand(0, X);
2861 // Pull the XOR out of the AND.
2862 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
2863 InsertNewInstBefore(NewAnd, TheAnd);
2864 NewAnd->takeName(Op);
2865 return BinaryOperator::createXor(NewAnd, AndRHS);
2872 case Instruction::Shl: {
2873 // We know that the AND will not produce any of the bits shifted in, so if
2874 // the anded constant includes them, clear them now!
2876 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2877 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2878 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2880 if (CI == ShlMask) { // Masking out bits that the shift already masks
2881 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2882 } else if (CI != AndRHS) { // Reducing bits set in and.
2883 TheAnd.setOperand(1, CI);
2888 case Instruction::LShr:
2890 // We know that the AND will not produce any of the bits shifted in, so if
2891 // the anded constant includes them, clear them now! This only applies to
2892 // unsigned shifts, because a signed shr may bring in set bits!
2894 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2895 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2896 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2898 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2899 return ReplaceInstUsesWith(TheAnd, Op);
2900 } else if (CI != AndRHS) {
2901 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2906 case Instruction::AShr:
2908 // See if this is shifting in some sign extension, then masking it out
2910 if (Op->hasOneUse()) {
2911 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2912 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2913 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
2914 if (C == AndRHS) { // Masking out bits shifted in.
2915 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2916 // Make the argument unsigned.
2917 Value *ShVal = Op->getOperand(0);
2918 ShVal = InsertNewInstBefore(
2919 BinaryOperator::createLShr(ShVal, OpRHS,
2920 Op->getName()), TheAnd);
2921 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
2930 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2931 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2932 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
2933 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
2934 /// insert new instructions.
2935 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2936 bool isSigned, bool Inside,
2938 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
2939 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
2940 "Lo is not <= Hi in range emission code!");
2943 if (Lo == Hi) // Trivially false.
2944 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
2946 // V >= Min && V < Hi --> V < Hi
2947 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2948 ICmpInst::Predicate pred = (isSigned ?
2949 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
2950 return new ICmpInst(pred, V, Hi);
2953 // Emit V-Lo <u Hi-Lo
2954 Constant *NegLo = ConstantExpr::getNeg(Lo);
2955 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2956 InsertNewInstBefore(Add, IB);
2957 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
2958 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
2961 if (Lo == Hi) // Trivially true.
2962 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
2964 // V < Min || V >= Hi ->'V > Hi-1'
2965 Hi = SubOne(cast<ConstantInt>(Hi));
2966 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2967 ICmpInst::Predicate pred = (isSigned ?
2968 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
2969 return new ICmpInst(pred, V, Hi);
2972 // Emit V-Lo > Hi-1-Lo
2973 Constant *NegLo = ConstantExpr::getNeg(Lo);
2974 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2975 InsertNewInstBefore(Add, IB);
2976 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
2977 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
2980 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2981 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2982 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2983 // not, since all 1s are not contiguous.
2984 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
2985 uint64_t V = Val->getZExtValue();
2986 if (!isShiftedMask_64(V)) return false;
2988 // look for the first zero bit after the run of ones
2989 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
2990 // look for the first non-zero bit
2991 ME = 64-CountLeadingZeros_64(V);
2997 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
2998 /// where isSub determines whether the operator is a sub. If we can fold one of
2999 /// the following xforms:
3001 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3002 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3003 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3005 /// return (A +/- B).
3007 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3008 ConstantInt *Mask, bool isSub,
3010 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3011 if (!LHSI || LHSI->getNumOperands() != 2 ||
3012 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3014 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3016 switch (LHSI->getOpcode()) {
3018 case Instruction::And:
3019 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3020 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3021 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3024 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3025 // part, we don't need any explicit masks to take them out of A. If that
3026 // is all N is, ignore it.
3028 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3029 uint64_t Mask = cast<IntegerType>(RHS->getType())->getBitMask();
3031 if (MaskedValueIsZero(RHS, Mask))
3036 case Instruction::Or:
3037 case Instruction::Xor:
3038 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3039 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3040 ConstantExpr::getAnd(N, Mask)->isNullValue())
3047 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3049 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3050 return InsertNewInstBefore(New, I);
3053 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3054 bool Changed = SimplifyCommutative(I);
3055 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3057 if (isa<UndefValue>(Op1)) // X & undef -> 0
3058 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3062 return ReplaceInstUsesWith(I, Op1);
3064 // See if we can simplify any instructions used by the instruction whose sole
3065 // purpose is to compute bits we don't care about.
3066 uint64_t KnownZero, KnownOne;
3067 if (!isa<VectorType>(I.getType())) {
3068 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3069 KnownZero, KnownOne))
3072 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3073 if (CP->isAllOnesValue())
3074 return ReplaceInstUsesWith(I, I.getOperand(0));
3078 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3079 uint64_t AndRHSMask = AndRHS->getZExtValue();
3080 uint64_t TypeMask = cast<IntegerType>(Op0->getType())->getBitMask();
3081 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3083 // Optimize a variety of ((val OP C1) & C2) combinations...
3084 if (isa<BinaryOperator>(Op0)) {
3085 Instruction *Op0I = cast<Instruction>(Op0);
3086 Value *Op0LHS = Op0I->getOperand(0);
3087 Value *Op0RHS = Op0I->getOperand(1);
3088 switch (Op0I->getOpcode()) {
3089 case Instruction::Xor:
3090 case Instruction::Or:
3091 // If the mask is only needed on one incoming arm, push it up.
3092 if (Op0I->hasOneUse()) {
3093 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3094 // Not masking anything out for the LHS, move to RHS.
3095 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3096 Op0RHS->getName()+".masked");
3097 InsertNewInstBefore(NewRHS, I);
3098 return BinaryOperator::create(
3099 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3101 if (!isa<Constant>(Op0RHS) &&
3102 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3103 // Not masking anything out for the RHS, move to LHS.
3104 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3105 Op0LHS->getName()+".masked");
3106 InsertNewInstBefore(NewLHS, I);
3107 return BinaryOperator::create(
3108 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3113 case Instruction::Add:
3114 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3115 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3116 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3117 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3118 return BinaryOperator::createAnd(V, AndRHS);
3119 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3120 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3123 case Instruction::Sub:
3124 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3125 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3126 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3127 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3128 return BinaryOperator::createAnd(V, AndRHS);
3132 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3133 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3135 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3136 // If this is an integer truncation or change from signed-to-unsigned, and
3137 // if the source is an and/or with immediate, transform it. This
3138 // frequently occurs for bitfield accesses.
3139 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3140 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3141 CastOp->getNumOperands() == 2)
3142 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3143 if (CastOp->getOpcode() == Instruction::And) {
3144 // Change: and (cast (and X, C1) to T), C2
3145 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3146 // This will fold the two constants together, which may allow
3147 // other simplifications.
3148 Instruction *NewCast = CastInst::createTruncOrBitCast(
3149 CastOp->getOperand(0), I.getType(),
3150 CastOp->getName()+".shrunk");
3151 NewCast = InsertNewInstBefore(NewCast, I);
3152 // trunc_or_bitcast(C1)&C2
3153 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3154 C3 = ConstantExpr::getAnd(C3, AndRHS);
3155 return BinaryOperator::createAnd(NewCast, C3);
3156 } else if (CastOp->getOpcode() == Instruction::Or) {
3157 // Change: and (cast (or X, C1) to T), C2
3158 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3159 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3160 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3161 return ReplaceInstUsesWith(I, AndRHS);
3166 // Try to fold constant and into select arguments.
3167 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3168 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3170 if (isa<PHINode>(Op0))
3171 if (Instruction *NV = FoldOpIntoPhi(I))
3175 Value *Op0NotVal = dyn_castNotVal(Op0);
3176 Value *Op1NotVal = dyn_castNotVal(Op1);
3178 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3179 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3181 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3182 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3183 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3184 I.getName()+".demorgan");
3185 InsertNewInstBefore(Or, I);
3186 return BinaryOperator::createNot(Or);
3190 Value *A = 0, *B = 0;
3191 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3192 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3193 return ReplaceInstUsesWith(I, Op1);
3194 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3195 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3196 return ReplaceInstUsesWith(I, Op0);
3198 if (Op0->hasOneUse() &&
3199 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3200 if (A == Op1) { // (A^B)&A -> A&(A^B)
3201 I.swapOperands(); // Simplify below
3202 std::swap(Op0, Op1);
3203 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3204 cast<BinaryOperator>(Op0)->swapOperands();
3205 I.swapOperands(); // Simplify below
3206 std::swap(Op0, Op1);
3209 if (Op1->hasOneUse() &&
3210 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3211 if (B == Op0) { // B&(A^B) -> B&(B^A)
3212 cast<BinaryOperator>(Op1)->swapOperands();
3215 if (A == Op0) { // A&(A^B) -> A & ~B
3216 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3217 InsertNewInstBefore(NotB, I);
3218 return BinaryOperator::createAnd(A, NotB);
3223 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3224 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3225 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3228 Value *LHSVal, *RHSVal;
3229 ConstantInt *LHSCst, *RHSCst;
3230 ICmpInst::Predicate LHSCC, RHSCC;
3231 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3232 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3233 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3234 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3235 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3236 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3237 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3238 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3239 // Ensure that the larger constant is on the RHS.
3240 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3241 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3242 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3243 ICmpInst *LHS = cast<ICmpInst>(Op0);
3244 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3245 std::swap(LHS, RHS);
3246 std::swap(LHSCst, RHSCst);
3247 std::swap(LHSCC, RHSCC);
3250 // At this point, we know we have have two icmp instructions
3251 // comparing a value against two constants and and'ing the result
3252 // together. Because of the above check, we know that we only have
3253 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3254 // (from the FoldICmpLogical check above), that the two constants
3255 // are not equal and that the larger constant is on the RHS
3256 assert(LHSCst != RHSCst && "Compares not folded above?");
3259 default: assert(0 && "Unknown integer condition code!");
3260 case ICmpInst::ICMP_EQ:
3262 default: assert(0 && "Unknown integer condition code!");
3263 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3264 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3265 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3266 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3267 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3268 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3269 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3270 return ReplaceInstUsesWith(I, LHS);
3272 case ICmpInst::ICMP_NE:
3274 default: assert(0 && "Unknown integer condition code!");
3275 case ICmpInst::ICMP_ULT:
3276 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3277 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3278 break; // (X != 13 & X u< 15) -> no change
3279 case ICmpInst::ICMP_SLT:
3280 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3281 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3282 break; // (X != 13 & X s< 15) -> no change
3283 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3284 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3285 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3286 return ReplaceInstUsesWith(I, RHS);
3287 case ICmpInst::ICMP_NE:
3288 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3289 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3290 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3291 LHSVal->getName()+".off");
3292 InsertNewInstBefore(Add, I);
3293 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3294 ConstantInt::get(Add->getType(), 1));
3296 break; // (X != 13 & X != 15) -> no change
3299 case ICmpInst::ICMP_ULT:
3301 default: assert(0 && "Unknown integer condition code!");
3302 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3303 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3304 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3305 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3307 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3308 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3309 return ReplaceInstUsesWith(I, LHS);
3310 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3314 case ICmpInst::ICMP_SLT:
3316 default: assert(0 && "Unknown integer condition code!");
3317 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3318 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3319 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3320 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3322 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3323 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3324 return ReplaceInstUsesWith(I, LHS);
3325 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3329 case ICmpInst::ICMP_UGT:
3331 default: assert(0 && "Unknown integer condition code!");
3332 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3333 return ReplaceInstUsesWith(I, LHS);
3334 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3335 return ReplaceInstUsesWith(I, RHS);
3336 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3338 case ICmpInst::ICMP_NE:
3339 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3340 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3341 break; // (X u> 13 & X != 15) -> no change
3342 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3343 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3345 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3349 case ICmpInst::ICMP_SGT:
3351 default: assert(0 && "Unknown integer condition code!");
3352 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3353 return ReplaceInstUsesWith(I, LHS);
3354 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3355 return ReplaceInstUsesWith(I, RHS);
3356 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3358 case ICmpInst::ICMP_NE:
3359 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3360 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3361 break; // (X s> 13 & X != 15) -> no change
3362 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3363 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3365 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3373 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3374 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3375 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3376 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3377 const Type *SrcTy = Op0C->getOperand(0)->getType();
3378 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3379 // Only do this if the casts both really cause code to be generated.
3380 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3382 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3384 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3385 Op1C->getOperand(0),
3387 InsertNewInstBefore(NewOp, I);
3388 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3392 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3393 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3394 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3395 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3396 SI0->getOperand(1) == SI1->getOperand(1) &&
3397 (SI0->hasOneUse() || SI1->hasOneUse())) {
3398 Instruction *NewOp =
3399 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3401 SI0->getName()), I);
3402 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3403 SI1->getOperand(1));
3407 return Changed ? &I : 0;
3410 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3411 /// in the result. If it does, and if the specified byte hasn't been filled in
3412 /// yet, fill it in and return false.
3413 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
3414 Instruction *I = dyn_cast<Instruction>(V);
3415 if (I == 0) return true;
3417 // If this is an or instruction, it is an inner node of the bswap.
3418 if (I->getOpcode() == Instruction::Or)
3419 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3420 CollectBSwapParts(I->getOperand(1), ByteValues);
3422 // If this is a shift by a constant int, and it is "24", then its operand
3423 // defines a byte. We only handle unsigned types here.
3424 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3425 // Not shifting the entire input by N-1 bytes?
3426 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3427 8*(ByteValues.size()-1))
3431 if (I->getOpcode() == Instruction::Shl) {
3432 // X << 24 defines the top byte with the lowest of the input bytes.
3433 DestNo = ByteValues.size()-1;
3435 // X >>u 24 defines the low byte with the highest of the input bytes.
3439 // If the destination byte value is already defined, the values are or'd
3440 // together, which isn't a bswap (unless it's an or of the same bits).
3441 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3443 ByteValues[DestNo] = I->getOperand(0);
3447 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3449 Value *Shift = 0, *ShiftLHS = 0;
3450 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3451 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3452 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3454 Instruction *SI = cast<Instruction>(Shift);
3456 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3457 if (ShiftAmt->getZExtValue() & 7 ||
3458 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3461 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3463 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3464 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3466 // Unknown mask for bswap.
3467 if (DestByte == ByteValues.size()) return true;
3469 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3471 if (SI->getOpcode() == Instruction::Shl)
3472 SrcByte = DestByte - ShiftBytes;
3474 SrcByte = DestByte + ShiftBytes;
3476 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3477 if (SrcByte != ByteValues.size()-DestByte-1)
3480 // If the destination byte value is already defined, the values are or'd
3481 // together, which isn't a bswap (unless it's an or of the same bits).
3482 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3484 ByteValues[DestByte] = SI->getOperand(0);
3488 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3489 /// If so, insert the new bswap intrinsic and return it.
3490 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3491 // We cannot bswap one byte.
3492 if (I.getType() == Type::Int8Ty)
3495 /// ByteValues - For each byte of the result, we keep track of which value
3496 /// defines each byte.
3497 std::vector<Value*> ByteValues;
3498 ByteValues.resize(TD->getTypeSize(I.getType()));
3500 // Try to find all the pieces corresponding to the bswap.
3501 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3502 CollectBSwapParts(I.getOperand(1), ByteValues))
3505 // Check to see if all of the bytes come from the same value.
3506 Value *V = ByteValues[0];
3507 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3509 // Check to make sure that all of the bytes come from the same value.
3510 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3511 if (ByteValues[i] != V)
3514 // If they do then *success* we can turn this into a bswap. Figure out what
3515 // bswap to make it into.
3516 Module *M = I.getParent()->getParent()->getParent();
3517 const char *FnName = 0;
3518 if (I.getType() == Type::Int16Ty)
3519 FnName = "llvm.bswap.i16";
3520 else if (I.getType() == Type::Int32Ty)
3521 FnName = "llvm.bswap.i32";
3522 else if (I.getType() == Type::Int64Ty)
3523 FnName = "llvm.bswap.i64";
3525 assert(0 && "Unknown integer type!");
3526 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3527 return new CallInst(F, V);
3531 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3532 bool Changed = SimplifyCommutative(I);
3533 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3535 if (isa<UndefValue>(Op1))
3536 return ReplaceInstUsesWith(I, // X | undef -> -1
3537 ConstantInt::getAllOnesValue(I.getType()));
3541 return ReplaceInstUsesWith(I, Op0);
3543 // See if we can simplify any instructions used by the instruction whose sole
3544 // purpose is to compute bits we don't care about.
3545 uint64_t KnownZero, KnownOne;
3546 if (!isa<VectorType>(I.getType()) &&
3547 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3548 KnownZero, KnownOne))
3552 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3553 ConstantInt *C1 = 0; Value *X = 0;
3554 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3555 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3556 Instruction *Or = BinaryOperator::createOr(X, RHS);
3557 InsertNewInstBefore(Or, I);
3559 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3562 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3563 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3564 Instruction *Or = BinaryOperator::createOr(X, RHS);
3565 InsertNewInstBefore(Or, I);
3567 return BinaryOperator::createXor(Or,
3568 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3571 // Try to fold constant and into select arguments.
3572 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3573 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3575 if (isa<PHINode>(Op0))
3576 if (Instruction *NV = FoldOpIntoPhi(I))
3580 Value *A = 0, *B = 0;
3581 ConstantInt *C1 = 0, *C2 = 0;
3583 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3584 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3585 return ReplaceInstUsesWith(I, Op1);
3586 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3587 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3588 return ReplaceInstUsesWith(I, Op0);
3590 // (A | B) | C and A | (B | C) -> bswap if possible.
3591 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3592 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3593 match(Op1, m_Or(m_Value(), m_Value())) ||
3594 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3595 match(Op1, m_Shift(m_Value(), m_Value())))) {
3596 if (Instruction *BSwap = MatchBSwap(I))
3600 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3601 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3602 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3603 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3604 InsertNewInstBefore(NOr, I);
3606 return BinaryOperator::createXor(NOr, C1);
3609 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3610 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3611 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3612 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3613 InsertNewInstBefore(NOr, I);
3615 return BinaryOperator::createXor(NOr, C1);
3618 // (A & C1)|(B & C2)
3619 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3620 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3622 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3623 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3626 // If we have: ((V + N) & C1) | (V & C2)
3627 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3628 // replace with V+N.
3629 if (C1 == ConstantExpr::getNot(C2)) {
3630 Value *V1 = 0, *V2 = 0;
3631 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3632 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3633 // Add commutes, try both ways.
3634 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3635 return ReplaceInstUsesWith(I, A);
3636 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3637 return ReplaceInstUsesWith(I, A);
3639 // Or commutes, try both ways.
3640 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3641 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3642 // Add commutes, try both ways.
3643 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3644 return ReplaceInstUsesWith(I, B);
3645 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3646 return ReplaceInstUsesWith(I, B);
3651 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3652 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3653 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3654 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3655 SI0->getOperand(1) == SI1->getOperand(1) &&
3656 (SI0->hasOneUse() || SI1->hasOneUse())) {
3657 Instruction *NewOp =
3658 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3660 SI0->getName()), I);
3661 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3662 SI1->getOperand(1));
3666 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3667 if (A == Op1) // ~A | A == -1
3668 return ReplaceInstUsesWith(I,
3669 ConstantInt::getAllOnesValue(I.getType()));
3673 // Note, A is still live here!
3674 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3676 return ReplaceInstUsesWith(I,
3677 ConstantInt::getAllOnesValue(I.getType()));
3679 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3680 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3681 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3682 I.getName()+".demorgan"), I);
3683 return BinaryOperator::createNot(And);
3687 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3688 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3689 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3692 Value *LHSVal, *RHSVal;
3693 ConstantInt *LHSCst, *RHSCst;
3694 ICmpInst::Predicate LHSCC, RHSCC;
3695 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3696 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3697 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3698 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3699 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3700 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3701 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3702 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3703 // Ensure that the larger constant is on the RHS.
3704 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3705 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3706 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3707 ICmpInst *LHS = cast<ICmpInst>(Op0);
3708 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3709 std::swap(LHS, RHS);
3710 std::swap(LHSCst, RHSCst);
3711 std::swap(LHSCC, RHSCC);
3714 // At this point, we know we have have two icmp instructions
3715 // comparing a value against two constants and or'ing the result
3716 // together. Because of the above check, we know that we only have
3717 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3718 // FoldICmpLogical check above), that the two constants are not
3720 assert(LHSCst != RHSCst && "Compares not folded above?");
3723 default: assert(0 && "Unknown integer condition code!");
3724 case ICmpInst::ICMP_EQ:
3726 default: assert(0 && "Unknown integer condition code!");
3727 case ICmpInst::ICMP_EQ:
3728 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3729 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3730 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3731 LHSVal->getName()+".off");
3732 InsertNewInstBefore(Add, I);
3733 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3734 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3736 break; // (X == 13 | X == 15) -> no change
3737 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3738 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3740 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3741 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3742 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3743 return ReplaceInstUsesWith(I, RHS);
3746 case ICmpInst::ICMP_NE:
3748 default: assert(0 && "Unknown integer condition code!");
3749 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3750 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3751 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3752 return ReplaceInstUsesWith(I, LHS);
3753 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3754 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3755 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3756 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3759 case ICmpInst::ICMP_ULT:
3761 default: assert(0 && "Unknown integer condition code!");
3762 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3764 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3765 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3767 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3769 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3770 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3771 return ReplaceInstUsesWith(I, RHS);
3772 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3776 case ICmpInst::ICMP_SLT:
3778 default: assert(0 && "Unknown integer condition code!");
3779 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3781 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3782 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3784 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3786 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3787 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3788 return ReplaceInstUsesWith(I, RHS);
3789 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3793 case ICmpInst::ICMP_UGT:
3795 default: assert(0 && "Unknown integer condition code!");
3796 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3797 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3798 return ReplaceInstUsesWith(I, LHS);
3799 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3801 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3802 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3803 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3804 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3808 case ICmpInst::ICMP_SGT:
3810 default: assert(0 && "Unknown integer condition code!");
3811 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3812 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3813 return ReplaceInstUsesWith(I, LHS);
3814 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3816 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3817 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3818 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3819 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3827 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3828 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3829 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3830 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3831 const Type *SrcTy = Op0C->getOperand(0)->getType();
3832 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3833 // Only do this if the casts both really cause code to be generated.
3834 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3836 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3838 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3839 Op1C->getOperand(0),
3841 InsertNewInstBefore(NewOp, I);
3842 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3847 return Changed ? &I : 0;
3850 // XorSelf - Implements: X ^ X --> 0
3853 XorSelf(Value *rhs) : RHS(rhs) {}
3854 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3855 Instruction *apply(BinaryOperator &Xor) const {
3861 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3862 bool Changed = SimplifyCommutative(I);
3863 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3865 if (isa<UndefValue>(Op1))
3866 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3868 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3869 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3870 assert(Result == &I && "AssociativeOpt didn't work?");
3871 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3874 // See if we can simplify any instructions used by the instruction whose sole
3875 // purpose is to compute bits we don't care about.
3876 uint64_t KnownZero, KnownOne;
3877 if (!isa<VectorType>(I.getType()) &&
3878 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3879 KnownZero, KnownOne))
3882 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3883 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
3884 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3885 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
3886 return new ICmpInst(ICI->getInversePredicate(),
3887 ICI->getOperand(0), ICI->getOperand(1));
3889 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3890 // ~(c-X) == X-c-1 == X+(-c-1)
3891 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3892 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3893 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3894 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3895 ConstantInt::get(I.getType(), 1));
3896 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3899 // ~(~X & Y) --> (X | ~Y)
3900 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3901 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3902 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3904 BinaryOperator::createNot(Op0I->getOperand(1),
3905 Op0I->getOperand(1)->getName()+".not");
3906 InsertNewInstBefore(NotY, I);
3907 return BinaryOperator::createOr(Op0NotVal, NotY);
3911 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3912 if (Op0I->getOpcode() == Instruction::Add) {
3913 // ~(X-c) --> (-c-1)-X
3914 if (RHS->isAllOnesValue()) {
3915 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3916 return BinaryOperator::createSub(
3917 ConstantExpr::getSub(NegOp0CI,
3918 ConstantInt::get(I.getType(), 1)),
3919 Op0I->getOperand(0));
3921 } else if (Op0I->getOpcode() == Instruction::Or) {
3922 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3923 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3924 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3925 // Anything in both C1 and C2 is known to be zero, remove it from
3927 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3928 NewRHS = ConstantExpr::getAnd(NewRHS,
3929 ConstantExpr::getNot(CommonBits));
3930 WorkList.push_back(Op0I);
3931 I.setOperand(0, Op0I->getOperand(0));
3932 I.setOperand(1, NewRHS);
3938 // Try to fold constant and into select arguments.
3939 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3940 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3942 if (isa<PHINode>(Op0))
3943 if (Instruction *NV = FoldOpIntoPhi(I))
3947 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3949 return ReplaceInstUsesWith(I,
3950 ConstantInt::getAllOnesValue(I.getType()));
3952 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3954 return ReplaceInstUsesWith(I,
3955 ConstantInt::getAllOnesValue(I.getType()));
3957 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3958 if (Op1I->getOpcode() == Instruction::Or) {
3959 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3960 Op1I->swapOperands();
3962 std::swap(Op0, Op1);
3963 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3964 I.swapOperands(); // Simplified below.
3965 std::swap(Op0, Op1);
3967 } else if (Op1I->getOpcode() == Instruction::Xor) {
3968 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3969 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3970 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3971 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3972 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3973 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3974 Op1I->swapOperands();
3975 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3976 I.swapOperands(); // Simplified below.
3977 std::swap(Op0, Op1);
3981 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3982 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
3983 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
3984 Op0I->swapOperands();
3985 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
3986 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
3987 InsertNewInstBefore(NotB, I);
3988 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
3990 } else if (Op0I->getOpcode() == Instruction::Xor) {
3991 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
3992 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3993 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
3994 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3995 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
3996 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
3997 Op0I->swapOperands();
3998 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
3999 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4000 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
4001 InsertNewInstBefore(N, I);
4002 return BinaryOperator::createAnd(N, Op1);
4006 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4007 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4008 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4011 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4012 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4013 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4014 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4015 const Type *SrcTy = Op0C->getOperand(0)->getType();
4016 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4017 // Only do this if the casts both really cause code to be generated.
4018 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4020 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4022 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4023 Op1C->getOperand(0),
4025 InsertNewInstBefore(NewOp, I);
4026 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4030 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4031 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4032 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4033 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4034 SI0->getOperand(1) == SI1->getOperand(1) &&
4035 (SI0->hasOneUse() || SI1->hasOneUse())) {
4036 Instruction *NewOp =
4037 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
4039 SI0->getName()), I);
4040 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4041 SI1->getOperand(1));
4045 return Changed ? &I : 0;
4048 static bool isPositive(ConstantInt *C) {
4049 return C->getSExtValue() >= 0;
4052 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4053 /// overflowed for this type.
4054 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4056 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4058 return cast<ConstantInt>(Result)->getZExtValue() <
4059 cast<ConstantInt>(In1)->getZExtValue();
4062 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4063 /// code necessary to compute the offset from the base pointer (without adding
4064 /// in the base pointer). Return the result as a signed integer of intptr size.
4065 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4066 TargetData &TD = IC.getTargetData();
4067 gep_type_iterator GTI = gep_type_begin(GEP);
4068 const Type *IntPtrTy = TD.getIntPtrType();
4069 Value *Result = Constant::getNullValue(IntPtrTy);
4071 // Build a mask for high order bits.
4072 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4074 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4075 Value *Op = GEP->getOperand(i);
4076 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4077 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4078 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4079 if (!OpC->isNullValue()) {
4080 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4081 Scale = ConstantExpr::getMul(OpC, Scale);
4082 if (Constant *RC = dyn_cast<Constant>(Result))
4083 Result = ConstantExpr::getAdd(RC, Scale);
4085 // Emit an add instruction.
4086 Result = IC.InsertNewInstBefore(
4087 BinaryOperator::createAdd(Result, Scale,
4088 GEP->getName()+".offs"), I);
4092 // Convert to correct type.
4093 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4094 Op->getName()+".c"), I);
4096 // We'll let instcombine(mul) convert this to a shl if possible.
4097 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4098 GEP->getName()+".idx"), I);
4100 // Emit an add instruction.
4101 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4102 GEP->getName()+".offs"), I);
4108 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4109 /// else. At this point we know that the GEP is on the LHS of the comparison.
4110 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4111 ICmpInst::Predicate Cond,
4113 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4115 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4116 if (isa<PointerType>(CI->getOperand(0)->getType()))
4117 RHS = CI->getOperand(0);
4119 Value *PtrBase = GEPLHS->getOperand(0);
4120 if (PtrBase == RHS) {
4121 // As an optimization, we don't actually have to compute the actual value of
4122 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4123 // each index is zero or not.
4124 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4125 Instruction *InVal = 0;
4126 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4127 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4129 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4130 if (isa<UndefValue>(C)) // undef index -> undef.
4131 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4132 if (C->isNullValue())
4134 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4135 EmitIt = false; // This is indexing into a zero sized array?
4136 } else if (isa<ConstantInt>(C))
4137 return ReplaceInstUsesWith(I, // No comparison is needed here.
4138 ConstantInt::get(Type::Int1Ty,
4139 Cond == ICmpInst::ICMP_NE));
4144 new ICmpInst(Cond, GEPLHS->getOperand(i),
4145 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4149 InVal = InsertNewInstBefore(InVal, I);
4150 InsertNewInstBefore(Comp, I);
4151 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4152 InVal = BinaryOperator::createOr(InVal, Comp);
4153 else // True if all are equal
4154 InVal = BinaryOperator::createAnd(InVal, Comp);
4162 // No comparison is needed here, all indexes = 0
4163 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4164 Cond == ICmpInst::ICMP_EQ));
4167 // Only lower this if the icmp is the only user of the GEP or if we expect
4168 // the result to fold to a constant!
4169 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4170 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4171 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4172 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4173 Constant::getNullValue(Offset->getType()));
4175 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4176 // If the base pointers are different, but the indices are the same, just
4177 // compare the base pointer.
4178 if (PtrBase != GEPRHS->getOperand(0)) {
4179 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4180 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4181 GEPRHS->getOperand(0)->getType();
4183 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4184 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4185 IndicesTheSame = false;
4189 // If all indices are the same, just compare the base pointers.
4191 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4192 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4194 // Otherwise, the base pointers are different and the indices are
4195 // different, bail out.
4199 // If one of the GEPs has all zero indices, recurse.
4200 bool AllZeros = true;
4201 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4202 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4203 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4208 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4209 ICmpInst::getSwappedPredicate(Cond), I);
4211 // If the other GEP has all zero indices, recurse.
4213 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4214 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4215 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4220 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4222 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4223 // If the GEPs only differ by one index, compare it.
4224 unsigned NumDifferences = 0; // Keep track of # differences.
4225 unsigned DiffOperand = 0; // The operand that differs.
4226 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4227 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4228 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4229 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4230 // Irreconcilable differences.
4234 if (NumDifferences++) break;
4239 if (NumDifferences == 0) // SAME GEP?
4240 return ReplaceInstUsesWith(I, // No comparison is needed here.
4241 ConstantInt::get(Type::Int1Ty,
4242 Cond == ICmpInst::ICMP_EQ));
4243 else if (NumDifferences == 1) {
4244 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4245 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4246 // Make sure we do a signed comparison here.
4247 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4251 // Only lower this if the icmp is the only user of the GEP or if we expect
4252 // the result to fold to a constant!
4253 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4254 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4255 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4256 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4257 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4258 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4264 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4265 bool Changed = SimplifyCompare(I);
4266 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4268 // Fold trivial predicates.
4269 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4270 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4271 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4272 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4274 // Simplify 'fcmp pred X, X'
4276 switch (I.getPredicate()) {
4277 default: assert(0 && "Unknown predicate!");
4278 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4279 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4280 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4281 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4282 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4283 case FCmpInst::FCMP_OLT: // True if ordered and less than
4284 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4285 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4287 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4288 case FCmpInst::FCMP_ULT: // True if unordered or less than
4289 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4290 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4291 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4292 I.setPredicate(FCmpInst::FCMP_UNO);
4293 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4296 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4297 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4298 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4299 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4300 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4301 I.setPredicate(FCmpInst::FCMP_ORD);
4302 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4307 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4308 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4310 // Handle fcmp with constant RHS
4311 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4312 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4313 switch (LHSI->getOpcode()) {
4314 case Instruction::PHI:
4315 if (Instruction *NV = FoldOpIntoPhi(I))
4318 case Instruction::Select:
4319 // If either operand of the select is a constant, we can fold the
4320 // comparison into the select arms, which will cause one to be
4321 // constant folded and the select turned into a bitwise or.
4322 Value *Op1 = 0, *Op2 = 0;
4323 if (LHSI->hasOneUse()) {
4324 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4325 // Fold the known value into the constant operand.
4326 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4327 // Insert a new FCmp of the other select operand.
4328 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4329 LHSI->getOperand(2), RHSC,
4331 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4332 // Fold the known value into the constant operand.
4333 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4334 // Insert a new FCmp of the other select operand.
4335 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4336 LHSI->getOperand(1), RHSC,
4342 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4347 return Changed ? &I : 0;
4350 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4351 bool Changed = SimplifyCompare(I);
4352 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4353 const Type *Ty = Op0->getType();
4357 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4358 isTrueWhenEqual(I)));
4360 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4361 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4363 // icmp of GlobalValues can never equal each other as long as they aren't
4364 // external weak linkage type.
4365 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4366 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4367 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4368 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4369 !isTrueWhenEqual(I)));
4371 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4372 // addresses never equal each other! We already know that Op0 != Op1.
4373 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4374 isa<ConstantPointerNull>(Op0)) &&
4375 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4376 isa<ConstantPointerNull>(Op1)))
4377 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4378 !isTrueWhenEqual(I)));
4380 // icmp's with boolean values can always be turned into bitwise operations
4381 if (Ty == Type::Int1Ty) {
4382 switch (I.getPredicate()) {
4383 default: assert(0 && "Invalid icmp instruction!");
4384 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4385 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4386 InsertNewInstBefore(Xor, I);
4387 return BinaryOperator::createNot(Xor);
4389 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4390 return BinaryOperator::createXor(Op0, Op1);
4392 case ICmpInst::ICMP_UGT:
4393 case ICmpInst::ICMP_SGT:
4394 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4396 case ICmpInst::ICMP_ULT:
4397 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4398 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4399 InsertNewInstBefore(Not, I);
4400 return BinaryOperator::createAnd(Not, Op1);
4402 case ICmpInst::ICMP_UGE:
4403 case ICmpInst::ICMP_SGE:
4404 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4406 case ICmpInst::ICMP_ULE:
4407 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4408 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4409 InsertNewInstBefore(Not, I);
4410 return BinaryOperator::createOr(Not, Op1);
4415 // See if we are doing a comparison between a constant and an instruction that
4416 // can be folded into the comparison.
4417 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4418 switch (I.getPredicate()) {
4420 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4421 if (CI->isMinValue(false))
4422 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4423 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4424 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4425 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4426 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4429 case ICmpInst::ICMP_SLT:
4430 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4431 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4432 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4433 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4434 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4435 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4438 case ICmpInst::ICMP_UGT:
4439 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4440 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4441 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4442 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4443 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4444 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4447 case ICmpInst::ICMP_SGT:
4448 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4449 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4450 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4451 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4452 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4453 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4456 case ICmpInst::ICMP_ULE:
4457 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4458 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4459 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4460 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4461 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4462 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4465 case ICmpInst::ICMP_SLE:
4466 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4467 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4468 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4469 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4470 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4471 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4474 case ICmpInst::ICMP_UGE:
4475 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4476 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4477 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4478 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4479 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4480 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4483 case ICmpInst::ICMP_SGE:
4484 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4485 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4486 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4487 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4488 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4489 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4493 // If we still have a icmp le or icmp ge instruction, turn it into the
4494 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4495 // already been handled above, this requires little checking.
4497 if (I.getPredicate() == ICmpInst::ICMP_ULE)
4498 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4499 if (I.getPredicate() == ICmpInst::ICMP_SLE)
4500 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4501 if (I.getPredicate() == ICmpInst::ICMP_UGE)
4502 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4503 if (I.getPredicate() == ICmpInst::ICMP_SGE)
4504 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4506 // See if we can fold the comparison based on bits known to be zero or one
4508 uint64_t KnownZero, KnownOne;
4509 if (SimplifyDemandedBits(Op0, cast<IntegerType>(Ty)->getBitMask(),
4510 KnownZero, KnownOne, 0))
4513 // Given the known and unknown bits, compute a range that the LHS could be
4515 if (KnownOne | KnownZero) {
4516 // Compute the Min, Max and RHS values based on the known bits. For the
4517 // EQ and NE we use unsigned values.
4518 uint64_t UMin = 0, UMax = 0, URHSVal = 0;
4519 int64_t SMin = 0, SMax = 0, SRHSVal = 0;
4520 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4521 SRHSVal = CI->getSExtValue();
4522 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
4525 URHSVal = CI->getZExtValue();
4526 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
4529 switch (I.getPredicate()) { // LE/GE have been folded already.
4530 default: assert(0 && "Unknown icmp opcode!");
4531 case ICmpInst::ICMP_EQ:
4532 if (UMax < URHSVal || UMin > URHSVal)
4533 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4535 case ICmpInst::ICMP_NE:
4536 if (UMax < URHSVal || UMin > URHSVal)
4537 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4539 case ICmpInst::ICMP_ULT:
4541 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4543 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4545 case ICmpInst::ICMP_UGT:
4547 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4549 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4551 case ICmpInst::ICMP_SLT:
4553 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4555 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4557 case ICmpInst::ICMP_SGT:
4559 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4561 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4566 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4567 // instruction, see if that instruction also has constants so that the
4568 // instruction can be folded into the icmp
4569 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4570 switch (LHSI->getOpcode()) {
4571 case Instruction::And:
4572 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4573 LHSI->getOperand(0)->hasOneUse()) {
4574 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4576 // If the LHS is an AND of a truncating cast, we can widen the
4577 // and/compare to be the input width without changing the value
4578 // produced, eliminating a cast.
4579 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4580 // We can do this transformation if either the AND constant does not
4581 // have its sign bit set or if it is an equality comparison.
4582 // Extending a relational comparison when we're checking the sign
4583 // bit would not work.
4584 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4586 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4587 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4588 ConstantInt *NewCST;
4590 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4591 AndCST->getZExtValue());
4592 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4593 CI->getZExtValue());
4594 Instruction *NewAnd =
4595 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4597 InsertNewInstBefore(NewAnd, I);
4598 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4602 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4603 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4604 // happens a LOT in code produced by the C front-end, for bitfield
4606 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
4607 if (Shift && !Shift->isShift())
4611 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4612 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4613 const Type *AndTy = AndCST->getType(); // Type of the and.
4615 // We can fold this as long as we can't shift unknown bits
4616 // into the mask. This can only happen with signed shift
4617 // rights, as they sign-extend.
4619 bool CanFold = Shift->isLogicalShift();
4621 // To test for the bad case of the signed shr, see if any
4622 // of the bits shifted in could be tested after the mask.
4623 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4624 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4626 Constant *OShAmt = ConstantInt::get(AndTy, ShAmtVal);
4628 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4630 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4636 if (Shift->getOpcode() == Instruction::Shl)
4637 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4639 NewCst = ConstantExpr::getShl(CI, ShAmt);
4641 // Check to see if we are shifting out any of the bits being
4643 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4644 // If we shifted bits out, the fold is not going to work out.
4645 // As a special case, check to see if this means that the
4646 // result is always true or false now.
4647 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4648 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4649 if (I.getPredicate() == ICmpInst::ICMP_NE)
4650 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4652 I.setOperand(1, NewCst);
4653 Constant *NewAndCST;
4654 if (Shift->getOpcode() == Instruction::Shl)
4655 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4657 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4658 LHSI->setOperand(1, NewAndCST);
4659 LHSI->setOperand(0, Shift->getOperand(0));
4660 WorkList.push_back(Shift); // Shift is dead.
4661 AddUsesToWorkList(I);
4667 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4668 // preferable because it allows the C<<Y expression to be hoisted out
4669 // of a loop if Y is invariant and X is not.
4670 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4671 I.isEquality() && !Shift->isArithmeticShift() &&
4672 isa<Instruction>(Shift->getOperand(0))) {
4675 if (Shift->getOpcode() == Instruction::LShr) {
4676 NS = BinaryOperator::createShl(AndCST,
4677 Shift->getOperand(1), "tmp");
4679 // Insert a logical shift.
4680 NS = BinaryOperator::createLShr(AndCST,
4681 Shift->getOperand(1), "tmp");
4683 InsertNewInstBefore(cast<Instruction>(NS), I);
4685 // Compute X & (C << Y).
4686 Instruction *NewAnd = BinaryOperator::createAnd(
4687 Shift->getOperand(0), NS, LHSI->getName());
4688 InsertNewInstBefore(NewAnd, I);
4690 I.setOperand(0, NewAnd);
4696 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4697 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4698 if (I.isEquality()) {
4699 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4701 // Check that the shift amount is in range. If not, don't perform
4702 // undefined shifts. When the shift is visited it will be
4704 if (ShAmt->getZExtValue() >= TypeBits)
4707 // If we are comparing against bits always shifted out, the
4708 // comparison cannot succeed.
4710 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4711 if (Comp != CI) {// Comparing against a bit that we know is zero.
4712 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4713 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4714 return ReplaceInstUsesWith(I, Cst);
4717 if (LHSI->hasOneUse()) {
4718 // Otherwise strength reduce the shift into an and.
4719 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4720 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4721 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4724 BinaryOperator::createAnd(LHSI->getOperand(0),
4725 Mask, LHSI->getName()+".mask");
4726 Value *And = InsertNewInstBefore(AndI, I);
4727 return new ICmpInst(I.getPredicate(), And,
4728 ConstantExpr::getLShr(CI, ShAmt));
4734 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4735 case Instruction::AShr:
4736 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4737 if (I.isEquality()) {
4738 // Check that the shift amount is in range. If not, don't perform
4739 // undefined shifts. When the shift is visited it will be
4741 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4742 if (ShAmt->getZExtValue() >= TypeBits)
4745 // If we are comparing against bits always shifted out, the
4746 // comparison cannot succeed.
4748 if (LHSI->getOpcode() == Instruction::LShr)
4749 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4752 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4755 if (Comp != CI) {// Comparing against a bit that we know is zero.
4756 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4757 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4758 return ReplaceInstUsesWith(I, Cst);
4761 if (LHSI->hasOneUse() || CI->isNullValue()) {
4762 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4764 // Otherwise strength reduce the shift into an and.
4765 uint64_t Val = ~0ULL; // All ones.
4766 Val <<= ShAmtVal; // Shift over to the right spot.
4767 Val &= ~0ULL >> (64-TypeBits);
4768 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4771 BinaryOperator::createAnd(LHSI->getOperand(0),
4772 Mask, LHSI->getName()+".mask");
4773 Value *And = InsertNewInstBefore(AndI, I);
4774 return new ICmpInst(I.getPredicate(), And,
4775 ConstantExpr::getShl(CI, ShAmt));
4781 case Instruction::SDiv:
4782 case Instruction::UDiv:
4783 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4784 // Fold this div into the comparison, producing a range check.
4785 // Determine, based on the divide type, what the range is being
4786 // checked. If there is an overflow on the low or high side, remember
4787 // it, otherwise compute the range [low, hi) bounding the new value.
4788 // See: InsertRangeTest above for the kinds of replacements possible.
4789 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4790 // FIXME: If the operand types don't match the type of the divide
4791 // then don't attempt this transform. The code below doesn't have the
4792 // logic to deal with a signed divide and an unsigned compare (and
4793 // vice versa). This is because (x /s C1) <s C2 produces different
4794 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4795 // (x /u C1) <u C2. Simply casting the operands and result won't
4796 // work. :( The if statement below tests that condition and bails
4798 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4799 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4802 // Initialize the variables that will indicate the nature of the
4804 bool LoOverflow = false, HiOverflow = false;
4805 ConstantInt *LoBound = 0, *HiBound = 0;
4807 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4808 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4809 // C2 (CI). By solving for X we can turn this into a range check
4810 // instead of computing a divide.
4812 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4814 // Determine if the product overflows by seeing if the product is
4815 // not equal to the divide. Make sure we do the same kind of divide
4816 // as in the LHS instruction that we're folding.
4817 bool ProdOV = !DivRHS->isNullValue() &&
4818 (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4819 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4821 // Get the ICmp opcode
4822 ICmpInst::Predicate predicate = I.getPredicate();
4824 if (DivRHS->isNullValue()) {
4825 // Don't hack on divide by zeros!
4826 } else if (!DivIsSigned) { // udiv
4828 LoOverflow = ProdOV;
4829 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4830 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4831 if (CI->isNullValue()) { // (X / pos) op 0
4833 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4835 } else if (isPositive(CI)) { // (X / pos) op pos
4837 LoOverflow = ProdOV;
4838 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4839 } else { // (X / pos) op neg
4840 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4841 LoOverflow = AddWithOverflow(LoBound, Prod,
4842 cast<ConstantInt>(DivRHSH));
4844 HiOverflow = ProdOV;
4846 } else { // Divisor is < 0.
4847 if (CI->isNullValue()) { // (X / neg) op 0
4848 LoBound = AddOne(DivRHS);
4849 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4850 if (HiBound == DivRHS)
4851 LoBound = 0; // - INTMIN = INTMIN
4852 } else if (isPositive(CI)) { // (X / neg) op pos
4853 HiOverflow = LoOverflow = ProdOV;
4855 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4856 HiBound = AddOne(Prod);
4857 } else { // (X / neg) op neg
4859 LoOverflow = HiOverflow = ProdOV;
4860 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4863 // Dividing by a negate swaps the condition.
4864 predicate = ICmpInst::getSwappedPredicate(predicate);
4868 Value *X = LHSI->getOperand(0);
4869 switch (predicate) {
4870 default: assert(0 && "Unhandled icmp opcode!");
4871 case ICmpInst::ICMP_EQ:
4872 if (LoOverflow && HiOverflow)
4873 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4874 else if (HiOverflow)
4875 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4876 ICmpInst::ICMP_UGE, X, LoBound);
4877 else if (LoOverflow)
4878 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4879 ICmpInst::ICMP_ULT, X, HiBound);
4881 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4883 case ICmpInst::ICMP_NE:
4884 if (LoOverflow && HiOverflow)
4885 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4886 else if (HiOverflow)
4887 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4888 ICmpInst::ICMP_ULT, X, LoBound);
4889 else if (LoOverflow)
4890 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4891 ICmpInst::ICMP_UGE, X, HiBound);
4893 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4895 case ICmpInst::ICMP_ULT:
4896 case ICmpInst::ICMP_SLT:
4898 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4899 return new ICmpInst(predicate, X, LoBound);
4900 case ICmpInst::ICMP_UGT:
4901 case ICmpInst::ICMP_SGT:
4903 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4904 if (predicate == ICmpInst::ICMP_UGT)
4905 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
4907 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
4914 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
4915 if (I.isEquality()) {
4916 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4918 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4919 // the second operand is a constant, simplify a bit.
4920 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4921 switch (BO->getOpcode()) {
4922 case Instruction::SRem:
4923 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4924 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4926 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4927 if (V > 1 && isPowerOf2_64(V)) {
4928 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4929 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4930 return new ICmpInst(I.getPredicate(), NewRem,
4931 Constant::getNullValue(BO->getType()));
4935 case Instruction::Add:
4936 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4937 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4938 if (BO->hasOneUse())
4939 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4940 ConstantExpr::getSub(CI, BOp1C));
4941 } else if (CI->isNullValue()) {
4942 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4943 // efficiently invertible, or if the add has just this one use.
4944 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4946 if (Value *NegVal = dyn_castNegVal(BOp1))
4947 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
4948 else if (Value *NegVal = dyn_castNegVal(BOp0))
4949 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
4950 else if (BO->hasOneUse()) {
4951 Instruction *Neg = BinaryOperator::createNeg(BOp1);
4952 InsertNewInstBefore(Neg, I);
4954 return new ICmpInst(I.getPredicate(), BOp0, Neg);
4958 case Instruction::Xor:
4959 // For the xor case, we can xor two constants together, eliminating
4960 // the explicit xor.
4961 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4962 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4963 ConstantExpr::getXor(CI, BOC));
4966 case Instruction::Sub:
4967 // Replace (([sub|xor] A, B) != 0) with (A != B)
4968 if (CI->isNullValue())
4969 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4973 case Instruction::Or:
4974 // If bits are being or'd in that are not present in the constant we
4975 // are comparing against, then the comparison could never succeed!
4976 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4977 Constant *NotCI = ConstantExpr::getNot(CI);
4978 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4979 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4984 case Instruction::And:
4985 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4986 // If bits are being compared against that are and'd out, then the
4987 // comparison can never succeed!
4988 if (!ConstantExpr::getAnd(CI,
4989 ConstantExpr::getNot(BOC))->isNullValue())
4990 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4993 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4994 if (CI == BOC && isOneBitSet(CI))
4995 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
4996 ICmpInst::ICMP_NE, Op0,
4997 Constant::getNullValue(CI->getType()));
4999 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5000 if (isSignBit(BOC)) {
5001 Value *X = BO->getOperand(0);
5002 Constant *Zero = Constant::getNullValue(X->getType());
5003 ICmpInst::Predicate pred = isICMP_NE ?
5004 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5005 return new ICmpInst(pred, X, Zero);
5008 // ((X & ~7) == 0) --> X < 8
5009 if (CI->isNullValue() && isHighOnes(BOC)) {
5010 Value *X = BO->getOperand(0);
5011 Constant *NegX = ConstantExpr::getNeg(BOC);
5012 ICmpInst::Predicate pred = isICMP_NE ?
5013 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5014 return new ICmpInst(pred, X, NegX);
5020 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5021 // Handle set{eq|ne} <intrinsic>, intcst.
5022 switch (II->getIntrinsicID()) {
5024 case Intrinsic::bswap_i16:
5025 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5026 WorkList.push_back(II); // Dead?
5027 I.setOperand(0, II->getOperand(1));
5028 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5029 ByteSwap_16(CI->getZExtValue())));
5031 case Intrinsic::bswap_i32:
5032 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5033 WorkList.push_back(II); // Dead?
5034 I.setOperand(0, II->getOperand(1));
5035 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5036 ByteSwap_32(CI->getZExtValue())));
5038 case Intrinsic::bswap_i64:
5039 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5040 WorkList.push_back(II); // Dead?
5041 I.setOperand(0, II->getOperand(1));
5042 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5043 ByteSwap_64(CI->getZExtValue())));
5047 } else { // Not a ICMP_EQ/ICMP_NE
5048 // If the LHS is a cast from an integral value of the same size, then
5049 // since we know the RHS is a constant, try to simlify.
5050 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5051 Value *CastOp = Cast->getOperand(0);
5052 const Type *SrcTy = CastOp->getType();
5053 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5054 if (SrcTy->isInteger() &&
5055 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5056 // If this is an unsigned comparison, try to make the comparison use
5057 // smaller constant values.
5058 switch (I.getPredicate()) {
5060 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5061 ConstantInt *CUI = cast<ConstantInt>(CI);
5062 if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
5063 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5064 ConstantInt::get(SrcTy, -1));
5067 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5068 ConstantInt *CUI = cast<ConstantInt>(CI);
5069 if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
5070 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5071 Constant::getNullValue(SrcTy));
5081 // Handle icmp with constant RHS
5082 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5083 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5084 switch (LHSI->getOpcode()) {
5085 case Instruction::GetElementPtr:
5086 if (RHSC->isNullValue()) {
5087 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5088 bool isAllZeros = true;
5089 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5090 if (!isa<Constant>(LHSI->getOperand(i)) ||
5091 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5096 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5097 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5101 case Instruction::PHI:
5102 if (Instruction *NV = FoldOpIntoPhi(I))
5105 case Instruction::Select:
5106 // If either operand of the select is a constant, we can fold the
5107 // comparison into the select arms, which will cause one to be
5108 // constant folded and the select turned into a bitwise or.
5109 Value *Op1 = 0, *Op2 = 0;
5110 if (LHSI->hasOneUse()) {
5111 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5112 // Fold the known value into the constant operand.
5113 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5114 // Insert a new ICmp of the other select operand.
5115 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5116 LHSI->getOperand(2), RHSC,
5118 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5119 // Fold the known value into the constant operand.
5120 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5121 // Insert a new ICmp of the other select operand.
5122 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5123 LHSI->getOperand(1), RHSC,
5129 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5134 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5135 if (User *GEP = dyn_castGetElementPtr(Op0))
5136 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5138 if (User *GEP = dyn_castGetElementPtr(Op1))
5139 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5140 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5143 // Test to see if the operands of the icmp are casted versions of other
5144 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5146 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5147 if (isa<PointerType>(Op0->getType()) &&
5148 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5149 // We keep moving the cast from the left operand over to the right
5150 // operand, where it can often be eliminated completely.
5151 Op0 = CI->getOperand(0);
5153 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5154 // so eliminate it as well.
5155 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5156 Op1 = CI2->getOperand(0);
5158 // If Op1 is a constant, we can fold the cast into the constant.
5159 if (Op0->getType() != Op1->getType())
5160 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5161 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5163 // Otherwise, cast the RHS right before the icmp
5164 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5166 return new ICmpInst(I.getPredicate(), Op0, Op1);
5170 if (isa<CastInst>(Op0)) {
5171 // Handle the special case of: icmp (cast bool to X), <cst>
5172 // This comes up when you have code like
5175 // For generality, we handle any zero-extension of any operand comparison
5176 // with a constant or another cast from the same type.
5177 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5178 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5182 if (I.isEquality()) {
5183 Value *A, *B, *C, *D;
5184 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5185 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5186 Value *OtherVal = A == Op1 ? B : A;
5187 return new ICmpInst(I.getPredicate(), OtherVal,
5188 Constant::getNullValue(A->getType()));
5191 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5192 // A^c1 == C^c2 --> A == C^(c1^c2)
5193 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5194 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5195 if (Op1->hasOneUse()) {
5196 Constant *NC = ConstantExpr::getXor(C1, C2);
5197 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5198 return new ICmpInst(I.getPredicate(), A,
5199 InsertNewInstBefore(Xor, I));
5202 // A^B == A^D -> B == D
5203 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5204 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5205 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5206 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5210 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5211 (A == Op0 || B == Op0)) {
5212 // A == (A^B) -> B == 0
5213 Value *OtherVal = A == Op0 ? B : A;
5214 return new ICmpInst(I.getPredicate(), OtherVal,
5215 Constant::getNullValue(A->getType()));
5217 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5218 // (A-B) == A -> B == 0
5219 return new ICmpInst(I.getPredicate(), B,
5220 Constant::getNullValue(B->getType()));
5222 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5223 // A == (A-B) -> B == 0
5224 return new ICmpInst(I.getPredicate(), B,
5225 Constant::getNullValue(B->getType()));
5228 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5229 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5230 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5231 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5232 Value *X = 0, *Y = 0, *Z = 0;
5235 X = B; Y = D; Z = A;
5236 } else if (A == D) {
5237 X = B; Y = C; Z = A;
5238 } else if (B == C) {
5239 X = A; Y = D; Z = B;
5240 } else if (B == D) {
5241 X = A; Y = C; Z = B;
5244 if (X) { // Build (X^Y) & Z
5245 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5246 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5247 I.setOperand(0, Op1);
5248 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5253 return Changed ? &I : 0;
5256 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5257 // We only handle extending casts so far.
5259 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5260 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5261 Value *LHSCIOp = LHSCI->getOperand(0);
5262 const Type *SrcTy = LHSCIOp->getType();
5263 const Type *DestTy = LHSCI->getType();
5266 // We only handle extension cast instructions, so far. Enforce this.
5267 if (LHSCI->getOpcode() != Instruction::ZExt &&
5268 LHSCI->getOpcode() != Instruction::SExt)
5271 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5272 bool isSignedCmp = ICI.isSignedPredicate();
5274 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5275 // Not an extension from the same type?
5276 RHSCIOp = CI->getOperand(0);
5277 if (RHSCIOp->getType() != LHSCIOp->getType())
5280 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5281 // and the other is a zext), then we can't handle this.
5282 if (CI->getOpcode() != LHSCI->getOpcode())
5285 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5286 // then we can't handle this.
5287 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5290 // Okay, just insert a compare of the reduced operands now!
5291 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5294 // If we aren't dealing with a constant on the RHS, exit early
5295 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5299 // Compute the constant that would happen if we truncated to SrcTy then
5300 // reextended to DestTy.
5301 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5302 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5304 // If the re-extended constant didn't change...
5306 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5307 // For example, we might have:
5308 // %A = sext short %X to uint
5309 // %B = icmp ugt uint %A, 1330
5310 // It is incorrect to transform this into
5311 // %B = icmp ugt short %X, 1330
5312 // because %A may have negative value.
5314 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5315 // OR operation is EQ/NE.
5316 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5317 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5322 // The re-extended constant changed so the constant cannot be represented
5323 // in the shorter type. Consequently, we cannot emit a simple comparison.
5325 // First, handle some easy cases. We know the result cannot be equal at this
5326 // point so handle the ICI.isEquality() cases
5327 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5328 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5329 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5330 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5332 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5333 // should have been folded away previously and not enter in here.
5336 // We're performing a signed comparison.
5337 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5338 Result = ConstantInt::getFalse(); // X < (small) --> false
5340 Result = ConstantInt::getTrue(); // X < (large) --> true
5342 // We're performing an unsigned comparison.
5344 // We're performing an unsigned comp with a sign extended value.
5345 // This is true if the input is >= 0. [aka >s -1]
5346 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5347 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5348 NegOne, ICI.getName()), ICI);
5350 // Unsigned extend & unsigned compare -> always true.
5351 Result = ConstantInt::getTrue();
5355 // Finally, return the value computed.
5356 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5357 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5358 return ReplaceInstUsesWith(ICI, Result);
5360 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5361 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5362 "ICmp should be folded!");
5363 if (Constant *CI = dyn_cast<Constant>(Result))
5364 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5366 return BinaryOperator::createNot(Result);
5370 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5371 return commonShiftTransforms(I);
5374 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5375 return commonShiftTransforms(I);
5378 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5379 return commonShiftTransforms(I);
5382 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5383 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5384 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5386 // shl X, 0 == X and shr X, 0 == X
5387 // shl 0, X == 0 and shr 0, X == 0
5388 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5389 Op0 == Constant::getNullValue(Op0->getType()))
5390 return ReplaceInstUsesWith(I, Op0);
5392 if (isa<UndefValue>(Op0)) {
5393 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5394 return ReplaceInstUsesWith(I, Op0);
5395 else // undef << X -> 0, undef >>u X -> 0
5396 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5398 if (isa<UndefValue>(Op1)) {
5399 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5400 return ReplaceInstUsesWith(I, Op0);
5401 else // X << undef, X >>u undef -> 0
5402 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5405 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5406 if (I.getOpcode() == Instruction::AShr)
5407 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5408 if (CSI->isAllOnesValue())
5409 return ReplaceInstUsesWith(I, CSI);
5411 // Try to fold constant and into select arguments.
5412 if (isa<Constant>(Op0))
5413 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5414 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5417 // See if we can turn a signed shr into an unsigned shr.
5418 if (I.isArithmeticShift()) {
5419 if (MaskedValueIsZero(Op0,
5420 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5421 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5425 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5426 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5431 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5432 BinaryOperator &I) {
5433 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5435 // See if we can simplify any instructions used by the instruction whose sole
5436 // purpose is to compute bits we don't care about.
5437 uint64_t KnownZero, KnownOne;
5438 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
5439 KnownZero, KnownOne))
5442 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5443 // of a signed value.
5445 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5446 if (Op1->getZExtValue() >= TypeBits) {
5447 if (I.getOpcode() != Instruction::AShr)
5448 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5450 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5455 // ((X*C1) << C2) == (X * (C1 << C2))
5456 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5457 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5458 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5459 return BinaryOperator::createMul(BO->getOperand(0),
5460 ConstantExpr::getShl(BOOp, Op1));
5462 // Try to fold constant and into select arguments.
5463 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5464 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5466 if (isa<PHINode>(Op0))
5467 if (Instruction *NV = FoldOpIntoPhi(I))
5470 if (Op0->hasOneUse()) {
5471 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5472 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5475 switch (Op0BO->getOpcode()) {
5477 case Instruction::Add:
5478 case Instruction::And:
5479 case Instruction::Or:
5480 case Instruction::Xor: {
5481 // These operators commute.
5482 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5483 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5484 match(Op0BO->getOperand(1),
5485 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5486 Instruction *YS = BinaryOperator::createShl(
5487 Op0BO->getOperand(0), Op1,
5489 InsertNewInstBefore(YS, I); // (Y << C)
5491 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5492 Op0BO->getOperand(1)->getName());
5493 InsertNewInstBefore(X, I); // (X + (Y << C))
5494 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5495 C2 = ConstantExpr::getShl(C2, Op1);
5496 return BinaryOperator::createAnd(X, C2);
5499 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5500 Value *Op0BOOp1 = Op0BO->getOperand(1);
5501 if (isLeftShift && Op0BOOp1->hasOneUse() && V2 == Op1 &&
5503 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5504 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)-> hasOneUse()) {
5505 Instruction *YS = BinaryOperator::createShl(
5506 Op0BO->getOperand(0), Op1,
5508 InsertNewInstBefore(YS, I); // (Y << C)
5510 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5511 V1->getName()+".mask");
5512 InsertNewInstBefore(XM, I); // X & (CC << C)
5514 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5519 case Instruction::Sub: {
5520 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5521 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5522 match(Op0BO->getOperand(0),
5523 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5524 Instruction *YS = BinaryOperator::createShl(
5525 Op0BO->getOperand(1), Op1,
5527 InsertNewInstBefore(YS, I); // (Y << C)
5529 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5530 Op0BO->getOperand(0)->getName());
5531 InsertNewInstBefore(X, I); // (X + (Y << C))
5532 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5533 C2 = ConstantExpr::getShl(C2, Op1);
5534 return BinaryOperator::createAnd(X, C2);
5537 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5538 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5539 match(Op0BO->getOperand(0),
5540 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5541 m_ConstantInt(CC))) && V2 == Op1 &&
5542 cast<BinaryOperator>(Op0BO->getOperand(0))
5543 ->getOperand(0)->hasOneUse()) {
5544 Instruction *YS = BinaryOperator::createShl(
5545 Op0BO->getOperand(1), Op1,
5547 InsertNewInstBefore(YS, I); // (Y << C)
5549 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5550 V1->getName()+".mask");
5551 InsertNewInstBefore(XM, I); // X & (CC << C)
5553 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5561 // If the operand is an bitwise operator with a constant RHS, and the
5562 // shift is the only use, we can pull it out of the shift.
5563 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5564 bool isValid = true; // Valid only for And, Or, Xor
5565 bool highBitSet = false; // Transform if high bit of constant set?
5567 switch (Op0BO->getOpcode()) {
5568 default: isValid = false; break; // Do not perform transform!
5569 case Instruction::Add:
5570 isValid = isLeftShift;
5572 case Instruction::Or:
5573 case Instruction::Xor:
5576 case Instruction::And:
5581 // If this is a signed shift right, and the high bit is modified
5582 // by the logical operation, do not perform the transformation.
5583 // The highBitSet boolean indicates the value of the high bit of
5584 // the constant which would cause it to be modified for this
5587 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
5588 uint64_t Val = Op0C->getZExtValue();
5589 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5593 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5595 Instruction *NewShift =
5596 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
5597 InsertNewInstBefore(NewShift, I);
5598 NewShift->takeName(Op0BO);
5600 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5607 // Find out if this is a shift of a shift by a constant.
5608 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
5609 if (ShiftOp && !ShiftOp->isShift())
5612 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5613 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5614 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5615 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5616 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
5617 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
5618 Value *X = ShiftOp->getOperand(0);
5620 unsigned AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5621 if (AmtSum > I.getType()->getPrimitiveSizeInBits())
5622 AmtSum = I.getType()->getPrimitiveSizeInBits();
5624 const IntegerType *Ty = cast<IntegerType>(I.getType());
5626 // Check for (X << c1) << c2 and (X >> c1) >> c2
5627 if (I.getOpcode() == ShiftOp->getOpcode()) {
5628 return BinaryOperator::create(I.getOpcode(), X,
5629 ConstantInt::get(Ty, AmtSum));
5630 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
5631 I.getOpcode() == Instruction::AShr) {
5632 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
5633 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
5634 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
5635 I.getOpcode() == Instruction::LShr) {
5636 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
5637 Instruction *Shift =
5638 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
5639 InsertNewInstBefore(Shift, I);
5641 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5642 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5645 // Okay, if we get here, one shift must be left, and the other shift must be
5646 // right. See if the amounts are equal.
5647 if (ShiftAmt1 == ShiftAmt2) {
5648 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
5649 if (I.getOpcode() == Instruction::Shl) {
5650 uint64_t Mask = Ty->getBitMask() << ShiftAmt1;
5651 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
5653 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
5654 if (I.getOpcode() == Instruction::LShr) {
5655 uint64_t Mask = Ty->getBitMask() >> ShiftAmt1;
5656 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
5658 // We can simplify ((X << C) >>s C) into a trunc + sext.
5659 // NOTE: we could do this for any C, but that would make 'unusual' integer
5660 // types. For now, just stick to ones well-supported by the code
5662 const Type *SExtType = 0;
5663 switch (Ty->getBitWidth() - ShiftAmt1) {
5664 case 8 : SExtType = Type::Int8Ty; break;
5665 case 16: SExtType = Type::Int16Ty; break;
5666 case 32: SExtType = Type::Int32Ty; break;
5670 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
5671 InsertNewInstBefore(NewTrunc, I);
5672 return new SExtInst(NewTrunc, Ty);
5674 // Otherwise, we can't handle it yet.
5675 } else if (ShiftAmt1 < ShiftAmt2) {
5676 unsigned ShiftDiff = ShiftAmt2-ShiftAmt1;
5678 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
5679 if (I.getOpcode() == Instruction::Shl) {
5680 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5681 ShiftOp->getOpcode() == Instruction::AShr);
5682 Instruction *Shift =
5683 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5684 InsertNewInstBefore(Shift, I);
5686 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
5687 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5690 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
5691 if (I.getOpcode() == Instruction::LShr) {
5692 assert(ShiftOp->getOpcode() == Instruction::Shl);
5693 Instruction *Shift =
5694 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
5695 InsertNewInstBefore(Shift, I);
5697 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5698 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5701 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
5703 assert(ShiftAmt2 < ShiftAmt1);
5704 unsigned ShiftDiff = ShiftAmt1-ShiftAmt2;
5706 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
5707 if (I.getOpcode() == Instruction::Shl) {
5708 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5709 ShiftOp->getOpcode() == Instruction::AShr);
5710 Instruction *Shift =
5711 BinaryOperator::create(ShiftOp->getOpcode(), X,
5712 ConstantInt::get(Ty, ShiftDiff));
5713 InsertNewInstBefore(Shift, I);
5715 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
5716 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5719 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
5720 if (I.getOpcode() == Instruction::LShr) {
5721 assert(ShiftOp->getOpcode() == Instruction::Shl);
5722 Instruction *Shift =
5723 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5724 InsertNewInstBefore(Shift, I);
5726 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5727 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5730 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
5737 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5738 /// expression. If so, decompose it, returning some value X, such that Val is
5741 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5743 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5744 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5745 Offset = CI->getZExtValue();
5747 return ConstantInt::get(Type::Int32Ty, 0);
5748 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5749 if (I->getNumOperands() == 2) {
5750 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5751 if (I->getOpcode() == Instruction::Shl) {
5752 // This is a value scaled by '1 << the shift amt'.
5753 Scale = 1U << CUI->getZExtValue();
5755 return I->getOperand(0);
5756 } else if (I->getOpcode() == Instruction::Mul) {
5757 // This value is scaled by 'CUI'.
5758 Scale = CUI->getZExtValue();
5760 return I->getOperand(0);
5761 } else if (I->getOpcode() == Instruction::Add) {
5762 // We have X+C. Check to see if we really have (X*C2)+C1,
5763 // where C1 is divisible by C2.
5766 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5767 Offset += CUI->getZExtValue();
5768 if (SubScale > 1 && (Offset % SubScale == 0)) {
5777 // Otherwise, we can't look past this.
5784 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5785 /// try to eliminate the cast by moving the type information into the alloc.
5786 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5787 AllocationInst &AI) {
5788 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5789 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5791 // Remove any uses of AI that are dead.
5792 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5793 std::vector<Instruction*> DeadUsers;
5794 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5795 Instruction *User = cast<Instruction>(*UI++);
5796 if (isInstructionTriviallyDead(User)) {
5797 while (UI != E && *UI == User)
5798 ++UI; // If this instruction uses AI more than once, don't break UI.
5800 // Add operands to the worklist.
5801 AddUsesToWorkList(*User);
5803 DOUT << "IC: DCE: " << *User;
5805 User->eraseFromParent();
5806 removeFromWorkList(User);
5810 // Get the type really allocated and the type casted to.
5811 const Type *AllocElTy = AI.getAllocatedType();
5812 const Type *CastElTy = PTy->getElementType();
5813 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5815 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
5816 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
5817 if (CastElTyAlign < AllocElTyAlign) return 0;
5819 // If the allocation has multiple uses, only promote it if we are strictly
5820 // increasing the alignment of the resultant allocation. If we keep it the
5821 // same, we open the door to infinite loops of various kinds.
5822 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5824 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5825 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5826 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5828 // See if we can satisfy the modulus by pulling a scale out of the array
5830 unsigned ArraySizeScale, ArrayOffset;
5831 Value *NumElements = // See if the array size is a decomposable linear expr.
5832 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5834 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5836 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5837 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5839 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5844 // If the allocation size is constant, form a constant mul expression
5845 Amt = ConstantInt::get(Type::Int32Ty, Scale);
5846 if (isa<ConstantInt>(NumElements))
5847 Amt = ConstantExpr::getMul(
5848 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5849 // otherwise multiply the amount and the number of elements
5850 else if (Scale != 1) {
5851 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5852 Amt = InsertNewInstBefore(Tmp, AI);
5856 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5857 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
5858 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5859 Amt = InsertNewInstBefore(Tmp, AI);
5862 AllocationInst *New;
5863 if (isa<MallocInst>(AI))
5864 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
5866 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
5867 InsertNewInstBefore(New, AI);
5870 // If the allocation has multiple uses, insert a cast and change all things
5871 // that used it to use the new cast. This will also hack on CI, but it will
5873 if (!AI.hasOneUse()) {
5874 AddUsesToWorkList(AI);
5875 // New is the allocation instruction, pointer typed. AI is the original
5876 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5877 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5878 InsertNewInstBefore(NewCast, AI);
5879 AI.replaceAllUsesWith(NewCast);
5881 return ReplaceInstUsesWith(CI, New);
5884 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5885 /// and return it without inserting any new casts. This is used by code that
5886 /// tries to decide whether promoting or shrinking integer operations to wider
5887 /// or smaller types will allow us to eliminate a truncate or extend.
5888 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5889 int &NumCastsRemoved) {
5890 if (isa<Constant>(V)) return true;
5892 Instruction *I = dyn_cast<Instruction>(V);
5893 if (!I || !I->hasOneUse()) return false;
5895 switch (I->getOpcode()) {
5896 case Instruction::And:
5897 case Instruction::Or:
5898 case Instruction::Xor:
5899 // These operators can all arbitrarily be extended or truncated.
5900 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5901 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5902 case Instruction::AShr:
5903 case Instruction::LShr:
5904 case Instruction::Shl:
5905 // If this is just a bitcast changing the sign of the operation, we can
5906 // convert if the operand can be converted.
5907 if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
5908 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5910 case Instruction::Trunc:
5911 case Instruction::ZExt:
5912 case Instruction::SExt:
5913 case Instruction::BitCast:
5914 // If this is a cast from the destination type, we can trivially eliminate
5915 // it, and this will remove a cast overall.
5916 if (I->getOperand(0)->getType() == Ty) {
5917 // If the first operand is itself a cast, and is eliminable, do not count
5918 // this as an eliminable cast. We would prefer to eliminate those two
5920 if (isa<CastInst>(I->getOperand(0)))
5928 // TODO: Can handle more cases here.
5935 /// EvaluateInDifferentType - Given an expression that
5936 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5937 /// evaluate the expression.
5938 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
5940 if (Constant *C = dyn_cast<Constant>(V))
5941 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
5943 // Otherwise, it must be an instruction.
5944 Instruction *I = cast<Instruction>(V);
5945 Instruction *Res = 0;
5946 switch (I->getOpcode()) {
5947 case Instruction::And:
5948 case Instruction::Or:
5949 case Instruction::Xor: {
5950 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5951 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
5952 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5953 LHS, RHS, I->getName());
5956 case Instruction::AShr:
5957 case Instruction::LShr:
5958 case Instruction::Shl: {
5959 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5960 Res = BinaryOperator::create(Instruction::BinaryOps(I->getOpcode()), LHS,
5961 I->getOperand(1), I->getName());
5964 case Instruction::Trunc:
5965 case Instruction::ZExt:
5966 case Instruction::SExt:
5967 case Instruction::BitCast:
5968 // If the source type of the cast is the type we're trying for then we can
5969 // just return the source. There's no need to insert it because its not new.
5970 if (I->getOperand(0)->getType() == Ty)
5971 return I->getOperand(0);
5973 // Some other kind of cast, which shouldn't happen, so just ..
5976 // TODO: Can handle more cases here.
5977 assert(0 && "Unreachable!");
5981 return InsertNewInstBefore(Res, *I);
5984 /// @brief Implement the transforms common to all CastInst visitors.
5985 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
5986 Value *Src = CI.getOperand(0);
5988 // Casting undef to anything results in undef so might as just replace it and
5989 // get rid of the cast.
5990 if (isa<UndefValue>(Src)) // cast undef -> undef
5991 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5993 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
5994 // eliminate it now.
5995 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5996 if (Instruction::CastOps opc =
5997 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
5998 // The first cast (CSrc) is eliminable so we need to fix up or replace
5999 // the second cast (CI). CSrc will then have a good chance of being dead.
6000 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6004 // If casting the result of a getelementptr instruction with no offset, turn
6005 // this into a cast of the original pointer!
6007 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6008 bool AllZeroOperands = true;
6009 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
6010 if (!isa<Constant>(GEP->getOperand(i)) ||
6011 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
6012 AllZeroOperands = false;
6015 if (AllZeroOperands) {
6016 // Changing the cast operand is usually not a good idea but it is safe
6017 // here because the pointer operand is being replaced with another
6018 // pointer operand so the opcode doesn't need to change.
6019 CI.setOperand(0, GEP->getOperand(0));
6024 // If we are casting a malloc or alloca to a pointer to a type of the same
6025 // size, rewrite the allocation instruction to allocate the "right" type.
6026 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6027 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6030 // If we are casting a select then fold the cast into the select
6031 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6032 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6035 // If we are casting a PHI then fold the cast into the PHI
6036 if (isa<PHINode>(Src))
6037 if (Instruction *NV = FoldOpIntoPhi(CI))
6043 /// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
6044 /// integers. This function implements the common transforms for all those
6046 /// @brief Implement the transforms common to CastInst with integer operands
6047 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6048 if (Instruction *Result = commonCastTransforms(CI))
6051 Value *Src = CI.getOperand(0);
6052 const Type *SrcTy = Src->getType();
6053 const Type *DestTy = CI.getType();
6054 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6055 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6057 // See if we can simplify any instructions used by the LHS whose sole
6058 // purpose is to compute bits we don't care about.
6059 uint64_t KnownZero = 0, KnownOne = 0;
6060 if (SimplifyDemandedBits(&CI, cast<IntegerType>(DestTy)->getBitMask(),
6061 KnownZero, KnownOne))
6064 // If the source isn't an instruction or has more than one use then we
6065 // can't do anything more.
6066 Instruction *SrcI = dyn_cast<Instruction>(Src);
6067 if (!SrcI || !Src->hasOneUse())
6070 // Attempt to propagate the cast into the instruction.
6071 int NumCastsRemoved = 0;
6072 if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
6073 // If this cast is a truncate, evaluting in a different type always
6074 // eliminates the cast, so it is always a win. If this is a noop-cast
6075 // this just removes a noop cast which isn't pointful, but simplifies
6076 // the code. If this is a zero-extension, we need to do an AND to
6077 // maintain the clear top-part of the computation, so we require that
6078 // the input have eliminated at least one cast. If this is a sign
6079 // extension, we insert two new casts (to do the extension) so we
6080 // require that two casts have been eliminated.
6081 bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
6083 switch (CI.getOpcode()) {
6084 case Instruction::Trunc:
6087 case Instruction::ZExt:
6088 DoXForm = NumCastsRemoved >= 1;
6090 case Instruction::SExt:
6091 DoXForm = NumCastsRemoved >= 2;
6093 case Instruction::BitCast:
6097 // All the others use floating point so we shouldn't actually
6098 // get here because of the check above.
6099 assert(!"Unknown cast type .. unreachable");
6105 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6106 CI.getOpcode() == Instruction::SExt);
6107 assert(Res->getType() == DestTy);
6108 switch (CI.getOpcode()) {
6109 default: assert(0 && "Unknown cast type!");
6110 case Instruction::Trunc:
6111 case Instruction::BitCast:
6112 // Just replace this cast with the result.
6113 return ReplaceInstUsesWith(CI, Res);
6114 case Instruction::ZExt: {
6115 // We need to emit an AND to clear the high bits.
6116 assert(SrcBitSize < DestBitSize && "Not a zext?");
6118 ConstantInt::get(Type::Int64Ty, (1ULL << SrcBitSize)-1);
6119 if (DestBitSize < 64)
6120 C = ConstantExpr::getTrunc(C, DestTy);
6121 return BinaryOperator::createAnd(Res, C);
6123 case Instruction::SExt:
6124 // We need to emit a cast to truncate, then a cast to sext.
6125 return CastInst::create(Instruction::SExt,
6126 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6132 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6133 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6135 switch (SrcI->getOpcode()) {
6136 case Instruction::Add:
6137 case Instruction::Mul:
6138 case Instruction::And:
6139 case Instruction::Or:
6140 case Instruction::Xor:
6141 // If we are discarding information, or just changing the sign,
6143 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6144 // Don't insert two casts if they cannot be eliminated. We allow
6145 // two casts to be inserted if the sizes are the same. This could
6146 // only be converting signedness, which is a noop.
6147 if (DestBitSize == SrcBitSize ||
6148 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6149 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6150 Instruction::CastOps opcode = CI.getOpcode();
6151 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6152 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6153 return BinaryOperator::create(
6154 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6158 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6159 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6160 SrcI->getOpcode() == Instruction::Xor &&
6161 Op1 == ConstantInt::getTrue() &&
6162 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6163 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6164 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6167 case Instruction::SDiv:
6168 case Instruction::UDiv:
6169 case Instruction::SRem:
6170 case Instruction::URem:
6171 // If we are just changing the sign, rewrite.
6172 if (DestBitSize == SrcBitSize) {
6173 // Don't insert two casts if they cannot be eliminated. We allow
6174 // two casts to be inserted if the sizes are the same. This could
6175 // only be converting signedness, which is a noop.
6176 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6177 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6178 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6180 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6182 return BinaryOperator::create(
6183 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6188 case Instruction::Shl:
6189 // Allow changing the sign of the source operand. Do not allow
6190 // changing the size of the shift, UNLESS the shift amount is a
6191 // constant. We must not change variable sized shifts to a smaller
6192 // size, because it is undefined to shift more bits out than exist
6194 if (DestBitSize == SrcBitSize ||
6195 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6196 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6197 Instruction::BitCast : Instruction::Trunc);
6198 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6199 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6200 return BinaryOperator::createShl(Op0c, Op1c);
6203 case Instruction::AShr:
6204 // If this is a signed shr, and if all bits shifted in are about to be
6205 // truncated off, turn it into an unsigned shr to allow greater
6207 if (DestBitSize < SrcBitSize &&
6208 isa<ConstantInt>(Op1)) {
6209 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6210 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6211 // Insert the new logical shift right.
6212 return BinaryOperator::createLShr(Op0, Op1);
6217 case Instruction::ICmp:
6218 // If we are just checking for a icmp eq of a single bit and casting it
6219 // to an integer, then shift the bit to the appropriate place and then
6220 // cast to integer to avoid the comparison.
6221 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6222 uint64_t Op1CV = Op1C->getZExtValue();
6223 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6224 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6225 // cast (X == 1) to int --> X iff X has only the low bit set.
6226 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6227 // cast (X != 0) to int --> X iff X has only the low bit set.
6228 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6229 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6230 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6231 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
6232 // If Op1C some other power of two, convert:
6233 uint64_t KnownZero, KnownOne;
6234 uint64_t TypeMask = Op1C->getType()->getBitMask();
6235 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6237 // This only works for EQ and NE
6238 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6239 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6242 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
6243 bool isNE = pred == ICmpInst::ICMP_NE;
6244 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
6245 // (X&4) == 2 --> false
6246 // (X&4) != 2 --> true
6247 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6248 Res = ConstantExpr::getZExt(Res, CI.getType());
6249 return ReplaceInstUsesWith(CI, Res);
6252 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
6255 // Perform a logical shr by shiftamt.
6256 // Insert the shift to put the result in the low bit.
6257 In = InsertNewInstBefore(
6258 BinaryOperator::createLShr(In,
6259 ConstantInt::get(In->getType(), ShiftAmt),
6260 In->getName()+".lobit"), CI);
6263 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6264 Constant *One = ConstantInt::get(In->getType(), 1);
6265 In = BinaryOperator::createXor(In, One, "tmp");
6266 InsertNewInstBefore(cast<Instruction>(In), CI);
6269 if (CI.getType() == In->getType())
6270 return ReplaceInstUsesWith(CI, In);
6272 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6281 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6282 if (Instruction *Result = commonIntCastTransforms(CI))
6285 Value *Src = CI.getOperand(0);
6286 const Type *Ty = CI.getType();
6287 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6289 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6290 switch (SrcI->getOpcode()) {
6292 case Instruction::LShr:
6293 // We can shrink lshr to something smaller if we know the bits shifted in
6294 // are already zeros.
6295 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6296 unsigned ShAmt = ShAmtV->getZExtValue();
6298 // Get a mask for the bits shifting in.
6299 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
6300 Value* SrcIOp0 = SrcI->getOperand(0);
6301 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6302 if (ShAmt >= DestBitWidth) // All zeros.
6303 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6305 // Okay, we can shrink this. Truncate the input, then return a new
6307 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6308 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6310 return BinaryOperator::createLShr(V1, V2);
6312 } else { // This is a variable shr.
6314 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6315 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6316 // loop-invariant and CSE'd.
6317 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6318 Value *One = ConstantInt::get(SrcI->getType(), 1);
6320 Value *V = InsertNewInstBefore(
6321 BinaryOperator::createShl(One, SrcI->getOperand(1),
6323 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6324 SrcI->getOperand(0),
6326 Value *Zero = Constant::getNullValue(V->getType());
6327 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6337 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6338 // If one of the common conversion will work ..
6339 if (Instruction *Result = commonIntCastTransforms(CI))
6342 Value *Src = CI.getOperand(0);
6344 // If this is a cast of a cast
6345 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6346 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6347 // types and if the sizes are just right we can convert this into a logical
6348 // 'and' which will be much cheaper than the pair of casts.
6349 if (isa<TruncInst>(CSrc)) {
6350 // Get the sizes of the types involved
6351 Value *A = CSrc->getOperand(0);
6352 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6353 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6354 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6355 // If we're actually extending zero bits and the trunc is a no-op
6356 if (MidSize < DstSize && SrcSize == DstSize) {
6357 // Replace both of the casts with an And of the type mask.
6358 uint64_t AndValue = cast<IntegerType>(CSrc->getType())->getBitMask();
6359 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6361 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6362 // Unfortunately, if the type changed, we need to cast it back.
6363 if (And->getType() != CI.getType()) {
6364 And->setName(CSrc->getName()+".mask");
6365 InsertNewInstBefore(And, CI);
6366 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6376 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6377 return commonIntCastTransforms(CI);
6380 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6381 return commonCastTransforms(CI);
6384 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6385 return commonCastTransforms(CI);
6388 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6389 return commonCastTransforms(CI);
6392 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6393 return commonCastTransforms(CI);
6396 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6397 return commonCastTransforms(CI);
6400 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6401 return commonCastTransforms(CI);
6404 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6405 return commonCastTransforms(CI);
6408 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6409 return commonCastTransforms(CI);
6412 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6414 // If the operands are integer typed then apply the integer transforms,
6415 // otherwise just apply the common ones.
6416 Value *Src = CI.getOperand(0);
6417 const Type *SrcTy = Src->getType();
6418 const Type *DestTy = CI.getType();
6420 if (SrcTy->isInteger() && DestTy->isInteger()) {
6421 if (Instruction *Result = commonIntCastTransforms(CI))
6424 if (Instruction *Result = commonCastTransforms(CI))
6429 // Get rid of casts from one type to the same type. These are useless and can
6430 // be replaced by the operand.
6431 if (DestTy == Src->getType())
6432 return ReplaceInstUsesWith(CI, Src);
6434 // If the source and destination are pointers, and this cast is equivalent to
6435 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6436 // This can enhance SROA and other transforms that want type-safe pointers.
6437 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6438 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6439 const Type *DstElTy = DstPTy->getElementType();
6440 const Type *SrcElTy = SrcPTy->getElementType();
6442 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6443 unsigned NumZeros = 0;
6444 while (SrcElTy != DstElTy &&
6445 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6446 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6447 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6451 // If we found a path from the src to dest, create the getelementptr now.
6452 if (SrcElTy == DstElTy) {
6453 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
6454 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
6459 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6460 if (SVI->hasOneUse()) {
6461 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6462 // a bitconvert to a vector with the same # elts.
6463 if (isa<VectorType>(DestTy) &&
6464 cast<VectorType>(DestTy)->getNumElements() ==
6465 SVI->getType()->getNumElements()) {
6467 // If either of the operands is a cast from CI.getType(), then
6468 // evaluating the shuffle in the casted destination's type will allow
6469 // us to eliminate at least one cast.
6470 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6471 Tmp->getOperand(0)->getType() == DestTy) ||
6472 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6473 Tmp->getOperand(0)->getType() == DestTy)) {
6474 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6475 SVI->getOperand(0), DestTy, &CI);
6476 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6477 SVI->getOperand(1), DestTy, &CI);
6478 // Return a new shuffle vector. Use the same element ID's, as we
6479 // know the vector types match #elts.
6480 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6488 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6490 /// %D = select %cond, %C, %A
6492 /// %C = select %cond, %B, 0
6495 /// Assuming that the specified instruction is an operand to the select, return
6496 /// a bitmask indicating which operands of this instruction are foldable if they
6497 /// equal the other incoming value of the select.
6499 static unsigned GetSelectFoldableOperands(Instruction *I) {
6500 switch (I->getOpcode()) {
6501 case Instruction::Add:
6502 case Instruction::Mul:
6503 case Instruction::And:
6504 case Instruction::Or:
6505 case Instruction::Xor:
6506 return 3; // Can fold through either operand.
6507 case Instruction::Sub: // Can only fold on the amount subtracted.
6508 case Instruction::Shl: // Can only fold on the shift amount.
6509 case Instruction::LShr:
6510 case Instruction::AShr:
6513 return 0; // Cannot fold
6517 /// GetSelectFoldableConstant - For the same transformation as the previous
6518 /// function, return the identity constant that goes into the select.
6519 static Constant *GetSelectFoldableConstant(Instruction *I) {
6520 switch (I->getOpcode()) {
6521 default: assert(0 && "This cannot happen!"); abort();
6522 case Instruction::Add:
6523 case Instruction::Sub:
6524 case Instruction::Or:
6525 case Instruction::Xor:
6526 case Instruction::Shl:
6527 case Instruction::LShr:
6528 case Instruction::AShr:
6529 return Constant::getNullValue(I->getType());
6530 case Instruction::And:
6531 return ConstantInt::getAllOnesValue(I->getType());
6532 case Instruction::Mul:
6533 return ConstantInt::get(I->getType(), 1);
6537 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6538 /// have the same opcode and only one use each. Try to simplify this.
6539 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6541 if (TI->getNumOperands() == 1) {
6542 // If this is a non-volatile load or a cast from the same type,
6545 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6548 return 0; // unknown unary op.
6551 // Fold this by inserting a select from the input values.
6552 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6553 FI->getOperand(0), SI.getName()+".v");
6554 InsertNewInstBefore(NewSI, SI);
6555 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6559 // Only handle binary operators here.
6560 if (!isa<BinaryOperator>(TI))
6563 // Figure out if the operations have any operands in common.
6564 Value *MatchOp, *OtherOpT, *OtherOpF;
6566 if (TI->getOperand(0) == FI->getOperand(0)) {
6567 MatchOp = TI->getOperand(0);
6568 OtherOpT = TI->getOperand(1);
6569 OtherOpF = FI->getOperand(1);
6570 MatchIsOpZero = true;
6571 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6572 MatchOp = TI->getOperand(1);
6573 OtherOpT = TI->getOperand(0);
6574 OtherOpF = FI->getOperand(0);
6575 MatchIsOpZero = false;
6576 } else if (!TI->isCommutative()) {
6578 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6579 MatchOp = TI->getOperand(0);
6580 OtherOpT = TI->getOperand(1);
6581 OtherOpF = FI->getOperand(0);
6582 MatchIsOpZero = true;
6583 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6584 MatchOp = TI->getOperand(1);
6585 OtherOpT = TI->getOperand(0);
6586 OtherOpF = FI->getOperand(1);
6587 MatchIsOpZero = true;
6592 // If we reach here, they do have operations in common.
6593 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6594 OtherOpF, SI.getName()+".v");
6595 InsertNewInstBefore(NewSI, SI);
6597 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6599 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6601 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6603 assert(0 && "Shouldn't get here");
6607 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6608 Value *CondVal = SI.getCondition();
6609 Value *TrueVal = SI.getTrueValue();
6610 Value *FalseVal = SI.getFalseValue();
6612 // select true, X, Y -> X
6613 // select false, X, Y -> Y
6614 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6615 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
6617 // select C, X, X -> X
6618 if (TrueVal == FalseVal)
6619 return ReplaceInstUsesWith(SI, TrueVal);
6621 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6622 return ReplaceInstUsesWith(SI, FalseVal);
6623 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6624 return ReplaceInstUsesWith(SI, TrueVal);
6625 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6626 if (isa<Constant>(TrueVal))
6627 return ReplaceInstUsesWith(SI, TrueVal);
6629 return ReplaceInstUsesWith(SI, FalseVal);
6632 if (SI.getType() == Type::Int1Ty) {
6633 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
6634 if (C->getZExtValue()) {
6635 // Change: A = select B, true, C --> A = or B, C
6636 return BinaryOperator::createOr(CondVal, FalseVal);
6638 // Change: A = select B, false, C --> A = and !B, C
6640 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6641 "not."+CondVal->getName()), SI);
6642 return BinaryOperator::createAnd(NotCond, FalseVal);
6644 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
6645 if (C->getZExtValue() == false) {
6646 // Change: A = select B, C, false --> A = and B, C
6647 return BinaryOperator::createAnd(CondVal, TrueVal);
6649 // Change: A = select B, C, true --> A = or !B, C
6651 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6652 "not."+CondVal->getName()), SI);
6653 return BinaryOperator::createOr(NotCond, TrueVal);
6658 // Selecting between two integer constants?
6659 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6660 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6661 // select C, 1, 0 -> cast C to int
6662 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6663 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6664 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6665 // select C, 0, 1 -> cast !C to int
6667 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6668 "not."+CondVal->getName()), SI);
6669 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6672 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6674 // (x <s 0) ? -1 : 0 -> ashr x, 31
6675 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6676 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6677 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6678 bool CanXForm = false;
6679 if (IC->isSignedPredicate())
6680 CanXForm = CmpCst->isNullValue() &&
6681 IC->getPredicate() == ICmpInst::ICMP_SLT;
6683 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6684 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6685 IC->getPredicate() == ICmpInst::ICMP_UGT;
6689 // The comparison constant and the result are not neccessarily the
6690 // same width. Make an all-ones value by inserting a AShr.
6691 Value *X = IC->getOperand(0);
6692 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6693 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
6694 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
6696 InsertNewInstBefore(SRA, SI);
6698 // Finally, convert to the type of the select RHS. We figure out
6699 // if this requires a SExt, Trunc or BitCast based on the sizes.
6700 Instruction::CastOps opc = Instruction::BitCast;
6701 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6702 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6703 if (SRASize < SISize)
6704 opc = Instruction::SExt;
6705 else if (SRASize > SISize)
6706 opc = Instruction::Trunc;
6707 return CastInst::create(opc, SRA, SI.getType());
6712 // If one of the constants is zero (we know they can't both be) and we
6713 // have a fcmp instruction with zero, and we have an 'and' with the
6714 // non-constant value, eliminate this whole mess. This corresponds to
6715 // cases like this: ((X & 27) ? 27 : 0)
6716 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6717 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6718 cast<Constant>(IC->getOperand(1))->isNullValue())
6719 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6720 if (ICA->getOpcode() == Instruction::And &&
6721 isa<ConstantInt>(ICA->getOperand(1)) &&
6722 (ICA->getOperand(1) == TrueValC ||
6723 ICA->getOperand(1) == FalseValC) &&
6724 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6725 // Okay, now we know that everything is set up, we just don't
6726 // know whether we have a icmp_ne or icmp_eq and whether the
6727 // true or false val is the zero.
6728 bool ShouldNotVal = !TrueValC->isNullValue();
6729 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6732 V = InsertNewInstBefore(BinaryOperator::create(
6733 Instruction::Xor, V, ICA->getOperand(1)), SI);
6734 return ReplaceInstUsesWith(SI, V);
6739 // See if we are selecting two values based on a comparison of the two values.
6740 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6741 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6742 // Transform (X == Y) ? X : Y -> Y
6743 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6744 return ReplaceInstUsesWith(SI, FalseVal);
6745 // Transform (X != Y) ? X : Y -> X
6746 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6747 return ReplaceInstUsesWith(SI, TrueVal);
6748 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6750 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6751 // Transform (X == Y) ? Y : X -> X
6752 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6753 return ReplaceInstUsesWith(SI, FalseVal);
6754 // Transform (X != Y) ? Y : X -> Y
6755 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6756 return ReplaceInstUsesWith(SI, TrueVal);
6757 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6761 // See if we are selecting two values based on a comparison of the two values.
6762 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6763 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6764 // Transform (X == Y) ? X : Y -> Y
6765 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6766 return ReplaceInstUsesWith(SI, FalseVal);
6767 // Transform (X != Y) ? X : Y -> X
6768 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6769 return ReplaceInstUsesWith(SI, TrueVal);
6770 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6772 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6773 // Transform (X == Y) ? Y : X -> X
6774 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6775 return ReplaceInstUsesWith(SI, FalseVal);
6776 // Transform (X != Y) ? Y : X -> Y
6777 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6778 return ReplaceInstUsesWith(SI, TrueVal);
6779 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6783 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6784 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6785 if (TI->hasOneUse() && FI->hasOneUse()) {
6786 Instruction *AddOp = 0, *SubOp = 0;
6788 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6789 if (TI->getOpcode() == FI->getOpcode())
6790 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6793 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6794 // even legal for FP.
6795 if (TI->getOpcode() == Instruction::Sub &&
6796 FI->getOpcode() == Instruction::Add) {
6797 AddOp = FI; SubOp = TI;
6798 } else if (FI->getOpcode() == Instruction::Sub &&
6799 TI->getOpcode() == Instruction::Add) {
6800 AddOp = TI; SubOp = FI;
6804 Value *OtherAddOp = 0;
6805 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6806 OtherAddOp = AddOp->getOperand(1);
6807 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6808 OtherAddOp = AddOp->getOperand(0);
6812 // So at this point we know we have (Y -> OtherAddOp):
6813 // select C, (add X, Y), (sub X, Z)
6814 Value *NegVal; // Compute -Z
6815 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6816 NegVal = ConstantExpr::getNeg(C);
6818 NegVal = InsertNewInstBefore(
6819 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6822 Value *NewTrueOp = OtherAddOp;
6823 Value *NewFalseOp = NegVal;
6825 std::swap(NewTrueOp, NewFalseOp);
6826 Instruction *NewSel =
6827 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6829 NewSel = InsertNewInstBefore(NewSel, SI);
6830 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6835 // See if we can fold the select into one of our operands.
6836 if (SI.getType()->isInteger()) {
6837 // See the comment above GetSelectFoldableOperands for a description of the
6838 // transformation we are doing here.
6839 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6840 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6841 !isa<Constant>(FalseVal))
6842 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6843 unsigned OpToFold = 0;
6844 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6846 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6851 Constant *C = GetSelectFoldableConstant(TVI);
6852 Instruction *NewSel =
6853 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
6854 InsertNewInstBefore(NewSel, SI);
6855 NewSel->takeName(TVI);
6856 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6857 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6859 assert(0 && "Unknown instruction!!");
6864 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6865 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6866 !isa<Constant>(TrueVal))
6867 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6868 unsigned OpToFold = 0;
6869 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6871 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6876 Constant *C = GetSelectFoldableConstant(FVI);
6877 Instruction *NewSel =
6878 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
6879 InsertNewInstBefore(NewSel, SI);
6880 NewSel->takeName(FVI);
6881 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6882 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6884 assert(0 && "Unknown instruction!!");
6889 if (BinaryOperator::isNot(CondVal)) {
6890 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6891 SI.setOperand(1, FalseVal);
6892 SI.setOperand(2, TrueVal);
6899 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6900 /// determine, return it, otherwise return 0.
6901 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6902 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6903 unsigned Align = GV->getAlignment();
6904 if (Align == 0 && TD)
6905 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
6907 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6908 unsigned Align = AI->getAlignment();
6909 if (Align == 0 && TD) {
6910 if (isa<AllocaInst>(AI))
6911 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
6912 else if (isa<MallocInst>(AI)) {
6913 // Malloc returns maximally aligned memory.
6914 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
6917 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
6920 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
6924 } else if (isa<BitCastInst>(V) ||
6925 (isa<ConstantExpr>(V) &&
6926 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
6927 User *CI = cast<User>(V);
6928 if (isa<PointerType>(CI->getOperand(0)->getType()))
6929 return GetKnownAlignment(CI->getOperand(0), TD);
6931 } else if (isa<GetElementPtrInst>(V) ||
6932 (isa<ConstantExpr>(V) &&
6933 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
6934 User *GEPI = cast<User>(V);
6935 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
6936 if (BaseAlignment == 0) return 0;
6938 // If all indexes are zero, it is just the alignment of the base pointer.
6939 bool AllZeroOperands = true;
6940 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
6941 if (!isa<Constant>(GEPI->getOperand(i)) ||
6942 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
6943 AllZeroOperands = false;
6946 if (AllZeroOperands)
6947 return BaseAlignment;
6949 // Otherwise, if the base alignment is >= the alignment we expect for the
6950 // base pointer type, then we know that the resultant pointer is aligned at
6951 // least as much as its type requires.
6954 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
6955 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
6956 if (TD->getABITypeAlignment(PtrTy->getElementType())
6958 const Type *GEPTy = GEPI->getType();
6959 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
6960 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
6968 /// visitCallInst - CallInst simplification. This mostly only handles folding
6969 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
6970 /// the heavy lifting.
6972 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
6973 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
6974 if (!II) return visitCallSite(&CI);
6976 // Intrinsics cannot occur in an invoke, so handle them here instead of in
6978 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
6979 bool Changed = false;
6981 // memmove/cpy/set of zero bytes is a noop.
6982 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
6983 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
6985 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
6986 if (CI->getZExtValue() == 1) {
6987 // Replace the instruction with just byte operations. We would
6988 // transform other cases to loads/stores, but we don't know if
6989 // alignment is sufficient.
6993 // If we have a memmove and the source operation is a constant global,
6994 // then the source and dest pointers can't alias, so we can change this
6995 // into a call to memcpy.
6996 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
6997 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
6998 if (GVSrc->isConstant()) {
6999 Module *M = CI.getParent()->getParent()->getParent();
7001 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7003 Name = "llvm.memcpy.i32";
7005 Name = "llvm.memcpy.i64";
7006 Constant *MemCpy = M->getOrInsertFunction(Name,
7007 CI.getCalledFunction()->getFunctionType());
7008 CI.setOperand(0, MemCpy);
7013 // If we can determine a pointer alignment that is bigger than currently
7014 // set, update the alignment.
7015 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7016 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7017 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7018 unsigned Align = std::min(Alignment1, Alignment2);
7019 if (MI->getAlignment()->getZExtValue() < Align) {
7020 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7023 } else if (isa<MemSetInst>(MI)) {
7024 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7025 if (MI->getAlignment()->getZExtValue() < Alignment) {
7026 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7031 if (Changed) return II;
7033 switch (II->getIntrinsicID()) {
7035 case Intrinsic::ppc_altivec_lvx:
7036 case Intrinsic::ppc_altivec_lvxl:
7037 case Intrinsic::x86_sse_loadu_ps:
7038 case Intrinsic::x86_sse2_loadu_pd:
7039 case Intrinsic::x86_sse2_loadu_dq:
7040 // Turn PPC lvx -> load if the pointer is known aligned.
7041 // Turn X86 loadups -> load if the pointer is known aligned.
7042 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7043 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7044 PointerType::get(II->getType()), CI);
7045 return new LoadInst(Ptr);
7048 case Intrinsic::ppc_altivec_stvx:
7049 case Intrinsic::ppc_altivec_stvxl:
7050 // Turn stvx -> store if the pointer is known aligned.
7051 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7052 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7053 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7055 return new StoreInst(II->getOperand(1), Ptr);
7058 case Intrinsic::x86_sse_storeu_ps:
7059 case Intrinsic::x86_sse2_storeu_pd:
7060 case Intrinsic::x86_sse2_storeu_dq:
7061 case Intrinsic::x86_sse2_storel_dq:
7062 // Turn X86 storeu -> store if the pointer is known aligned.
7063 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7064 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7065 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7067 return new StoreInst(II->getOperand(2), Ptr);
7071 case Intrinsic::x86_sse_cvttss2si: {
7072 // These intrinsics only demands the 0th element of its input vector. If
7073 // we can simplify the input based on that, do so now.
7075 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7077 II->setOperand(1, V);
7083 case Intrinsic::ppc_altivec_vperm:
7084 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7085 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7086 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7088 // Check that all of the elements are integer constants or undefs.
7089 bool AllEltsOk = true;
7090 for (unsigned i = 0; i != 16; ++i) {
7091 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7092 !isa<UndefValue>(Mask->getOperand(i))) {
7099 // Cast the input vectors to byte vectors.
7100 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7101 II->getOperand(1), Mask->getType(), CI);
7102 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7103 II->getOperand(2), Mask->getType(), CI);
7104 Value *Result = UndefValue::get(Op0->getType());
7106 // Only extract each element once.
7107 Value *ExtractedElts[32];
7108 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7110 for (unsigned i = 0; i != 16; ++i) {
7111 if (isa<UndefValue>(Mask->getOperand(i)))
7113 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7114 Idx &= 31; // Match the hardware behavior.
7116 if (ExtractedElts[Idx] == 0) {
7118 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7119 InsertNewInstBefore(Elt, CI);
7120 ExtractedElts[Idx] = Elt;
7123 // Insert this value into the result vector.
7124 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7125 InsertNewInstBefore(cast<Instruction>(Result), CI);
7127 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7132 case Intrinsic::stackrestore: {
7133 // If the save is right next to the restore, remove the restore. This can
7134 // happen when variable allocas are DCE'd.
7135 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7136 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7137 BasicBlock::iterator BI = SS;
7139 return EraseInstFromFunction(CI);
7143 // If the stack restore is in a return/unwind block and if there are no
7144 // allocas or calls between the restore and the return, nuke the restore.
7145 TerminatorInst *TI = II->getParent()->getTerminator();
7146 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7147 BasicBlock::iterator BI = II;
7148 bool CannotRemove = false;
7149 for (++BI; &*BI != TI; ++BI) {
7150 if (isa<AllocaInst>(BI) ||
7151 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7152 CannotRemove = true;
7157 return EraseInstFromFunction(CI);
7164 return visitCallSite(II);
7167 // InvokeInst simplification
7169 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7170 return visitCallSite(&II);
7173 // visitCallSite - Improvements for call and invoke instructions.
7175 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7176 bool Changed = false;
7178 // If the callee is a constexpr cast of a function, attempt to move the cast
7179 // to the arguments of the call/invoke.
7180 if (transformConstExprCastCall(CS)) return 0;
7182 Value *Callee = CS.getCalledValue();
7184 if (Function *CalleeF = dyn_cast<Function>(Callee))
7185 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7186 Instruction *OldCall = CS.getInstruction();
7187 // If the call and callee calling conventions don't match, this call must
7188 // be unreachable, as the call is undefined.
7189 new StoreInst(ConstantInt::getTrue(),
7190 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7191 if (!OldCall->use_empty())
7192 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7193 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7194 return EraseInstFromFunction(*OldCall);
7198 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7199 // This instruction is not reachable, just remove it. We insert a store to
7200 // undef so that we know that this code is not reachable, despite the fact
7201 // that we can't modify the CFG here.
7202 new StoreInst(ConstantInt::getTrue(),
7203 UndefValue::get(PointerType::get(Type::Int1Ty)),
7204 CS.getInstruction());
7206 if (!CS.getInstruction()->use_empty())
7207 CS.getInstruction()->
7208 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7210 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7211 // Don't break the CFG, insert a dummy cond branch.
7212 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7213 ConstantInt::getTrue(), II);
7215 return EraseInstFromFunction(*CS.getInstruction());
7218 const PointerType *PTy = cast<PointerType>(Callee->getType());
7219 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7220 if (FTy->isVarArg()) {
7221 // See if we can optimize any arguments passed through the varargs area of
7223 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7224 E = CS.arg_end(); I != E; ++I)
7225 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7226 // If this cast does not effect the value passed through the varargs
7227 // area, we can eliminate the use of the cast.
7228 Value *Op = CI->getOperand(0);
7229 if (CI->isLosslessCast()) {
7236 return Changed ? CS.getInstruction() : 0;
7239 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7240 // attempt to move the cast to the arguments of the call/invoke.
7242 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7243 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7244 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7245 if (CE->getOpcode() != Instruction::BitCast ||
7246 !isa<Function>(CE->getOperand(0)))
7248 Function *Callee = cast<Function>(CE->getOperand(0));
7249 Instruction *Caller = CS.getInstruction();
7251 // Okay, this is a cast from a function to a different type. Unless doing so
7252 // would cause a type conversion of one of our arguments, change this call to
7253 // be a direct call with arguments casted to the appropriate types.
7255 const FunctionType *FT = Callee->getFunctionType();
7256 const Type *OldRetTy = Caller->getType();
7258 // Check to see if we are changing the return type...
7259 if (OldRetTy != FT->getReturnType()) {
7260 if (Callee->isDeclaration() && !Caller->use_empty() &&
7261 OldRetTy != FT->getReturnType() &&
7262 // Conversion is ok if changing from pointer to int of same size.
7263 !(isa<PointerType>(FT->getReturnType()) &&
7264 TD->getIntPtrType() == OldRetTy))
7265 return false; // Cannot transform this return value.
7267 // If the callsite is an invoke instruction, and the return value is used by
7268 // a PHI node in a successor, we cannot change the return type of the call
7269 // because there is no place to put the cast instruction (without breaking
7270 // the critical edge). Bail out in this case.
7271 if (!Caller->use_empty())
7272 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7273 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7275 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7276 if (PN->getParent() == II->getNormalDest() ||
7277 PN->getParent() == II->getUnwindDest())
7281 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7282 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7284 CallSite::arg_iterator AI = CS.arg_begin();
7285 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7286 const Type *ParamTy = FT->getParamType(i);
7287 const Type *ActTy = (*AI)->getType();
7288 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7289 //Either we can cast directly, or we can upconvert the argument
7290 bool isConvertible = ActTy == ParamTy ||
7291 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7292 (ParamTy->isInteger() && ActTy->isInteger() &&
7293 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7294 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7295 && c->getSExtValue() > 0);
7296 if (Callee->isDeclaration() && !isConvertible) return false;
7299 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7300 Callee->isDeclaration())
7301 return false; // Do not delete arguments unless we have a function body...
7303 // Okay, we decided that this is a safe thing to do: go ahead and start
7304 // inserting cast instructions as necessary...
7305 std::vector<Value*> Args;
7306 Args.reserve(NumActualArgs);
7308 AI = CS.arg_begin();
7309 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7310 const Type *ParamTy = FT->getParamType(i);
7311 if ((*AI)->getType() == ParamTy) {
7312 Args.push_back(*AI);
7314 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7315 false, ParamTy, false);
7316 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7317 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7321 // If the function takes more arguments than the call was taking, add them
7323 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7324 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7326 // If we are removing arguments to the function, emit an obnoxious warning...
7327 if (FT->getNumParams() < NumActualArgs)
7328 if (!FT->isVarArg()) {
7329 cerr << "WARNING: While resolving call to function '"
7330 << Callee->getName() << "' arguments were dropped!\n";
7332 // Add all of the arguments in their promoted form to the arg list...
7333 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7334 const Type *PTy = getPromotedType((*AI)->getType());
7335 if (PTy != (*AI)->getType()) {
7336 // Must promote to pass through va_arg area!
7337 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7339 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7340 InsertNewInstBefore(Cast, *Caller);
7341 Args.push_back(Cast);
7343 Args.push_back(*AI);
7348 if (FT->getReturnType() == Type::VoidTy)
7349 Caller->setName(""); // Void type should not have a name.
7352 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7353 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7354 &Args[0], Args.size(), Caller->getName(), Caller);
7355 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7357 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
7358 if (cast<CallInst>(Caller)->isTailCall())
7359 cast<CallInst>(NC)->setTailCall();
7360 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7363 // Insert a cast of the return type as necessary.
7365 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7366 if (NV->getType() != Type::VoidTy) {
7367 const Type *CallerTy = Caller->getType();
7368 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7370 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7372 // If this is an invoke instruction, we should insert it after the first
7373 // non-phi, instruction in the normal successor block.
7374 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7375 BasicBlock::iterator I = II->getNormalDest()->begin();
7376 while (isa<PHINode>(I)) ++I;
7377 InsertNewInstBefore(NC, *I);
7379 // Otherwise, it's a call, just insert cast right after the call instr
7380 InsertNewInstBefore(NC, *Caller);
7382 AddUsersToWorkList(*Caller);
7384 NV = UndefValue::get(Caller->getType());
7388 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7389 Caller->replaceAllUsesWith(NV);
7390 Caller->getParent()->getInstList().erase(Caller);
7391 removeFromWorkList(Caller);
7395 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7396 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7397 /// and a single binop.
7398 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7399 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7400 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
7401 isa<CmpInst>(FirstInst));
7402 unsigned Opc = FirstInst->getOpcode();
7403 Value *LHSVal = FirstInst->getOperand(0);
7404 Value *RHSVal = FirstInst->getOperand(1);
7406 const Type *LHSType = LHSVal->getType();
7407 const Type *RHSType = RHSVal->getType();
7409 // Scan to see if all operands are the same opcode, all have one use, and all
7410 // kill their operands (i.e. the operands have one use).
7411 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7412 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7413 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7414 // Verify type of the LHS matches so we don't fold cmp's of different
7415 // types or GEP's with different index types.
7416 I->getOperand(0)->getType() != LHSType ||
7417 I->getOperand(1)->getType() != RHSType)
7420 // If they are CmpInst instructions, check their predicates
7421 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7422 if (cast<CmpInst>(I)->getPredicate() !=
7423 cast<CmpInst>(FirstInst)->getPredicate())
7426 // Keep track of which operand needs a phi node.
7427 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7428 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7431 // Otherwise, this is safe to transform, determine if it is profitable.
7433 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7434 // Indexes are often folded into load/store instructions, so we don't want to
7435 // hide them behind a phi.
7436 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7439 Value *InLHS = FirstInst->getOperand(0);
7440 Value *InRHS = FirstInst->getOperand(1);
7441 PHINode *NewLHS = 0, *NewRHS = 0;
7443 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7444 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7445 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7446 InsertNewInstBefore(NewLHS, PN);
7451 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7452 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7453 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7454 InsertNewInstBefore(NewRHS, PN);
7458 // Add all operands to the new PHIs.
7459 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7461 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7462 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7465 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7466 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7470 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7471 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7472 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7473 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7476 assert(isa<GetElementPtrInst>(FirstInst));
7477 return new GetElementPtrInst(LHSVal, RHSVal);
7481 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7482 /// of the block that defines it. This means that it must be obvious the value
7483 /// of the load is not changed from the point of the load to the end of the
7486 /// Finally, it is safe, but not profitable, to sink a load targetting a
7487 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
7489 static bool isSafeToSinkLoad(LoadInst *L) {
7490 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7492 for (++BBI; BBI != E; ++BBI)
7493 if (BBI->mayWriteToMemory())
7496 // Check for non-address taken alloca. If not address-taken already, it isn't
7497 // profitable to do this xform.
7498 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
7499 bool isAddressTaken = false;
7500 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
7502 if (isa<LoadInst>(UI)) continue;
7503 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
7504 // If storing TO the alloca, then the address isn't taken.
7505 if (SI->getOperand(1) == AI) continue;
7507 isAddressTaken = true;
7511 if (!isAddressTaken)
7519 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7520 // operator and they all are only used by the PHI, PHI together their
7521 // inputs, and do the operation once, to the result of the PHI.
7522 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7523 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7525 // Scan the instruction, looking for input operations that can be folded away.
7526 // If all input operands to the phi are the same instruction (e.g. a cast from
7527 // the same type or "+42") we can pull the operation through the PHI, reducing
7528 // code size and simplifying code.
7529 Constant *ConstantOp = 0;
7530 const Type *CastSrcTy = 0;
7531 bool isVolatile = false;
7532 if (isa<CastInst>(FirstInst)) {
7533 CastSrcTy = FirstInst->getOperand(0)->getType();
7534 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
7535 // Can fold binop, compare or shift here if the RHS is a constant,
7536 // otherwise call FoldPHIArgBinOpIntoPHI.
7537 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7538 if (ConstantOp == 0)
7539 return FoldPHIArgBinOpIntoPHI(PN);
7540 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7541 isVolatile = LI->isVolatile();
7542 // We can't sink the load if the loaded value could be modified between the
7543 // load and the PHI.
7544 if (LI->getParent() != PN.getIncomingBlock(0) ||
7545 !isSafeToSinkLoad(LI))
7547 } else if (isa<GetElementPtrInst>(FirstInst)) {
7548 if (FirstInst->getNumOperands() == 2)
7549 return FoldPHIArgBinOpIntoPHI(PN);
7550 // Can't handle general GEPs yet.
7553 return 0; // Cannot fold this operation.
7556 // Check to see if all arguments are the same operation.
7557 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7558 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7559 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7560 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7563 if (I->getOperand(0)->getType() != CastSrcTy)
7564 return 0; // Cast operation must match.
7565 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7566 // We can't sink the load if the loaded value could be modified between
7567 // the load and the PHI.
7568 if (LI->isVolatile() != isVolatile ||
7569 LI->getParent() != PN.getIncomingBlock(i) ||
7570 !isSafeToSinkLoad(LI))
7572 } else if (I->getOperand(1) != ConstantOp) {
7577 // Okay, they are all the same operation. Create a new PHI node of the
7578 // correct type, and PHI together all of the LHS's of the instructions.
7579 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7580 PN.getName()+".in");
7581 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7583 Value *InVal = FirstInst->getOperand(0);
7584 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7586 // Add all operands to the new PHI.
7587 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7588 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7589 if (NewInVal != InVal)
7591 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7596 // The new PHI unions all of the same values together. This is really
7597 // common, so we handle it intelligently here for compile-time speed.
7601 InsertNewInstBefore(NewPN, PN);
7605 // Insert and return the new operation.
7606 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7607 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7608 else if (isa<LoadInst>(FirstInst))
7609 return new LoadInst(PhiVal, "", isVolatile);
7610 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7611 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7612 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7613 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7614 PhiVal, ConstantOp);
7616 assert(0 && "Unknown operation");
7619 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7621 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7622 if (PN->use_empty()) return true;
7623 if (!PN->hasOneUse()) return false;
7625 // Remember this node, and if we find the cycle, return.
7626 if (!PotentiallyDeadPHIs.insert(PN).second)
7629 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7630 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7635 // PHINode simplification
7637 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7638 // If LCSSA is around, don't mess with Phi nodes
7639 if (mustPreserveAnalysisID(LCSSAID)) return 0;
7641 if (Value *V = PN.hasConstantValue())
7642 return ReplaceInstUsesWith(PN, V);
7644 // If all PHI operands are the same operation, pull them through the PHI,
7645 // reducing code size.
7646 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7647 PN.getIncomingValue(0)->hasOneUse())
7648 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7651 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7652 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7653 // PHI)... break the cycle.
7654 if (PN.hasOneUse()) {
7655 Instruction *PHIUser = cast<Instruction>(PN.use_back());
7656 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
7657 std::set<PHINode*> PotentiallyDeadPHIs;
7658 PotentiallyDeadPHIs.insert(&PN);
7659 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7660 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7663 // If this phi has a single use, and if that use just computes a value for
7664 // the next iteration of a loop, delete the phi. This occurs with unused
7665 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
7666 // common case here is good because the only other things that catch this
7667 // are induction variable analysis (sometimes) and ADCE, which is only run
7669 if (PHIUser->hasOneUse() &&
7670 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
7671 PHIUser->use_back() == &PN) {
7672 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7679 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7680 Instruction *InsertPoint,
7682 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7683 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7684 // We must cast correctly to the pointer type. Ensure that we
7685 // sign extend the integer value if it is smaller as this is
7686 // used for address computation.
7687 Instruction::CastOps opcode =
7688 (VTySize < PtrSize ? Instruction::SExt :
7689 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7690 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7694 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7695 Value *PtrOp = GEP.getOperand(0);
7696 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7697 // If so, eliminate the noop.
7698 if (GEP.getNumOperands() == 1)
7699 return ReplaceInstUsesWith(GEP, PtrOp);
7701 if (isa<UndefValue>(GEP.getOperand(0)))
7702 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7704 bool HasZeroPointerIndex = false;
7705 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7706 HasZeroPointerIndex = C->isNullValue();
7708 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7709 return ReplaceInstUsesWith(GEP, PtrOp);
7711 // Eliminate unneeded casts for indices.
7712 bool MadeChange = false;
7713 gep_type_iterator GTI = gep_type_begin(GEP);
7714 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7715 if (isa<SequentialType>(*GTI)) {
7716 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7717 if (CI->getOpcode() == Instruction::ZExt ||
7718 CI->getOpcode() == Instruction::SExt) {
7719 const Type *SrcTy = CI->getOperand(0)->getType();
7720 // We can eliminate a cast from i32 to i64 iff the target
7721 // is a 32-bit pointer target.
7722 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7724 GEP.setOperand(i, CI->getOperand(0));
7728 // If we are using a wider index than needed for this platform, shrink it
7729 // to what we need. If the incoming value needs a cast instruction,
7730 // insert it. This explicit cast can make subsequent optimizations more
7732 Value *Op = GEP.getOperand(i);
7733 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
7734 if (Constant *C = dyn_cast<Constant>(Op)) {
7735 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7738 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7740 GEP.setOperand(i, Op);
7744 if (MadeChange) return &GEP;
7746 // Combine Indices - If the source pointer to this getelementptr instruction
7747 // is a getelementptr instruction, combine the indices of the two
7748 // getelementptr instructions into a single instruction.
7750 std::vector<Value*> SrcGEPOperands;
7751 if (User *Src = dyn_castGetElementPtr(PtrOp))
7752 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
7754 if (!SrcGEPOperands.empty()) {
7755 // Note that if our source is a gep chain itself that we wait for that
7756 // chain to be resolved before we perform this transformation. This
7757 // avoids us creating a TON of code in some cases.
7759 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7760 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7761 return 0; // Wait until our source is folded to completion.
7763 std::vector<Value *> Indices;
7765 // Find out whether the last index in the source GEP is a sequential idx.
7766 bool EndsWithSequential = false;
7767 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7768 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7769 EndsWithSequential = !isa<StructType>(*I);
7771 // Can we combine the two pointer arithmetics offsets?
7772 if (EndsWithSequential) {
7773 // Replace: gep (gep %P, long B), long A, ...
7774 // With: T = long A+B; gep %P, T, ...
7776 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7777 if (SO1 == Constant::getNullValue(SO1->getType())) {
7779 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7782 // If they aren't the same type, convert both to an integer of the
7783 // target's pointer size.
7784 if (SO1->getType() != GO1->getType()) {
7785 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7786 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7787 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7788 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7790 unsigned PS = TD->getPointerSize();
7791 if (TD->getTypeSize(SO1->getType()) == PS) {
7792 // Convert GO1 to SO1's type.
7793 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
7795 } else if (TD->getTypeSize(GO1->getType()) == PS) {
7796 // Convert SO1 to GO1's type.
7797 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
7799 const Type *PT = TD->getIntPtrType();
7800 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
7801 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
7805 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7806 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7808 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7809 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7813 // Recycle the GEP we already have if possible.
7814 if (SrcGEPOperands.size() == 2) {
7815 GEP.setOperand(0, SrcGEPOperands[0]);
7816 GEP.setOperand(1, Sum);
7819 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7820 SrcGEPOperands.end()-1);
7821 Indices.push_back(Sum);
7822 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7824 } else if (isa<Constant>(*GEP.idx_begin()) &&
7825 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7826 SrcGEPOperands.size() != 1) {
7827 // Otherwise we can do the fold if the first index of the GEP is a zero
7828 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7829 SrcGEPOperands.end());
7830 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7833 if (!Indices.empty())
7834 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
7835 Indices.size(), GEP.getName());
7837 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7838 // GEP of global variable. If all of the indices for this GEP are
7839 // constants, we can promote this to a constexpr instead of an instruction.
7841 // Scan for nonconstants...
7842 SmallVector<Constant*, 8> Indices;
7843 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7844 for (; I != E && isa<Constant>(*I); ++I)
7845 Indices.push_back(cast<Constant>(*I));
7847 if (I == E) { // If they are all constants...
7848 Constant *CE = ConstantExpr::getGetElementPtr(GV,
7849 &Indices[0],Indices.size());
7851 // Replace all uses of the GEP with the new constexpr...
7852 return ReplaceInstUsesWith(GEP, CE);
7854 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7855 if (!isa<PointerType>(X->getType())) {
7856 // Not interesting. Source pointer must be a cast from pointer.
7857 } else if (HasZeroPointerIndex) {
7858 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7859 // into : GEP [10 x ubyte]* X, long 0, ...
7861 // This occurs when the program declares an array extern like "int X[];"
7863 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7864 const PointerType *XTy = cast<PointerType>(X->getType());
7865 if (const ArrayType *XATy =
7866 dyn_cast<ArrayType>(XTy->getElementType()))
7867 if (const ArrayType *CATy =
7868 dyn_cast<ArrayType>(CPTy->getElementType()))
7869 if (CATy->getElementType() == XATy->getElementType()) {
7870 // At this point, we know that the cast source type is a pointer
7871 // to an array of the same type as the destination pointer
7872 // array. Because the array type is never stepped over (there
7873 // is a leading zero) we can fold the cast into this GEP.
7874 GEP.setOperand(0, X);
7877 } else if (GEP.getNumOperands() == 2) {
7878 // Transform things like:
7879 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7880 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7881 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7882 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7883 if (isa<ArrayType>(SrcElTy) &&
7884 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7885 TD->getTypeSize(ResElTy)) {
7886 Value *V = InsertNewInstBefore(
7887 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7888 GEP.getOperand(1), GEP.getName()), GEP);
7889 // V and GEP are both pointer types --> BitCast
7890 return new BitCastInst(V, GEP.getType());
7893 // Transform things like:
7894 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7895 // (where tmp = 8*tmp2) into:
7896 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7898 if (isa<ArrayType>(SrcElTy) &&
7899 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
7900 uint64_t ArrayEltSize =
7901 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7903 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7904 // allow either a mul, shift, or constant here.
7906 ConstantInt *Scale = 0;
7907 if (ArrayEltSize == 1) {
7908 NewIdx = GEP.getOperand(1);
7909 Scale = ConstantInt::get(NewIdx->getType(), 1);
7910 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7911 NewIdx = ConstantInt::get(CI->getType(), 1);
7913 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7914 if (Inst->getOpcode() == Instruction::Shl &&
7915 isa<ConstantInt>(Inst->getOperand(1))) {
7917 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7918 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7919 NewIdx = Inst->getOperand(0);
7920 } else if (Inst->getOpcode() == Instruction::Mul &&
7921 isa<ConstantInt>(Inst->getOperand(1))) {
7922 Scale = cast<ConstantInt>(Inst->getOperand(1));
7923 NewIdx = Inst->getOperand(0);
7927 // If the index will be to exactly the right offset with the scale taken
7928 // out, perform the transformation.
7929 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7930 if (isa<ConstantInt>(Scale))
7931 Scale = ConstantInt::get(Scale->getType(),
7932 Scale->getZExtValue() / ArrayEltSize);
7933 if (Scale->getZExtValue() != 1) {
7934 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
7936 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7937 NewIdx = InsertNewInstBefore(Sc, GEP);
7940 // Insert the new GEP instruction.
7941 Instruction *NewGEP =
7942 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7943 NewIdx, GEP.getName());
7944 NewGEP = InsertNewInstBefore(NewGEP, GEP);
7945 // The NewGEP must be pointer typed, so must the old one -> BitCast
7946 return new BitCastInst(NewGEP, GEP.getType());
7955 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7956 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7957 if (AI.isArrayAllocation()) // Check C != 1
7958 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7960 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7961 AllocationInst *New = 0;
7963 // Create and insert the replacement instruction...
7964 if (isa<MallocInst>(AI))
7965 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7967 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7968 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7971 InsertNewInstBefore(New, AI);
7973 // Scan to the end of the allocation instructions, to skip over a block of
7974 // allocas if possible...
7976 BasicBlock::iterator It = New;
7977 while (isa<AllocationInst>(*It)) ++It;
7979 // Now that I is pointing to the first non-allocation-inst in the block,
7980 // insert our getelementptr instruction...
7982 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
7983 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
7984 New->getName()+".sub", It);
7986 // Now make everything use the getelementptr instead of the original
7988 return ReplaceInstUsesWith(AI, V);
7989 } else if (isa<UndefValue>(AI.getArraySize())) {
7990 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7993 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7994 // Note that we only do this for alloca's, because malloc should allocate and
7995 // return a unique pointer, even for a zero byte allocation.
7996 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
7997 TD->getTypeSize(AI.getAllocatedType()) == 0)
7998 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8003 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8004 Value *Op = FI.getOperand(0);
8006 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8007 if (CastInst *CI = dyn_cast<CastInst>(Op))
8008 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8009 FI.setOperand(0, CI->getOperand(0));
8013 // free undef -> unreachable.
8014 if (isa<UndefValue>(Op)) {
8015 // Insert a new store to null because we cannot modify the CFG here.
8016 new StoreInst(ConstantInt::getTrue(),
8017 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8018 return EraseInstFromFunction(FI);
8021 // If we have 'free null' delete the instruction. This can happen in stl code
8022 // when lots of inlining happens.
8023 if (isa<ConstantPointerNull>(Op))
8024 return EraseInstFromFunction(FI);
8030 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8031 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8032 User *CI = cast<User>(LI.getOperand(0));
8033 Value *CastOp = CI->getOperand(0);
8035 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8036 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8037 const Type *SrcPTy = SrcTy->getElementType();
8039 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8040 isa<VectorType>(DestPTy)) {
8041 // If the source is an array, the code below will not succeed. Check to
8042 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8044 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8045 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8046 if (ASrcTy->getNumElements() != 0) {
8048 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8049 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8050 SrcTy = cast<PointerType>(CastOp->getType());
8051 SrcPTy = SrcTy->getElementType();
8054 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8055 isa<VectorType>(SrcPTy)) &&
8056 // Do not allow turning this into a load of an integer, which is then
8057 // casted to a pointer, this pessimizes pointer analysis a lot.
8058 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8059 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8060 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8062 // Okay, we are casting from one integer or pointer type to another of
8063 // the same size. Instead of casting the pointer before the load, cast
8064 // the result of the loaded value.
8065 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8067 LI.isVolatile()),LI);
8068 // Now cast the result of the load.
8069 return new BitCastInst(NewLoad, LI.getType());
8076 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8077 /// from this value cannot trap. If it is not obviously safe to load from the
8078 /// specified pointer, we do a quick local scan of the basic block containing
8079 /// ScanFrom, to determine if the address is already accessed.
8080 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8081 // If it is an alloca or global variable, it is always safe to load from.
8082 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8084 // Otherwise, be a little bit agressive by scanning the local block where we
8085 // want to check to see if the pointer is already being loaded or stored
8086 // from/to. If so, the previous load or store would have already trapped,
8087 // so there is no harm doing an extra load (also, CSE will later eliminate
8088 // the load entirely).
8089 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8094 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8095 if (LI->getOperand(0) == V) return true;
8096 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8097 if (SI->getOperand(1) == V) return true;
8103 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8104 Value *Op = LI.getOperand(0);
8106 // load (cast X) --> cast (load X) iff safe
8107 if (isa<CastInst>(Op))
8108 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8111 // None of the following transforms are legal for volatile loads.
8112 if (LI.isVolatile()) return 0;
8114 if (&LI.getParent()->front() != &LI) {
8115 BasicBlock::iterator BBI = &LI; --BBI;
8116 // If the instruction immediately before this is a store to the same
8117 // address, do a simple form of store->load forwarding.
8118 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8119 if (SI->getOperand(1) == LI.getOperand(0))
8120 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8121 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8122 if (LIB->getOperand(0) == LI.getOperand(0))
8123 return ReplaceInstUsesWith(LI, LIB);
8126 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8127 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8128 isa<UndefValue>(GEPI->getOperand(0))) {
8129 // Insert a new store to null instruction before the load to indicate
8130 // that this code is not reachable. We do this instead of inserting
8131 // an unreachable instruction directly because we cannot modify the
8133 new StoreInst(UndefValue::get(LI.getType()),
8134 Constant::getNullValue(Op->getType()), &LI);
8135 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8138 if (Constant *C = dyn_cast<Constant>(Op)) {
8139 // load null/undef -> undef
8140 if ((C->isNullValue() || isa<UndefValue>(C))) {
8141 // Insert a new store to null instruction before the load to indicate that
8142 // this code is not reachable. We do this instead of inserting an
8143 // unreachable instruction directly because we cannot modify the CFG.
8144 new StoreInst(UndefValue::get(LI.getType()),
8145 Constant::getNullValue(Op->getType()), &LI);
8146 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8149 // Instcombine load (constant global) into the value loaded.
8150 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8151 if (GV->isConstant() && !GV->isDeclaration())
8152 return ReplaceInstUsesWith(LI, GV->getInitializer());
8154 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8155 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8156 if (CE->getOpcode() == Instruction::GetElementPtr) {
8157 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8158 if (GV->isConstant() && !GV->isDeclaration())
8160 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8161 return ReplaceInstUsesWith(LI, V);
8162 if (CE->getOperand(0)->isNullValue()) {
8163 // Insert a new store to null instruction before the load to indicate
8164 // that this code is not reachable. We do this instead of inserting
8165 // an unreachable instruction directly because we cannot modify the
8167 new StoreInst(UndefValue::get(LI.getType()),
8168 Constant::getNullValue(Op->getType()), &LI);
8169 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8172 } else if (CE->isCast()) {
8173 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8178 if (Op->hasOneUse()) {
8179 // Change select and PHI nodes to select values instead of addresses: this
8180 // helps alias analysis out a lot, allows many others simplifications, and
8181 // exposes redundancy in the code.
8183 // Note that we cannot do the transformation unless we know that the
8184 // introduced loads cannot trap! Something like this is valid as long as
8185 // the condition is always false: load (select bool %C, int* null, int* %G),
8186 // but it would not be valid if we transformed it to load from null
8189 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8190 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8191 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8192 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8193 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8194 SI->getOperand(1)->getName()+".val"), LI);
8195 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8196 SI->getOperand(2)->getName()+".val"), LI);
8197 return new SelectInst(SI->getCondition(), V1, V2);
8200 // load (select (cond, null, P)) -> load P
8201 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8202 if (C->isNullValue()) {
8203 LI.setOperand(0, SI->getOperand(2));
8207 // load (select (cond, P, null)) -> load P
8208 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8209 if (C->isNullValue()) {
8210 LI.setOperand(0, SI->getOperand(1));
8218 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8220 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8221 User *CI = cast<User>(SI.getOperand(1));
8222 Value *CastOp = CI->getOperand(0);
8224 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8225 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8226 const Type *SrcPTy = SrcTy->getElementType();
8228 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8229 // If the source is an array, the code below will not succeed. Check to
8230 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8232 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8233 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8234 if (ASrcTy->getNumElements() != 0) {
8236 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8237 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8238 SrcTy = cast<PointerType>(CastOp->getType());
8239 SrcPTy = SrcTy->getElementType();
8242 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8243 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8244 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8246 // Okay, we are casting from one integer or pointer type to another of
8247 // the same size. Instead of casting the pointer before
8248 // the store, cast the value to be stored.
8250 Value *SIOp0 = SI.getOperand(0);
8251 Instruction::CastOps opcode = Instruction::BitCast;
8252 const Type* CastSrcTy = SIOp0->getType();
8253 const Type* CastDstTy = SrcPTy;
8254 if (isa<PointerType>(CastDstTy)) {
8255 if (CastSrcTy->isInteger())
8256 opcode = Instruction::IntToPtr;
8257 } else if (isa<IntegerType>(CastDstTy)) {
8258 if (isa<PointerType>(SIOp0->getType()))
8259 opcode = Instruction::PtrToInt;
8261 if (Constant *C = dyn_cast<Constant>(SIOp0))
8262 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8264 NewCast = IC.InsertNewInstBefore(
8265 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8267 return new StoreInst(NewCast, CastOp);
8274 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8275 Value *Val = SI.getOperand(0);
8276 Value *Ptr = SI.getOperand(1);
8278 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8279 EraseInstFromFunction(SI);
8284 // If the RHS is an alloca with a single use, zapify the store, making the
8286 if (Ptr->hasOneUse()) {
8287 if (isa<AllocaInst>(Ptr)) {
8288 EraseInstFromFunction(SI);
8293 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8294 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8295 GEP->getOperand(0)->hasOneUse()) {
8296 EraseInstFromFunction(SI);
8302 // Do really simple DSE, to catch cases where there are several consequtive
8303 // stores to the same location, separated by a few arithmetic operations. This
8304 // situation often occurs with bitfield accesses.
8305 BasicBlock::iterator BBI = &SI;
8306 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8310 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8311 // Prev store isn't volatile, and stores to the same location?
8312 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8315 EraseInstFromFunction(*PrevSI);
8321 // If this is a load, we have to stop. However, if the loaded value is from
8322 // the pointer we're loading and is producing the pointer we're storing,
8323 // then *this* store is dead (X = load P; store X -> P).
8324 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8325 if (LI == Val && LI->getOperand(0) == Ptr) {
8326 EraseInstFromFunction(SI);
8330 // Otherwise, this is a load from some other location. Stores before it
8335 // Don't skip over loads or things that can modify memory.
8336 if (BBI->mayWriteToMemory())
8341 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8343 // store X, null -> turns into 'unreachable' in SimplifyCFG
8344 if (isa<ConstantPointerNull>(Ptr)) {
8345 if (!isa<UndefValue>(Val)) {
8346 SI.setOperand(0, UndefValue::get(Val->getType()));
8347 if (Instruction *U = dyn_cast<Instruction>(Val))
8348 WorkList.push_back(U); // Dropped a use.
8351 return 0; // Do not modify these!
8354 // store undef, Ptr -> noop
8355 if (isa<UndefValue>(Val)) {
8356 EraseInstFromFunction(SI);
8361 // If the pointer destination is a cast, see if we can fold the cast into the
8363 if (isa<CastInst>(Ptr))
8364 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8366 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8368 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8372 // If this store is the last instruction in the basic block, and if the block
8373 // ends with an unconditional branch, try to move it to the successor block.
8375 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8376 if (BI->isUnconditional()) {
8377 // Check to see if the successor block has exactly two incoming edges. If
8378 // so, see if the other predecessor contains a store to the same location.
8379 // if so, insert a PHI node (if needed) and move the stores down.
8380 BasicBlock *Dest = BI->getSuccessor(0);
8382 pred_iterator PI = pred_begin(Dest);
8383 BasicBlock *Other = 0;
8384 if (*PI != BI->getParent())
8387 if (PI != pred_end(Dest)) {
8388 if (*PI != BI->getParent())
8393 if (++PI != pred_end(Dest))
8396 if (Other) { // If only one other pred...
8397 BBI = Other->getTerminator();
8398 // Make sure this other block ends in an unconditional branch and that
8399 // there is an instruction before the branch.
8400 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8401 BBI != Other->begin()) {
8403 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8405 // If this instruction is a store to the same location.
8406 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8407 // Okay, we know we can perform this transformation. Insert a PHI
8408 // node now if we need it.
8409 Value *MergedVal = OtherStore->getOperand(0);
8410 if (MergedVal != SI.getOperand(0)) {
8411 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8412 PN->reserveOperandSpace(2);
8413 PN->addIncoming(SI.getOperand(0), SI.getParent());
8414 PN->addIncoming(OtherStore->getOperand(0), Other);
8415 MergedVal = InsertNewInstBefore(PN, Dest->front());
8418 // Advance to a place where it is safe to insert the new store and
8420 BBI = Dest->begin();
8421 while (isa<PHINode>(BBI)) ++BBI;
8422 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8423 OtherStore->isVolatile()), *BBI);
8425 // Nuke the old stores.
8426 EraseInstFromFunction(SI);
8427 EraseInstFromFunction(*OtherStore);
8439 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8440 // Change br (not X), label True, label False to: br X, label False, True
8442 BasicBlock *TrueDest;
8443 BasicBlock *FalseDest;
8444 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8445 !isa<Constant>(X)) {
8446 // Swap Destinations and condition...
8448 BI.setSuccessor(0, FalseDest);
8449 BI.setSuccessor(1, TrueDest);
8453 // Cannonicalize fcmp_one -> fcmp_oeq
8454 FCmpInst::Predicate FPred; Value *Y;
8455 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8456 TrueDest, FalseDest)))
8457 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8458 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8459 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8460 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8461 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
8462 NewSCC->takeName(I);
8463 // Swap Destinations and condition...
8464 BI.setCondition(NewSCC);
8465 BI.setSuccessor(0, FalseDest);
8466 BI.setSuccessor(1, TrueDest);
8467 removeFromWorkList(I);
8468 I->eraseFromParent();
8469 WorkList.push_back(NewSCC);
8473 // Cannonicalize icmp_ne -> icmp_eq
8474 ICmpInst::Predicate IPred;
8475 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8476 TrueDest, FalseDest)))
8477 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8478 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8479 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8480 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8481 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8482 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
8483 NewSCC->takeName(I);
8484 // Swap Destinations and condition...
8485 BI.setCondition(NewSCC);
8486 BI.setSuccessor(0, FalseDest);
8487 BI.setSuccessor(1, TrueDest);
8488 removeFromWorkList(I);
8489 I->eraseFromParent();;
8490 WorkList.push_back(NewSCC);
8497 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8498 Value *Cond = SI.getCondition();
8499 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8500 if (I->getOpcode() == Instruction::Add)
8501 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8502 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8503 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8504 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8506 SI.setOperand(0, I->getOperand(0));
8507 WorkList.push_back(I);
8514 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8515 /// is to leave as a vector operation.
8516 static bool CheapToScalarize(Value *V, bool isConstant) {
8517 if (isa<ConstantAggregateZero>(V))
8519 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
8520 if (isConstant) return true;
8521 // If all elts are the same, we can extract.
8522 Constant *Op0 = C->getOperand(0);
8523 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8524 if (C->getOperand(i) != Op0)
8528 Instruction *I = dyn_cast<Instruction>(V);
8529 if (!I) return false;
8531 // Insert element gets simplified to the inserted element or is deleted if
8532 // this is constant idx extract element and its a constant idx insertelt.
8533 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8534 isa<ConstantInt>(I->getOperand(2)))
8536 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8538 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8539 if (BO->hasOneUse() &&
8540 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8541 CheapToScalarize(BO->getOperand(1), isConstant)))
8543 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8544 if (CI->hasOneUse() &&
8545 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8546 CheapToScalarize(CI->getOperand(1), isConstant)))
8552 /// Read and decode a shufflevector mask.
8554 /// It turns undef elements into values that are larger than the number of
8555 /// elements in the input.
8556 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8557 unsigned NElts = SVI->getType()->getNumElements();
8558 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8559 return std::vector<unsigned>(NElts, 0);
8560 if (isa<UndefValue>(SVI->getOperand(2)))
8561 return std::vector<unsigned>(NElts, 2*NElts);
8563 std::vector<unsigned> Result;
8564 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
8565 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8566 if (isa<UndefValue>(CP->getOperand(i)))
8567 Result.push_back(NElts*2); // undef -> 8
8569 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8573 /// FindScalarElement - Given a vector and an element number, see if the scalar
8574 /// value is already around as a register, for example if it were inserted then
8575 /// extracted from the vector.
8576 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8577 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
8578 const VectorType *PTy = cast<VectorType>(V->getType());
8579 unsigned Width = PTy->getNumElements();
8580 if (EltNo >= Width) // Out of range access.
8581 return UndefValue::get(PTy->getElementType());
8583 if (isa<UndefValue>(V))
8584 return UndefValue::get(PTy->getElementType());
8585 else if (isa<ConstantAggregateZero>(V))
8586 return Constant::getNullValue(PTy->getElementType());
8587 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
8588 return CP->getOperand(EltNo);
8589 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8590 // If this is an insert to a variable element, we don't know what it is.
8591 if (!isa<ConstantInt>(III->getOperand(2)))
8593 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8595 // If this is an insert to the element we are looking for, return the
8598 return III->getOperand(1);
8600 // Otherwise, the insertelement doesn't modify the value, recurse on its
8602 return FindScalarElement(III->getOperand(0), EltNo);
8603 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8604 unsigned InEl = getShuffleMask(SVI)[EltNo];
8606 return FindScalarElement(SVI->getOperand(0), InEl);
8607 else if (InEl < Width*2)
8608 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8610 return UndefValue::get(PTy->getElementType());
8613 // Otherwise, we don't know.
8617 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8619 // If packed val is undef, replace extract with scalar undef.
8620 if (isa<UndefValue>(EI.getOperand(0)))
8621 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8623 // If packed val is constant 0, replace extract with scalar 0.
8624 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8625 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8627 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
8628 // If packed val is constant with uniform operands, replace EI
8629 // with that operand
8630 Constant *op0 = C->getOperand(0);
8631 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8632 if (C->getOperand(i) != op0) {
8637 return ReplaceInstUsesWith(EI, op0);
8640 // If extracting a specified index from the vector, see if we can recursively
8641 // find a previously computed scalar that was inserted into the vector.
8642 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8643 // This instruction only demands the single element from the input vector.
8644 // If the input vector has a single use, simplify it based on this use
8646 uint64_t IndexVal = IdxC->getZExtValue();
8647 if (EI.getOperand(0)->hasOneUse()) {
8649 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8652 EI.setOperand(0, V);
8657 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8658 return ReplaceInstUsesWith(EI, Elt);
8661 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8662 if (I->hasOneUse()) {
8663 // Push extractelement into predecessor operation if legal and
8664 // profitable to do so
8665 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8666 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8667 if (CheapToScalarize(BO, isConstantElt)) {
8668 ExtractElementInst *newEI0 =
8669 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8670 EI.getName()+".lhs");
8671 ExtractElementInst *newEI1 =
8672 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8673 EI.getName()+".rhs");
8674 InsertNewInstBefore(newEI0, EI);
8675 InsertNewInstBefore(newEI1, EI);
8676 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8678 } else if (isa<LoadInst>(I)) {
8679 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8680 PointerType::get(EI.getType()), EI);
8681 GetElementPtrInst *GEP =
8682 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8683 InsertNewInstBefore(GEP, EI);
8684 return new LoadInst(GEP);
8687 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8688 // Extracting the inserted element?
8689 if (IE->getOperand(2) == EI.getOperand(1))
8690 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8691 // If the inserted and extracted elements are constants, they must not
8692 // be the same value, extract from the pre-inserted value instead.
8693 if (isa<Constant>(IE->getOperand(2)) &&
8694 isa<Constant>(EI.getOperand(1))) {
8695 AddUsesToWorkList(EI);
8696 EI.setOperand(0, IE->getOperand(0));
8699 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8700 // If this is extracting an element from a shufflevector, figure out where
8701 // it came from and extract from the appropriate input element instead.
8702 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8703 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8705 if (SrcIdx < SVI->getType()->getNumElements())
8706 Src = SVI->getOperand(0);
8707 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8708 SrcIdx -= SVI->getType()->getNumElements();
8709 Src = SVI->getOperand(1);
8711 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8713 return new ExtractElementInst(Src, SrcIdx);
8720 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8721 /// elements from either LHS or RHS, return the shuffle mask and true.
8722 /// Otherwise, return false.
8723 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8724 std::vector<Constant*> &Mask) {
8725 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8726 "Invalid CollectSingleShuffleElements");
8727 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8729 if (isa<UndefValue>(V)) {
8730 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8732 } else if (V == LHS) {
8733 for (unsigned i = 0; i != NumElts; ++i)
8734 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8736 } else if (V == RHS) {
8737 for (unsigned i = 0; i != NumElts; ++i)
8738 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8740 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8741 // If this is an insert of an extract from some other vector, include it.
8742 Value *VecOp = IEI->getOperand(0);
8743 Value *ScalarOp = IEI->getOperand(1);
8744 Value *IdxOp = IEI->getOperand(2);
8746 if (!isa<ConstantInt>(IdxOp))
8748 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8750 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8751 // Okay, we can handle this if the vector we are insertinting into is
8753 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8754 // If so, update the mask to reflect the inserted undef.
8755 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
8758 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8759 if (isa<ConstantInt>(EI->getOperand(1)) &&
8760 EI->getOperand(0)->getType() == V->getType()) {
8761 unsigned ExtractedIdx =
8762 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8764 // This must be extracting from either LHS or RHS.
8765 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8766 // Okay, we can handle this if the vector we are insertinting into is
8768 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8769 // If so, update the mask to reflect the inserted value.
8770 if (EI->getOperand(0) == LHS) {
8771 Mask[InsertedIdx & (NumElts-1)] =
8772 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8774 assert(EI->getOperand(0) == RHS);
8775 Mask[InsertedIdx & (NumElts-1)] =
8776 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
8785 // TODO: Handle shufflevector here!
8790 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8791 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8792 /// that computes V and the LHS value of the shuffle.
8793 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8795 assert(isa<VectorType>(V->getType()) &&
8796 (RHS == 0 || V->getType() == RHS->getType()) &&
8797 "Invalid shuffle!");
8798 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8800 if (isa<UndefValue>(V)) {
8801 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8803 } else if (isa<ConstantAggregateZero>(V)) {
8804 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
8806 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8807 // If this is an insert of an extract from some other vector, include it.
8808 Value *VecOp = IEI->getOperand(0);
8809 Value *ScalarOp = IEI->getOperand(1);
8810 Value *IdxOp = IEI->getOperand(2);
8812 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8813 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8814 EI->getOperand(0)->getType() == V->getType()) {
8815 unsigned ExtractedIdx =
8816 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8817 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8819 // Either the extracted from or inserted into vector must be RHSVec,
8820 // otherwise we'd end up with a shuffle of three inputs.
8821 if (EI->getOperand(0) == RHS || RHS == 0) {
8822 RHS = EI->getOperand(0);
8823 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8824 Mask[InsertedIdx & (NumElts-1)] =
8825 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
8830 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8831 // Everything but the extracted element is replaced with the RHS.
8832 for (unsigned i = 0; i != NumElts; ++i) {
8833 if (i != InsertedIdx)
8834 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
8839 // If this insertelement is a chain that comes from exactly these two
8840 // vectors, return the vector and the effective shuffle.
8841 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8842 return EI->getOperand(0);
8847 // TODO: Handle shufflevector here!
8849 // Otherwise, can't do anything fancy. Return an identity vector.
8850 for (unsigned i = 0; i != NumElts; ++i)
8851 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8855 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8856 Value *VecOp = IE.getOperand(0);
8857 Value *ScalarOp = IE.getOperand(1);
8858 Value *IdxOp = IE.getOperand(2);
8860 // If the inserted element was extracted from some other vector, and if the
8861 // indexes are constant, try to turn this into a shufflevector operation.
8862 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8863 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8864 EI->getOperand(0)->getType() == IE.getType()) {
8865 unsigned NumVectorElts = IE.getType()->getNumElements();
8866 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8867 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8869 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8870 return ReplaceInstUsesWith(IE, VecOp);
8872 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8873 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8875 // If we are extracting a value from a vector, then inserting it right
8876 // back into the same place, just use the input vector.
8877 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8878 return ReplaceInstUsesWith(IE, VecOp);
8880 // We could theoretically do this for ANY input. However, doing so could
8881 // turn chains of insertelement instructions into a chain of shufflevector
8882 // instructions, and right now we do not merge shufflevectors. As such,
8883 // only do this in a situation where it is clear that there is benefit.
8884 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8885 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8886 // the values of VecOp, except then one read from EIOp0.
8887 // Build a new shuffle mask.
8888 std::vector<Constant*> Mask;
8889 if (isa<UndefValue>(VecOp))
8890 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
8892 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8893 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
8896 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8897 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8898 ConstantVector::get(Mask));
8901 // If this insertelement isn't used by some other insertelement, turn it
8902 // (and any insertelements it points to), into one big shuffle.
8903 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8904 std::vector<Constant*> Mask;
8906 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8907 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8908 // We now have a shuffle of LHS, RHS, Mask.
8909 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
8918 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8919 Value *LHS = SVI.getOperand(0);
8920 Value *RHS = SVI.getOperand(1);
8921 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8923 bool MadeChange = false;
8925 // Undefined shuffle mask -> undefined value.
8926 if (isa<UndefValue>(SVI.getOperand(2)))
8927 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8929 // If we have shuffle(x, undef, mask) and any elements of mask refer to
8930 // the undef, change them to undefs.
8931 if (isa<UndefValue>(SVI.getOperand(1))) {
8932 // Scan to see if there are any references to the RHS. If so, replace them
8933 // with undef element refs and set MadeChange to true.
8934 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8935 if (Mask[i] >= e && Mask[i] != 2*e) {
8942 // Remap any references to RHS to use LHS.
8943 std::vector<Constant*> Elts;
8944 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8946 Elts.push_back(UndefValue::get(Type::Int32Ty));
8948 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8950 SVI.setOperand(2, ConstantVector::get(Elts));
8954 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8955 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8956 if (LHS == RHS || isa<UndefValue>(LHS)) {
8957 if (isa<UndefValue>(LHS) && LHS == RHS) {
8958 // shuffle(undef,undef,mask) -> undef.
8959 return ReplaceInstUsesWith(SVI, LHS);
8962 // Remap any references to RHS to use LHS.
8963 std::vector<Constant*> Elts;
8964 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8966 Elts.push_back(UndefValue::get(Type::Int32Ty));
8968 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8969 (Mask[i] < e && isa<UndefValue>(LHS)))
8970 Mask[i] = 2*e; // Turn into undef.
8972 Mask[i] &= (e-1); // Force to LHS.
8973 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8976 SVI.setOperand(0, SVI.getOperand(1));
8977 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8978 SVI.setOperand(2, ConstantVector::get(Elts));
8979 LHS = SVI.getOperand(0);
8980 RHS = SVI.getOperand(1);
8984 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8985 bool isLHSID = true, isRHSID = true;
8987 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8988 if (Mask[i] >= e*2) continue; // Ignore undef values.
8989 // Is this an identity shuffle of the LHS value?
8990 isLHSID &= (Mask[i] == i);
8992 // Is this an identity shuffle of the RHS value?
8993 isRHSID &= (Mask[i]-e == i);
8996 // Eliminate identity shuffles.
8997 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
8998 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9000 // If the LHS is a shufflevector itself, see if we can combine it with this
9001 // one without producing an unusual shuffle. Here we are really conservative:
9002 // we are absolutely afraid of producing a shuffle mask not in the input
9003 // program, because the code gen may not be smart enough to turn a merged
9004 // shuffle into two specific shuffles: it may produce worse code. As such,
9005 // we only merge two shuffles if the result is one of the two input shuffle
9006 // masks. In this case, merging the shuffles just removes one instruction,
9007 // which we know is safe. This is good for things like turning:
9008 // (splat(splat)) -> splat.
9009 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9010 if (isa<UndefValue>(RHS)) {
9011 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9013 std::vector<unsigned> NewMask;
9014 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9016 NewMask.push_back(2*e);
9018 NewMask.push_back(LHSMask[Mask[i]]);
9020 // If the result mask is equal to the src shuffle or this shuffle mask, do
9022 if (NewMask == LHSMask || NewMask == Mask) {
9023 std::vector<Constant*> Elts;
9024 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9025 if (NewMask[i] >= e*2) {
9026 Elts.push_back(UndefValue::get(Type::Int32Ty));
9028 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9031 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9032 LHSSVI->getOperand(1),
9033 ConstantVector::get(Elts));
9038 return MadeChange ? &SVI : 0;
9043 void InstCombiner::removeFromWorkList(Instruction *I) {
9044 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
9049 /// TryToSinkInstruction - Try to move the specified instruction from its
9050 /// current block into the beginning of DestBlock, which can only happen if it's
9051 /// safe to move the instruction past all of the instructions between it and the
9052 /// end of its block.
9053 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9054 assert(I->hasOneUse() && "Invariants didn't hold!");
9056 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9057 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9059 // Do not sink alloca instructions out of the entry block.
9060 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
9063 // We can only sink load instructions if there is nothing between the load and
9064 // the end of block that could change the value.
9065 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9066 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9068 if (Scan->mayWriteToMemory())
9072 BasicBlock::iterator InsertPos = DestBlock->begin();
9073 while (isa<PHINode>(InsertPos)) ++InsertPos;
9075 I->moveBefore(InsertPos);
9081 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9082 /// all reachable code to the worklist.
9084 /// This has a couple of tricks to make the code faster and more powerful. In
9085 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9086 /// them to the worklist (this significantly speeds up instcombine on code where
9087 /// many instructions are dead or constant). Additionally, if we find a branch
9088 /// whose condition is a known constant, we only visit the reachable successors.
9090 static void AddReachableCodeToWorklist(BasicBlock *BB,
9091 std::set<BasicBlock*> &Visited,
9092 std::vector<Instruction*> &WorkList,
9093 const TargetData *TD) {
9094 // We have now visited this block! If we've already been here, bail out.
9095 if (!Visited.insert(BB).second) return;
9097 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9098 Instruction *Inst = BBI++;
9100 // DCE instruction if trivially dead.
9101 if (isInstructionTriviallyDead(Inst)) {
9103 DOUT << "IC: DCE: " << *Inst;
9104 Inst->eraseFromParent();
9108 // ConstantProp instruction if trivially constant.
9109 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9110 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9111 Inst->replaceAllUsesWith(C);
9113 Inst->eraseFromParent();
9117 WorkList.push_back(Inst);
9120 // Recursively visit successors. If this is a branch or switch on a constant,
9121 // only visit the reachable successor.
9122 TerminatorInst *TI = BB->getTerminator();
9123 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9124 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9125 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9126 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
9130 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9131 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9132 // See if this is an explicit destination.
9133 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9134 if (SI->getCaseValue(i) == Cond) {
9135 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
9139 // Otherwise it is the default destination.
9140 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
9145 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9146 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
9149 bool InstCombiner::runOnFunction(Function &F) {
9150 bool Changed = false;
9151 TD = &getAnalysis<TargetData>();
9154 // Do a depth-first traversal of the function, populate the worklist with
9155 // the reachable instructions. Ignore blocks that are not reachable. Keep
9156 // track of which blocks we visit.
9157 std::set<BasicBlock*> Visited;
9158 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
9160 // Do a quick scan over the function. If we find any blocks that are
9161 // unreachable, remove any instructions inside of them. This prevents
9162 // the instcombine code from having to deal with some bad special cases.
9163 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9164 if (!Visited.count(BB)) {
9165 Instruction *Term = BB->getTerminator();
9166 while (Term != BB->begin()) { // Remove instrs bottom-up
9167 BasicBlock::iterator I = Term; --I;
9169 DOUT << "IC: DCE: " << *I;
9172 if (!I->use_empty())
9173 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9174 I->eraseFromParent();
9179 while (!WorkList.empty()) {
9180 Instruction *I = WorkList.back(); // Get an instruction from the worklist
9181 WorkList.pop_back();
9183 // Check to see if we can DCE the instruction.
9184 if (isInstructionTriviallyDead(I)) {
9185 // Add operands to the worklist.
9186 if (I->getNumOperands() < 4)
9187 AddUsesToWorkList(*I);
9190 DOUT << "IC: DCE: " << *I;
9192 I->eraseFromParent();
9193 removeFromWorkList(I);
9197 // Instruction isn't dead, see if we can constant propagate it.
9198 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9199 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9201 // Add operands to the worklist.
9202 AddUsesToWorkList(*I);
9203 ReplaceInstUsesWith(*I, C);
9206 I->eraseFromParent();
9207 removeFromWorkList(I);
9211 // See if we can trivially sink this instruction to a successor basic block.
9212 if (I->hasOneUse()) {
9213 BasicBlock *BB = I->getParent();
9214 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9215 if (UserParent != BB) {
9216 bool UserIsSuccessor = false;
9217 // See if the user is one of our successors.
9218 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9219 if (*SI == UserParent) {
9220 UserIsSuccessor = true;
9224 // If the user is one of our immediate successors, and if that successor
9225 // only has us as a predecessors (we'd have to split the critical edge
9226 // otherwise), we can keep going.
9227 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9228 next(pred_begin(UserParent)) == pred_end(UserParent))
9229 // Okay, the CFG is simple enough, try to sink this instruction.
9230 Changed |= TryToSinkInstruction(I, UserParent);
9234 // Now that we have an instruction, try combining it to simplify it...
9235 if (Instruction *Result = visit(*I)) {
9237 // Should we replace the old instruction with a new one?
9239 DOUT << "IC: Old = " << *I
9240 << " New = " << *Result;
9242 // Everything uses the new instruction now.
9243 I->replaceAllUsesWith(Result);
9245 // Push the new instruction and any users onto the worklist.
9246 WorkList.push_back(Result);
9247 AddUsersToWorkList(*Result);
9249 // Move the name to the new instruction first.
9250 Result->takeName(I);
9252 // Insert the new instruction into the basic block...
9253 BasicBlock *InstParent = I->getParent();
9254 BasicBlock::iterator InsertPos = I;
9256 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9257 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9260 InstParent->getInstList().insert(InsertPos, Result);
9262 // Make sure that we reprocess all operands now that we reduced their
9264 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9265 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9266 WorkList.push_back(OpI);
9268 // Instructions can end up on the worklist more than once. Make sure
9269 // we do not process an instruction that has been deleted.
9270 removeFromWorkList(I);
9272 // Erase the old instruction.
9273 InstParent->getInstList().erase(I);
9275 DOUT << "IC: MOD = " << *I;
9277 // If the instruction was modified, it's possible that it is now dead.
9278 // if so, remove it.
9279 if (isInstructionTriviallyDead(I)) {
9280 // Make sure we process all operands now that we are reducing their
9282 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9283 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9284 WorkList.push_back(OpI);
9286 // Instructions may end up in the worklist more than once. Erase all
9287 // occurrences of this instruction.
9288 removeFromWorkList(I);
9289 I->eraseFromParent();
9291 WorkList.push_back(Result);
9292 AddUsersToWorkList(*Result);
9302 FunctionPass *llvm::createInstructionCombiningPass() {
9303 return new InstCombiner();