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/SmallPtrSet.h"
55 #include "llvm/ADT/Statistic.h"
56 #include "llvm/ADT/STLExtras.h"
60 using namespace llvm::PatternMatch;
62 STATISTIC(NumCombined , "Number of insts combined");
63 STATISTIC(NumConstProp, "Number of constant folds");
64 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
65 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
66 STATISTIC(NumSunkInst , "Number of instructions sunk");
69 class VISIBILITY_HIDDEN InstCombiner
70 : public FunctionPass,
71 public InstVisitor<InstCombiner, Instruction*> {
72 // Worklist of all of the instructions that need to be simplified.
73 std::vector<Instruction*> WorkList;
76 /// AddUsersToWorkList - When an instruction is simplified, add all users of
77 /// the instruction to the work lists because they might get more simplified
80 void AddUsersToWorkList(Value &I) {
81 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
83 WorkList.push_back(cast<Instruction>(*UI));
86 /// AddUsesToWorkList - When an instruction is simplified, add operands to
87 /// the work lists because they might get more simplified now.
89 void AddUsesToWorkList(Instruction &I) {
90 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
91 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
92 WorkList.push_back(Op);
95 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
96 /// dead. Add all of its operands to the worklist, turning them into
97 /// undef's to reduce the number of uses of those instructions.
99 /// Return the specified operand before it is turned into an undef.
101 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
102 Value *R = I.getOperand(op);
104 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
105 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
106 WorkList.push_back(Op);
107 // Set the operand to undef to drop the use.
108 I.setOperand(i, UndefValue::get(Op->getType()));
114 // removeFromWorkList - remove all instances of I from the worklist.
115 void removeFromWorkList(Instruction *I);
117 virtual bool runOnFunction(Function &F);
119 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
120 AU.addRequired<TargetData>();
121 AU.addPreservedID(LCSSAID);
122 AU.setPreservesCFG();
125 TargetData &getTargetData() const { return *TD; }
127 // Visitation implementation - Implement instruction combining for different
128 // instruction types. The semantics are as follows:
130 // null - No change was made
131 // I - Change was made, I is still valid, I may be dead though
132 // otherwise - Change was made, replace I with returned instruction
134 Instruction *visitAdd(BinaryOperator &I);
135 Instruction *visitSub(BinaryOperator &I);
136 Instruction *visitMul(BinaryOperator &I);
137 Instruction *visitURem(BinaryOperator &I);
138 Instruction *visitSRem(BinaryOperator &I);
139 Instruction *visitFRem(BinaryOperator &I);
140 Instruction *commonRemTransforms(BinaryOperator &I);
141 Instruction *commonIRemTransforms(BinaryOperator &I);
142 Instruction *commonDivTransforms(BinaryOperator &I);
143 Instruction *commonIDivTransforms(BinaryOperator &I);
144 Instruction *visitUDiv(BinaryOperator &I);
145 Instruction *visitSDiv(BinaryOperator &I);
146 Instruction *visitFDiv(BinaryOperator &I);
147 Instruction *visitAnd(BinaryOperator &I);
148 Instruction *visitOr (BinaryOperator &I);
149 Instruction *visitXor(BinaryOperator &I);
150 Instruction *visitShl(BinaryOperator &I);
151 Instruction *visitAShr(BinaryOperator &I);
152 Instruction *visitLShr(BinaryOperator &I);
153 Instruction *commonShiftTransforms(BinaryOperator &I);
154 Instruction *visitFCmpInst(FCmpInst &I);
155 Instruction *visitICmpInst(ICmpInst &I);
156 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
158 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
159 ICmpInst::Predicate Cond, Instruction &I);
160 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
162 Instruction *commonCastTransforms(CastInst &CI);
163 Instruction *commonIntCastTransforms(CastInst &CI);
164 Instruction *visitTrunc(CastInst &CI);
165 Instruction *visitZExt(CastInst &CI);
166 Instruction *visitSExt(CastInst &CI);
167 Instruction *visitFPTrunc(CastInst &CI);
168 Instruction *visitFPExt(CastInst &CI);
169 Instruction *visitFPToUI(CastInst &CI);
170 Instruction *visitFPToSI(CastInst &CI);
171 Instruction *visitUIToFP(CastInst &CI);
172 Instruction *visitSIToFP(CastInst &CI);
173 Instruction *visitPtrToInt(CastInst &CI);
174 Instruction *visitIntToPtr(CastInst &CI);
175 Instruction *visitBitCast(CastInst &CI);
176 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
178 Instruction *visitSelectInst(SelectInst &CI);
179 Instruction *visitCallInst(CallInst &CI);
180 Instruction *visitInvokeInst(InvokeInst &II);
181 Instruction *visitPHINode(PHINode &PN);
182 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
183 Instruction *visitAllocationInst(AllocationInst &AI);
184 Instruction *visitFreeInst(FreeInst &FI);
185 Instruction *visitLoadInst(LoadInst &LI);
186 Instruction *visitStoreInst(StoreInst &SI);
187 Instruction *visitBranchInst(BranchInst &BI);
188 Instruction *visitSwitchInst(SwitchInst &SI);
189 Instruction *visitInsertElementInst(InsertElementInst &IE);
190 Instruction *visitExtractElementInst(ExtractElementInst &EI);
191 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
193 // visitInstruction - Specify what to return for unhandled instructions...
194 Instruction *visitInstruction(Instruction &I) { return 0; }
197 Instruction *visitCallSite(CallSite CS);
198 bool transformConstExprCastCall(CallSite CS);
201 // InsertNewInstBefore - insert an instruction New before instruction Old
202 // in the program. Add the new instruction to the worklist.
204 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
205 assert(New && New->getParent() == 0 &&
206 "New instruction already inserted into a basic block!");
207 BasicBlock *BB = Old.getParent();
208 BB->getInstList().insert(&Old, New); // Insert inst
209 WorkList.push_back(New); // Add to worklist
213 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
214 /// This also adds the cast to the worklist. Finally, this returns the
216 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
218 if (V->getType() == Ty) return V;
220 if (Constant *CV = dyn_cast<Constant>(V))
221 return ConstantExpr::getCast(opc, CV, Ty);
223 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
224 WorkList.push_back(C);
228 // ReplaceInstUsesWith - This method is to be used when an instruction is
229 // found to be dead, replacable with another preexisting expression. Here
230 // we add all uses of I to the worklist, replace all uses of I with the new
231 // value, then return I, so that the inst combiner will know that I was
234 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
235 AddUsersToWorkList(I); // Add all modified instrs to worklist
237 I.replaceAllUsesWith(V);
240 // If we are replacing the instruction with itself, this must be in a
241 // segment of unreachable code, so just clobber the instruction.
242 I.replaceAllUsesWith(UndefValue::get(I.getType()));
247 // UpdateValueUsesWith - This method is to be used when an value is
248 // found to be replacable with another preexisting expression or was
249 // updated. Here we add all uses of I to the worklist, replace all uses of
250 // I with the new value (unless the instruction was just updated), then
251 // return true, so that the inst combiner will know that I was modified.
253 bool UpdateValueUsesWith(Value *Old, Value *New) {
254 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
256 Old->replaceAllUsesWith(New);
257 if (Instruction *I = dyn_cast<Instruction>(Old))
258 WorkList.push_back(I);
259 if (Instruction *I = dyn_cast<Instruction>(New))
260 WorkList.push_back(I);
264 // EraseInstFromFunction - When dealing with an instruction that has side
265 // effects or produces a void value, we can't rely on DCE to delete the
266 // instruction. Instead, visit methods should return the value returned by
268 Instruction *EraseInstFromFunction(Instruction &I) {
269 assert(I.use_empty() && "Cannot erase instruction that is used!");
270 AddUsesToWorkList(I);
271 removeFromWorkList(&I);
273 return 0; // Don't do anything with FI
277 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
278 /// InsertBefore instruction. This is specialized a bit to avoid inserting
279 /// casts that are known to not do anything...
281 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
282 Value *V, const Type *DestTy,
283 Instruction *InsertBefore);
285 /// SimplifyCommutative - This performs a few simplifications for
286 /// commutative operators.
287 bool SimplifyCommutative(BinaryOperator &I);
289 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
290 /// most-complex to least-complex order.
291 bool SimplifyCompare(CmpInst &I);
293 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
294 uint64_t &KnownZero, uint64_t &KnownOne,
297 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
298 uint64_t &UndefElts, unsigned Depth = 0);
300 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
301 // PHI node as operand #0, see if we can fold the instruction into the PHI
302 // (which is only possible if all operands to the PHI are constants).
303 Instruction *FoldOpIntoPhi(Instruction &I);
305 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
306 // operator and they all are only used by the PHI, PHI together their
307 // inputs, and do the operation once, to the result of the PHI.
308 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
309 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
312 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
313 ConstantInt *AndRHS, BinaryOperator &TheAnd);
315 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
316 bool isSub, Instruction &I);
317 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
318 bool isSigned, bool Inside, Instruction &IB);
319 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
320 Instruction *MatchBSwap(BinaryOperator &I);
322 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
325 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
328 // getComplexity: Assign a complexity or rank value to LLVM Values...
329 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
330 static unsigned getComplexity(Value *V) {
331 if (isa<Instruction>(V)) {
332 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
336 if (isa<Argument>(V)) return 3;
337 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
340 // isOnlyUse - Return true if this instruction will be deleted if we stop using
342 static bool isOnlyUse(Value *V) {
343 return V->hasOneUse() || isa<Constant>(V);
346 // getPromotedType - Return the specified type promoted as it would be to pass
347 // though a va_arg area...
348 static const Type *getPromotedType(const Type *Ty) {
349 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
350 if (ITy->getBitWidth() < 32)
351 return Type::Int32Ty;
352 } else if (Ty == Type::FloatTy)
353 return Type::DoubleTy;
357 /// getBitCastOperand - If the specified operand is a CastInst or a constant
358 /// expression bitcast, return the operand value, otherwise return null.
359 static Value *getBitCastOperand(Value *V) {
360 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
361 return I->getOperand(0);
362 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
363 if (CE->getOpcode() == Instruction::BitCast)
364 return CE->getOperand(0);
368 /// This function is a wrapper around CastInst::isEliminableCastPair. It
369 /// simply extracts arguments and returns what that function returns.
370 static Instruction::CastOps
371 isEliminableCastPair(
372 const CastInst *CI, ///< The first cast instruction
373 unsigned opcode, ///< The opcode of the second cast instruction
374 const Type *DstTy, ///< The target type for the second cast instruction
375 TargetData *TD ///< The target data for pointer size
378 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
379 const Type *MidTy = CI->getType(); // B from above
381 // Get the opcodes of the two Cast instructions
382 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
383 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
385 return Instruction::CastOps(
386 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
387 DstTy, TD->getIntPtrType()));
390 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
391 /// in any code being generated. It does not require codegen if V is simple
392 /// enough or if the cast can be folded into other casts.
393 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
394 const Type *Ty, TargetData *TD) {
395 if (V->getType() == Ty || isa<Constant>(V)) return false;
397 // If this is another cast that can be eliminated, it isn't codegen either.
398 if (const CastInst *CI = dyn_cast<CastInst>(V))
399 if (isEliminableCastPair(CI, opcode, Ty, TD))
404 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
405 /// InsertBefore instruction. This is specialized a bit to avoid inserting
406 /// casts that are known to not do anything...
408 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
409 Value *V, const Type *DestTy,
410 Instruction *InsertBefore) {
411 if (V->getType() == DestTy) return V;
412 if (Constant *C = dyn_cast<Constant>(V))
413 return ConstantExpr::getCast(opcode, C, DestTy);
415 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
418 // SimplifyCommutative - This performs a few simplifications for commutative
421 // 1. Order operands such that they are listed from right (least complex) to
422 // left (most complex). This puts constants before unary operators before
425 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
426 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
428 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
429 bool Changed = false;
430 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
431 Changed = !I.swapOperands();
433 if (!I.isAssociative()) return Changed;
434 Instruction::BinaryOps Opcode = I.getOpcode();
435 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
436 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
437 if (isa<Constant>(I.getOperand(1))) {
438 Constant *Folded = ConstantExpr::get(I.getOpcode(),
439 cast<Constant>(I.getOperand(1)),
440 cast<Constant>(Op->getOperand(1)));
441 I.setOperand(0, Op->getOperand(0));
442 I.setOperand(1, Folded);
444 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
445 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
446 isOnlyUse(Op) && isOnlyUse(Op1)) {
447 Constant *C1 = cast<Constant>(Op->getOperand(1));
448 Constant *C2 = cast<Constant>(Op1->getOperand(1));
450 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
451 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
452 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
455 WorkList.push_back(New);
456 I.setOperand(0, New);
457 I.setOperand(1, Folded);
464 /// SimplifyCompare - For a CmpInst this function just orders the operands
465 /// so that theyare listed from right (least complex) to left (most complex).
466 /// This puts constants before unary operators before binary operators.
467 bool InstCombiner::SimplifyCompare(CmpInst &I) {
468 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
471 // Compare instructions are not associative so there's nothing else we can do.
475 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
476 // if the LHS is a constant zero (which is the 'negate' form).
478 static inline Value *dyn_castNegVal(Value *V) {
479 if (BinaryOperator::isNeg(V))
480 return BinaryOperator::getNegArgument(V);
482 // Constants can be considered to be negated values if they can be folded.
483 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
484 return ConstantExpr::getNeg(C);
488 static inline Value *dyn_castNotVal(Value *V) {
489 if (BinaryOperator::isNot(V))
490 return BinaryOperator::getNotArgument(V);
492 // Constants can be considered to be not'ed values...
493 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
494 return ConstantExpr::getNot(C);
498 // dyn_castFoldableMul - If this value is a multiply that can be folded into
499 // other computations (because it has a constant operand), return the
500 // non-constant operand of the multiply, and set CST to point to the multiplier.
501 // Otherwise, return null.
503 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
504 if (V->hasOneUse() && V->getType()->isInteger())
505 if (Instruction *I = dyn_cast<Instruction>(V)) {
506 if (I->getOpcode() == Instruction::Mul)
507 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
508 return I->getOperand(0);
509 if (I->getOpcode() == Instruction::Shl)
510 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
511 // The multiplier is really 1 << CST.
512 Constant *One = ConstantInt::get(V->getType(), 1);
513 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
514 return I->getOperand(0);
520 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
521 /// expression, return it.
522 static User *dyn_castGetElementPtr(Value *V) {
523 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
524 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
525 if (CE->getOpcode() == Instruction::GetElementPtr)
526 return cast<User>(V);
530 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
531 static ConstantInt *AddOne(ConstantInt *C) {
532 return cast<ConstantInt>(ConstantExpr::getAdd(C,
533 ConstantInt::get(C->getType(), 1)));
535 static ConstantInt *SubOne(ConstantInt *C) {
536 return cast<ConstantInt>(ConstantExpr::getSub(C,
537 ConstantInt::get(C->getType(), 1)));
540 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
541 /// known to be either zero or one and return them in the KnownZero/KnownOne
542 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
544 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
545 uint64_t &KnownOne, unsigned Depth = 0) {
546 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
547 // we cannot optimize based on the assumption that it is zero without changing
548 // it to be an explicit zero. If we don't change it to zero, other code could
549 // optimized based on the contradictory assumption that it is non-zero.
550 // Because instcombine aggressively folds operations with undef args anyway,
551 // this won't lose us code quality.
552 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
553 // We know all of the bits for a constant!
554 KnownOne = CI->getZExtValue() & Mask;
555 KnownZero = ~KnownOne & Mask;
559 KnownZero = KnownOne = 0; // Don't know anything.
560 if (Depth == 6 || Mask == 0)
561 return; // Limit search depth.
563 uint64_t KnownZero2, KnownOne2;
564 Instruction *I = dyn_cast<Instruction>(V);
567 Mask &= cast<IntegerType>(V->getType())->getBitMask();
569 switch (I->getOpcode()) {
570 case Instruction::And:
571 // If either the LHS or the RHS are Zero, the result is zero.
572 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
574 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
575 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
576 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
578 // Output known-1 bits are only known if set in both the LHS & RHS.
579 KnownOne &= KnownOne2;
580 // Output known-0 are known to be clear if zero in either the LHS | RHS.
581 KnownZero |= KnownZero2;
583 case Instruction::Or:
584 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
586 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
587 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
588 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
590 // Output known-0 bits are only known if clear in both the LHS & RHS.
591 KnownZero &= KnownZero2;
592 // Output known-1 are known to be set if set in either the LHS | RHS.
593 KnownOne |= KnownOne2;
595 case Instruction::Xor: {
596 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
597 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
598 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
599 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
601 // Output known-0 bits are known if clear or set in both the LHS & RHS.
602 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
603 // Output known-1 are known to be set if set in only one of the LHS, RHS.
604 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
605 KnownZero = KnownZeroOut;
608 case Instruction::Select:
609 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
610 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
611 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
612 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
614 // Only known if known in both the LHS and RHS.
615 KnownOne &= KnownOne2;
616 KnownZero &= KnownZero2;
618 case Instruction::FPTrunc:
619 case Instruction::FPExt:
620 case Instruction::FPToUI:
621 case Instruction::FPToSI:
622 case Instruction::SIToFP:
623 case Instruction::PtrToInt:
624 case Instruction::UIToFP:
625 case Instruction::IntToPtr:
626 return; // Can't work with floating point or pointers
627 case Instruction::Trunc:
628 // All these have integer operands
629 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
631 case Instruction::BitCast: {
632 const Type *SrcTy = I->getOperand(0)->getType();
633 if (SrcTy->isInteger()) {
634 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
639 case Instruction::ZExt: {
640 // Compute the bits in the result that are not present in the input.
641 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
642 uint64_t NotIn = ~SrcTy->getBitMask();
643 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
645 Mask &= SrcTy->getBitMask();
646 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
647 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
648 // The top bits are known to be zero.
649 KnownZero |= NewBits;
652 case Instruction::SExt: {
653 // Compute the bits in the result that are not present in the input.
654 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
655 uint64_t NotIn = ~SrcTy->getBitMask();
656 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
658 Mask &= SrcTy->getBitMask();
659 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
660 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
662 // If the sign bit of the input is known set or clear, then we know the
663 // top bits of the result.
664 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
665 if (KnownZero & InSignBit) { // Input sign bit known zero
666 KnownZero |= NewBits;
667 KnownOne &= ~NewBits;
668 } else if (KnownOne & InSignBit) { // Input sign bit known set
670 KnownZero &= ~NewBits;
671 } else { // Input sign bit unknown
672 KnownZero &= ~NewBits;
673 KnownOne &= ~NewBits;
677 case Instruction::Shl:
678 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
679 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
680 uint64_t ShiftAmt = SA->getZExtValue();
682 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
683 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
684 KnownZero <<= ShiftAmt;
685 KnownOne <<= ShiftAmt;
686 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
690 case Instruction::LShr:
691 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
692 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
693 // Compute the new bits that are at the top now.
694 uint64_t ShiftAmt = SA->getZExtValue();
695 uint64_t HighBits = (1ULL << ShiftAmt)-1;
696 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
698 // Unsigned shift right.
700 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
701 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
702 KnownZero >>= ShiftAmt;
703 KnownOne >>= ShiftAmt;
704 KnownZero |= HighBits; // high bits known zero.
708 case Instruction::AShr:
709 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
710 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
711 // Compute the new bits that are at the top now.
712 uint64_t ShiftAmt = SA->getZExtValue();
713 uint64_t HighBits = (1ULL << ShiftAmt)-1;
714 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
716 // Signed shift right.
718 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
719 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
720 KnownZero >>= ShiftAmt;
721 KnownOne >>= ShiftAmt;
723 // Handle the sign bits.
724 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
725 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
727 if (KnownZero & SignBit) { // New bits are known zero.
728 KnownZero |= HighBits;
729 } else if (KnownOne & SignBit) { // New bits are known one.
730 KnownOne |= HighBits;
738 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
739 /// this predicate to simplify operations downstream. Mask is known to be zero
740 /// for bits that V cannot have.
741 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
742 uint64_t KnownZero, KnownOne;
743 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
744 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
745 return (KnownZero & Mask) == Mask;
748 /// ShrinkDemandedConstant - Check to see if the specified operand of the
749 /// specified instruction is a constant integer. If so, check to see if there
750 /// are any bits set in the constant that are not demanded. If so, shrink the
751 /// constant and return true.
752 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
754 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
755 if (!OpC) return false;
757 // If there are no bits set that aren't demanded, nothing to do.
758 if ((~Demanded & OpC->getZExtValue()) == 0)
761 // This is producing any bits that are not needed, shrink the RHS.
762 uint64_t Val = Demanded & OpC->getZExtValue();
763 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Val));
767 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
768 // set of known zero and one bits, compute the maximum and minimum values that
769 // could have the specified known zero and known one bits, returning them in
771 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
774 int64_t &Min, int64_t &Max) {
775 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
776 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
778 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
780 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
781 // bit if it is unknown.
783 Max = KnownOne|UnknownBits;
785 if (SignBit & UnknownBits) { // Sign bit is unknown
790 // Sign extend the min/max values.
791 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
792 Min = (Min << ShAmt) >> ShAmt;
793 Max = (Max << ShAmt) >> ShAmt;
796 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
797 // a set of known zero and one bits, compute the maximum and minimum values that
798 // could have the specified known zero and known one bits, returning them in
800 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
805 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
806 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
808 // The minimum value is when the unknown bits are all zeros.
810 // The maximum value is when the unknown bits are all ones.
811 Max = KnownOne|UnknownBits;
815 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
816 /// DemandedMask bits of the result of V are ever used downstream. If we can
817 /// use this information to simplify V, do so and return true. Otherwise,
818 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
819 /// the expression (used to simplify the caller). The KnownZero/One bits may
820 /// only be accurate for those bits in the DemandedMask.
821 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
822 uint64_t &KnownZero, uint64_t &KnownOne,
824 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
825 // We know all of the bits for a constant!
826 KnownOne = CI->getZExtValue() & DemandedMask;
827 KnownZero = ~KnownOne & DemandedMask;
831 KnownZero = KnownOne = 0;
832 if (!V->hasOneUse()) { // Other users may use these bits.
833 if (Depth != 0) { // Not at the root.
834 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
835 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
838 // If this is the root being simplified, allow it to have multiple uses,
839 // just set the DemandedMask to all bits.
840 DemandedMask = cast<IntegerType>(V->getType())->getBitMask();
841 } else if (DemandedMask == 0) { // Not demanding any bits from V.
842 if (V != UndefValue::get(V->getType()))
843 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
845 } else if (Depth == 6) { // Limit search depth.
849 Instruction *I = dyn_cast<Instruction>(V);
850 if (!I) return false; // Only analyze instructions.
852 DemandedMask &= cast<IntegerType>(V->getType())->getBitMask();
854 uint64_t KnownZero2 = 0, KnownOne2 = 0;
855 switch (I->getOpcode()) {
857 case Instruction::And:
858 // If either the LHS or the RHS are Zero, the result is zero.
859 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
860 KnownZero, KnownOne, Depth+1))
862 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
864 // If something is known zero on the RHS, the bits aren't demanded on the
866 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
867 KnownZero2, KnownOne2, Depth+1))
869 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
871 // If all of the demanded bits are known 1 on one side, return the other.
872 // These bits cannot contribute to the result of the 'and'.
873 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
874 return UpdateValueUsesWith(I, I->getOperand(0));
875 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
876 return UpdateValueUsesWith(I, I->getOperand(1));
878 // If all of the demanded bits in the inputs are known zeros, return zero.
879 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
880 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
882 // If the RHS is a constant, see if we can simplify it.
883 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
884 return UpdateValueUsesWith(I, I);
886 // Output known-1 bits are only known if set in both the LHS & RHS.
887 KnownOne &= KnownOne2;
888 // Output known-0 are known to be clear if zero in either the LHS | RHS.
889 KnownZero |= KnownZero2;
891 case Instruction::Or:
892 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
893 KnownZero, KnownOne, Depth+1))
895 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
896 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
897 KnownZero2, KnownOne2, Depth+1))
899 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
901 // If all of the demanded bits are known zero on one side, return the other.
902 // These bits cannot contribute to the result of the 'or'.
903 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
904 return UpdateValueUsesWith(I, I->getOperand(0));
905 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
906 return UpdateValueUsesWith(I, I->getOperand(1));
908 // If all of the potentially set bits on one side are known to be set on
909 // the other side, just use the 'other' side.
910 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
911 (DemandedMask & (~KnownZero)))
912 return UpdateValueUsesWith(I, I->getOperand(0));
913 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
914 (DemandedMask & (~KnownZero2)))
915 return UpdateValueUsesWith(I, I->getOperand(1));
917 // If the RHS is a constant, see if we can simplify it.
918 if (ShrinkDemandedConstant(I, 1, DemandedMask))
919 return UpdateValueUsesWith(I, I);
921 // Output known-0 bits are only known if clear in both the LHS & RHS.
922 KnownZero &= KnownZero2;
923 // Output known-1 are known to be set if set in either the LHS | RHS.
924 KnownOne |= KnownOne2;
926 case Instruction::Xor: {
927 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
928 KnownZero, KnownOne, Depth+1))
930 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
931 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
932 KnownZero2, KnownOne2, Depth+1))
934 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
936 // If all of the demanded bits are known zero on one side, return the other.
937 // These bits cannot contribute to the result of the 'xor'.
938 if ((DemandedMask & KnownZero) == DemandedMask)
939 return UpdateValueUsesWith(I, I->getOperand(0));
940 if ((DemandedMask & KnownZero2) == DemandedMask)
941 return UpdateValueUsesWith(I, I->getOperand(1));
943 // Output known-0 bits are known if clear or set in both the LHS & RHS.
944 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
945 // Output known-1 are known to be set if set in only one of the LHS, RHS.
946 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
948 // If all of the demanded bits are known to be zero on one side or the
949 // other, turn this into an *inclusive* or.
950 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
951 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
953 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
955 InsertNewInstBefore(Or, *I);
956 return UpdateValueUsesWith(I, Or);
959 // If all of the demanded bits on one side are known, and all of the set
960 // bits on that side are also known to be set on the other side, turn this
961 // into an AND, as we know the bits will be cleared.
962 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
963 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
964 if ((KnownOne & KnownOne2) == KnownOne) {
965 Constant *AndC = ConstantInt::get(I->getType(),
966 ~KnownOne & DemandedMask);
968 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
969 InsertNewInstBefore(And, *I);
970 return UpdateValueUsesWith(I, And);
974 // If the RHS is a constant, see if we can simplify it.
975 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
976 if (ShrinkDemandedConstant(I, 1, DemandedMask))
977 return UpdateValueUsesWith(I, I);
979 KnownZero = KnownZeroOut;
980 KnownOne = KnownOneOut;
983 case Instruction::Select:
984 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
985 KnownZero, KnownOne, Depth+1))
987 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
988 KnownZero2, KnownOne2, Depth+1))
990 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
991 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
993 // If the operands are constants, see if we can simplify them.
994 if (ShrinkDemandedConstant(I, 1, DemandedMask))
995 return UpdateValueUsesWith(I, I);
996 if (ShrinkDemandedConstant(I, 2, DemandedMask))
997 return UpdateValueUsesWith(I, I);
999 // Only known if known in both the LHS and RHS.
1000 KnownOne &= KnownOne2;
1001 KnownZero &= KnownZero2;
1003 case Instruction::Trunc:
1004 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1005 KnownZero, KnownOne, Depth+1))
1007 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1009 case Instruction::BitCast:
1010 if (!I->getOperand(0)->getType()->isInteger())
1013 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1014 KnownZero, KnownOne, Depth+1))
1016 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1018 case Instruction::ZExt: {
1019 // Compute the bits in the result that are not present in the input.
1020 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1021 uint64_t NotIn = ~SrcTy->getBitMask();
1022 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
1024 DemandedMask &= SrcTy->getBitMask();
1025 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1026 KnownZero, KnownOne, Depth+1))
1028 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1029 // The top bits are known to be zero.
1030 KnownZero |= NewBits;
1033 case Instruction::SExt: {
1034 // Compute the bits in the result that are not present in the input.
1035 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1036 uint64_t NotIn = ~SrcTy->getBitMask();
1037 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
1039 // Get the sign bit for the source type
1040 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1041 int64_t InputDemandedBits = DemandedMask & SrcTy->getBitMask();
1043 // If any of the sign extended bits are demanded, we know that the sign
1045 if (NewBits & DemandedMask)
1046 InputDemandedBits |= InSignBit;
1048 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1049 KnownZero, KnownOne, Depth+1))
1051 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1053 // If the sign bit of the input is known set or clear, then we know the
1054 // top bits of the result.
1056 // If the input sign bit is known zero, or if the NewBits are not demanded
1057 // convert this into a zero extension.
1058 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1059 // Convert to ZExt cast
1060 CastInst *NewCast = CastInst::create(
1061 Instruction::ZExt, I->getOperand(0), I->getType(), I->getName(), I);
1062 return UpdateValueUsesWith(I, NewCast);
1063 } else if (KnownOne & InSignBit) { // Input sign bit known set
1064 KnownOne |= NewBits;
1065 KnownZero &= ~NewBits;
1066 } else { // Input sign bit unknown
1067 KnownZero &= ~NewBits;
1068 KnownOne &= ~NewBits;
1072 case Instruction::Add:
1073 // If there is a constant on the RHS, there are a variety of xformations
1075 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1076 // If null, this should be simplified elsewhere. Some of the xforms here
1077 // won't work if the RHS is zero.
1078 if (RHS->isNullValue())
1081 // Figure out what the input bits are. If the top bits of the and result
1082 // are not demanded, then the add doesn't demand them from its input
1085 // Shift the demanded mask up so that it's at the top of the uint64_t.
1086 unsigned BitWidth = I->getType()->getPrimitiveSizeInBits();
1087 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1089 // If the top bit of the output is demanded, demand everything from the
1090 // input. Otherwise, we demand all the input bits except NLZ top bits.
1091 uint64_t InDemandedBits = ~0ULL >> (64-BitWidth+NLZ);
1093 // Find information about known zero/one bits in the input.
1094 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1095 KnownZero2, KnownOne2, Depth+1))
1098 // If the RHS of the add has bits set that can't affect the input, reduce
1100 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1101 return UpdateValueUsesWith(I, I);
1103 // Avoid excess work.
1104 if (KnownZero2 == 0 && KnownOne2 == 0)
1107 // Turn it into OR if input bits are zero.
1108 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1110 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1112 InsertNewInstBefore(Or, *I);
1113 return UpdateValueUsesWith(I, Or);
1116 // We can say something about the output known-zero and known-one bits,
1117 // depending on potential carries from the input constant and the
1118 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1119 // bits set and the RHS constant is 0x01001, then we know we have a known
1120 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1122 // To compute this, we first compute the potential carry bits. These are
1123 // the bits which may be modified. I'm not aware of a better way to do
1125 uint64_t RHSVal = RHS->getZExtValue();
1127 bool CarryIn = false;
1128 uint64_t CarryBits = 0;
1129 uint64_t CurBit = 1;
1130 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1131 // Record the current carry in.
1132 if (CarryIn) CarryBits |= CurBit;
1136 // This bit has a carry out unless it is "zero + zero" or
1137 // "zero + anything" with no carry in.
1138 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1139 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1140 } else if (!CarryIn &&
1141 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1142 CarryOut = false; // 0 + anything has no carry out if no carry in.
1144 // Otherwise, we have to assume we have a carry out.
1148 // This stage's carry out becomes the next stage's carry-in.
1152 // Now that we know which bits have carries, compute the known-1/0 sets.
1154 // Bits are known one if they are known zero in one operand and one in the
1155 // other, and there is no input carry.
1156 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1158 // Bits are known zero if they are known zero in both operands and there
1159 // is no input carry.
1160 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1163 case Instruction::Shl:
1164 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1165 uint64_t ShiftAmt = SA->getZExtValue();
1166 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1167 KnownZero, KnownOne, Depth+1))
1169 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1170 KnownZero <<= ShiftAmt;
1171 KnownOne <<= ShiftAmt;
1172 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1175 case Instruction::LShr:
1176 // For a logical shift right
1177 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1178 unsigned ShiftAmt = SA->getZExtValue();
1180 // Compute the new bits that are at the top now.
1181 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1182 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1183 uint64_t TypeMask = cast<IntegerType>(I->getType())->getBitMask();
1184 // Unsigned shift right.
1185 if (SimplifyDemandedBits(I->getOperand(0),
1186 (DemandedMask << ShiftAmt) & TypeMask,
1187 KnownZero, KnownOne, Depth+1))
1189 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1190 KnownZero &= TypeMask;
1191 KnownOne &= TypeMask;
1192 KnownZero >>= ShiftAmt;
1193 KnownOne >>= ShiftAmt;
1194 KnownZero |= HighBits; // high bits known zero.
1197 case Instruction::AShr:
1198 // If this is an arithmetic shift right and only the low-bit is set, we can
1199 // always convert this into a logical shr, even if the shift amount is
1200 // variable. The low bit of the shift cannot be an input sign bit unless
1201 // the shift amount is >= the size of the datatype, which is undefined.
1202 if (DemandedMask == 1) {
1203 // Perform the logical shift right.
1204 Value *NewVal = BinaryOperator::createLShr(
1205 I->getOperand(0), I->getOperand(1), I->getName());
1206 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1207 return UpdateValueUsesWith(I, NewVal);
1210 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1211 unsigned ShiftAmt = SA->getZExtValue();
1213 // Compute the new bits that are at the top now.
1214 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1215 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1216 uint64_t TypeMask = cast<IntegerType>(I->getType())->getBitMask();
1217 // Signed shift right.
1218 if (SimplifyDemandedBits(I->getOperand(0),
1219 (DemandedMask << ShiftAmt) & TypeMask,
1220 KnownZero, KnownOne, Depth+1))
1222 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1223 KnownZero &= TypeMask;
1224 KnownOne &= TypeMask;
1225 KnownZero >>= ShiftAmt;
1226 KnownOne >>= ShiftAmt;
1228 // Handle the sign bits.
1229 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1230 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1232 // If the input sign bit is known to be zero, or if none of the top bits
1233 // are demanded, turn this into an unsigned shift right.
1234 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1235 // Perform the logical shift right.
1236 Value *NewVal = BinaryOperator::createLShr(
1237 I->getOperand(0), SA, I->getName());
1238 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1239 return UpdateValueUsesWith(I, NewVal);
1240 } else if (KnownOne & SignBit) { // New bits are known one.
1241 KnownOne |= HighBits;
1247 // If the client is only demanding bits that we know, return the known
1249 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1250 return UpdateValueUsesWith(I, ConstantInt::get(I->getType(), KnownOne));
1255 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1256 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1257 /// actually used by the caller. This method analyzes which elements of the
1258 /// operand are undef and returns that information in UndefElts.
1260 /// If the information about demanded elements can be used to simplify the
1261 /// operation, the operation is simplified, then the resultant value is
1262 /// returned. This returns null if no change was made.
1263 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1264 uint64_t &UndefElts,
1266 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1267 assert(VWidth <= 64 && "Vector too wide to analyze!");
1268 uint64_t EltMask = ~0ULL >> (64-VWidth);
1269 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1270 "Invalid DemandedElts!");
1272 if (isa<UndefValue>(V)) {
1273 // If the entire vector is undefined, just return this info.
1274 UndefElts = EltMask;
1276 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1277 UndefElts = EltMask;
1278 return UndefValue::get(V->getType());
1282 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1283 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1284 Constant *Undef = UndefValue::get(EltTy);
1286 std::vector<Constant*> Elts;
1287 for (unsigned i = 0; i != VWidth; ++i)
1288 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1289 Elts.push_back(Undef);
1290 UndefElts |= (1ULL << i);
1291 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1292 Elts.push_back(Undef);
1293 UndefElts |= (1ULL << i);
1294 } else { // Otherwise, defined.
1295 Elts.push_back(CP->getOperand(i));
1298 // If we changed the constant, return it.
1299 Constant *NewCP = ConstantVector::get(Elts);
1300 return NewCP != CP ? NewCP : 0;
1301 } else if (isa<ConstantAggregateZero>(V)) {
1302 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1304 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1305 Constant *Zero = Constant::getNullValue(EltTy);
1306 Constant *Undef = UndefValue::get(EltTy);
1307 std::vector<Constant*> Elts;
1308 for (unsigned i = 0; i != VWidth; ++i)
1309 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1310 UndefElts = DemandedElts ^ EltMask;
1311 return ConstantVector::get(Elts);
1314 if (!V->hasOneUse()) { // Other users may use these bits.
1315 if (Depth != 0) { // Not at the root.
1316 // TODO: Just compute the UndefElts information recursively.
1320 } else if (Depth == 10) { // Limit search depth.
1324 Instruction *I = dyn_cast<Instruction>(V);
1325 if (!I) return false; // Only analyze instructions.
1327 bool MadeChange = false;
1328 uint64_t UndefElts2;
1330 switch (I->getOpcode()) {
1333 case Instruction::InsertElement: {
1334 // If this is a variable index, we don't know which element it overwrites.
1335 // demand exactly the same input as we produce.
1336 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1338 // Note that we can't propagate undef elt info, because we don't know
1339 // which elt is getting updated.
1340 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1341 UndefElts2, Depth+1);
1342 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1346 // If this is inserting an element that isn't demanded, remove this
1348 unsigned IdxNo = Idx->getZExtValue();
1349 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1350 return AddSoonDeadInstToWorklist(*I, 0);
1352 // Otherwise, the element inserted overwrites whatever was there, so the
1353 // input demanded set is simpler than the output set.
1354 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1355 DemandedElts & ~(1ULL << IdxNo),
1356 UndefElts, Depth+1);
1357 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1359 // The inserted element is defined.
1360 UndefElts |= 1ULL << IdxNo;
1364 case Instruction::And:
1365 case Instruction::Or:
1366 case Instruction::Xor:
1367 case Instruction::Add:
1368 case Instruction::Sub:
1369 case Instruction::Mul:
1370 // div/rem demand all inputs, because they don't want divide by zero.
1371 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1372 UndefElts, Depth+1);
1373 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1374 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1375 UndefElts2, Depth+1);
1376 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1378 // Output elements are undefined if both are undefined. Consider things
1379 // like undef&0. The result is known zero, not undef.
1380 UndefElts &= UndefElts2;
1383 case Instruction::Call: {
1384 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1386 switch (II->getIntrinsicID()) {
1389 // Binary vector operations that work column-wise. A dest element is a
1390 // function of the corresponding input elements from the two inputs.
1391 case Intrinsic::x86_sse_sub_ss:
1392 case Intrinsic::x86_sse_mul_ss:
1393 case Intrinsic::x86_sse_min_ss:
1394 case Intrinsic::x86_sse_max_ss:
1395 case Intrinsic::x86_sse2_sub_sd:
1396 case Intrinsic::x86_sse2_mul_sd:
1397 case Intrinsic::x86_sse2_min_sd:
1398 case Intrinsic::x86_sse2_max_sd:
1399 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1400 UndefElts, Depth+1);
1401 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1402 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1403 UndefElts2, Depth+1);
1404 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1406 // If only the low elt is demanded and this is a scalarizable intrinsic,
1407 // scalarize it now.
1408 if (DemandedElts == 1) {
1409 switch (II->getIntrinsicID()) {
1411 case Intrinsic::x86_sse_sub_ss:
1412 case Intrinsic::x86_sse_mul_ss:
1413 case Intrinsic::x86_sse2_sub_sd:
1414 case Intrinsic::x86_sse2_mul_sd:
1415 // TODO: Lower MIN/MAX/ABS/etc
1416 Value *LHS = II->getOperand(1);
1417 Value *RHS = II->getOperand(2);
1418 // Extract the element as scalars.
1419 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1420 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1422 switch (II->getIntrinsicID()) {
1423 default: assert(0 && "Case stmts out of sync!");
1424 case Intrinsic::x86_sse_sub_ss:
1425 case Intrinsic::x86_sse2_sub_sd:
1426 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1427 II->getName()), *II);
1429 case Intrinsic::x86_sse_mul_ss:
1430 case Intrinsic::x86_sse2_mul_sd:
1431 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1432 II->getName()), *II);
1437 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1439 InsertNewInstBefore(New, *II);
1440 AddSoonDeadInstToWorklist(*II, 0);
1445 // Output elements are undefined if both are undefined. Consider things
1446 // like undef&0. The result is known zero, not undef.
1447 UndefElts &= UndefElts2;
1453 return MadeChange ? I : 0;
1456 /// @returns true if the specified compare instruction is
1457 /// true when both operands are equal...
1458 /// @brief Determine if the ICmpInst returns true if both operands are equal
1459 static bool isTrueWhenEqual(ICmpInst &ICI) {
1460 ICmpInst::Predicate pred = ICI.getPredicate();
1461 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1462 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1463 pred == ICmpInst::ICMP_SLE;
1466 /// AssociativeOpt - Perform an optimization on an associative operator. This
1467 /// function is designed to check a chain of associative operators for a
1468 /// potential to apply a certain optimization. Since the optimization may be
1469 /// applicable if the expression was reassociated, this checks the chain, then
1470 /// reassociates the expression as necessary to expose the optimization
1471 /// opportunity. This makes use of a special Functor, which must define
1472 /// 'shouldApply' and 'apply' methods.
1474 template<typename Functor>
1475 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1476 unsigned Opcode = Root.getOpcode();
1477 Value *LHS = Root.getOperand(0);
1479 // Quick check, see if the immediate LHS matches...
1480 if (F.shouldApply(LHS))
1481 return F.apply(Root);
1483 // Otherwise, if the LHS is not of the same opcode as the root, return.
1484 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1485 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1486 // Should we apply this transform to the RHS?
1487 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1489 // If not to the RHS, check to see if we should apply to the LHS...
1490 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1491 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1495 // If the functor wants to apply the optimization to the RHS of LHSI,
1496 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1498 BasicBlock *BB = Root.getParent();
1500 // Now all of the instructions are in the current basic block, go ahead
1501 // and perform the reassociation.
1502 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1504 // First move the selected RHS to the LHS of the root...
1505 Root.setOperand(0, LHSI->getOperand(1));
1507 // Make what used to be the LHS of the root be the user of the root...
1508 Value *ExtraOperand = TmpLHSI->getOperand(1);
1509 if (&Root == TmpLHSI) {
1510 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1513 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1514 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1515 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1516 BasicBlock::iterator ARI = &Root; ++ARI;
1517 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1520 // Now propagate the ExtraOperand down the chain of instructions until we
1522 while (TmpLHSI != LHSI) {
1523 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1524 // Move the instruction to immediately before the chain we are
1525 // constructing to avoid breaking dominance properties.
1526 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1527 BB->getInstList().insert(ARI, NextLHSI);
1530 Value *NextOp = NextLHSI->getOperand(1);
1531 NextLHSI->setOperand(1, ExtraOperand);
1533 ExtraOperand = NextOp;
1536 // Now that the instructions are reassociated, have the functor perform
1537 // the transformation...
1538 return F.apply(Root);
1541 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1547 // AddRHS - Implements: X + X --> X << 1
1550 AddRHS(Value *rhs) : RHS(rhs) {}
1551 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1552 Instruction *apply(BinaryOperator &Add) const {
1553 return BinaryOperator::createShl(Add.getOperand(0),
1554 ConstantInt::get(Add.getType(), 1));
1558 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1560 struct AddMaskingAnd {
1562 AddMaskingAnd(Constant *c) : C2(c) {}
1563 bool shouldApply(Value *LHS) const {
1565 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1566 ConstantExpr::getAnd(C1, C2)->isNullValue();
1568 Instruction *apply(BinaryOperator &Add) const {
1569 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1573 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1575 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1576 if (Constant *SOC = dyn_cast<Constant>(SO))
1577 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1579 return IC->InsertNewInstBefore(CastInst::create(
1580 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1583 // Figure out if the constant is the left or the right argument.
1584 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1585 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1587 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1589 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1590 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1593 Value *Op0 = SO, *Op1 = ConstOperand;
1595 std::swap(Op0, Op1);
1597 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1598 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1599 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1600 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1601 SO->getName()+".cmp");
1603 assert(0 && "Unknown binary instruction type!");
1606 return IC->InsertNewInstBefore(New, I);
1609 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1610 // constant as the other operand, try to fold the binary operator into the
1611 // select arguments. This also works for Cast instructions, which obviously do
1612 // not have a second operand.
1613 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1615 // Don't modify shared select instructions
1616 if (!SI->hasOneUse()) return 0;
1617 Value *TV = SI->getOperand(1);
1618 Value *FV = SI->getOperand(2);
1620 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1621 // Bool selects with constant operands can be folded to logical ops.
1622 if (SI->getType() == Type::Int1Ty) return 0;
1624 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1625 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1627 return new SelectInst(SI->getCondition(), SelectTrueVal,
1634 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1635 /// node as operand #0, see if we can fold the instruction into the PHI (which
1636 /// is only possible if all operands to the PHI are constants).
1637 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1638 PHINode *PN = cast<PHINode>(I.getOperand(0));
1639 unsigned NumPHIValues = PN->getNumIncomingValues();
1640 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1642 // Check to see if all of the operands of the PHI are constants. If there is
1643 // one non-constant value, remember the BB it is. If there is more than one
1644 // or if *it* is a PHI, bail out.
1645 BasicBlock *NonConstBB = 0;
1646 for (unsigned i = 0; i != NumPHIValues; ++i)
1647 if (!isa<Constant>(PN->getIncomingValue(i))) {
1648 if (NonConstBB) return 0; // More than one non-const value.
1649 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1650 NonConstBB = PN->getIncomingBlock(i);
1652 // If the incoming non-constant value is in I's block, we have an infinite
1654 if (NonConstBB == I.getParent())
1658 // If there is exactly one non-constant value, we can insert a copy of the
1659 // operation in that block. However, if this is a critical edge, we would be
1660 // inserting the computation one some other paths (e.g. inside a loop). Only
1661 // do this if the pred block is unconditionally branching into the phi block.
1663 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1664 if (!BI || !BI->isUnconditional()) return 0;
1667 // Okay, we can do the transformation: create the new PHI node.
1668 PHINode *NewPN = new PHINode(I.getType(), "");
1669 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1670 InsertNewInstBefore(NewPN, *PN);
1671 NewPN->takeName(PN);
1673 // Next, add all of the operands to the PHI.
1674 if (I.getNumOperands() == 2) {
1675 Constant *C = cast<Constant>(I.getOperand(1));
1676 for (unsigned i = 0; i != NumPHIValues; ++i) {
1678 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1679 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1680 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1682 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1684 assert(PN->getIncomingBlock(i) == NonConstBB);
1685 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1686 InV = BinaryOperator::create(BO->getOpcode(),
1687 PN->getIncomingValue(i), C, "phitmp",
1688 NonConstBB->getTerminator());
1689 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1690 InV = CmpInst::create(CI->getOpcode(),
1692 PN->getIncomingValue(i), C, "phitmp",
1693 NonConstBB->getTerminator());
1695 assert(0 && "Unknown binop!");
1697 WorkList.push_back(cast<Instruction>(InV));
1699 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1702 CastInst *CI = cast<CastInst>(&I);
1703 const Type *RetTy = CI->getType();
1704 for (unsigned i = 0; i != NumPHIValues; ++i) {
1706 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1707 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1709 assert(PN->getIncomingBlock(i) == NonConstBB);
1710 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1711 I.getType(), "phitmp",
1712 NonConstBB->getTerminator());
1713 WorkList.push_back(cast<Instruction>(InV));
1715 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1718 return ReplaceInstUsesWith(I, NewPN);
1721 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1722 bool Changed = SimplifyCommutative(I);
1723 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1725 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1726 // X + undef -> undef
1727 if (isa<UndefValue>(RHS))
1728 return ReplaceInstUsesWith(I, RHS);
1731 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1732 if (RHSC->isNullValue())
1733 return ReplaceInstUsesWith(I, LHS);
1734 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1735 if (CFP->isExactlyValue(-0.0))
1736 return ReplaceInstUsesWith(I, LHS);
1739 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1740 // X + (signbit) --> X ^ signbit
1741 uint64_t Val = CI->getZExtValue();
1742 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1743 return BinaryOperator::createXor(LHS, RHS);
1745 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1746 // (X & 254)+1 -> (X&254)|1
1747 uint64_t KnownZero, KnownOne;
1748 if (!isa<VectorType>(I.getType()) &&
1749 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
1750 KnownZero, KnownOne))
1754 if (isa<PHINode>(LHS))
1755 if (Instruction *NV = FoldOpIntoPhi(I))
1758 ConstantInt *XorRHS = 0;
1760 if (isa<ConstantInt>(RHSC) &&
1761 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1762 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1763 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1764 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1766 uint64_t C0080Val = 1ULL << 31;
1767 int64_t CFF80Val = -C0080Val;
1770 if (TySizeBits > Size) {
1772 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1773 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1774 if (RHSSExt == CFF80Val) {
1775 if (XorRHS->getZExtValue() == C0080Val)
1777 } else if (RHSZExt == C0080Val) {
1778 if (XorRHS->getSExtValue() == CFF80Val)
1782 // This is a sign extend if the top bits are known zero.
1783 uint64_t Mask = ~0ULL;
1784 Mask <<= 64-(TySizeBits-Size);
1785 Mask &= cast<IntegerType>(XorLHS->getType())->getBitMask();
1786 if (!MaskedValueIsZero(XorLHS, Mask))
1787 Size = 0; // Not a sign ext, but can't be any others either.
1794 } while (Size >= 8);
1797 const Type *MiddleType = 0;
1800 case 32: MiddleType = Type::Int32Ty; break;
1801 case 16: MiddleType = Type::Int16Ty; break;
1802 case 8: MiddleType = Type::Int8Ty; break;
1805 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1806 InsertNewInstBefore(NewTrunc, I);
1807 return new SExtInst(NewTrunc, I.getType());
1813 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
1814 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1816 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1817 if (RHSI->getOpcode() == Instruction::Sub)
1818 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1819 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1821 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1822 if (LHSI->getOpcode() == Instruction::Sub)
1823 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1824 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1829 if (Value *V = dyn_castNegVal(LHS))
1830 return BinaryOperator::createSub(RHS, V);
1833 if (!isa<Constant>(RHS))
1834 if (Value *V = dyn_castNegVal(RHS))
1835 return BinaryOperator::createSub(LHS, V);
1839 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1840 if (X == RHS) // X*C + X --> X * (C+1)
1841 return BinaryOperator::createMul(RHS, AddOne(C2));
1843 // X*C1 + X*C2 --> X * (C1+C2)
1845 if (X == dyn_castFoldableMul(RHS, C1))
1846 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1849 // X + X*C --> X * (C+1)
1850 if (dyn_castFoldableMul(RHS, C2) == LHS)
1851 return BinaryOperator::createMul(LHS, AddOne(C2));
1853 // X + ~X --> -1 since ~X = -X-1
1854 if (dyn_castNotVal(LHS) == RHS ||
1855 dyn_castNotVal(RHS) == LHS)
1856 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
1859 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1860 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1861 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
1864 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1866 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1867 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1868 return BinaryOperator::createSub(C, X);
1871 // (X & FF00) + xx00 -> (X+xx00) & FF00
1872 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1873 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1874 if (Anded == CRHS) {
1875 // See if all bits from the first bit set in the Add RHS up are included
1876 // in the mask. First, get the rightmost bit.
1877 uint64_t AddRHSV = CRHS->getZExtValue();
1879 // Form a mask of all bits from the lowest bit added through the top.
1880 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1881 AddRHSHighBits &= C2->getType()->getBitMask();
1883 // See if the and mask includes all of these bits.
1884 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1886 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1887 // Okay, the xform is safe. Insert the new add pronto.
1888 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1889 LHS->getName()), I);
1890 return BinaryOperator::createAnd(NewAdd, C2);
1895 // Try to fold constant add into select arguments.
1896 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1897 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1901 // add (cast *A to intptrtype) B ->
1902 // cast (GEP (cast *A to sbyte*) B) ->
1905 CastInst *CI = dyn_cast<CastInst>(LHS);
1908 CI = dyn_cast<CastInst>(RHS);
1911 if (CI && CI->getType()->isSized() &&
1912 (CI->getType()->getPrimitiveSizeInBits() ==
1913 TD->getIntPtrType()->getPrimitiveSizeInBits())
1914 && isa<PointerType>(CI->getOperand(0)->getType())) {
1915 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
1916 PointerType::get(Type::Int8Ty), I);
1917 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1918 return new PtrToIntInst(I2, CI->getType());
1922 return Changed ? &I : 0;
1925 // isSignBit - Return true if the value represented by the constant only has the
1926 // highest order bit set.
1927 static bool isSignBit(ConstantInt *CI) {
1928 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1929 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1932 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1933 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1935 if (Op0 == Op1) // sub X, X -> 0
1936 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1938 // If this is a 'B = x-(-A)', change to B = x+A...
1939 if (Value *V = dyn_castNegVal(Op1))
1940 return BinaryOperator::createAdd(Op0, V);
1942 if (isa<UndefValue>(Op0))
1943 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1944 if (isa<UndefValue>(Op1))
1945 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1947 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1948 // Replace (-1 - A) with (~A)...
1949 if (C->isAllOnesValue())
1950 return BinaryOperator::createNot(Op1);
1952 // C - ~X == X + (1+C)
1954 if (match(Op1, m_Not(m_Value(X))))
1955 return BinaryOperator::createAdd(X,
1956 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1957 // -(X >>u 31) -> (X >>s 31)
1958 // -(X >>s 31) -> (X >>u 31)
1959 if (C->isNullValue()) {
1960 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
1961 if (SI->getOpcode() == Instruction::LShr) {
1962 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1963 // Check to see if we are shifting out everything but the sign bit.
1964 if (CU->getZExtValue() ==
1965 SI->getType()->getPrimitiveSizeInBits()-1) {
1966 // Ok, the transformation is safe. Insert AShr.
1967 return BinaryOperator::create(Instruction::AShr,
1968 SI->getOperand(0), CU, SI->getName());
1972 else if (SI->getOpcode() == Instruction::AShr) {
1973 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1974 // Check to see if we are shifting out everything but the sign bit.
1975 if (CU->getZExtValue() ==
1976 SI->getType()->getPrimitiveSizeInBits()-1) {
1977 // Ok, the transformation is safe. Insert LShr.
1978 return BinaryOperator::createLShr(
1979 SI->getOperand(0), CU, SI->getName());
1985 // Try to fold constant sub into select arguments.
1986 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1987 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1990 if (isa<PHINode>(Op0))
1991 if (Instruction *NV = FoldOpIntoPhi(I))
1995 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1996 if (Op1I->getOpcode() == Instruction::Add &&
1997 !Op0->getType()->isFPOrFPVector()) {
1998 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1999 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2000 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2001 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2002 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2003 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2004 // C1-(X+C2) --> (C1-C2)-X
2005 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2006 Op1I->getOperand(0));
2010 if (Op1I->hasOneUse()) {
2011 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2012 // is not used by anyone else...
2014 if (Op1I->getOpcode() == Instruction::Sub &&
2015 !Op1I->getType()->isFPOrFPVector()) {
2016 // Swap the two operands of the subexpr...
2017 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2018 Op1I->setOperand(0, IIOp1);
2019 Op1I->setOperand(1, IIOp0);
2021 // Create the new top level add instruction...
2022 return BinaryOperator::createAdd(Op0, Op1);
2025 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2027 if (Op1I->getOpcode() == Instruction::And &&
2028 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2029 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2032 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2033 return BinaryOperator::createAnd(Op0, NewNot);
2036 // 0 - (X sdiv C) -> (X sdiv -C)
2037 if (Op1I->getOpcode() == Instruction::SDiv)
2038 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2039 if (CSI->isNullValue())
2040 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2041 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2042 ConstantExpr::getNeg(DivRHS));
2044 // X - X*C --> X * (1-C)
2045 ConstantInt *C2 = 0;
2046 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2048 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2049 return BinaryOperator::createMul(Op0, CP1);
2054 if (!Op0->getType()->isFPOrFPVector())
2055 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2056 if (Op0I->getOpcode() == Instruction::Add) {
2057 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2058 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2059 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2060 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2061 } else if (Op0I->getOpcode() == Instruction::Sub) {
2062 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2063 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2067 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2068 if (X == Op1) { // X*C - X --> X * (C-1)
2069 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2070 return BinaryOperator::createMul(Op1, CP1);
2073 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2074 if (X == dyn_castFoldableMul(Op1, C2))
2075 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2080 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2081 /// really just returns true if the most significant (sign) bit is set.
2082 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2084 case ICmpInst::ICMP_SLT:
2085 // True if LHS s< RHS and RHS == 0
2086 return RHS->isNullValue();
2087 case ICmpInst::ICMP_SLE:
2088 // True if LHS s<= RHS and RHS == -1
2089 return RHS->isAllOnesValue();
2090 case ICmpInst::ICMP_UGE:
2091 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2092 return RHS->getZExtValue() == (1ULL <<
2093 (RHS->getType()->getPrimitiveSizeInBits()-1));
2094 case ICmpInst::ICMP_UGT:
2095 // True if LHS u> RHS and RHS == high-bit-mask - 1
2096 return RHS->getZExtValue() ==
2097 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2103 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2104 bool Changed = SimplifyCommutative(I);
2105 Value *Op0 = I.getOperand(0);
2107 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2108 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2110 // Simplify mul instructions with a constant RHS...
2111 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2112 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2114 // ((X << C1)*C2) == (X * (C2 << C1))
2115 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2116 if (SI->getOpcode() == Instruction::Shl)
2117 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2118 return BinaryOperator::createMul(SI->getOperand(0),
2119 ConstantExpr::getShl(CI, ShOp));
2121 if (CI->isNullValue())
2122 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2123 if (CI->equalsInt(1)) // X * 1 == X
2124 return ReplaceInstUsesWith(I, Op0);
2125 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2126 return BinaryOperator::createNeg(Op0, I.getName());
2128 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2129 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2130 uint64_t C = Log2_64(Val);
2131 return BinaryOperator::createShl(Op0,
2132 ConstantInt::get(Op0->getType(), C));
2134 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2135 if (Op1F->isNullValue())
2136 return ReplaceInstUsesWith(I, Op1);
2138 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2139 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2140 if (Op1F->getValue() == 1.0)
2141 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2144 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2145 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2146 isa<ConstantInt>(Op0I->getOperand(1))) {
2147 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2148 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2150 InsertNewInstBefore(Add, I);
2151 Value *C1C2 = ConstantExpr::getMul(Op1,
2152 cast<Constant>(Op0I->getOperand(1)));
2153 return BinaryOperator::createAdd(Add, C1C2);
2157 // Try to fold constant mul into select arguments.
2158 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2159 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2162 if (isa<PHINode>(Op0))
2163 if (Instruction *NV = FoldOpIntoPhi(I))
2167 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2168 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2169 return BinaryOperator::createMul(Op0v, Op1v);
2171 // If one of the operands of the multiply is a cast from a boolean value, then
2172 // we know the bool is either zero or one, so this is a 'masking' multiply.
2173 // See if we can simplify things based on how the boolean was originally
2175 CastInst *BoolCast = 0;
2176 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2177 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2180 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2181 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2184 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2185 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2186 const Type *SCOpTy = SCIOp0->getType();
2188 // If the icmp is true iff the sign bit of X is set, then convert this
2189 // multiply into a shift/and combination.
2190 if (isa<ConstantInt>(SCIOp1) &&
2191 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2192 // Shift the X value right to turn it into "all signbits".
2193 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2194 SCOpTy->getPrimitiveSizeInBits()-1);
2196 InsertNewInstBefore(
2197 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2198 BoolCast->getOperand(0)->getName()+
2201 // If the multiply type is not the same as the source type, sign extend
2202 // or truncate to the multiply type.
2203 if (I.getType() != V->getType()) {
2204 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2205 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2206 Instruction::CastOps opcode =
2207 (SrcBits == DstBits ? Instruction::BitCast :
2208 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2209 V = InsertCastBefore(opcode, V, I.getType(), I);
2212 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2213 return BinaryOperator::createAnd(V, OtherOp);
2218 return Changed ? &I : 0;
2221 /// This function implements the transforms on div instructions that work
2222 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2223 /// used by the visitors to those instructions.
2224 /// @brief Transforms common to all three div instructions
2225 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2226 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2229 if (isa<UndefValue>(Op0))
2230 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2232 // X / undef -> undef
2233 if (isa<UndefValue>(Op1))
2234 return ReplaceInstUsesWith(I, Op1);
2236 // Handle cases involving: div X, (select Cond, Y, Z)
2237 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2238 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2239 // same basic block, then we replace the select with Y, and the condition
2240 // of the select with false (if the cond value is in the same BB). If the
2241 // select has uses other than the div, this allows them to be simplified
2242 // also. Note that div X, Y is just as good as div X, 0 (undef)
2243 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2244 if (ST->isNullValue()) {
2245 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2246 if (CondI && CondI->getParent() == I.getParent())
2247 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2248 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2249 I.setOperand(1, SI->getOperand(2));
2251 UpdateValueUsesWith(SI, SI->getOperand(2));
2255 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2256 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2257 if (ST->isNullValue()) {
2258 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2259 if (CondI && CondI->getParent() == I.getParent())
2260 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2261 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2262 I.setOperand(1, SI->getOperand(1));
2264 UpdateValueUsesWith(SI, SI->getOperand(1));
2272 /// This function implements the transforms common to both integer division
2273 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2274 /// division instructions.
2275 /// @brief Common integer divide transforms
2276 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2277 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2279 if (Instruction *Common = commonDivTransforms(I))
2282 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2284 if (RHS->equalsInt(1))
2285 return ReplaceInstUsesWith(I, Op0);
2287 // (X / C1) / C2 -> X / (C1*C2)
2288 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2289 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2290 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2291 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2292 ConstantExpr::getMul(RHS, LHSRHS));
2295 if (!RHS->isNullValue()) { // avoid X udiv 0
2296 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2297 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2299 if (isa<PHINode>(Op0))
2300 if (Instruction *NV = FoldOpIntoPhi(I))
2305 // 0 / X == 0, we don't need to preserve faults!
2306 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2307 if (LHS->equalsInt(0))
2308 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2313 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2314 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2316 // Handle the integer div common cases
2317 if (Instruction *Common = commonIDivTransforms(I))
2320 // X udiv C^2 -> X >> C
2321 // Check to see if this is an unsigned division with an exact power of 2,
2322 // if so, convert to a right shift.
2323 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2324 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2325 if (isPowerOf2_64(Val)) {
2326 uint64_t ShiftAmt = Log2_64(Val);
2327 return BinaryOperator::createLShr(Op0,
2328 ConstantInt::get(Op0->getType(), ShiftAmt));
2332 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2333 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2334 if (RHSI->getOpcode() == Instruction::Shl &&
2335 isa<ConstantInt>(RHSI->getOperand(0))) {
2336 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2337 if (isPowerOf2_64(C1)) {
2338 Value *N = RHSI->getOperand(1);
2339 const Type *NTy = N->getType();
2340 if (uint64_t C2 = Log2_64(C1)) {
2341 Constant *C2V = ConstantInt::get(NTy, C2);
2342 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2344 return BinaryOperator::createLShr(Op0, N);
2349 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2350 // where C1&C2 are powers of two.
2351 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2352 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2353 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2354 if (!STO->isNullValue() && !STO->isNullValue()) {
2355 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2356 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2357 // Compute the shift amounts
2358 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2359 // Construct the "on true" case of the select
2360 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2361 Instruction *TSI = BinaryOperator::createLShr(
2362 Op0, TC, SI->getName()+".t");
2363 TSI = InsertNewInstBefore(TSI, I);
2365 // Construct the "on false" case of the select
2366 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2367 Instruction *FSI = BinaryOperator::createLShr(
2368 Op0, FC, SI->getName()+".f");
2369 FSI = InsertNewInstBefore(FSI, I);
2371 // construct the select instruction and return it.
2372 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2379 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2380 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2382 // Handle the integer div common cases
2383 if (Instruction *Common = commonIDivTransforms(I))
2386 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2388 if (RHS->isAllOnesValue())
2389 return BinaryOperator::createNeg(Op0);
2392 if (Value *LHSNeg = dyn_castNegVal(Op0))
2393 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2396 // If the sign bits of both operands are zero (i.e. we can prove they are
2397 // unsigned inputs), turn this into a udiv.
2398 if (I.getType()->isInteger()) {
2399 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2400 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2401 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2408 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2409 return commonDivTransforms(I);
2412 /// GetFactor - If we can prove that the specified value is at least a multiple
2413 /// of some factor, return that factor.
2414 static Constant *GetFactor(Value *V) {
2415 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2418 // Unless we can be tricky, we know this is a multiple of 1.
2419 Constant *Result = ConstantInt::get(V->getType(), 1);
2421 Instruction *I = dyn_cast<Instruction>(V);
2422 if (!I) return Result;
2424 if (I->getOpcode() == Instruction::Mul) {
2425 // Handle multiplies by a constant, etc.
2426 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2427 GetFactor(I->getOperand(1)));
2428 } else if (I->getOpcode() == Instruction::Shl) {
2429 // (X<<C) -> X * (1 << C)
2430 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2431 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2432 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2434 } else if (I->getOpcode() == Instruction::And) {
2435 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2436 // X & 0xFFF0 is known to be a multiple of 16.
2437 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2438 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2439 return ConstantExpr::getShl(Result,
2440 ConstantInt::get(Result->getType(), Zeros));
2442 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2443 // Only handle int->int casts.
2444 if (!CI->isIntegerCast())
2446 Value *Op = CI->getOperand(0);
2447 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2452 /// This function implements the transforms on rem instructions that work
2453 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2454 /// is used by the visitors to those instructions.
2455 /// @brief Transforms common to all three rem instructions
2456 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2457 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2459 // 0 % X == 0, we don't need to preserve faults!
2460 if (Constant *LHS = dyn_cast<Constant>(Op0))
2461 if (LHS->isNullValue())
2462 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2464 if (isa<UndefValue>(Op0)) // undef % X -> 0
2465 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2466 if (isa<UndefValue>(Op1))
2467 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2469 // Handle cases involving: rem X, (select Cond, Y, Z)
2470 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2471 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2472 // the same basic block, then we replace the select with Y, and the
2473 // condition of the select with false (if the cond value is in the same
2474 // BB). If the select has uses other than the div, this allows them to be
2476 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2477 if (ST->isNullValue()) {
2478 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2479 if (CondI && CondI->getParent() == I.getParent())
2480 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2481 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2482 I.setOperand(1, SI->getOperand(2));
2484 UpdateValueUsesWith(SI, SI->getOperand(2));
2487 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2488 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2489 if (ST->isNullValue()) {
2490 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2491 if (CondI && CondI->getParent() == I.getParent())
2492 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2493 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2494 I.setOperand(1, SI->getOperand(1));
2496 UpdateValueUsesWith(SI, SI->getOperand(1));
2504 /// This function implements the transforms common to both integer remainder
2505 /// instructions (urem and srem). It is called by the visitors to those integer
2506 /// remainder instructions.
2507 /// @brief Common integer remainder transforms
2508 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2509 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2511 if (Instruction *common = commonRemTransforms(I))
2514 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2515 // X % 0 == undef, we don't need to preserve faults!
2516 if (RHS->equalsInt(0))
2517 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2519 if (RHS->equalsInt(1)) // X % 1 == 0
2520 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2522 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2523 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2524 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2526 } else if (isa<PHINode>(Op0I)) {
2527 if (Instruction *NV = FoldOpIntoPhi(I))
2530 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2531 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2532 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2539 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2540 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2542 if (Instruction *common = commonIRemTransforms(I))
2545 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2546 // X urem C^2 -> X and C
2547 // Check to see if this is an unsigned remainder with an exact power of 2,
2548 // if so, convert to a bitwise and.
2549 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2550 if (isPowerOf2_64(C->getZExtValue()))
2551 return BinaryOperator::createAnd(Op0, SubOne(C));
2554 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2555 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2556 if (RHSI->getOpcode() == Instruction::Shl &&
2557 isa<ConstantInt>(RHSI->getOperand(0))) {
2558 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2559 if (isPowerOf2_64(C1)) {
2560 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2561 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2563 return BinaryOperator::createAnd(Op0, Add);
2568 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2569 // where C1&C2 are powers of two.
2570 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2571 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2572 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2573 // STO == 0 and SFO == 0 handled above.
2574 if (isPowerOf2_64(STO->getZExtValue()) &&
2575 isPowerOf2_64(SFO->getZExtValue())) {
2576 Value *TrueAnd = InsertNewInstBefore(
2577 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2578 Value *FalseAnd = InsertNewInstBefore(
2579 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2580 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2588 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2589 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2591 if (Instruction *common = commonIRemTransforms(I))
2594 if (Value *RHSNeg = dyn_castNegVal(Op1))
2595 if (!isa<ConstantInt>(RHSNeg) ||
2596 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2598 AddUsesToWorkList(I);
2599 I.setOperand(1, RHSNeg);
2603 // If the top bits of both operands are zero (i.e. we can prove they are
2604 // unsigned inputs), turn this into a urem.
2605 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2606 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2607 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2608 return BinaryOperator::createURem(Op0, Op1, I.getName());
2614 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2615 return commonRemTransforms(I);
2618 // isMaxValueMinusOne - return true if this is Max-1
2619 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2621 // Calculate 0111111111..11111
2622 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2623 int64_t Val = INT64_MAX; // All ones
2624 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2625 return C->getSExtValue() == Val-1;
2627 return C->getZExtValue() == C->getType()->getBitMask()-1;
2630 // isMinValuePlusOne - return true if this is Min+1
2631 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2633 // Calculate 1111111111000000000000
2634 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2635 int64_t Val = -1; // All ones
2636 Val <<= TypeBits-1; // Shift over to the right spot
2637 return C->getSExtValue() == Val+1;
2639 return C->getZExtValue() == 1; // unsigned
2642 // isOneBitSet - Return true if there is exactly one bit set in the specified
2644 static bool isOneBitSet(const ConstantInt *CI) {
2645 uint64_t V = CI->getZExtValue();
2646 return V && (V & (V-1)) == 0;
2649 #if 0 // Currently unused
2650 // isLowOnes - Return true if the constant is of the form 0+1+.
2651 static bool isLowOnes(const ConstantInt *CI) {
2652 uint64_t V = CI->getZExtValue();
2654 // There won't be bits set in parts that the type doesn't contain.
2655 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2657 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2658 return U && V && (U & V) == 0;
2662 // isHighOnes - Return true if the constant is of the form 1+0+.
2663 // This is the same as lowones(~X).
2664 static bool isHighOnes(const ConstantInt *CI) {
2665 uint64_t V = ~CI->getZExtValue();
2666 if (~V == 0) return false; // 0's does not match "1+"
2668 // There won't be bits set in parts that the type doesn't contain.
2669 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2671 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2672 return U && V && (U & V) == 0;
2675 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2676 /// are carefully arranged to allow folding of expressions such as:
2678 /// (A < B) | (A > B) --> (A != B)
2680 /// Note that this is only valid if the first and second predicates have the
2681 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2683 /// Three bits are used to represent the condition, as follows:
2688 /// <=> Value Definition
2689 /// 000 0 Always false
2696 /// 111 7 Always true
2698 static unsigned getICmpCode(const ICmpInst *ICI) {
2699 switch (ICI->getPredicate()) {
2701 case ICmpInst::ICMP_UGT: return 1; // 001
2702 case ICmpInst::ICMP_SGT: return 1; // 001
2703 case ICmpInst::ICMP_EQ: return 2; // 010
2704 case ICmpInst::ICMP_UGE: return 3; // 011
2705 case ICmpInst::ICMP_SGE: return 3; // 011
2706 case ICmpInst::ICMP_ULT: return 4; // 100
2707 case ICmpInst::ICMP_SLT: return 4; // 100
2708 case ICmpInst::ICMP_NE: return 5; // 101
2709 case ICmpInst::ICMP_ULE: return 6; // 110
2710 case ICmpInst::ICMP_SLE: return 6; // 110
2713 assert(0 && "Invalid ICmp predicate!");
2718 /// getICmpValue - This is the complement of getICmpCode, which turns an
2719 /// opcode and two operands into either a constant true or false, or a brand
2720 /// new /// ICmp instruction. The sign is passed in to determine which kind
2721 /// of predicate to use in new icmp instructions.
2722 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2724 default: assert(0 && "Illegal ICmp code!");
2725 case 0: return ConstantInt::getFalse();
2728 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2730 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2731 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2734 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2736 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2739 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2741 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2742 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2745 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2747 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2748 case 7: return ConstantInt::getTrue();
2752 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2753 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2754 (ICmpInst::isSignedPredicate(p1) &&
2755 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2756 (ICmpInst::isSignedPredicate(p2) &&
2757 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2761 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2762 struct FoldICmpLogical {
2765 ICmpInst::Predicate pred;
2766 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2767 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2768 pred(ICI->getPredicate()) {}
2769 bool shouldApply(Value *V) const {
2770 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2771 if (PredicatesFoldable(pred, ICI->getPredicate()))
2772 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2773 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2776 Instruction *apply(Instruction &Log) const {
2777 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2778 if (ICI->getOperand(0) != LHS) {
2779 assert(ICI->getOperand(1) == LHS);
2780 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2783 unsigned LHSCode = getICmpCode(ICI);
2784 unsigned RHSCode = getICmpCode(cast<ICmpInst>(Log.getOperand(1)));
2786 switch (Log.getOpcode()) {
2787 case Instruction::And: Code = LHSCode & RHSCode; break;
2788 case Instruction::Or: Code = LHSCode | RHSCode; break;
2789 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2790 default: assert(0 && "Illegal logical opcode!"); return 0;
2793 Value *RV = getICmpValue(ICmpInst::isSignedPredicate(pred), Code, LHS, RHS);
2794 if (Instruction *I = dyn_cast<Instruction>(RV))
2796 // Otherwise, it's a constant boolean value...
2797 return IC.ReplaceInstUsesWith(Log, RV);
2800 } // end anonymous namespace
2802 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2803 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2804 // guaranteed to be a binary operator.
2805 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2807 ConstantInt *AndRHS,
2808 BinaryOperator &TheAnd) {
2809 Value *X = Op->getOperand(0);
2810 Constant *Together = 0;
2812 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2814 switch (Op->getOpcode()) {
2815 case Instruction::Xor:
2816 if (Op->hasOneUse()) {
2817 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2818 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2819 InsertNewInstBefore(And, TheAnd);
2821 return BinaryOperator::createXor(And, Together);
2824 case Instruction::Or:
2825 if (Together == AndRHS) // (X | C) & C --> C
2826 return ReplaceInstUsesWith(TheAnd, AndRHS);
2828 if (Op->hasOneUse() && Together != OpRHS) {
2829 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2830 Instruction *Or = BinaryOperator::createOr(X, Together);
2831 InsertNewInstBefore(Or, TheAnd);
2833 return BinaryOperator::createAnd(Or, AndRHS);
2836 case Instruction::Add:
2837 if (Op->hasOneUse()) {
2838 // Adding a one to a single bit bit-field should be turned into an XOR
2839 // of the bit. First thing to check is to see if this AND is with a
2840 // single bit constant.
2841 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2843 // Clear bits that are not part of the constant.
2844 AndRHSV &= AndRHS->getType()->getBitMask();
2846 // If there is only one bit set...
2847 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2848 // Ok, at this point, we know that we are masking the result of the
2849 // ADD down to exactly one bit. If the constant we are adding has
2850 // no bits set below this bit, then we can eliminate the ADD.
2851 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2853 // Check to see if any bits below the one bit set in AndRHSV are set.
2854 if ((AddRHS & (AndRHSV-1)) == 0) {
2855 // If not, the only thing that can effect the output of the AND is
2856 // the bit specified by AndRHSV. If that bit is set, the effect of
2857 // the XOR is to toggle the bit. If it is clear, then the ADD has
2859 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2860 TheAnd.setOperand(0, X);
2863 // Pull the XOR out of the AND.
2864 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
2865 InsertNewInstBefore(NewAnd, TheAnd);
2866 NewAnd->takeName(Op);
2867 return BinaryOperator::createXor(NewAnd, AndRHS);
2874 case Instruction::Shl: {
2875 // We know that the AND will not produce any of the bits shifted in, so if
2876 // the anded constant includes them, clear them now!
2878 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2879 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2880 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2882 if (CI == ShlMask) { // Masking out bits that the shift already masks
2883 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2884 } else if (CI != AndRHS) { // Reducing bits set in and.
2885 TheAnd.setOperand(1, CI);
2890 case Instruction::LShr:
2892 // We know that the AND will not produce any of the bits shifted in, so if
2893 // the anded constant includes them, clear them now! This only applies to
2894 // unsigned shifts, because a signed shr may bring in set bits!
2896 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2897 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2898 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2900 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2901 return ReplaceInstUsesWith(TheAnd, Op);
2902 } else if (CI != AndRHS) {
2903 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2908 case Instruction::AShr:
2910 // See if this is shifting in some sign extension, then masking it out
2912 if (Op->hasOneUse()) {
2913 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2914 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2915 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
2916 if (C == AndRHS) { // Masking out bits shifted in.
2917 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2918 // Make the argument unsigned.
2919 Value *ShVal = Op->getOperand(0);
2920 ShVal = InsertNewInstBefore(
2921 BinaryOperator::createLShr(ShVal, OpRHS,
2922 Op->getName()), TheAnd);
2923 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
2932 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2933 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2934 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
2935 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
2936 /// insert new instructions.
2937 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2938 bool isSigned, bool Inside,
2940 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
2941 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
2942 "Lo is not <= Hi in range emission code!");
2945 if (Lo == Hi) // Trivially false.
2946 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
2948 // V >= Min && V < Hi --> V < Hi
2949 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2950 ICmpInst::Predicate pred = (isSigned ?
2951 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
2952 return new ICmpInst(pred, V, Hi);
2955 // Emit V-Lo <u Hi-Lo
2956 Constant *NegLo = ConstantExpr::getNeg(Lo);
2957 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2958 InsertNewInstBefore(Add, IB);
2959 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
2960 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
2963 if (Lo == Hi) // Trivially true.
2964 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
2966 // V < Min || V >= Hi ->'V > Hi-1'
2967 Hi = SubOne(cast<ConstantInt>(Hi));
2968 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2969 ICmpInst::Predicate pred = (isSigned ?
2970 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
2971 return new ICmpInst(pred, V, Hi);
2974 // Emit V-Lo > Hi-1-Lo
2975 Constant *NegLo = ConstantExpr::getNeg(Lo);
2976 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2977 InsertNewInstBefore(Add, IB);
2978 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
2979 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
2982 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2983 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2984 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2985 // not, since all 1s are not contiguous.
2986 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
2987 uint64_t V = Val->getZExtValue();
2988 if (!isShiftedMask_64(V)) return false;
2990 // look for the first zero bit after the run of ones
2991 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
2992 // look for the first non-zero bit
2993 ME = 64-CountLeadingZeros_64(V);
2999 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3000 /// where isSub determines whether the operator is a sub. If we can fold one of
3001 /// the following xforms:
3003 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3004 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3005 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3007 /// return (A +/- B).
3009 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3010 ConstantInt *Mask, bool isSub,
3012 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3013 if (!LHSI || LHSI->getNumOperands() != 2 ||
3014 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3016 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3018 switch (LHSI->getOpcode()) {
3020 case Instruction::And:
3021 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3022 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3023 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3026 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3027 // part, we don't need any explicit masks to take them out of A. If that
3028 // is all N is, ignore it.
3030 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3031 uint64_t Mask = cast<IntegerType>(RHS->getType())->getBitMask();
3033 if (MaskedValueIsZero(RHS, Mask))
3038 case Instruction::Or:
3039 case Instruction::Xor:
3040 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3041 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3042 ConstantExpr::getAnd(N, Mask)->isNullValue())
3049 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3051 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3052 return InsertNewInstBefore(New, I);
3055 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3056 bool Changed = SimplifyCommutative(I);
3057 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3059 if (isa<UndefValue>(Op1)) // X & undef -> 0
3060 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3064 return ReplaceInstUsesWith(I, Op1);
3066 // See if we can simplify any instructions used by the instruction whose sole
3067 // purpose is to compute bits we don't care about.
3068 uint64_t KnownZero, KnownOne;
3069 if (!isa<VectorType>(I.getType())) {
3070 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3071 KnownZero, KnownOne))
3074 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3075 if (CP->isAllOnesValue())
3076 return ReplaceInstUsesWith(I, I.getOperand(0));
3080 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3081 uint64_t AndRHSMask = AndRHS->getZExtValue();
3082 uint64_t TypeMask = cast<IntegerType>(Op0->getType())->getBitMask();
3083 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3085 // Optimize a variety of ((val OP C1) & C2) combinations...
3086 if (isa<BinaryOperator>(Op0)) {
3087 Instruction *Op0I = cast<Instruction>(Op0);
3088 Value *Op0LHS = Op0I->getOperand(0);
3089 Value *Op0RHS = Op0I->getOperand(1);
3090 switch (Op0I->getOpcode()) {
3091 case Instruction::Xor:
3092 case Instruction::Or:
3093 // If the mask is only needed on one incoming arm, push it up.
3094 if (Op0I->hasOneUse()) {
3095 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3096 // Not masking anything out for the LHS, move to RHS.
3097 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3098 Op0RHS->getName()+".masked");
3099 InsertNewInstBefore(NewRHS, I);
3100 return BinaryOperator::create(
3101 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3103 if (!isa<Constant>(Op0RHS) &&
3104 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3105 // Not masking anything out for the RHS, move to LHS.
3106 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3107 Op0LHS->getName()+".masked");
3108 InsertNewInstBefore(NewLHS, I);
3109 return BinaryOperator::create(
3110 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3115 case Instruction::Add:
3116 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3117 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3118 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3119 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3120 return BinaryOperator::createAnd(V, AndRHS);
3121 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3122 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3125 case Instruction::Sub:
3126 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3127 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3128 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3129 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3130 return BinaryOperator::createAnd(V, AndRHS);
3134 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3135 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3137 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3138 // If this is an integer truncation or change from signed-to-unsigned, and
3139 // if the source is an and/or with immediate, transform it. This
3140 // frequently occurs for bitfield accesses.
3141 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3142 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3143 CastOp->getNumOperands() == 2)
3144 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3145 if (CastOp->getOpcode() == Instruction::And) {
3146 // Change: and (cast (and X, C1) to T), C2
3147 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3148 // This will fold the two constants together, which may allow
3149 // other simplifications.
3150 Instruction *NewCast = CastInst::createTruncOrBitCast(
3151 CastOp->getOperand(0), I.getType(),
3152 CastOp->getName()+".shrunk");
3153 NewCast = InsertNewInstBefore(NewCast, I);
3154 // trunc_or_bitcast(C1)&C2
3155 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3156 C3 = ConstantExpr::getAnd(C3, AndRHS);
3157 return BinaryOperator::createAnd(NewCast, C3);
3158 } else if (CastOp->getOpcode() == Instruction::Or) {
3159 // Change: and (cast (or X, C1) to T), C2
3160 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3161 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3162 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3163 return ReplaceInstUsesWith(I, AndRHS);
3168 // Try to fold constant and into select arguments.
3169 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3170 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3172 if (isa<PHINode>(Op0))
3173 if (Instruction *NV = FoldOpIntoPhi(I))
3177 Value *Op0NotVal = dyn_castNotVal(Op0);
3178 Value *Op1NotVal = dyn_castNotVal(Op1);
3180 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3181 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3183 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3184 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3185 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3186 I.getName()+".demorgan");
3187 InsertNewInstBefore(Or, I);
3188 return BinaryOperator::createNot(Or);
3192 Value *A = 0, *B = 0;
3193 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3194 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3195 return ReplaceInstUsesWith(I, Op1);
3196 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3197 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3198 return ReplaceInstUsesWith(I, Op0);
3200 if (Op0->hasOneUse() &&
3201 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3202 if (A == Op1) { // (A^B)&A -> A&(A^B)
3203 I.swapOperands(); // Simplify below
3204 std::swap(Op0, Op1);
3205 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3206 cast<BinaryOperator>(Op0)->swapOperands();
3207 I.swapOperands(); // Simplify below
3208 std::swap(Op0, Op1);
3211 if (Op1->hasOneUse() &&
3212 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3213 if (B == Op0) { // B&(A^B) -> B&(B^A)
3214 cast<BinaryOperator>(Op1)->swapOperands();
3217 if (A == Op0) { // A&(A^B) -> A & ~B
3218 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3219 InsertNewInstBefore(NotB, I);
3220 return BinaryOperator::createAnd(A, NotB);
3225 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3226 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3227 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3230 Value *LHSVal, *RHSVal;
3231 ConstantInt *LHSCst, *RHSCst;
3232 ICmpInst::Predicate LHSCC, RHSCC;
3233 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3234 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3235 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3236 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3237 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3238 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3239 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3240 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3241 // Ensure that the larger constant is on the RHS.
3242 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3243 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3244 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3245 ICmpInst *LHS = cast<ICmpInst>(Op0);
3246 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3247 std::swap(LHS, RHS);
3248 std::swap(LHSCst, RHSCst);
3249 std::swap(LHSCC, RHSCC);
3252 // At this point, we know we have have two icmp instructions
3253 // comparing a value against two constants and and'ing the result
3254 // together. Because of the above check, we know that we only have
3255 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3256 // (from the FoldICmpLogical check above), that the two constants
3257 // are not equal and that the larger constant is on the RHS
3258 assert(LHSCst != RHSCst && "Compares not folded above?");
3261 default: assert(0 && "Unknown integer condition code!");
3262 case ICmpInst::ICMP_EQ:
3264 default: assert(0 && "Unknown integer condition code!");
3265 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3266 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3267 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3268 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3269 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3270 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3271 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3272 return ReplaceInstUsesWith(I, LHS);
3274 case ICmpInst::ICMP_NE:
3276 default: assert(0 && "Unknown integer condition code!");
3277 case ICmpInst::ICMP_ULT:
3278 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3279 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3280 break; // (X != 13 & X u< 15) -> no change
3281 case ICmpInst::ICMP_SLT:
3282 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3283 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3284 break; // (X != 13 & X s< 15) -> no change
3285 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3286 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3287 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3288 return ReplaceInstUsesWith(I, RHS);
3289 case ICmpInst::ICMP_NE:
3290 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3291 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3292 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3293 LHSVal->getName()+".off");
3294 InsertNewInstBefore(Add, I);
3295 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3296 ConstantInt::get(Add->getType(), 1));
3298 break; // (X != 13 & X != 15) -> no change
3301 case ICmpInst::ICMP_ULT:
3303 default: assert(0 && "Unknown integer condition code!");
3304 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3305 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3306 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3307 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3309 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3310 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3311 return ReplaceInstUsesWith(I, LHS);
3312 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3316 case ICmpInst::ICMP_SLT:
3318 default: assert(0 && "Unknown integer condition code!");
3319 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3320 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3321 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3322 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3324 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3325 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3326 return ReplaceInstUsesWith(I, LHS);
3327 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3331 case ICmpInst::ICMP_UGT:
3333 default: assert(0 && "Unknown integer condition code!");
3334 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3335 return ReplaceInstUsesWith(I, LHS);
3336 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3337 return ReplaceInstUsesWith(I, RHS);
3338 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3340 case ICmpInst::ICMP_NE:
3341 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3342 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3343 break; // (X u> 13 & X != 15) -> no change
3344 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3345 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3347 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3351 case ICmpInst::ICMP_SGT:
3353 default: assert(0 && "Unknown integer condition code!");
3354 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3355 return ReplaceInstUsesWith(I, LHS);
3356 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3357 return ReplaceInstUsesWith(I, RHS);
3358 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3360 case ICmpInst::ICMP_NE:
3361 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3362 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3363 break; // (X s> 13 & X != 15) -> no change
3364 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3365 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3367 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3375 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3376 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3377 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3378 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3379 const Type *SrcTy = Op0C->getOperand(0)->getType();
3380 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3381 // Only do this if the casts both really cause code to be generated.
3382 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3384 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3386 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3387 Op1C->getOperand(0),
3389 InsertNewInstBefore(NewOp, I);
3390 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3394 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3395 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3396 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3397 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3398 SI0->getOperand(1) == SI1->getOperand(1) &&
3399 (SI0->hasOneUse() || SI1->hasOneUse())) {
3400 Instruction *NewOp =
3401 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3403 SI0->getName()), I);
3404 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3405 SI1->getOperand(1));
3409 return Changed ? &I : 0;
3412 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3413 /// in the result. If it does, and if the specified byte hasn't been filled in
3414 /// yet, fill it in and return false.
3415 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3416 Instruction *I = dyn_cast<Instruction>(V);
3417 if (I == 0) return true;
3419 // If this is an or instruction, it is an inner node of the bswap.
3420 if (I->getOpcode() == Instruction::Or)
3421 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3422 CollectBSwapParts(I->getOperand(1), ByteValues);
3424 // If this is a shift by a constant int, and it is "24", then its operand
3425 // defines a byte. We only handle unsigned types here.
3426 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3427 // Not shifting the entire input by N-1 bytes?
3428 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3429 8*(ByteValues.size()-1))
3433 if (I->getOpcode() == Instruction::Shl) {
3434 // X << 24 defines the top byte with the lowest of the input bytes.
3435 DestNo = ByteValues.size()-1;
3437 // X >>u 24 defines the low byte with the highest of the input bytes.
3441 // If the destination byte value is already defined, the values are or'd
3442 // together, which isn't a bswap (unless it's an or of the same bits).
3443 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3445 ByteValues[DestNo] = I->getOperand(0);
3449 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3451 Value *Shift = 0, *ShiftLHS = 0;
3452 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3453 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3454 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3456 Instruction *SI = cast<Instruction>(Shift);
3458 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3459 if (ShiftAmt->getZExtValue() & 7 ||
3460 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3463 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3465 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3466 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3468 // Unknown mask for bswap.
3469 if (DestByte == ByteValues.size()) return true;
3471 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3473 if (SI->getOpcode() == Instruction::Shl)
3474 SrcByte = DestByte - ShiftBytes;
3476 SrcByte = DestByte + ShiftBytes;
3478 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3479 if (SrcByte != ByteValues.size()-DestByte-1)
3482 // If the destination byte value is already defined, the values are or'd
3483 // together, which isn't a bswap (unless it's an or of the same bits).
3484 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3486 ByteValues[DestByte] = SI->getOperand(0);
3490 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3491 /// If so, insert the new bswap intrinsic and return it.
3492 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3493 // We cannot bswap one byte.
3494 if (I.getType() == Type::Int8Ty)
3497 /// ByteValues - For each byte of the result, we keep track of which value
3498 /// defines each byte.
3499 SmallVector<Value*, 8> ByteValues;
3500 ByteValues.resize(TD->getTypeSize(I.getType()));
3502 // Try to find all the pieces corresponding to the bswap.
3503 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3504 CollectBSwapParts(I.getOperand(1), ByteValues))
3507 // Check to see if all of the bytes come from the same value.
3508 Value *V = ByteValues[0];
3509 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3511 // Check to make sure that all of the bytes come from the same value.
3512 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3513 if (ByteValues[i] != V)
3516 // If they do then *success* we can turn this into a bswap. Figure out what
3517 // bswap to make it into.
3518 Module *M = I.getParent()->getParent()->getParent();
3519 const char *FnName = 0;
3520 if (I.getType() == Type::Int16Ty)
3521 FnName = "llvm.bswap.i16";
3522 else if (I.getType() == Type::Int32Ty)
3523 FnName = "llvm.bswap.i32";
3524 else if (I.getType() == Type::Int64Ty)
3525 FnName = "llvm.bswap.i64";
3527 assert(0 && "Unknown integer type!");
3528 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3529 return new CallInst(F, V);
3533 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3534 bool Changed = SimplifyCommutative(I);
3535 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3537 if (isa<UndefValue>(Op1))
3538 return ReplaceInstUsesWith(I, // X | undef -> -1
3539 ConstantInt::getAllOnesValue(I.getType()));
3543 return ReplaceInstUsesWith(I, Op0);
3545 // See if we can simplify any instructions used by the instruction whose sole
3546 // purpose is to compute bits we don't care about.
3547 uint64_t KnownZero, KnownOne;
3548 if (!isa<VectorType>(I.getType()) &&
3549 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3550 KnownZero, KnownOne))
3554 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3555 ConstantInt *C1 = 0; Value *X = 0;
3556 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3557 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3558 Instruction *Or = BinaryOperator::createOr(X, RHS);
3559 InsertNewInstBefore(Or, I);
3561 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3564 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3565 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3566 Instruction *Or = BinaryOperator::createOr(X, RHS);
3567 InsertNewInstBefore(Or, I);
3569 return BinaryOperator::createXor(Or,
3570 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3573 // Try to fold constant and into select arguments.
3574 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3575 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3577 if (isa<PHINode>(Op0))
3578 if (Instruction *NV = FoldOpIntoPhi(I))
3582 Value *A = 0, *B = 0;
3583 ConstantInt *C1 = 0, *C2 = 0;
3585 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3586 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3587 return ReplaceInstUsesWith(I, Op1);
3588 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3589 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3590 return ReplaceInstUsesWith(I, Op0);
3592 // (A | B) | C and A | (B | C) -> bswap if possible.
3593 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3594 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3595 match(Op1, m_Or(m_Value(), m_Value())) ||
3596 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3597 match(Op1, m_Shift(m_Value(), m_Value())))) {
3598 if (Instruction *BSwap = MatchBSwap(I))
3602 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3603 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3604 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3605 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3606 InsertNewInstBefore(NOr, I);
3608 return BinaryOperator::createXor(NOr, C1);
3611 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3612 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3613 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3614 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3615 InsertNewInstBefore(NOr, I);
3617 return BinaryOperator::createXor(NOr, C1);
3620 // (A & C1)|(B & C2)
3621 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3622 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3624 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3625 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3628 // If we have: ((V + N) & C1) | (V & C2)
3629 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3630 // replace with V+N.
3631 if (C1 == ConstantExpr::getNot(C2)) {
3632 Value *V1 = 0, *V2 = 0;
3633 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3634 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3635 // Add commutes, try both ways.
3636 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3637 return ReplaceInstUsesWith(I, A);
3638 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3639 return ReplaceInstUsesWith(I, A);
3641 // Or commutes, try both ways.
3642 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3643 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3644 // Add commutes, try both ways.
3645 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3646 return ReplaceInstUsesWith(I, B);
3647 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3648 return ReplaceInstUsesWith(I, B);
3653 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3654 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3655 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3656 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3657 SI0->getOperand(1) == SI1->getOperand(1) &&
3658 (SI0->hasOneUse() || SI1->hasOneUse())) {
3659 Instruction *NewOp =
3660 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3662 SI0->getName()), I);
3663 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3664 SI1->getOperand(1));
3668 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3669 if (A == Op1) // ~A | A == -1
3670 return ReplaceInstUsesWith(I,
3671 ConstantInt::getAllOnesValue(I.getType()));
3675 // Note, A is still live here!
3676 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3678 return ReplaceInstUsesWith(I,
3679 ConstantInt::getAllOnesValue(I.getType()));
3681 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3682 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3683 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3684 I.getName()+".demorgan"), I);
3685 return BinaryOperator::createNot(And);
3689 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3690 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3691 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3694 Value *LHSVal, *RHSVal;
3695 ConstantInt *LHSCst, *RHSCst;
3696 ICmpInst::Predicate LHSCC, RHSCC;
3697 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3698 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3699 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3700 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3701 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3702 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3703 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3704 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3705 // Ensure that the larger constant is on the RHS.
3706 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3707 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3708 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3709 ICmpInst *LHS = cast<ICmpInst>(Op0);
3710 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3711 std::swap(LHS, RHS);
3712 std::swap(LHSCst, RHSCst);
3713 std::swap(LHSCC, RHSCC);
3716 // At this point, we know we have have two icmp instructions
3717 // comparing a value against two constants and or'ing the result
3718 // together. Because of the above check, we know that we only have
3719 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3720 // FoldICmpLogical check above), that the two constants are not
3722 assert(LHSCst != RHSCst && "Compares not folded above?");
3725 default: assert(0 && "Unknown integer condition code!");
3726 case ICmpInst::ICMP_EQ:
3728 default: assert(0 && "Unknown integer condition code!");
3729 case ICmpInst::ICMP_EQ:
3730 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3731 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3732 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3733 LHSVal->getName()+".off");
3734 InsertNewInstBefore(Add, I);
3735 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3736 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3738 break; // (X == 13 | X == 15) -> no change
3739 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3740 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3742 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3743 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3744 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3745 return ReplaceInstUsesWith(I, RHS);
3748 case ICmpInst::ICMP_NE:
3750 default: assert(0 && "Unknown integer condition code!");
3751 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3752 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3753 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3754 return ReplaceInstUsesWith(I, LHS);
3755 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3756 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3757 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3758 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3761 case ICmpInst::ICMP_ULT:
3763 default: assert(0 && "Unknown integer condition code!");
3764 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3766 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3767 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3769 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3771 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3772 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3773 return ReplaceInstUsesWith(I, RHS);
3774 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3778 case ICmpInst::ICMP_SLT:
3780 default: assert(0 && "Unknown integer condition code!");
3781 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3783 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3784 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3786 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3788 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3789 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3790 return ReplaceInstUsesWith(I, RHS);
3791 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3795 case ICmpInst::ICMP_UGT:
3797 default: assert(0 && "Unknown integer condition code!");
3798 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3799 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3800 return ReplaceInstUsesWith(I, LHS);
3801 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3803 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3804 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3805 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3806 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3810 case ICmpInst::ICMP_SGT:
3812 default: assert(0 && "Unknown integer condition code!");
3813 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3814 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3815 return ReplaceInstUsesWith(I, LHS);
3816 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3818 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3819 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3820 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3821 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3829 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3830 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3831 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3832 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3833 const Type *SrcTy = Op0C->getOperand(0)->getType();
3834 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3835 // Only do this if the casts both really cause code to be generated.
3836 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3838 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3840 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3841 Op1C->getOperand(0),
3843 InsertNewInstBefore(NewOp, I);
3844 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3849 return Changed ? &I : 0;
3852 // XorSelf - Implements: X ^ X --> 0
3855 XorSelf(Value *rhs) : RHS(rhs) {}
3856 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3857 Instruction *apply(BinaryOperator &Xor) const {
3863 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3864 bool Changed = SimplifyCommutative(I);
3865 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3867 if (isa<UndefValue>(Op1))
3868 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3870 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3871 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3872 assert(Result == &I && "AssociativeOpt didn't work?");
3873 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3876 // See if we can simplify any instructions used by the instruction whose sole
3877 // purpose is to compute bits we don't care about.
3878 uint64_t KnownZero, KnownOne;
3879 if (!isa<VectorType>(I.getType()) &&
3880 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3881 KnownZero, KnownOne))
3884 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3885 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
3886 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3887 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
3888 return new ICmpInst(ICI->getInversePredicate(),
3889 ICI->getOperand(0), ICI->getOperand(1));
3891 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3892 // ~(c-X) == X-c-1 == X+(-c-1)
3893 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3894 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3895 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3896 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3897 ConstantInt::get(I.getType(), 1));
3898 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3901 // ~(~X & Y) --> (X | ~Y)
3902 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3903 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3904 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3906 BinaryOperator::createNot(Op0I->getOperand(1),
3907 Op0I->getOperand(1)->getName()+".not");
3908 InsertNewInstBefore(NotY, I);
3909 return BinaryOperator::createOr(Op0NotVal, NotY);
3913 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3914 if (Op0I->getOpcode() == Instruction::Add) {
3915 // ~(X-c) --> (-c-1)-X
3916 if (RHS->isAllOnesValue()) {
3917 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3918 return BinaryOperator::createSub(
3919 ConstantExpr::getSub(NegOp0CI,
3920 ConstantInt::get(I.getType(), 1)),
3921 Op0I->getOperand(0));
3923 } else if (Op0I->getOpcode() == Instruction::Or) {
3924 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3925 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3926 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3927 // Anything in both C1 and C2 is known to be zero, remove it from
3929 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3930 NewRHS = ConstantExpr::getAnd(NewRHS,
3931 ConstantExpr::getNot(CommonBits));
3932 WorkList.push_back(Op0I);
3933 I.setOperand(0, Op0I->getOperand(0));
3934 I.setOperand(1, NewRHS);
3940 // Try to fold constant and into select arguments.
3941 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3942 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3944 if (isa<PHINode>(Op0))
3945 if (Instruction *NV = FoldOpIntoPhi(I))
3949 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3951 return ReplaceInstUsesWith(I,
3952 ConstantInt::getAllOnesValue(I.getType()));
3954 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3956 return ReplaceInstUsesWith(I,
3957 ConstantInt::getAllOnesValue(I.getType()));
3959 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3960 if (Op1I->getOpcode() == Instruction::Or) {
3961 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3962 Op1I->swapOperands();
3964 std::swap(Op0, Op1);
3965 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3966 I.swapOperands(); // Simplified below.
3967 std::swap(Op0, Op1);
3969 } else if (Op1I->getOpcode() == Instruction::Xor) {
3970 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3971 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3972 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3973 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3974 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3975 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3976 Op1I->swapOperands();
3977 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3978 I.swapOperands(); // Simplified below.
3979 std::swap(Op0, Op1);
3983 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3984 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
3985 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
3986 Op0I->swapOperands();
3987 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
3988 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
3989 InsertNewInstBefore(NotB, I);
3990 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
3992 } else if (Op0I->getOpcode() == Instruction::Xor) {
3993 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
3994 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3995 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
3996 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3997 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
3998 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
3999 Op0I->swapOperands();
4000 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
4001 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4002 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
4003 InsertNewInstBefore(N, I);
4004 return BinaryOperator::createAnd(N, Op1);
4008 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4009 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4010 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4013 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4014 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4015 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4016 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4017 const Type *SrcTy = Op0C->getOperand(0)->getType();
4018 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4019 // Only do this if the casts both really cause code to be generated.
4020 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4022 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4024 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4025 Op1C->getOperand(0),
4027 InsertNewInstBefore(NewOp, I);
4028 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4032 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4033 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4034 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4035 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4036 SI0->getOperand(1) == SI1->getOperand(1) &&
4037 (SI0->hasOneUse() || SI1->hasOneUse())) {
4038 Instruction *NewOp =
4039 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
4041 SI0->getName()), I);
4042 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4043 SI1->getOperand(1));
4047 return Changed ? &I : 0;
4050 static bool isPositive(ConstantInt *C) {
4051 return C->getSExtValue() >= 0;
4054 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4055 /// overflowed for this type.
4056 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4058 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4060 return cast<ConstantInt>(Result)->getZExtValue() <
4061 cast<ConstantInt>(In1)->getZExtValue();
4064 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4065 /// code necessary to compute the offset from the base pointer (without adding
4066 /// in the base pointer). Return the result as a signed integer of intptr size.
4067 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4068 TargetData &TD = IC.getTargetData();
4069 gep_type_iterator GTI = gep_type_begin(GEP);
4070 const Type *IntPtrTy = TD.getIntPtrType();
4071 Value *Result = Constant::getNullValue(IntPtrTy);
4073 // Build a mask for high order bits.
4074 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4076 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4077 Value *Op = GEP->getOperand(i);
4078 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4079 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4080 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4081 if (!OpC->isNullValue()) {
4082 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4083 Scale = ConstantExpr::getMul(OpC, Scale);
4084 if (Constant *RC = dyn_cast<Constant>(Result))
4085 Result = ConstantExpr::getAdd(RC, Scale);
4087 // Emit an add instruction.
4088 Result = IC.InsertNewInstBefore(
4089 BinaryOperator::createAdd(Result, Scale,
4090 GEP->getName()+".offs"), I);
4094 // Convert to correct type.
4095 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4096 Op->getName()+".c"), I);
4098 // We'll let instcombine(mul) convert this to a shl if possible.
4099 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4100 GEP->getName()+".idx"), I);
4102 // Emit an add instruction.
4103 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4104 GEP->getName()+".offs"), I);
4110 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4111 /// else. At this point we know that the GEP is on the LHS of the comparison.
4112 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4113 ICmpInst::Predicate Cond,
4115 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4117 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4118 if (isa<PointerType>(CI->getOperand(0)->getType()))
4119 RHS = CI->getOperand(0);
4121 Value *PtrBase = GEPLHS->getOperand(0);
4122 if (PtrBase == RHS) {
4123 // As an optimization, we don't actually have to compute the actual value of
4124 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4125 // each index is zero or not.
4126 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4127 Instruction *InVal = 0;
4128 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4129 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4131 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4132 if (isa<UndefValue>(C)) // undef index -> undef.
4133 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4134 if (C->isNullValue())
4136 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4137 EmitIt = false; // This is indexing into a zero sized array?
4138 } else if (isa<ConstantInt>(C))
4139 return ReplaceInstUsesWith(I, // No comparison is needed here.
4140 ConstantInt::get(Type::Int1Ty,
4141 Cond == ICmpInst::ICMP_NE));
4146 new ICmpInst(Cond, GEPLHS->getOperand(i),
4147 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4151 InVal = InsertNewInstBefore(InVal, I);
4152 InsertNewInstBefore(Comp, I);
4153 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4154 InVal = BinaryOperator::createOr(InVal, Comp);
4155 else // True if all are equal
4156 InVal = BinaryOperator::createAnd(InVal, Comp);
4164 // No comparison is needed here, all indexes = 0
4165 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4166 Cond == ICmpInst::ICMP_EQ));
4169 // Only lower this if the icmp is the only user of the GEP or if we expect
4170 // the result to fold to a constant!
4171 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4172 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4173 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4174 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4175 Constant::getNullValue(Offset->getType()));
4177 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4178 // If the base pointers are different, but the indices are the same, just
4179 // compare the base pointer.
4180 if (PtrBase != GEPRHS->getOperand(0)) {
4181 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4182 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4183 GEPRHS->getOperand(0)->getType();
4185 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4186 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4187 IndicesTheSame = false;
4191 // If all indices are the same, just compare the base pointers.
4193 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4194 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4196 // Otherwise, the base pointers are different and the indices are
4197 // different, bail out.
4201 // If one of the GEPs has all zero indices, recurse.
4202 bool AllZeros = true;
4203 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4204 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4205 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4210 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4211 ICmpInst::getSwappedPredicate(Cond), I);
4213 // If the other GEP has all zero indices, recurse.
4215 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4216 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4217 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4222 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4224 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4225 // If the GEPs only differ by one index, compare it.
4226 unsigned NumDifferences = 0; // Keep track of # differences.
4227 unsigned DiffOperand = 0; // The operand that differs.
4228 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4229 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4230 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4231 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4232 // Irreconcilable differences.
4236 if (NumDifferences++) break;
4241 if (NumDifferences == 0) // SAME GEP?
4242 return ReplaceInstUsesWith(I, // No comparison is needed here.
4243 ConstantInt::get(Type::Int1Ty,
4244 Cond == ICmpInst::ICMP_EQ));
4245 else if (NumDifferences == 1) {
4246 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4247 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4248 // Make sure we do a signed comparison here.
4249 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4253 // Only lower this if the icmp is the only user of the GEP or if we expect
4254 // the result to fold to a constant!
4255 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4256 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4257 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4258 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4259 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4260 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4266 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4267 bool Changed = SimplifyCompare(I);
4268 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4270 // Fold trivial predicates.
4271 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4272 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4273 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4274 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4276 // Simplify 'fcmp pred X, X'
4278 switch (I.getPredicate()) {
4279 default: assert(0 && "Unknown predicate!");
4280 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4281 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4282 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4283 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4284 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4285 case FCmpInst::FCMP_OLT: // True if ordered and less than
4286 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4287 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4289 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4290 case FCmpInst::FCMP_ULT: // True if unordered or less than
4291 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4292 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4293 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4294 I.setPredicate(FCmpInst::FCMP_UNO);
4295 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4298 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4299 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4300 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4301 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4302 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4303 I.setPredicate(FCmpInst::FCMP_ORD);
4304 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4309 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4310 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4312 // Handle fcmp with constant RHS
4313 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4314 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4315 switch (LHSI->getOpcode()) {
4316 case Instruction::PHI:
4317 if (Instruction *NV = FoldOpIntoPhi(I))
4320 case Instruction::Select:
4321 // If either operand of the select is a constant, we can fold the
4322 // comparison into the select arms, which will cause one to be
4323 // constant folded and the select turned into a bitwise or.
4324 Value *Op1 = 0, *Op2 = 0;
4325 if (LHSI->hasOneUse()) {
4326 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4327 // Fold the known value into the constant operand.
4328 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4329 // Insert a new FCmp of the other select operand.
4330 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4331 LHSI->getOperand(2), RHSC,
4333 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4334 // Fold the known value into the constant operand.
4335 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4336 // Insert a new FCmp of the other select operand.
4337 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4338 LHSI->getOperand(1), RHSC,
4344 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4349 return Changed ? &I : 0;
4352 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4353 bool Changed = SimplifyCompare(I);
4354 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4355 const Type *Ty = Op0->getType();
4359 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4360 isTrueWhenEqual(I)));
4362 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4363 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4365 // icmp of GlobalValues can never equal each other as long as they aren't
4366 // external weak linkage type.
4367 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4368 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4369 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4370 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4371 !isTrueWhenEqual(I)));
4373 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4374 // addresses never equal each other! We already know that Op0 != Op1.
4375 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4376 isa<ConstantPointerNull>(Op0)) &&
4377 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4378 isa<ConstantPointerNull>(Op1)))
4379 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4380 !isTrueWhenEqual(I)));
4382 // icmp's with boolean values can always be turned into bitwise operations
4383 if (Ty == Type::Int1Ty) {
4384 switch (I.getPredicate()) {
4385 default: assert(0 && "Invalid icmp instruction!");
4386 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4387 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4388 InsertNewInstBefore(Xor, I);
4389 return BinaryOperator::createNot(Xor);
4391 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4392 return BinaryOperator::createXor(Op0, Op1);
4394 case ICmpInst::ICMP_UGT:
4395 case ICmpInst::ICMP_SGT:
4396 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4398 case ICmpInst::ICMP_ULT:
4399 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4400 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4401 InsertNewInstBefore(Not, I);
4402 return BinaryOperator::createAnd(Not, Op1);
4404 case ICmpInst::ICMP_UGE:
4405 case ICmpInst::ICMP_SGE:
4406 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4408 case ICmpInst::ICMP_ULE:
4409 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4410 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4411 InsertNewInstBefore(Not, I);
4412 return BinaryOperator::createOr(Not, Op1);
4417 // See if we are doing a comparison between a constant and an instruction that
4418 // can be folded into the comparison.
4419 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4420 switch (I.getPredicate()) {
4422 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4423 if (CI->isMinValue(false))
4424 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4425 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4426 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4427 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4428 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4431 case ICmpInst::ICMP_SLT:
4432 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4433 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4434 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4435 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4436 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4437 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4440 case ICmpInst::ICMP_UGT:
4441 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4442 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4443 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4444 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4445 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4446 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4449 case ICmpInst::ICMP_SGT:
4450 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4451 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4452 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4453 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4454 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4455 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4458 case ICmpInst::ICMP_ULE:
4459 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4460 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4461 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4462 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4463 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4464 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4467 case ICmpInst::ICMP_SLE:
4468 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4469 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4470 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4471 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4472 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4473 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4476 case ICmpInst::ICMP_UGE:
4477 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4478 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4479 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4480 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4481 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4482 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4485 case ICmpInst::ICMP_SGE:
4486 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4487 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4488 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4489 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4490 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4491 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4495 // If we still have a icmp le or icmp ge instruction, turn it into the
4496 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4497 // already been handled above, this requires little checking.
4499 if (I.getPredicate() == ICmpInst::ICMP_ULE)
4500 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4501 if (I.getPredicate() == ICmpInst::ICMP_SLE)
4502 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4503 if (I.getPredicate() == ICmpInst::ICMP_UGE)
4504 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4505 if (I.getPredicate() == ICmpInst::ICMP_SGE)
4506 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4508 // See if we can fold the comparison based on bits known to be zero or one
4510 uint64_t KnownZero, KnownOne;
4511 if (SimplifyDemandedBits(Op0, cast<IntegerType>(Ty)->getBitMask(),
4512 KnownZero, KnownOne, 0))
4515 // Given the known and unknown bits, compute a range that the LHS could be
4517 if (KnownOne | KnownZero) {
4518 // Compute the Min, Max and RHS values based on the known bits. For the
4519 // EQ and NE we use unsigned values.
4520 uint64_t UMin = 0, UMax = 0, URHSVal = 0;
4521 int64_t SMin = 0, SMax = 0, SRHSVal = 0;
4522 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4523 SRHSVal = CI->getSExtValue();
4524 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
4527 URHSVal = CI->getZExtValue();
4528 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
4531 switch (I.getPredicate()) { // LE/GE have been folded already.
4532 default: assert(0 && "Unknown icmp opcode!");
4533 case ICmpInst::ICMP_EQ:
4534 if (UMax < URHSVal || UMin > URHSVal)
4535 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4537 case ICmpInst::ICMP_NE:
4538 if (UMax < URHSVal || UMin > URHSVal)
4539 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4541 case ICmpInst::ICMP_ULT:
4543 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4545 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4547 case ICmpInst::ICMP_UGT:
4549 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4551 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4553 case ICmpInst::ICMP_SLT:
4555 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4557 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4559 case ICmpInst::ICMP_SGT:
4561 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4563 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4568 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4569 // instruction, see if that instruction also has constants so that the
4570 // instruction can be folded into the icmp
4571 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4572 switch (LHSI->getOpcode()) {
4573 case Instruction::And:
4574 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4575 LHSI->getOperand(0)->hasOneUse()) {
4576 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4578 // If the LHS is an AND of a truncating cast, we can widen the
4579 // and/compare to be the input width without changing the value
4580 // produced, eliminating a cast.
4581 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4582 // We can do this transformation if either the AND constant does not
4583 // have its sign bit set or if it is an equality comparison.
4584 // Extending a relational comparison when we're checking the sign
4585 // bit would not work.
4586 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4588 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4589 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4590 ConstantInt *NewCST;
4592 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4593 AndCST->getZExtValue());
4594 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4595 CI->getZExtValue());
4596 Instruction *NewAnd =
4597 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4599 InsertNewInstBefore(NewAnd, I);
4600 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4604 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4605 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4606 // happens a LOT in code produced by the C front-end, for bitfield
4608 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
4609 if (Shift && !Shift->isShift())
4613 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4614 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4615 const Type *AndTy = AndCST->getType(); // Type of the and.
4617 // We can fold this as long as we can't shift unknown bits
4618 // into the mask. This can only happen with signed shift
4619 // rights, as they sign-extend.
4621 bool CanFold = Shift->isLogicalShift();
4623 // To test for the bad case of the signed shr, see if any
4624 // of the bits shifted in could be tested after the mask.
4625 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4626 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4628 Constant *OShAmt = ConstantInt::get(AndTy, ShAmtVal);
4630 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4632 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4638 if (Shift->getOpcode() == Instruction::Shl)
4639 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4641 NewCst = ConstantExpr::getShl(CI, ShAmt);
4643 // Check to see if we are shifting out any of the bits being
4645 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4646 // If we shifted bits out, the fold is not going to work out.
4647 // As a special case, check to see if this means that the
4648 // result is always true or false now.
4649 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4650 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4651 if (I.getPredicate() == ICmpInst::ICMP_NE)
4652 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4654 I.setOperand(1, NewCst);
4655 Constant *NewAndCST;
4656 if (Shift->getOpcode() == Instruction::Shl)
4657 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4659 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4660 LHSI->setOperand(1, NewAndCST);
4661 LHSI->setOperand(0, Shift->getOperand(0));
4662 WorkList.push_back(Shift); // Shift is dead.
4663 AddUsesToWorkList(I);
4669 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4670 // preferable because it allows the C<<Y expression to be hoisted out
4671 // of a loop if Y is invariant and X is not.
4672 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4673 I.isEquality() && !Shift->isArithmeticShift() &&
4674 isa<Instruction>(Shift->getOperand(0))) {
4677 if (Shift->getOpcode() == Instruction::LShr) {
4678 NS = BinaryOperator::createShl(AndCST,
4679 Shift->getOperand(1), "tmp");
4681 // Insert a logical shift.
4682 NS = BinaryOperator::createLShr(AndCST,
4683 Shift->getOperand(1), "tmp");
4685 InsertNewInstBefore(cast<Instruction>(NS), I);
4687 // Compute X & (C << Y).
4688 Instruction *NewAnd = BinaryOperator::createAnd(
4689 Shift->getOperand(0), NS, LHSI->getName());
4690 InsertNewInstBefore(NewAnd, I);
4692 I.setOperand(0, NewAnd);
4698 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4699 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4700 if (I.isEquality()) {
4701 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4703 // Check that the shift amount is in range. If not, don't perform
4704 // undefined shifts. When the shift is visited it will be
4706 if (ShAmt->getZExtValue() >= TypeBits)
4709 // If we are comparing against bits always shifted out, the
4710 // comparison cannot succeed.
4712 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4713 if (Comp != CI) {// Comparing against a bit that we know is zero.
4714 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4715 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4716 return ReplaceInstUsesWith(I, Cst);
4719 if (LHSI->hasOneUse()) {
4720 // Otherwise strength reduce the shift into an and.
4721 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4722 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4723 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4726 BinaryOperator::createAnd(LHSI->getOperand(0),
4727 Mask, LHSI->getName()+".mask");
4728 Value *And = InsertNewInstBefore(AndI, I);
4729 return new ICmpInst(I.getPredicate(), And,
4730 ConstantExpr::getLShr(CI, ShAmt));
4736 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4737 case Instruction::AShr:
4738 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4739 if (I.isEquality()) {
4740 // Check that the shift amount is in range. If not, don't perform
4741 // undefined shifts. When the shift is visited it will be
4743 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4744 if (ShAmt->getZExtValue() >= TypeBits)
4747 // If we are comparing against bits always shifted out, the
4748 // comparison cannot succeed.
4750 if (LHSI->getOpcode() == Instruction::LShr)
4751 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4754 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4757 if (Comp != CI) {// Comparing against a bit that we know is zero.
4758 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4759 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4760 return ReplaceInstUsesWith(I, Cst);
4763 if (LHSI->hasOneUse() || CI->isNullValue()) {
4764 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4766 // Otherwise strength reduce the shift into an and.
4767 uint64_t Val = ~0ULL; // All ones.
4768 Val <<= ShAmtVal; // Shift over to the right spot.
4769 Val &= ~0ULL >> (64-TypeBits);
4770 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4773 BinaryOperator::createAnd(LHSI->getOperand(0),
4774 Mask, LHSI->getName()+".mask");
4775 Value *And = InsertNewInstBefore(AndI, I);
4776 return new ICmpInst(I.getPredicate(), And,
4777 ConstantExpr::getShl(CI, ShAmt));
4783 case Instruction::SDiv:
4784 case Instruction::UDiv:
4785 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4786 // Fold this div into the comparison, producing a range check.
4787 // Determine, based on the divide type, what the range is being
4788 // checked. If there is an overflow on the low or high side, remember
4789 // it, otherwise compute the range [low, hi) bounding the new value.
4790 // See: InsertRangeTest above for the kinds of replacements possible.
4791 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4792 // FIXME: If the operand types don't match the type of the divide
4793 // then don't attempt this transform. The code below doesn't have the
4794 // logic to deal with a signed divide and an unsigned compare (and
4795 // vice versa). This is because (x /s C1) <s C2 produces different
4796 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4797 // (x /u C1) <u C2. Simply casting the operands and result won't
4798 // work. :( The if statement below tests that condition and bails
4800 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4801 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4804 // Initialize the variables that will indicate the nature of the
4806 bool LoOverflow = false, HiOverflow = false;
4807 ConstantInt *LoBound = 0, *HiBound = 0;
4809 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4810 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4811 // C2 (CI). By solving for X we can turn this into a range check
4812 // instead of computing a divide.
4814 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4816 // Determine if the product overflows by seeing if the product is
4817 // not equal to the divide. Make sure we do the same kind of divide
4818 // as in the LHS instruction that we're folding.
4819 bool ProdOV = !DivRHS->isNullValue() &&
4820 (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4821 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4823 // Get the ICmp opcode
4824 ICmpInst::Predicate predicate = I.getPredicate();
4826 if (DivRHS->isNullValue()) {
4827 // Don't hack on divide by zeros!
4828 } else if (!DivIsSigned) { // udiv
4830 LoOverflow = ProdOV;
4831 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4832 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4833 if (CI->isNullValue()) { // (X / pos) op 0
4835 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4837 } else if (isPositive(CI)) { // (X / pos) op pos
4839 LoOverflow = ProdOV;
4840 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4841 } else { // (X / pos) op neg
4842 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4843 LoOverflow = AddWithOverflow(LoBound, Prod,
4844 cast<ConstantInt>(DivRHSH));
4846 HiOverflow = ProdOV;
4848 } else { // Divisor is < 0.
4849 if (CI->isNullValue()) { // (X / neg) op 0
4850 LoBound = AddOne(DivRHS);
4851 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4852 if (HiBound == DivRHS)
4853 LoBound = 0; // - INTMIN = INTMIN
4854 } else if (isPositive(CI)) { // (X / neg) op pos
4855 HiOverflow = LoOverflow = ProdOV;
4857 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4858 HiBound = AddOne(Prod);
4859 } else { // (X / neg) op neg
4861 LoOverflow = HiOverflow = ProdOV;
4862 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4865 // Dividing by a negate swaps the condition.
4866 predicate = ICmpInst::getSwappedPredicate(predicate);
4870 Value *X = LHSI->getOperand(0);
4871 switch (predicate) {
4872 default: assert(0 && "Unhandled icmp opcode!");
4873 case ICmpInst::ICMP_EQ:
4874 if (LoOverflow && HiOverflow)
4875 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4876 else if (HiOverflow)
4877 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4878 ICmpInst::ICMP_UGE, X, LoBound);
4879 else if (LoOverflow)
4880 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4881 ICmpInst::ICMP_ULT, X, HiBound);
4883 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4885 case ICmpInst::ICMP_NE:
4886 if (LoOverflow && HiOverflow)
4887 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4888 else if (HiOverflow)
4889 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4890 ICmpInst::ICMP_ULT, X, LoBound);
4891 else if (LoOverflow)
4892 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4893 ICmpInst::ICMP_UGE, X, HiBound);
4895 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4897 case ICmpInst::ICMP_ULT:
4898 case ICmpInst::ICMP_SLT:
4900 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4901 return new ICmpInst(predicate, X, LoBound);
4902 case ICmpInst::ICMP_UGT:
4903 case ICmpInst::ICMP_SGT:
4905 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4906 if (predicate == ICmpInst::ICMP_UGT)
4907 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
4909 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
4916 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
4917 if (I.isEquality()) {
4918 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4920 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4921 // the second operand is a constant, simplify a bit.
4922 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4923 switch (BO->getOpcode()) {
4924 case Instruction::SRem:
4925 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4926 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4928 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4929 if (V > 1 && isPowerOf2_64(V)) {
4930 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4931 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4932 return new ICmpInst(I.getPredicate(), NewRem,
4933 Constant::getNullValue(BO->getType()));
4937 case Instruction::Add:
4938 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4939 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4940 if (BO->hasOneUse())
4941 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4942 ConstantExpr::getSub(CI, BOp1C));
4943 } else if (CI->isNullValue()) {
4944 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4945 // efficiently invertible, or if the add has just this one use.
4946 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4948 if (Value *NegVal = dyn_castNegVal(BOp1))
4949 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
4950 else if (Value *NegVal = dyn_castNegVal(BOp0))
4951 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
4952 else if (BO->hasOneUse()) {
4953 Instruction *Neg = BinaryOperator::createNeg(BOp1);
4954 InsertNewInstBefore(Neg, I);
4956 return new ICmpInst(I.getPredicate(), BOp0, Neg);
4960 case Instruction::Xor:
4961 // For the xor case, we can xor two constants together, eliminating
4962 // the explicit xor.
4963 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4964 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4965 ConstantExpr::getXor(CI, BOC));
4968 case Instruction::Sub:
4969 // Replace (([sub|xor] A, B) != 0) with (A != B)
4970 if (CI->isNullValue())
4971 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4975 case Instruction::Or:
4976 // If bits are being or'd in that are not present in the constant we
4977 // are comparing against, then the comparison could never succeed!
4978 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4979 Constant *NotCI = ConstantExpr::getNot(CI);
4980 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4981 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4986 case Instruction::And:
4987 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4988 // If bits are being compared against that are and'd out, then the
4989 // comparison can never succeed!
4990 if (!ConstantExpr::getAnd(CI,
4991 ConstantExpr::getNot(BOC))->isNullValue())
4992 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4995 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4996 if (CI == BOC && isOneBitSet(CI))
4997 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
4998 ICmpInst::ICMP_NE, Op0,
4999 Constant::getNullValue(CI->getType()));
5001 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5002 if (isSignBit(BOC)) {
5003 Value *X = BO->getOperand(0);
5004 Constant *Zero = Constant::getNullValue(X->getType());
5005 ICmpInst::Predicate pred = isICMP_NE ?
5006 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5007 return new ICmpInst(pred, X, Zero);
5010 // ((X & ~7) == 0) --> X < 8
5011 if (CI->isNullValue() && isHighOnes(BOC)) {
5012 Value *X = BO->getOperand(0);
5013 Constant *NegX = ConstantExpr::getNeg(BOC);
5014 ICmpInst::Predicate pred = isICMP_NE ?
5015 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5016 return new ICmpInst(pred, X, NegX);
5022 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5023 // Handle set{eq|ne} <intrinsic>, intcst.
5024 switch (II->getIntrinsicID()) {
5026 case Intrinsic::bswap_i16:
5027 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5028 WorkList.push_back(II); // Dead?
5029 I.setOperand(0, II->getOperand(1));
5030 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5031 ByteSwap_16(CI->getZExtValue())));
5033 case Intrinsic::bswap_i32:
5034 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5035 WorkList.push_back(II); // Dead?
5036 I.setOperand(0, II->getOperand(1));
5037 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5038 ByteSwap_32(CI->getZExtValue())));
5040 case Intrinsic::bswap_i64:
5041 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5042 WorkList.push_back(II); // Dead?
5043 I.setOperand(0, II->getOperand(1));
5044 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5045 ByteSwap_64(CI->getZExtValue())));
5049 } else { // Not a ICMP_EQ/ICMP_NE
5050 // If the LHS is a cast from an integral value of the same size, then
5051 // since we know the RHS is a constant, try to simlify.
5052 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5053 Value *CastOp = Cast->getOperand(0);
5054 const Type *SrcTy = CastOp->getType();
5055 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5056 if (SrcTy->isInteger() &&
5057 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5058 // If this is an unsigned comparison, try to make the comparison use
5059 // smaller constant values.
5060 switch (I.getPredicate()) {
5062 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5063 ConstantInt *CUI = cast<ConstantInt>(CI);
5064 if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
5065 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5066 ConstantInt::get(SrcTy, -1ULL));
5069 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5070 ConstantInt *CUI = cast<ConstantInt>(CI);
5071 if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
5072 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5073 Constant::getNullValue(SrcTy));
5083 // Handle icmp with constant RHS
5084 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5085 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5086 switch (LHSI->getOpcode()) {
5087 case Instruction::GetElementPtr:
5088 if (RHSC->isNullValue()) {
5089 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5090 bool isAllZeros = true;
5091 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5092 if (!isa<Constant>(LHSI->getOperand(i)) ||
5093 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5098 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5099 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5103 case Instruction::PHI:
5104 if (Instruction *NV = FoldOpIntoPhi(I))
5107 case Instruction::Select:
5108 // If either operand of the select is a constant, we can fold the
5109 // comparison into the select arms, which will cause one to be
5110 // constant folded and the select turned into a bitwise or.
5111 Value *Op1 = 0, *Op2 = 0;
5112 if (LHSI->hasOneUse()) {
5113 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5114 // Fold the known value into the constant operand.
5115 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5116 // Insert a new ICmp of the other select operand.
5117 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5118 LHSI->getOperand(2), RHSC,
5120 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5121 // Fold the known value into the constant operand.
5122 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5123 // Insert a new ICmp of the other select operand.
5124 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5125 LHSI->getOperand(1), RHSC,
5131 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5136 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5137 if (User *GEP = dyn_castGetElementPtr(Op0))
5138 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5140 if (User *GEP = dyn_castGetElementPtr(Op1))
5141 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5142 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5145 // Test to see if the operands of the icmp are casted versions of other
5146 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5148 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5149 if (isa<PointerType>(Op0->getType()) &&
5150 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5151 // We keep moving the cast from the left operand over to the right
5152 // operand, where it can often be eliminated completely.
5153 Op0 = CI->getOperand(0);
5155 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5156 // so eliminate it as well.
5157 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5158 Op1 = CI2->getOperand(0);
5160 // If Op1 is a constant, we can fold the cast into the constant.
5161 if (Op0->getType() != Op1->getType())
5162 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5163 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5165 // Otherwise, cast the RHS right before the icmp
5166 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5168 return new ICmpInst(I.getPredicate(), Op0, Op1);
5172 if (isa<CastInst>(Op0)) {
5173 // Handle the special case of: icmp (cast bool to X), <cst>
5174 // This comes up when you have code like
5177 // For generality, we handle any zero-extension of any operand comparison
5178 // with a constant or another cast from the same type.
5179 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5180 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5184 if (I.isEquality()) {
5185 Value *A, *B, *C, *D;
5186 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5187 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5188 Value *OtherVal = A == Op1 ? B : A;
5189 return new ICmpInst(I.getPredicate(), OtherVal,
5190 Constant::getNullValue(A->getType()));
5193 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5194 // A^c1 == C^c2 --> A == C^(c1^c2)
5195 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5196 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5197 if (Op1->hasOneUse()) {
5198 Constant *NC = ConstantExpr::getXor(C1, C2);
5199 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5200 return new ICmpInst(I.getPredicate(), A,
5201 InsertNewInstBefore(Xor, I));
5204 // A^B == A^D -> B == D
5205 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5206 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5207 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5208 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5212 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5213 (A == Op0 || B == Op0)) {
5214 // A == (A^B) -> B == 0
5215 Value *OtherVal = A == Op0 ? B : A;
5216 return new ICmpInst(I.getPredicate(), OtherVal,
5217 Constant::getNullValue(A->getType()));
5219 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5220 // (A-B) == A -> B == 0
5221 return new ICmpInst(I.getPredicate(), B,
5222 Constant::getNullValue(B->getType()));
5224 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5225 // A == (A-B) -> B == 0
5226 return new ICmpInst(I.getPredicate(), B,
5227 Constant::getNullValue(B->getType()));
5230 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5231 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5232 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5233 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5234 Value *X = 0, *Y = 0, *Z = 0;
5237 X = B; Y = D; Z = A;
5238 } else if (A == D) {
5239 X = B; Y = C; Z = A;
5240 } else if (B == C) {
5241 X = A; Y = D; Z = B;
5242 } else if (B == D) {
5243 X = A; Y = C; Z = B;
5246 if (X) { // Build (X^Y) & Z
5247 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5248 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5249 I.setOperand(0, Op1);
5250 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5255 return Changed ? &I : 0;
5258 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5259 // We only handle extending casts so far.
5261 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5262 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5263 Value *LHSCIOp = LHSCI->getOperand(0);
5264 const Type *SrcTy = LHSCIOp->getType();
5265 const Type *DestTy = LHSCI->getType();
5268 // We only handle extension cast instructions, so far. Enforce this.
5269 if (LHSCI->getOpcode() != Instruction::ZExt &&
5270 LHSCI->getOpcode() != Instruction::SExt)
5273 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5274 bool isSignedCmp = ICI.isSignedPredicate();
5276 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5277 // Not an extension from the same type?
5278 RHSCIOp = CI->getOperand(0);
5279 if (RHSCIOp->getType() != LHSCIOp->getType())
5282 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5283 // and the other is a zext), then we can't handle this.
5284 if (CI->getOpcode() != LHSCI->getOpcode())
5287 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5288 // then we can't handle this.
5289 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5292 // Okay, just insert a compare of the reduced operands now!
5293 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5296 // If we aren't dealing with a constant on the RHS, exit early
5297 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5301 // Compute the constant that would happen if we truncated to SrcTy then
5302 // reextended to DestTy.
5303 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5304 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5306 // If the re-extended constant didn't change...
5308 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5309 // For example, we might have:
5310 // %A = sext short %X to uint
5311 // %B = icmp ugt uint %A, 1330
5312 // It is incorrect to transform this into
5313 // %B = icmp ugt short %X, 1330
5314 // because %A may have negative value.
5316 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5317 // OR operation is EQ/NE.
5318 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5319 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5324 // The re-extended constant changed so the constant cannot be represented
5325 // in the shorter type. Consequently, we cannot emit a simple comparison.
5327 // First, handle some easy cases. We know the result cannot be equal at this
5328 // point so handle the ICI.isEquality() cases
5329 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5330 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5331 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5332 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5334 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5335 // should have been folded away previously and not enter in here.
5338 // We're performing a signed comparison.
5339 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5340 Result = ConstantInt::getFalse(); // X < (small) --> false
5342 Result = ConstantInt::getTrue(); // X < (large) --> true
5344 // We're performing an unsigned comparison.
5346 // We're performing an unsigned comp with a sign extended value.
5347 // This is true if the input is >= 0. [aka >s -1]
5348 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5349 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5350 NegOne, ICI.getName()), ICI);
5352 // Unsigned extend & unsigned compare -> always true.
5353 Result = ConstantInt::getTrue();
5357 // Finally, return the value computed.
5358 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5359 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5360 return ReplaceInstUsesWith(ICI, Result);
5362 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5363 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5364 "ICmp should be folded!");
5365 if (Constant *CI = dyn_cast<Constant>(Result))
5366 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5368 return BinaryOperator::createNot(Result);
5372 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5373 return commonShiftTransforms(I);
5376 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5377 return commonShiftTransforms(I);
5380 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5381 return commonShiftTransforms(I);
5384 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5385 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5386 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5388 // shl X, 0 == X and shr X, 0 == X
5389 // shl 0, X == 0 and shr 0, X == 0
5390 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5391 Op0 == Constant::getNullValue(Op0->getType()))
5392 return ReplaceInstUsesWith(I, Op0);
5394 if (isa<UndefValue>(Op0)) {
5395 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5396 return ReplaceInstUsesWith(I, Op0);
5397 else // undef << X -> 0, undef >>u X -> 0
5398 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5400 if (isa<UndefValue>(Op1)) {
5401 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5402 return ReplaceInstUsesWith(I, Op0);
5403 else // X << undef, X >>u undef -> 0
5404 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5407 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5408 if (I.getOpcode() == Instruction::AShr)
5409 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5410 if (CSI->isAllOnesValue())
5411 return ReplaceInstUsesWith(I, CSI);
5413 // Try to fold constant and into select arguments.
5414 if (isa<Constant>(Op0))
5415 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5416 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5419 // See if we can turn a signed shr into an unsigned shr.
5420 if (I.isArithmeticShift()) {
5421 if (MaskedValueIsZero(Op0,
5422 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5423 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5427 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5428 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5433 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5434 BinaryOperator &I) {
5435 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5437 // See if we can simplify any instructions used by the instruction whose sole
5438 // purpose is to compute bits we don't care about.
5439 uint64_t KnownZero, KnownOne;
5440 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
5441 KnownZero, KnownOne))
5444 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5445 // of a signed value.
5447 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5448 if (Op1->getZExtValue() >= TypeBits) {
5449 if (I.getOpcode() != Instruction::AShr)
5450 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5452 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5457 // ((X*C1) << C2) == (X * (C1 << C2))
5458 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5459 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5460 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5461 return BinaryOperator::createMul(BO->getOperand(0),
5462 ConstantExpr::getShl(BOOp, Op1));
5464 // Try to fold constant and into select arguments.
5465 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5466 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5468 if (isa<PHINode>(Op0))
5469 if (Instruction *NV = FoldOpIntoPhi(I))
5472 if (Op0->hasOneUse()) {
5473 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5474 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5477 switch (Op0BO->getOpcode()) {
5479 case Instruction::Add:
5480 case Instruction::And:
5481 case Instruction::Or:
5482 case Instruction::Xor: {
5483 // These operators commute.
5484 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5485 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5486 match(Op0BO->getOperand(1),
5487 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5488 Instruction *YS = BinaryOperator::createShl(
5489 Op0BO->getOperand(0), Op1,
5491 InsertNewInstBefore(YS, I); // (Y << C)
5493 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5494 Op0BO->getOperand(1)->getName());
5495 InsertNewInstBefore(X, I); // (X + (Y << C))
5496 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5497 C2 = ConstantExpr::getShl(C2, Op1);
5498 return BinaryOperator::createAnd(X, C2);
5501 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5502 Value *Op0BOOp1 = Op0BO->getOperand(1);
5503 if (isLeftShift && Op0BOOp1->hasOneUse() && V2 == Op1 &&
5505 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5506 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)-> hasOneUse()) {
5507 Instruction *YS = BinaryOperator::createShl(
5508 Op0BO->getOperand(0), Op1,
5510 InsertNewInstBefore(YS, I); // (Y << C)
5512 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5513 V1->getName()+".mask");
5514 InsertNewInstBefore(XM, I); // X & (CC << C)
5516 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5521 case Instruction::Sub: {
5522 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5523 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5524 match(Op0BO->getOperand(0),
5525 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5526 Instruction *YS = BinaryOperator::createShl(
5527 Op0BO->getOperand(1), Op1,
5529 InsertNewInstBefore(YS, I); // (Y << C)
5531 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5532 Op0BO->getOperand(0)->getName());
5533 InsertNewInstBefore(X, I); // (X + (Y << C))
5534 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5535 C2 = ConstantExpr::getShl(C2, Op1);
5536 return BinaryOperator::createAnd(X, C2);
5539 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5540 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5541 match(Op0BO->getOperand(0),
5542 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5543 m_ConstantInt(CC))) && V2 == Op1 &&
5544 cast<BinaryOperator>(Op0BO->getOperand(0))
5545 ->getOperand(0)->hasOneUse()) {
5546 Instruction *YS = BinaryOperator::createShl(
5547 Op0BO->getOperand(1), Op1,
5549 InsertNewInstBefore(YS, I); // (Y << C)
5551 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5552 V1->getName()+".mask");
5553 InsertNewInstBefore(XM, I); // X & (CC << C)
5555 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5563 // If the operand is an bitwise operator with a constant RHS, and the
5564 // shift is the only use, we can pull it out of the shift.
5565 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5566 bool isValid = true; // Valid only for And, Or, Xor
5567 bool highBitSet = false; // Transform if high bit of constant set?
5569 switch (Op0BO->getOpcode()) {
5570 default: isValid = false; break; // Do not perform transform!
5571 case Instruction::Add:
5572 isValid = isLeftShift;
5574 case Instruction::Or:
5575 case Instruction::Xor:
5578 case Instruction::And:
5583 // If this is a signed shift right, and the high bit is modified
5584 // by the logical operation, do not perform the transformation.
5585 // The highBitSet boolean indicates the value of the high bit of
5586 // the constant which would cause it to be modified for this
5589 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
5590 uint64_t Val = Op0C->getZExtValue();
5591 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5595 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5597 Instruction *NewShift =
5598 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
5599 InsertNewInstBefore(NewShift, I);
5600 NewShift->takeName(Op0BO);
5602 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5609 // Find out if this is a shift of a shift by a constant.
5610 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
5611 if (ShiftOp && !ShiftOp->isShift())
5614 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5615 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5616 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5617 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5618 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
5619 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
5620 Value *X = ShiftOp->getOperand(0);
5622 unsigned AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5623 if (AmtSum > I.getType()->getPrimitiveSizeInBits())
5624 AmtSum = I.getType()->getPrimitiveSizeInBits();
5626 const IntegerType *Ty = cast<IntegerType>(I.getType());
5628 // Check for (X << c1) << c2 and (X >> c1) >> c2
5629 if (I.getOpcode() == ShiftOp->getOpcode()) {
5630 return BinaryOperator::create(I.getOpcode(), X,
5631 ConstantInt::get(Ty, AmtSum));
5632 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
5633 I.getOpcode() == Instruction::AShr) {
5634 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
5635 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
5636 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
5637 I.getOpcode() == Instruction::LShr) {
5638 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
5639 Instruction *Shift =
5640 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
5641 InsertNewInstBefore(Shift, I);
5643 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5644 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5647 // Okay, if we get here, one shift must be left, and the other shift must be
5648 // right. See if the amounts are equal.
5649 if (ShiftAmt1 == ShiftAmt2) {
5650 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
5651 if (I.getOpcode() == Instruction::Shl) {
5652 uint64_t Mask = Ty->getBitMask() << ShiftAmt1;
5653 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
5655 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
5656 if (I.getOpcode() == Instruction::LShr) {
5657 uint64_t Mask = Ty->getBitMask() >> ShiftAmt1;
5658 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
5660 // We can simplify ((X << C) >>s C) into a trunc + sext.
5661 // NOTE: we could do this for any C, but that would make 'unusual' integer
5662 // types. For now, just stick to ones well-supported by the code
5664 const Type *SExtType = 0;
5665 switch (Ty->getBitWidth() - ShiftAmt1) {
5666 case 8 : SExtType = Type::Int8Ty; break;
5667 case 16: SExtType = Type::Int16Ty; break;
5668 case 32: SExtType = Type::Int32Ty; break;
5672 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
5673 InsertNewInstBefore(NewTrunc, I);
5674 return new SExtInst(NewTrunc, Ty);
5676 // Otherwise, we can't handle it yet.
5677 } else if (ShiftAmt1 < ShiftAmt2) {
5678 unsigned ShiftDiff = ShiftAmt2-ShiftAmt1;
5680 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
5681 if (I.getOpcode() == Instruction::Shl) {
5682 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5683 ShiftOp->getOpcode() == Instruction::AShr);
5684 Instruction *Shift =
5685 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5686 InsertNewInstBefore(Shift, I);
5688 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
5689 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5692 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
5693 if (I.getOpcode() == Instruction::LShr) {
5694 assert(ShiftOp->getOpcode() == Instruction::Shl);
5695 Instruction *Shift =
5696 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
5697 InsertNewInstBefore(Shift, I);
5699 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5700 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5703 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
5705 assert(ShiftAmt2 < ShiftAmt1);
5706 unsigned ShiftDiff = ShiftAmt1-ShiftAmt2;
5708 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
5709 if (I.getOpcode() == Instruction::Shl) {
5710 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5711 ShiftOp->getOpcode() == Instruction::AShr);
5712 Instruction *Shift =
5713 BinaryOperator::create(ShiftOp->getOpcode(), X,
5714 ConstantInt::get(Ty, ShiftDiff));
5715 InsertNewInstBefore(Shift, I);
5717 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
5718 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5721 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
5722 if (I.getOpcode() == Instruction::LShr) {
5723 assert(ShiftOp->getOpcode() == Instruction::Shl);
5724 Instruction *Shift =
5725 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5726 InsertNewInstBefore(Shift, I);
5728 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
5729 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
5732 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
5739 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5740 /// expression. If so, decompose it, returning some value X, such that Val is
5743 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5745 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5746 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5747 Offset = CI->getZExtValue();
5749 return ConstantInt::get(Type::Int32Ty, 0);
5750 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5751 if (I->getNumOperands() == 2) {
5752 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5753 if (I->getOpcode() == Instruction::Shl) {
5754 // This is a value scaled by '1 << the shift amt'.
5755 Scale = 1U << CUI->getZExtValue();
5757 return I->getOperand(0);
5758 } else if (I->getOpcode() == Instruction::Mul) {
5759 // This value is scaled by 'CUI'.
5760 Scale = CUI->getZExtValue();
5762 return I->getOperand(0);
5763 } else if (I->getOpcode() == Instruction::Add) {
5764 // We have X+C. Check to see if we really have (X*C2)+C1,
5765 // where C1 is divisible by C2.
5768 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5769 Offset += CUI->getZExtValue();
5770 if (SubScale > 1 && (Offset % SubScale == 0)) {
5779 // Otherwise, we can't look past this.
5786 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5787 /// try to eliminate the cast by moving the type information into the alloc.
5788 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5789 AllocationInst &AI) {
5790 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5791 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5793 // Remove any uses of AI that are dead.
5794 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5796 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5797 Instruction *User = cast<Instruction>(*UI++);
5798 if (isInstructionTriviallyDead(User)) {
5799 while (UI != E && *UI == User)
5800 ++UI; // If this instruction uses AI more than once, don't break UI.
5802 // Add operands to the worklist.
5803 AddUsesToWorkList(*User);
5805 DOUT << "IC: DCE: " << *User;
5807 User->eraseFromParent();
5808 removeFromWorkList(User);
5812 // Get the type really allocated and the type casted to.
5813 const Type *AllocElTy = AI.getAllocatedType();
5814 const Type *CastElTy = PTy->getElementType();
5815 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5817 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
5818 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
5819 if (CastElTyAlign < AllocElTyAlign) return 0;
5821 // If the allocation has multiple uses, only promote it if we are strictly
5822 // increasing the alignment of the resultant allocation. If we keep it the
5823 // same, we open the door to infinite loops of various kinds.
5824 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5826 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5827 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5828 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5830 // See if we can satisfy the modulus by pulling a scale out of the array
5832 unsigned ArraySizeScale, ArrayOffset;
5833 Value *NumElements = // See if the array size is a decomposable linear expr.
5834 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5836 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5838 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5839 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5841 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5846 // If the allocation size is constant, form a constant mul expression
5847 Amt = ConstantInt::get(Type::Int32Ty, Scale);
5848 if (isa<ConstantInt>(NumElements))
5849 Amt = ConstantExpr::getMul(
5850 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5851 // otherwise multiply the amount and the number of elements
5852 else if (Scale != 1) {
5853 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5854 Amt = InsertNewInstBefore(Tmp, AI);
5858 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5859 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
5860 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5861 Amt = InsertNewInstBefore(Tmp, AI);
5864 AllocationInst *New;
5865 if (isa<MallocInst>(AI))
5866 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
5868 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
5869 InsertNewInstBefore(New, AI);
5872 // If the allocation has multiple uses, insert a cast and change all things
5873 // that used it to use the new cast. This will also hack on CI, but it will
5875 if (!AI.hasOneUse()) {
5876 AddUsesToWorkList(AI);
5877 // New is the allocation instruction, pointer typed. AI is the original
5878 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5879 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5880 InsertNewInstBefore(NewCast, AI);
5881 AI.replaceAllUsesWith(NewCast);
5883 return ReplaceInstUsesWith(CI, New);
5886 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5887 /// and return it without inserting any new casts. This is used by code that
5888 /// tries to decide whether promoting or shrinking integer operations to wider
5889 /// or smaller types will allow us to eliminate a truncate or extend.
5890 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5891 int &NumCastsRemoved) {
5892 if (isa<Constant>(V)) return true;
5894 Instruction *I = dyn_cast<Instruction>(V);
5895 if (!I || !I->hasOneUse()) return false;
5897 switch (I->getOpcode()) {
5898 case Instruction::And:
5899 case Instruction::Or:
5900 case Instruction::Xor:
5901 // These operators can all arbitrarily be extended or truncated.
5902 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5903 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5904 case Instruction::AShr:
5905 case Instruction::LShr:
5906 case Instruction::Shl:
5907 // If this is just a bitcast changing the sign of the operation, we can
5908 // convert if the operand can be converted.
5909 if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
5910 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5912 case Instruction::Trunc:
5913 case Instruction::ZExt:
5914 case Instruction::SExt:
5915 case Instruction::BitCast:
5916 // If this is a cast from the destination type, we can trivially eliminate
5917 // it, and this will remove a cast overall.
5918 if (I->getOperand(0)->getType() == Ty) {
5919 // If the first operand is itself a cast, and is eliminable, do not count
5920 // this as an eliminable cast. We would prefer to eliminate those two
5922 if (isa<CastInst>(I->getOperand(0)))
5930 // TODO: Can handle more cases here.
5937 /// EvaluateInDifferentType - Given an expression that
5938 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5939 /// evaluate the expression.
5940 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
5942 if (Constant *C = dyn_cast<Constant>(V))
5943 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
5945 // Otherwise, it must be an instruction.
5946 Instruction *I = cast<Instruction>(V);
5947 Instruction *Res = 0;
5948 switch (I->getOpcode()) {
5949 case Instruction::And:
5950 case Instruction::Or:
5951 case Instruction::Xor: {
5952 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5953 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
5954 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5955 LHS, RHS, I->getName());
5958 case Instruction::AShr:
5959 case Instruction::LShr:
5960 case Instruction::Shl: {
5961 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5962 Res = BinaryOperator::create(Instruction::BinaryOps(I->getOpcode()), LHS,
5963 I->getOperand(1), I->getName());
5966 case Instruction::Trunc:
5967 case Instruction::ZExt:
5968 case Instruction::SExt:
5969 case Instruction::BitCast:
5970 // If the source type of the cast is the type we're trying for then we can
5971 // just return the source. There's no need to insert it because its not new.
5972 if (I->getOperand(0)->getType() == Ty)
5973 return I->getOperand(0);
5975 // Some other kind of cast, which shouldn't happen, so just ..
5978 // TODO: Can handle more cases here.
5979 assert(0 && "Unreachable!");
5983 return InsertNewInstBefore(Res, *I);
5986 /// @brief Implement the transforms common to all CastInst visitors.
5987 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
5988 Value *Src = CI.getOperand(0);
5990 // Casting undef to anything results in undef so might as just replace it and
5991 // get rid of the cast.
5992 if (isa<UndefValue>(Src)) // cast undef -> undef
5993 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5995 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
5996 // eliminate it now.
5997 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5998 if (Instruction::CastOps opc =
5999 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6000 // The first cast (CSrc) is eliminable so we need to fix up or replace
6001 // the second cast (CI). CSrc will then have a good chance of being dead.
6002 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6006 // If casting the result of a getelementptr instruction with no offset, turn
6007 // this into a cast of the original pointer!
6009 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6010 bool AllZeroOperands = true;
6011 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
6012 if (!isa<Constant>(GEP->getOperand(i)) ||
6013 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
6014 AllZeroOperands = false;
6017 if (AllZeroOperands) {
6018 // Changing the cast operand is usually not a good idea but it is safe
6019 // here because the pointer operand is being replaced with another
6020 // pointer operand so the opcode doesn't need to change.
6021 CI.setOperand(0, GEP->getOperand(0));
6026 // If we are casting a malloc or alloca to a pointer to a type of the same
6027 // size, rewrite the allocation instruction to allocate the "right" type.
6028 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6029 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6032 // If we are casting a select then fold the cast into the select
6033 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6034 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6037 // If we are casting a PHI then fold the cast into the PHI
6038 if (isa<PHINode>(Src))
6039 if (Instruction *NV = FoldOpIntoPhi(CI))
6045 /// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
6046 /// integers. This function implements the common transforms for all those
6048 /// @brief Implement the transforms common to CastInst with integer operands
6049 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6050 if (Instruction *Result = commonCastTransforms(CI))
6053 Value *Src = CI.getOperand(0);
6054 const Type *SrcTy = Src->getType();
6055 const Type *DestTy = CI.getType();
6056 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6057 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6059 // See if we can simplify any instructions used by the LHS whose sole
6060 // purpose is to compute bits we don't care about.
6061 uint64_t KnownZero = 0, KnownOne = 0;
6062 if (SimplifyDemandedBits(&CI, cast<IntegerType>(DestTy)->getBitMask(),
6063 KnownZero, KnownOne))
6066 // If the source isn't an instruction or has more than one use then we
6067 // can't do anything more.
6068 Instruction *SrcI = dyn_cast<Instruction>(Src);
6069 if (!SrcI || !Src->hasOneUse())
6072 // Attempt to propagate the cast into the instruction.
6073 int NumCastsRemoved = 0;
6074 if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
6075 // If this cast is a truncate, evaluting in a different type always
6076 // eliminates the cast, so it is always a win. If this is a noop-cast
6077 // this just removes a noop cast which isn't pointful, but simplifies
6078 // the code. If this is a zero-extension, we need to do an AND to
6079 // maintain the clear top-part of the computation, so we require that
6080 // the input have eliminated at least one cast. If this is a sign
6081 // extension, we insert two new casts (to do the extension) so we
6082 // require that two casts have been eliminated.
6083 bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
6085 switch (CI.getOpcode()) {
6086 case Instruction::Trunc:
6089 case Instruction::ZExt:
6090 DoXForm = NumCastsRemoved >= 1;
6092 case Instruction::SExt:
6093 DoXForm = NumCastsRemoved >= 2;
6095 case Instruction::BitCast:
6099 // All the others use floating point so we shouldn't actually
6100 // get here because of the check above.
6101 assert(!"Unknown cast type .. unreachable");
6107 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6108 CI.getOpcode() == Instruction::SExt);
6109 assert(Res->getType() == DestTy);
6110 switch (CI.getOpcode()) {
6111 default: assert(0 && "Unknown cast type!");
6112 case Instruction::Trunc:
6113 case Instruction::BitCast:
6114 // Just replace this cast with the result.
6115 return ReplaceInstUsesWith(CI, Res);
6116 case Instruction::ZExt: {
6117 // We need to emit an AND to clear the high bits.
6118 assert(SrcBitSize < DestBitSize && "Not a zext?");
6120 ConstantInt::get(Type::Int64Ty, (1ULL << SrcBitSize)-1);
6121 if (DestBitSize < 64)
6122 C = ConstantExpr::getTrunc(C, DestTy);
6123 return BinaryOperator::createAnd(Res, C);
6125 case Instruction::SExt:
6126 // We need to emit a cast to truncate, then a cast to sext.
6127 return CastInst::create(Instruction::SExt,
6128 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6134 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6135 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6137 switch (SrcI->getOpcode()) {
6138 case Instruction::Add:
6139 case Instruction::Mul:
6140 case Instruction::And:
6141 case Instruction::Or:
6142 case Instruction::Xor:
6143 // If we are discarding information, or just changing the sign,
6145 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6146 // Don't insert two casts if they cannot be eliminated. We allow
6147 // two casts to be inserted if the sizes are the same. This could
6148 // only be converting signedness, which is a noop.
6149 if (DestBitSize == SrcBitSize ||
6150 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6151 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6152 Instruction::CastOps opcode = CI.getOpcode();
6153 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6154 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6155 return BinaryOperator::create(
6156 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6160 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6161 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6162 SrcI->getOpcode() == Instruction::Xor &&
6163 Op1 == ConstantInt::getTrue() &&
6164 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6165 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6166 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6169 case Instruction::SDiv:
6170 case Instruction::UDiv:
6171 case Instruction::SRem:
6172 case Instruction::URem:
6173 // If we are just changing the sign, rewrite.
6174 if (DestBitSize == SrcBitSize) {
6175 // Don't insert two casts if they cannot be eliminated. We allow
6176 // two casts to be inserted if the sizes are the same. This could
6177 // only be converting signedness, which is a noop.
6178 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6179 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6180 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6182 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6184 return BinaryOperator::create(
6185 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6190 case Instruction::Shl:
6191 // Allow changing the sign of the source operand. Do not allow
6192 // changing the size of the shift, UNLESS the shift amount is a
6193 // constant. We must not change variable sized shifts to a smaller
6194 // size, because it is undefined to shift more bits out than exist
6196 if (DestBitSize == SrcBitSize ||
6197 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6198 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6199 Instruction::BitCast : Instruction::Trunc);
6200 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6201 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6202 return BinaryOperator::createShl(Op0c, Op1c);
6205 case Instruction::AShr:
6206 // If this is a signed shr, and if all bits shifted in are about to be
6207 // truncated off, turn it into an unsigned shr to allow greater
6209 if (DestBitSize < SrcBitSize &&
6210 isa<ConstantInt>(Op1)) {
6211 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6212 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6213 // Insert the new logical shift right.
6214 return BinaryOperator::createLShr(Op0, Op1);
6219 case Instruction::ICmp:
6220 // If we are just checking for a icmp eq of a single bit and casting it
6221 // to an integer, then shift the bit to the appropriate place and then
6222 // cast to integer to avoid the comparison.
6223 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6224 uint64_t Op1CV = Op1C->getZExtValue();
6225 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6226 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6227 // cast (X == 1) to int --> X iff X has only the low bit set.
6228 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6229 // cast (X != 0) to int --> X iff X has only the low bit set.
6230 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6231 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6232 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6233 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
6234 // If Op1C some other power of two, convert:
6235 uint64_t KnownZero, KnownOne;
6236 uint64_t TypeMask = Op1C->getType()->getBitMask();
6237 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6239 // This only works for EQ and NE
6240 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6241 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6244 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
6245 bool isNE = pred == ICmpInst::ICMP_NE;
6246 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
6247 // (X&4) == 2 --> false
6248 // (X&4) != 2 --> true
6249 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6250 Res = ConstantExpr::getZExt(Res, CI.getType());
6251 return ReplaceInstUsesWith(CI, Res);
6254 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
6257 // Perform a logical shr by shiftamt.
6258 // Insert the shift to put the result in the low bit.
6259 In = InsertNewInstBefore(
6260 BinaryOperator::createLShr(In,
6261 ConstantInt::get(In->getType(), ShiftAmt),
6262 In->getName()+".lobit"), CI);
6265 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6266 Constant *One = ConstantInt::get(In->getType(), 1);
6267 In = BinaryOperator::createXor(In, One, "tmp");
6268 InsertNewInstBefore(cast<Instruction>(In), CI);
6271 if (CI.getType() == In->getType())
6272 return ReplaceInstUsesWith(CI, In);
6274 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6283 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6284 if (Instruction *Result = commonIntCastTransforms(CI))
6287 Value *Src = CI.getOperand(0);
6288 const Type *Ty = CI.getType();
6289 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6291 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6292 switch (SrcI->getOpcode()) {
6294 case Instruction::LShr:
6295 // We can shrink lshr to something smaller if we know the bits shifted in
6296 // are already zeros.
6297 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6298 unsigned ShAmt = ShAmtV->getZExtValue();
6300 // Get a mask for the bits shifting in.
6301 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
6302 Value* SrcIOp0 = SrcI->getOperand(0);
6303 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6304 if (ShAmt >= DestBitWidth) // All zeros.
6305 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6307 // Okay, we can shrink this. Truncate the input, then return a new
6309 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6310 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6312 return BinaryOperator::createLShr(V1, V2);
6314 } else { // This is a variable shr.
6316 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6317 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6318 // loop-invariant and CSE'd.
6319 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6320 Value *One = ConstantInt::get(SrcI->getType(), 1);
6322 Value *V = InsertNewInstBefore(
6323 BinaryOperator::createShl(One, SrcI->getOperand(1),
6325 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6326 SrcI->getOperand(0),
6328 Value *Zero = Constant::getNullValue(V->getType());
6329 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6339 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6340 // If one of the common conversion will work ..
6341 if (Instruction *Result = commonIntCastTransforms(CI))
6344 Value *Src = CI.getOperand(0);
6346 // If this is a cast of a cast
6347 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6348 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6349 // types and if the sizes are just right we can convert this into a logical
6350 // 'and' which will be much cheaper than the pair of casts.
6351 if (isa<TruncInst>(CSrc)) {
6352 // Get the sizes of the types involved
6353 Value *A = CSrc->getOperand(0);
6354 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6355 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6356 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6357 // If we're actually extending zero bits and the trunc is a no-op
6358 if (MidSize < DstSize && SrcSize == DstSize) {
6359 // Replace both of the casts with an And of the type mask.
6360 uint64_t AndValue = cast<IntegerType>(CSrc->getType())->getBitMask();
6361 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6363 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6364 // Unfortunately, if the type changed, we need to cast it back.
6365 if (And->getType() != CI.getType()) {
6366 And->setName(CSrc->getName()+".mask");
6367 InsertNewInstBefore(And, CI);
6368 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6378 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6379 return commonIntCastTransforms(CI);
6382 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6383 return commonCastTransforms(CI);
6386 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6387 return commonCastTransforms(CI);
6390 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6391 return commonCastTransforms(CI);
6394 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6395 return commonCastTransforms(CI);
6398 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6399 return commonCastTransforms(CI);
6402 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6403 return commonCastTransforms(CI);
6406 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6407 return commonCastTransforms(CI);
6410 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6411 return commonCastTransforms(CI);
6414 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6416 // If the operands are integer typed then apply the integer transforms,
6417 // otherwise just apply the common ones.
6418 Value *Src = CI.getOperand(0);
6419 const Type *SrcTy = Src->getType();
6420 const Type *DestTy = CI.getType();
6422 if (SrcTy->isInteger() && DestTy->isInteger()) {
6423 if (Instruction *Result = commonIntCastTransforms(CI))
6426 if (Instruction *Result = commonCastTransforms(CI))
6431 // Get rid of casts from one type to the same type. These are useless and can
6432 // be replaced by the operand.
6433 if (DestTy == Src->getType())
6434 return ReplaceInstUsesWith(CI, Src);
6436 // If the source and destination are pointers, and this cast is equivalent to
6437 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6438 // This can enhance SROA and other transforms that want type-safe pointers.
6439 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6440 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6441 const Type *DstElTy = DstPTy->getElementType();
6442 const Type *SrcElTy = SrcPTy->getElementType();
6444 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6445 unsigned NumZeros = 0;
6446 while (SrcElTy != DstElTy &&
6447 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6448 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6449 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6453 // If we found a path from the src to dest, create the getelementptr now.
6454 if (SrcElTy == DstElTy) {
6455 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
6456 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
6461 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6462 if (SVI->hasOneUse()) {
6463 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6464 // a bitconvert to a vector with the same # elts.
6465 if (isa<VectorType>(DestTy) &&
6466 cast<VectorType>(DestTy)->getNumElements() ==
6467 SVI->getType()->getNumElements()) {
6469 // If either of the operands is a cast from CI.getType(), then
6470 // evaluating the shuffle in the casted destination's type will allow
6471 // us to eliminate at least one cast.
6472 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6473 Tmp->getOperand(0)->getType() == DestTy) ||
6474 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6475 Tmp->getOperand(0)->getType() == DestTy)) {
6476 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6477 SVI->getOperand(0), DestTy, &CI);
6478 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6479 SVI->getOperand(1), DestTy, &CI);
6480 // Return a new shuffle vector. Use the same element ID's, as we
6481 // know the vector types match #elts.
6482 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6490 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6492 /// %D = select %cond, %C, %A
6494 /// %C = select %cond, %B, 0
6497 /// Assuming that the specified instruction is an operand to the select, return
6498 /// a bitmask indicating which operands of this instruction are foldable if they
6499 /// equal the other incoming value of the select.
6501 static unsigned GetSelectFoldableOperands(Instruction *I) {
6502 switch (I->getOpcode()) {
6503 case Instruction::Add:
6504 case Instruction::Mul:
6505 case Instruction::And:
6506 case Instruction::Or:
6507 case Instruction::Xor:
6508 return 3; // Can fold through either operand.
6509 case Instruction::Sub: // Can only fold on the amount subtracted.
6510 case Instruction::Shl: // Can only fold on the shift amount.
6511 case Instruction::LShr:
6512 case Instruction::AShr:
6515 return 0; // Cannot fold
6519 /// GetSelectFoldableConstant - For the same transformation as the previous
6520 /// function, return the identity constant that goes into the select.
6521 static Constant *GetSelectFoldableConstant(Instruction *I) {
6522 switch (I->getOpcode()) {
6523 default: assert(0 && "This cannot happen!"); abort();
6524 case Instruction::Add:
6525 case Instruction::Sub:
6526 case Instruction::Or:
6527 case Instruction::Xor:
6528 case Instruction::Shl:
6529 case Instruction::LShr:
6530 case Instruction::AShr:
6531 return Constant::getNullValue(I->getType());
6532 case Instruction::And:
6533 return ConstantInt::getAllOnesValue(I->getType());
6534 case Instruction::Mul:
6535 return ConstantInt::get(I->getType(), 1);
6539 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6540 /// have the same opcode and only one use each. Try to simplify this.
6541 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6543 if (TI->getNumOperands() == 1) {
6544 // If this is a non-volatile load or a cast from the same type,
6547 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6550 return 0; // unknown unary op.
6553 // Fold this by inserting a select from the input values.
6554 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6555 FI->getOperand(0), SI.getName()+".v");
6556 InsertNewInstBefore(NewSI, SI);
6557 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6561 // Only handle binary operators here.
6562 if (!isa<BinaryOperator>(TI))
6565 // Figure out if the operations have any operands in common.
6566 Value *MatchOp, *OtherOpT, *OtherOpF;
6568 if (TI->getOperand(0) == FI->getOperand(0)) {
6569 MatchOp = TI->getOperand(0);
6570 OtherOpT = TI->getOperand(1);
6571 OtherOpF = FI->getOperand(1);
6572 MatchIsOpZero = true;
6573 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6574 MatchOp = TI->getOperand(1);
6575 OtherOpT = TI->getOperand(0);
6576 OtherOpF = FI->getOperand(0);
6577 MatchIsOpZero = false;
6578 } else if (!TI->isCommutative()) {
6580 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6581 MatchOp = TI->getOperand(0);
6582 OtherOpT = TI->getOperand(1);
6583 OtherOpF = FI->getOperand(0);
6584 MatchIsOpZero = true;
6585 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6586 MatchOp = TI->getOperand(1);
6587 OtherOpT = TI->getOperand(0);
6588 OtherOpF = FI->getOperand(1);
6589 MatchIsOpZero = true;
6594 // If we reach here, they do have operations in common.
6595 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6596 OtherOpF, SI.getName()+".v");
6597 InsertNewInstBefore(NewSI, SI);
6599 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6601 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6603 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6605 assert(0 && "Shouldn't get here");
6609 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6610 Value *CondVal = SI.getCondition();
6611 Value *TrueVal = SI.getTrueValue();
6612 Value *FalseVal = SI.getFalseValue();
6614 // select true, X, Y -> X
6615 // select false, X, Y -> Y
6616 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6617 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
6619 // select C, X, X -> X
6620 if (TrueVal == FalseVal)
6621 return ReplaceInstUsesWith(SI, TrueVal);
6623 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6624 return ReplaceInstUsesWith(SI, FalseVal);
6625 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6626 return ReplaceInstUsesWith(SI, TrueVal);
6627 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6628 if (isa<Constant>(TrueVal))
6629 return ReplaceInstUsesWith(SI, TrueVal);
6631 return ReplaceInstUsesWith(SI, FalseVal);
6634 if (SI.getType() == Type::Int1Ty) {
6635 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
6636 if (C->getZExtValue()) {
6637 // Change: A = select B, true, C --> A = or B, C
6638 return BinaryOperator::createOr(CondVal, FalseVal);
6640 // Change: A = select B, false, C --> A = and !B, C
6642 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6643 "not."+CondVal->getName()), SI);
6644 return BinaryOperator::createAnd(NotCond, FalseVal);
6646 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
6647 if (C->getZExtValue() == false) {
6648 // Change: A = select B, C, false --> A = and B, C
6649 return BinaryOperator::createAnd(CondVal, TrueVal);
6651 // Change: A = select B, C, true --> A = or !B, C
6653 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6654 "not."+CondVal->getName()), SI);
6655 return BinaryOperator::createOr(NotCond, TrueVal);
6660 // Selecting between two integer constants?
6661 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6662 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6663 // select C, 1, 0 -> cast C to int
6664 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6665 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6666 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6667 // select C, 0, 1 -> cast !C to int
6669 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6670 "not."+CondVal->getName()), SI);
6671 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6674 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6676 // (x <s 0) ? -1 : 0 -> ashr x, 31
6677 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6678 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6679 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6680 bool CanXForm = false;
6681 if (IC->isSignedPredicate())
6682 CanXForm = CmpCst->isNullValue() &&
6683 IC->getPredicate() == ICmpInst::ICMP_SLT;
6685 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6686 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6687 IC->getPredicate() == ICmpInst::ICMP_UGT;
6691 // The comparison constant and the result are not neccessarily the
6692 // same width. Make an all-ones value by inserting a AShr.
6693 Value *X = IC->getOperand(0);
6694 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6695 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
6696 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
6698 InsertNewInstBefore(SRA, SI);
6700 // Finally, convert to the type of the select RHS. We figure out
6701 // if this requires a SExt, Trunc or BitCast based on the sizes.
6702 Instruction::CastOps opc = Instruction::BitCast;
6703 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6704 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6705 if (SRASize < SISize)
6706 opc = Instruction::SExt;
6707 else if (SRASize > SISize)
6708 opc = Instruction::Trunc;
6709 return CastInst::create(opc, SRA, SI.getType());
6714 // If one of the constants is zero (we know they can't both be) and we
6715 // have a fcmp instruction with zero, and we have an 'and' with the
6716 // non-constant value, eliminate this whole mess. This corresponds to
6717 // cases like this: ((X & 27) ? 27 : 0)
6718 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6719 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6720 cast<Constant>(IC->getOperand(1))->isNullValue())
6721 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6722 if (ICA->getOpcode() == Instruction::And &&
6723 isa<ConstantInt>(ICA->getOperand(1)) &&
6724 (ICA->getOperand(1) == TrueValC ||
6725 ICA->getOperand(1) == FalseValC) &&
6726 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6727 // Okay, now we know that everything is set up, we just don't
6728 // know whether we have a icmp_ne or icmp_eq and whether the
6729 // true or false val is the zero.
6730 bool ShouldNotVal = !TrueValC->isNullValue();
6731 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6734 V = InsertNewInstBefore(BinaryOperator::create(
6735 Instruction::Xor, V, ICA->getOperand(1)), SI);
6736 return ReplaceInstUsesWith(SI, V);
6741 // See if we are selecting two values based on a comparison of the two values.
6742 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6743 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6744 // Transform (X == Y) ? X : Y -> Y
6745 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6746 return ReplaceInstUsesWith(SI, FalseVal);
6747 // Transform (X != Y) ? X : Y -> X
6748 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6749 return ReplaceInstUsesWith(SI, TrueVal);
6750 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6752 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6753 // Transform (X == Y) ? Y : X -> X
6754 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6755 return ReplaceInstUsesWith(SI, FalseVal);
6756 // Transform (X != Y) ? Y : X -> Y
6757 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6758 return ReplaceInstUsesWith(SI, TrueVal);
6759 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6763 // See if we are selecting two values based on a comparison of the two values.
6764 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6765 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6766 // Transform (X == Y) ? X : Y -> Y
6767 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6768 return ReplaceInstUsesWith(SI, FalseVal);
6769 // Transform (X != Y) ? X : Y -> X
6770 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6771 return ReplaceInstUsesWith(SI, TrueVal);
6772 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6774 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6775 // Transform (X == Y) ? Y : X -> X
6776 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6777 return ReplaceInstUsesWith(SI, FalseVal);
6778 // Transform (X != Y) ? Y : X -> Y
6779 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6780 return ReplaceInstUsesWith(SI, TrueVal);
6781 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6785 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6786 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6787 if (TI->hasOneUse() && FI->hasOneUse()) {
6788 Instruction *AddOp = 0, *SubOp = 0;
6790 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6791 if (TI->getOpcode() == FI->getOpcode())
6792 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6795 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6796 // even legal for FP.
6797 if (TI->getOpcode() == Instruction::Sub &&
6798 FI->getOpcode() == Instruction::Add) {
6799 AddOp = FI; SubOp = TI;
6800 } else if (FI->getOpcode() == Instruction::Sub &&
6801 TI->getOpcode() == Instruction::Add) {
6802 AddOp = TI; SubOp = FI;
6806 Value *OtherAddOp = 0;
6807 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6808 OtherAddOp = AddOp->getOperand(1);
6809 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6810 OtherAddOp = AddOp->getOperand(0);
6814 // So at this point we know we have (Y -> OtherAddOp):
6815 // select C, (add X, Y), (sub X, Z)
6816 Value *NegVal; // Compute -Z
6817 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6818 NegVal = ConstantExpr::getNeg(C);
6820 NegVal = InsertNewInstBefore(
6821 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6824 Value *NewTrueOp = OtherAddOp;
6825 Value *NewFalseOp = NegVal;
6827 std::swap(NewTrueOp, NewFalseOp);
6828 Instruction *NewSel =
6829 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6831 NewSel = InsertNewInstBefore(NewSel, SI);
6832 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6837 // See if we can fold the select into one of our operands.
6838 if (SI.getType()->isInteger()) {
6839 // See the comment above GetSelectFoldableOperands for a description of the
6840 // transformation we are doing here.
6841 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6842 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6843 !isa<Constant>(FalseVal))
6844 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6845 unsigned OpToFold = 0;
6846 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6848 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6853 Constant *C = GetSelectFoldableConstant(TVI);
6854 Instruction *NewSel =
6855 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
6856 InsertNewInstBefore(NewSel, SI);
6857 NewSel->takeName(TVI);
6858 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6859 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6861 assert(0 && "Unknown instruction!!");
6866 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6867 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6868 !isa<Constant>(TrueVal))
6869 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6870 unsigned OpToFold = 0;
6871 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6873 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6878 Constant *C = GetSelectFoldableConstant(FVI);
6879 Instruction *NewSel =
6880 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
6881 InsertNewInstBefore(NewSel, SI);
6882 NewSel->takeName(FVI);
6883 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6884 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6886 assert(0 && "Unknown instruction!!");
6891 if (BinaryOperator::isNot(CondVal)) {
6892 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6893 SI.setOperand(1, FalseVal);
6894 SI.setOperand(2, TrueVal);
6901 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6902 /// determine, return it, otherwise return 0.
6903 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6904 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6905 unsigned Align = GV->getAlignment();
6906 if (Align == 0 && TD)
6907 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
6909 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6910 unsigned Align = AI->getAlignment();
6911 if (Align == 0 && TD) {
6912 if (isa<AllocaInst>(AI))
6913 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
6914 else if (isa<MallocInst>(AI)) {
6915 // Malloc returns maximally aligned memory.
6916 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
6919 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
6922 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
6926 } else if (isa<BitCastInst>(V) ||
6927 (isa<ConstantExpr>(V) &&
6928 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
6929 User *CI = cast<User>(V);
6930 if (isa<PointerType>(CI->getOperand(0)->getType()))
6931 return GetKnownAlignment(CI->getOperand(0), TD);
6933 } else if (isa<GetElementPtrInst>(V) ||
6934 (isa<ConstantExpr>(V) &&
6935 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
6936 User *GEPI = cast<User>(V);
6937 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
6938 if (BaseAlignment == 0) return 0;
6940 // If all indexes are zero, it is just the alignment of the base pointer.
6941 bool AllZeroOperands = true;
6942 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
6943 if (!isa<Constant>(GEPI->getOperand(i)) ||
6944 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
6945 AllZeroOperands = false;
6948 if (AllZeroOperands)
6949 return BaseAlignment;
6951 // Otherwise, if the base alignment is >= the alignment we expect for the
6952 // base pointer type, then we know that the resultant pointer is aligned at
6953 // least as much as its type requires.
6956 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
6957 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
6958 if (TD->getABITypeAlignment(PtrTy->getElementType())
6960 const Type *GEPTy = GEPI->getType();
6961 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
6962 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
6970 /// visitCallInst - CallInst simplification. This mostly only handles folding
6971 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
6972 /// the heavy lifting.
6974 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
6975 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
6976 if (!II) return visitCallSite(&CI);
6978 // Intrinsics cannot occur in an invoke, so handle them here instead of in
6980 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
6981 bool Changed = false;
6983 // memmove/cpy/set of zero bytes is a noop.
6984 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
6985 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
6987 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
6988 if (CI->getZExtValue() == 1) {
6989 // Replace the instruction with just byte operations. We would
6990 // transform other cases to loads/stores, but we don't know if
6991 // alignment is sufficient.
6995 // If we have a memmove and the source operation is a constant global,
6996 // then the source and dest pointers can't alias, so we can change this
6997 // into a call to memcpy.
6998 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
6999 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7000 if (GVSrc->isConstant()) {
7001 Module *M = CI.getParent()->getParent()->getParent();
7003 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7005 Name = "llvm.memcpy.i32";
7007 Name = "llvm.memcpy.i64";
7008 Constant *MemCpy = M->getOrInsertFunction(Name,
7009 CI.getCalledFunction()->getFunctionType());
7010 CI.setOperand(0, MemCpy);
7015 // If we can determine a pointer alignment that is bigger than currently
7016 // set, update the alignment.
7017 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7018 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7019 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7020 unsigned Align = std::min(Alignment1, Alignment2);
7021 if (MI->getAlignment()->getZExtValue() < Align) {
7022 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7025 } else if (isa<MemSetInst>(MI)) {
7026 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7027 if (MI->getAlignment()->getZExtValue() < Alignment) {
7028 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7033 if (Changed) return II;
7035 switch (II->getIntrinsicID()) {
7037 case Intrinsic::ppc_altivec_lvx:
7038 case Intrinsic::ppc_altivec_lvxl:
7039 case Intrinsic::x86_sse_loadu_ps:
7040 case Intrinsic::x86_sse2_loadu_pd:
7041 case Intrinsic::x86_sse2_loadu_dq:
7042 // Turn PPC lvx -> load if the pointer is known aligned.
7043 // Turn X86 loadups -> load if the pointer is known aligned.
7044 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7045 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7046 PointerType::get(II->getType()), CI);
7047 return new LoadInst(Ptr);
7050 case Intrinsic::ppc_altivec_stvx:
7051 case Intrinsic::ppc_altivec_stvxl:
7052 // Turn stvx -> store if the pointer is known aligned.
7053 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7054 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7055 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7057 return new StoreInst(II->getOperand(1), Ptr);
7060 case Intrinsic::x86_sse_storeu_ps:
7061 case Intrinsic::x86_sse2_storeu_pd:
7062 case Intrinsic::x86_sse2_storeu_dq:
7063 case Intrinsic::x86_sse2_storel_dq:
7064 // Turn X86 storeu -> store if the pointer is known aligned.
7065 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7066 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7067 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7069 return new StoreInst(II->getOperand(2), Ptr);
7073 case Intrinsic::x86_sse_cvttss2si: {
7074 // These intrinsics only demands the 0th element of its input vector. If
7075 // we can simplify the input based on that, do so now.
7077 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7079 II->setOperand(1, V);
7085 case Intrinsic::ppc_altivec_vperm:
7086 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7087 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7088 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7090 // Check that all of the elements are integer constants or undefs.
7091 bool AllEltsOk = true;
7092 for (unsigned i = 0; i != 16; ++i) {
7093 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7094 !isa<UndefValue>(Mask->getOperand(i))) {
7101 // Cast the input vectors to byte vectors.
7102 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7103 II->getOperand(1), Mask->getType(), CI);
7104 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7105 II->getOperand(2), Mask->getType(), CI);
7106 Value *Result = UndefValue::get(Op0->getType());
7108 // Only extract each element once.
7109 Value *ExtractedElts[32];
7110 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7112 for (unsigned i = 0; i != 16; ++i) {
7113 if (isa<UndefValue>(Mask->getOperand(i)))
7115 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7116 Idx &= 31; // Match the hardware behavior.
7118 if (ExtractedElts[Idx] == 0) {
7120 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7121 InsertNewInstBefore(Elt, CI);
7122 ExtractedElts[Idx] = Elt;
7125 // Insert this value into the result vector.
7126 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7127 InsertNewInstBefore(cast<Instruction>(Result), CI);
7129 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7134 case Intrinsic::stackrestore: {
7135 // If the save is right next to the restore, remove the restore. This can
7136 // happen when variable allocas are DCE'd.
7137 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7138 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7139 BasicBlock::iterator BI = SS;
7141 return EraseInstFromFunction(CI);
7145 // If the stack restore is in a return/unwind block and if there are no
7146 // allocas or calls between the restore and the return, nuke the restore.
7147 TerminatorInst *TI = II->getParent()->getTerminator();
7148 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7149 BasicBlock::iterator BI = II;
7150 bool CannotRemove = false;
7151 for (++BI; &*BI != TI; ++BI) {
7152 if (isa<AllocaInst>(BI) ||
7153 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7154 CannotRemove = true;
7159 return EraseInstFromFunction(CI);
7166 return visitCallSite(II);
7169 // InvokeInst simplification
7171 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7172 return visitCallSite(&II);
7175 // visitCallSite - Improvements for call and invoke instructions.
7177 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7178 bool Changed = false;
7180 // If the callee is a constexpr cast of a function, attempt to move the cast
7181 // to the arguments of the call/invoke.
7182 if (transformConstExprCastCall(CS)) return 0;
7184 Value *Callee = CS.getCalledValue();
7186 if (Function *CalleeF = dyn_cast<Function>(Callee))
7187 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7188 Instruction *OldCall = CS.getInstruction();
7189 // If the call and callee calling conventions don't match, this call must
7190 // be unreachable, as the call is undefined.
7191 new StoreInst(ConstantInt::getTrue(),
7192 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7193 if (!OldCall->use_empty())
7194 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7195 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7196 return EraseInstFromFunction(*OldCall);
7200 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7201 // This instruction is not reachable, just remove it. We insert a store to
7202 // undef so that we know that this code is not reachable, despite the fact
7203 // that we can't modify the CFG here.
7204 new StoreInst(ConstantInt::getTrue(),
7205 UndefValue::get(PointerType::get(Type::Int1Ty)),
7206 CS.getInstruction());
7208 if (!CS.getInstruction()->use_empty())
7209 CS.getInstruction()->
7210 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7212 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7213 // Don't break the CFG, insert a dummy cond branch.
7214 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7215 ConstantInt::getTrue(), II);
7217 return EraseInstFromFunction(*CS.getInstruction());
7220 const PointerType *PTy = cast<PointerType>(Callee->getType());
7221 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7222 if (FTy->isVarArg()) {
7223 // See if we can optimize any arguments passed through the varargs area of
7225 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7226 E = CS.arg_end(); I != E; ++I)
7227 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7228 // If this cast does not effect the value passed through the varargs
7229 // area, we can eliminate the use of the cast.
7230 Value *Op = CI->getOperand(0);
7231 if (CI->isLosslessCast()) {
7238 return Changed ? CS.getInstruction() : 0;
7241 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7242 // attempt to move the cast to the arguments of the call/invoke.
7244 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7245 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7246 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7247 if (CE->getOpcode() != Instruction::BitCast ||
7248 !isa<Function>(CE->getOperand(0)))
7250 Function *Callee = cast<Function>(CE->getOperand(0));
7251 Instruction *Caller = CS.getInstruction();
7253 // Okay, this is a cast from a function to a different type. Unless doing so
7254 // would cause a type conversion of one of our arguments, change this call to
7255 // be a direct call with arguments casted to the appropriate types.
7257 const FunctionType *FT = Callee->getFunctionType();
7258 const Type *OldRetTy = Caller->getType();
7260 // Check to see if we are changing the return type...
7261 if (OldRetTy != FT->getReturnType()) {
7262 if (Callee->isDeclaration() && !Caller->use_empty() &&
7263 OldRetTy != FT->getReturnType() &&
7264 // Conversion is ok if changing from pointer to int of same size.
7265 !(isa<PointerType>(FT->getReturnType()) &&
7266 TD->getIntPtrType() == OldRetTy))
7267 return false; // Cannot transform this return value.
7269 // If the callsite is an invoke instruction, and the return value is used by
7270 // a PHI node in a successor, we cannot change the return type of the call
7271 // because there is no place to put the cast instruction (without breaking
7272 // the critical edge). Bail out in this case.
7273 if (!Caller->use_empty())
7274 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7275 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7277 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7278 if (PN->getParent() == II->getNormalDest() ||
7279 PN->getParent() == II->getUnwindDest())
7283 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7284 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7286 CallSite::arg_iterator AI = CS.arg_begin();
7287 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7288 const Type *ParamTy = FT->getParamType(i);
7289 const Type *ActTy = (*AI)->getType();
7290 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7291 //Either we can cast directly, or we can upconvert the argument
7292 bool isConvertible = ActTy == ParamTy ||
7293 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7294 (ParamTy->isInteger() && ActTy->isInteger() &&
7295 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7296 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7297 && c->getSExtValue() > 0);
7298 if (Callee->isDeclaration() && !isConvertible) return false;
7301 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7302 Callee->isDeclaration())
7303 return false; // Do not delete arguments unless we have a function body...
7305 // Okay, we decided that this is a safe thing to do: go ahead and start
7306 // inserting cast instructions as necessary...
7307 std::vector<Value*> Args;
7308 Args.reserve(NumActualArgs);
7310 AI = CS.arg_begin();
7311 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7312 const Type *ParamTy = FT->getParamType(i);
7313 if ((*AI)->getType() == ParamTy) {
7314 Args.push_back(*AI);
7316 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7317 false, ParamTy, false);
7318 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7319 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7323 // If the function takes more arguments than the call was taking, add them
7325 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7326 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7328 // If we are removing arguments to the function, emit an obnoxious warning...
7329 if (FT->getNumParams() < NumActualArgs)
7330 if (!FT->isVarArg()) {
7331 cerr << "WARNING: While resolving call to function '"
7332 << Callee->getName() << "' arguments were dropped!\n";
7334 // Add all of the arguments in their promoted form to the arg list...
7335 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7336 const Type *PTy = getPromotedType((*AI)->getType());
7337 if (PTy != (*AI)->getType()) {
7338 // Must promote to pass through va_arg area!
7339 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7341 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7342 InsertNewInstBefore(Cast, *Caller);
7343 Args.push_back(Cast);
7345 Args.push_back(*AI);
7350 if (FT->getReturnType() == Type::VoidTy)
7351 Caller->setName(""); // Void type should not have a name.
7354 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7355 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7356 &Args[0], Args.size(), Caller->getName(), Caller);
7357 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7359 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
7360 if (cast<CallInst>(Caller)->isTailCall())
7361 cast<CallInst>(NC)->setTailCall();
7362 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7365 // Insert a cast of the return type as necessary.
7367 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7368 if (NV->getType() != Type::VoidTy) {
7369 const Type *CallerTy = Caller->getType();
7370 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7372 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7374 // If this is an invoke instruction, we should insert it after the first
7375 // non-phi, instruction in the normal successor block.
7376 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7377 BasicBlock::iterator I = II->getNormalDest()->begin();
7378 while (isa<PHINode>(I)) ++I;
7379 InsertNewInstBefore(NC, *I);
7381 // Otherwise, it's a call, just insert cast right after the call instr
7382 InsertNewInstBefore(NC, *Caller);
7384 AddUsersToWorkList(*Caller);
7386 NV = UndefValue::get(Caller->getType());
7390 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7391 Caller->replaceAllUsesWith(NV);
7392 Caller->getParent()->getInstList().erase(Caller);
7393 removeFromWorkList(Caller);
7397 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7398 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7399 /// and a single binop.
7400 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7401 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7402 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
7403 isa<CmpInst>(FirstInst));
7404 unsigned Opc = FirstInst->getOpcode();
7405 Value *LHSVal = FirstInst->getOperand(0);
7406 Value *RHSVal = FirstInst->getOperand(1);
7408 const Type *LHSType = LHSVal->getType();
7409 const Type *RHSType = RHSVal->getType();
7411 // Scan to see if all operands are the same opcode, all have one use, and all
7412 // kill their operands (i.e. the operands have one use).
7413 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7414 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7415 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7416 // Verify type of the LHS matches so we don't fold cmp's of different
7417 // types or GEP's with different index types.
7418 I->getOperand(0)->getType() != LHSType ||
7419 I->getOperand(1)->getType() != RHSType)
7422 // If they are CmpInst instructions, check their predicates
7423 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7424 if (cast<CmpInst>(I)->getPredicate() !=
7425 cast<CmpInst>(FirstInst)->getPredicate())
7428 // Keep track of which operand needs a phi node.
7429 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7430 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7433 // Otherwise, this is safe to transform, determine if it is profitable.
7435 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7436 // Indexes are often folded into load/store instructions, so we don't want to
7437 // hide them behind a phi.
7438 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7441 Value *InLHS = FirstInst->getOperand(0);
7442 Value *InRHS = FirstInst->getOperand(1);
7443 PHINode *NewLHS = 0, *NewRHS = 0;
7445 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7446 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7447 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7448 InsertNewInstBefore(NewLHS, PN);
7453 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7454 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7455 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7456 InsertNewInstBefore(NewRHS, PN);
7460 // Add all operands to the new PHIs.
7461 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7463 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7464 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7467 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7468 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7472 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7473 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7474 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7475 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7478 assert(isa<GetElementPtrInst>(FirstInst));
7479 return new GetElementPtrInst(LHSVal, RHSVal);
7483 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7484 /// of the block that defines it. This means that it must be obvious the value
7485 /// of the load is not changed from the point of the load to the end of the
7488 /// Finally, it is safe, but not profitable, to sink a load targetting a
7489 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
7491 static bool isSafeToSinkLoad(LoadInst *L) {
7492 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7494 for (++BBI; BBI != E; ++BBI)
7495 if (BBI->mayWriteToMemory())
7498 // Check for non-address taken alloca. If not address-taken already, it isn't
7499 // profitable to do this xform.
7500 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
7501 bool isAddressTaken = false;
7502 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
7504 if (isa<LoadInst>(UI)) continue;
7505 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
7506 // If storing TO the alloca, then the address isn't taken.
7507 if (SI->getOperand(1) == AI) continue;
7509 isAddressTaken = true;
7513 if (!isAddressTaken)
7521 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7522 // operator and they all are only used by the PHI, PHI together their
7523 // inputs, and do the operation once, to the result of the PHI.
7524 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7525 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7527 // Scan the instruction, looking for input operations that can be folded away.
7528 // If all input operands to the phi are the same instruction (e.g. a cast from
7529 // the same type or "+42") we can pull the operation through the PHI, reducing
7530 // code size and simplifying code.
7531 Constant *ConstantOp = 0;
7532 const Type *CastSrcTy = 0;
7533 bool isVolatile = false;
7534 if (isa<CastInst>(FirstInst)) {
7535 CastSrcTy = FirstInst->getOperand(0)->getType();
7536 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
7537 // Can fold binop, compare or shift here if the RHS is a constant,
7538 // otherwise call FoldPHIArgBinOpIntoPHI.
7539 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7540 if (ConstantOp == 0)
7541 return FoldPHIArgBinOpIntoPHI(PN);
7542 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7543 isVolatile = LI->isVolatile();
7544 // We can't sink the load if the loaded value could be modified between the
7545 // load and the PHI.
7546 if (LI->getParent() != PN.getIncomingBlock(0) ||
7547 !isSafeToSinkLoad(LI))
7549 } else if (isa<GetElementPtrInst>(FirstInst)) {
7550 if (FirstInst->getNumOperands() == 2)
7551 return FoldPHIArgBinOpIntoPHI(PN);
7552 // Can't handle general GEPs yet.
7555 return 0; // Cannot fold this operation.
7558 // Check to see if all arguments are the same operation.
7559 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7560 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7561 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7562 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7565 if (I->getOperand(0)->getType() != CastSrcTy)
7566 return 0; // Cast operation must match.
7567 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7568 // We can't sink the load if the loaded value could be modified between
7569 // the load and the PHI.
7570 if (LI->isVolatile() != isVolatile ||
7571 LI->getParent() != PN.getIncomingBlock(i) ||
7572 !isSafeToSinkLoad(LI))
7574 } else if (I->getOperand(1) != ConstantOp) {
7579 // Okay, they are all the same operation. Create a new PHI node of the
7580 // correct type, and PHI together all of the LHS's of the instructions.
7581 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7582 PN.getName()+".in");
7583 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7585 Value *InVal = FirstInst->getOperand(0);
7586 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7588 // Add all operands to the new PHI.
7589 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7590 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7591 if (NewInVal != InVal)
7593 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7598 // The new PHI unions all of the same values together. This is really
7599 // common, so we handle it intelligently here for compile-time speed.
7603 InsertNewInstBefore(NewPN, PN);
7607 // Insert and return the new operation.
7608 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7609 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7610 else if (isa<LoadInst>(FirstInst))
7611 return new LoadInst(PhiVal, "", isVolatile);
7612 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7613 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7614 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7615 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7616 PhiVal, ConstantOp);
7618 assert(0 && "Unknown operation");
7621 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7623 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7624 if (PN->use_empty()) return true;
7625 if (!PN->hasOneUse()) return false;
7627 // Remember this node, and if we find the cycle, return.
7628 if (!PotentiallyDeadPHIs.insert(PN).second)
7631 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7632 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7637 // PHINode simplification
7639 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7640 // If LCSSA is around, don't mess with Phi nodes
7641 if (mustPreserveAnalysisID(LCSSAID)) return 0;
7643 if (Value *V = PN.hasConstantValue())
7644 return ReplaceInstUsesWith(PN, V);
7646 // If all PHI operands are the same operation, pull them through the PHI,
7647 // reducing code size.
7648 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7649 PN.getIncomingValue(0)->hasOneUse())
7650 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7653 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7654 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7655 // PHI)... break the cycle.
7656 if (PN.hasOneUse()) {
7657 Instruction *PHIUser = cast<Instruction>(PN.use_back());
7658 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
7659 std::set<PHINode*> PotentiallyDeadPHIs;
7660 PotentiallyDeadPHIs.insert(&PN);
7661 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7662 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7665 // If this phi has a single use, and if that use just computes a value for
7666 // the next iteration of a loop, delete the phi. This occurs with unused
7667 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
7668 // common case here is good because the only other things that catch this
7669 // are induction variable analysis (sometimes) and ADCE, which is only run
7671 if (PHIUser->hasOneUse() &&
7672 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
7673 PHIUser->use_back() == &PN) {
7674 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7681 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7682 Instruction *InsertPoint,
7684 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7685 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7686 // We must cast correctly to the pointer type. Ensure that we
7687 // sign extend the integer value if it is smaller as this is
7688 // used for address computation.
7689 Instruction::CastOps opcode =
7690 (VTySize < PtrSize ? Instruction::SExt :
7691 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7692 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7696 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7697 Value *PtrOp = GEP.getOperand(0);
7698 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7699 // If so, eliminate the noop.
7700 if (GEP.getNumOperands() == 1)
7701 return ReplaceInstUsesWith(GEP, PtrOp);
7703 if (isa<UndefValue>(GEP.getOperand(0)))
7704 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7706 bool HasZeroPointerIndex = false;
7707 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7708 HasZeroPointerIndex = C->isNullValue();
7710 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7711 return ReplaceInstUsesWith(GEP, PtrOp);
7713 // Eliminate unneeded casts for indices.
7714 bool MadeChange = false;
7715 gep_type_iterator GTI = gep_type_begin(GEP);
7716 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7717 if (isa<SequentialType>(*GTI)) {
7718 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7719 if (CI->getOpcode() == Instruction::ZExt ||
7720 CI->getOpcode() == Instruction::SExt) {
7721 const Type *SrcTy = CI->getOperand(0)->getType();
7722 // We can eliminate a cast from i32 to i64 iff the target
7723 // is a 32-bit pointer target.
7724 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7726 GEP.setOperand(i, CI->getOperand(0));
7730 // If we are using a wider index than needed for this platform, shrink it
7731 // to what we need. If the incoming value needs a cast instruction,
7732 // insert it. This explicit cast can make subsequent optimizations more
7734 Value *Op = GEP.getOperand(i);
7735 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
7736 if (Constant *C = dyn_cast<Constant>(Op)) {
7737 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7740 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7742 GEP.setOperand(i, Op);
7746 if (MadeChange) return &GEP;
7748 // Combine Indices - If the source pointer to this getelementptr instruction
7749 // is a getelementptr instruction, combine the indices of the two
7750 // getelementptr instructions into a single instruction.
7752 SmallVector<Value*, 8> SrcGEPOperands;
7753 if (User *Src = dyn_castGetElementPtr(PtrOp))
7754 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
7756 if (!SrcGEPOperands.empty()) {
7757 // Note that if our source is a gep chain itself that we wait for that
7758 // chain to be resolved before we perform this transformation. This
7759 // avoids us creating a TON of code in some cases.
7761 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7762 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7763 return 0; // Wait until our source is folded to completion.
7765 SmallVector<Value*, 8> Indices;
7767 // Find out whether the last index in the source GEP is a sequential idx.
7768 bool EndsWithSequential = false;
7769 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7770 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7771 EndsWithSequential = !isa<StructType>(*I);
7773 // Can we combine the two pointer arithmetics offsets?
7774 if (EndsWithSequential) {
7775 // Replace: gep (gep %P, long B), long A, ...
7776 // With: T = long A+B; gep %P, T, ...
7778 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7779 if (SO1 == Constant::getNullValue(SO1->getType())) {
7781 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7784 // If they aren't the same type, convert both to an integer of the
7785 // target's pointer size.
7786 if (SO1->getType() != GO1->getType()) {
7787 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7788 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7789 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7790 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7792 unsigned PS = TD->getPointerSize();
7793 if (TD->getTypeSize(SO1->getType()) == PS) {
7794 // Convert GO1 to SO1's type.
7795 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
7797 } else if (TD->getTypeSize(GO1->getType()) == PS) {
7798 // Convert SO1 to GO1's type.
7799 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
7801 const Type *PT = TD->getIntPtrType();
7802 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
7803 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
7807 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7808 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7810 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7811 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7815 // Recycle the GEP we already have if possible.
7816 if (SrcGEPOperands.size() == 2) {
7817 GEP.setOperand(0, SrcGEPOperands[0]);
7818 GEP.setOperand(1, Sum);
7821 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7822 SrcGEPOperands.end()-1);
7823 Indices.push_back(Sum);
7824 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7826 } else if (isa<Constant>(*GEP.idx_begin()) &&
7827 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7828 SrcGEPOperands.size() != 1) {
7829 // Otherwise we can do the fold if the first index of the GEP is a zero
7830 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7831 SrcGEPOperands.end());
7832 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7835 if (!Indices.empty())
7836 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
7837 Indices.size(), GEP.getName());
7839 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7840 // GEP of global variable. If all of the indices for this GEP are
7841 // constants, we can promote this to a constexpr instead of an instruction.
7843 // Scan for nonconstants...
7844 SmallVector<Constant*, 8> Indices;
7845 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7846 for (; I != E && isa<Constant>(*I); ++I)
7847 Indices.push_back(cast<Constant>(*I));
7849 if (I == E) { // If they are all constants...
7850 Constant *CE = ConstantExpr::getGetElementPtr(GV,
7851 &Indices[0],Indices.size());
7853 // Replace all uses of the GEP with the new constexpr...
7854 return ReplaceInstUsesWith(GEP, CE);
7856 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7857 if (!isa<PointerType>(X->getType())) {
7858 // Not interesting. Source pointer must be a cast from pointer.
7859 } else if (HasZeroPointerIndex) {
7860 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7861 // into : GEP [10 x ubyte]* X, long 0, ...
7863 // This occurs when the program declares an array extern like "int X[];"
7865 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7866 const PointerType *XTy = cast<PointerType>(X->getType());
7867 if (const ArrayType *XATy =
7868 dyn_cast<ArrayType>(XTy->getElementType()))
7869 if (const ArrayType *CATy =
7870 dyn_cast<ArrayType>(CPTy->getElementType()))
7871 if (CATy->getElementType() == XATy->getElementType()) {
7872 // At this point, we know that the cast source type is a pointer
7873 // to an array of the same type as the destination pointer
7874 // array. Because the array type is never stepped over (there
7875 // is a leading zero) we can fold the cast into this GEP.
7876 GEP.setOperand(0, X);
7879 } else if (GEP.getNumOperands() == 2) {
7880 // Transform things like:
7881 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7882 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7883 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7884 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7885 if (isa<ArrayType>(SrcElTy) &&
7886 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7887 TD->getTypeSize(ResElTy)) {
7888 Value *V = InsertNewInstBefore(
7889 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7890 GEP.getOperand(1), GEP.getName()), GEP);
7891 // V and GEP are both pointer types --> BitCast
7892 return new BitCastInst(V, GEP.getType());
7895 // Transform things like:
7896 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7897 // (where tmp = 8*tmp2) into:
7898 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7900 if (isa<ArrayType>(SrcElTy) &&
7901 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
7902 uint64_t ArrayEltSize =
7903 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7905 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7906 // allow either a mul, shift, or constant here.
7908 ConstantInt *Scale = 0;
7909 if (ArrayEltSize == 1) {
7910 NewIdx = GEP.getOperand(1);
7911 Scale = ConstantInt::get(NewIdx->getType(), 1);
7912 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7913 NewIdx = ConstantInt::get(CI->getType(), 1);
7915 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7916 if (Inst->getOpcode() == Instruction::Shl &&
7917 isa<ConstantInt>(Inst->getOperand(1))) {
7919 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7920 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7921 NewIdx = Inst->getOperand(0);
7922 } else if (Inst->getOpcode() == Instruction::Mul &&
7923 isa<ConstantInt>(Inst->getOperand(1))) {
7924 Scale = cast<ConstantInt>(Inst->getOperand(1));
7925 NewIdx = Inst->getOperand(0);
7929 // If the index will be to exactly the right offset with the scale taken
7930 // out, perform the transformation.
7931 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7932 if (isa<ConstantInt>(Scale))
7933 Scale = ConstantInt::get(Scale->getType(),
7934 Scale->getZExtValue() / ArrayEltSize);
7935 if (Scale->getZExtValue() != 1) {
7936 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
7938 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7939 NewIdx = InsertNewInstBefore(Sc, GEP);
7942 // Insert the new GEP instruction.
7943 Instruction *NewGEP =
7944 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7945 NewIdx, GEP.getName());
7946 NewGEP = InsertNewInstBefore(NewGEP, GEP);
7947 // The NewGEP must be pointer typed, so must the old one -> BitCast
7948 return new BitCastInst(NewGEP, GEP.getType());
7957 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7958 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7959 if (AI.isArrayAllocation()) // Check C != 1
7960 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7962 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7963 AllocationInst *New = 0;
7965 // Create and insert the replacement instruction...
7966 if (isa<MallocInst>(AI))
7967 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7969 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7970 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7973 InsertNewInstBefore(New, AI);
7975 // Scan to the end of the allocation instructions, to skip over a block of
7976 // allocas if possible...
7978 BasicBlock::iterator It = New;
7979 while (isa<AllocationInst>(*It)) ++It;
7981 // Now that I is pointing to the first non-allocation-inst in the block,
7982 // insert our getelementptr instruction...
7984 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
7985 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
7986 New->getName()+".sub", It);
7988 // Now make everything use the getelementptr instead of the original
7990 return ReplaceInstUsesWith(AI, V);
7991 } else if (isa<UndefValue>(AI.getArraySize())) {
7992 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7995 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7996 // Note that we only do this for alloca's, because malloc should allocate and
7997 // return a unique pointer, even for a zero byte allocation.
7998 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
7999 TD->getTypeSize(AI.getAllocatedType()) == 0)
8000 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8005 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8006 Value *Op = FI.getOperand(0);
8008 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8009 if (CastInst *CI = dyn_cast<CastInst>(Op))
8010 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8011 FI.setOperand(0, CI->getOperand(0));
8015 // free undef -> unreachable.
8016 if (isa<UndefValue>(Op)) {
8017 // Insert a new store to null because we cannot modify the CFG here.
8018 new StoreInst(ConstantInt::getTrue(),
8019 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8020 return EraseInstFromFunction(FI);
8023 // If we have 'free null' delete the instruction. This can happen in stl code
8024 // when lots of inlining happens.
8025 if (isa<ConstantPointerNull>(Op))
8026 return EraseInstFromFunction(FI);
8032 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8033 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8034 User *CI = cast<User>(LI.getOperand(0));
8035 Value *CastOp = CI->getOperand(0);
8037 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8038 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8039 const Type *SrcPTy = SrcTy->getElementType();
8041 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8042 isa<VectorType>(DestPTy)) {
8043 // If the source is an array, the code below will not succeed. Check to
8044 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8046 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8047 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8048 if (ASrcTy->getNumElements() != 0) {
8050 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8051 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8052 SrcTy = cast<PointerType>(CastOp->getType());
8053 SrcPTy = SrcTy->getElementType();
8056 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8057 isa<VectorType>(SrcPTy)) &&
8058 // Do not allow turning this into a load of an integer, which is then
8059 // casted to a pointer, this pessimizes pointer analysis a lot.
8060 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8061 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8062 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8064 // Okay, we are casting from one integer or pointer type to another of
8065 // the same size. Instead of casting the pointer before the load, cast
8066 // the result of the loaded value.
8067 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8069 LI.isVolatile()),LI);
8070 // Now cast the result of the load.
8071 return new BitCastInst(NewLoad, LI.getType());
8078 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8079 /// from this value cannot trap. If it is not obviously safe to load from the
8080 /// specified pointer, we do a quick local scan of the basic block containing
8081 /// ScanFrom, to determine if the address is already accessed.
8082 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8083 // If it is an alloca or global variable, it is always safe to load from.
8084 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8086 // Otherwise, be a little bit agressive by scanning the local block where we
8087 // want to check to see if the pointer is already being loaded or stored
8088 // from/to. If so, the previous load or store would have already trapped,
8089 // so there is no harm doing an extra load (also, CSE will later eliminate
8090 // the load entirely).
8091 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8096 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8097 if (LI->getOperand(0) == V) return true;
8098 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8099 if (SI->getOperand(1) == V) return true;
8105 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8106 Value *Op = LI.getOperand(0);
8108 // load (cast X) --> cast (load X) iff safe
8109 if (isa<CastInst>(Op))
8110 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8113 // None of the following transforms are legal for volatile loads.
8114 if (LI.isVolatile()) return 0;
8116 if (&LI.getParent()->front() != &LI) {
8117 BasicBlock::iterator BBI = &LI; --BBI;
8118 // If the instruction immediately before this is a store to the same
8119 // address, do a simple form of store->load forwarding.
8120 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8121 if (SI->getOperand(1) == LI.getOperand(0))
8122 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8123 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8124 if (LIB->getOperand(0) == LI.getOperand(0))
8125 return ReplaceInstUsesWith(LI, LIB);
8128 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8129 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8130 isa<UndefValue>(GEPI->getOperand(0))) {
8131 // Insert a new store to null instruction before the load to indicate
8132 // that this code is not reachable. We do this instead of inserting
8133 // an unreachable instruction directly because we cannot modify the
8135 new StoreInst(UndefValue::get(LI.getType()),
8136 Constant::getNullValue(Op->getType()), &LI);
8137 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8140 if (Constant *C = dyn_cast<Constant>(Op)) {
8141 // load null/undef -> undef
8142 if ((C->isNullValue() || isa<UndefValue>(C))) {
8143 // Insert a new store to null instruction before the load to indicate that
8144 // this code is not reachable. We do this instead of inserting an
8145 // unreachable instruction directly because we cannot modify the CFG.
8146 new StoreInst(UndefValue::get(LI.getType()),
8147 Constant::getNullValue(Op->getType()), &LI);
8148 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8151 // Instcombine load (constant global) into the value loaded.
8152 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8153 if (GV->isConstant() && !GV->isDeclaration())
8154 return ReplaceInstUsesWith(LI, GV->getInitializer());
8156 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8157 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8158 if (CE->getOpcode() == Instruction::GetElementPtr) {
8159 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8160 if (GV->isConstant() && !GV->isDeclaration())
8162 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8163 return ReplaceInstUsesWith(LI, V);
8164 if (CE->getOperand(0)->isNullValue()) {
8165 // Insert a new store to null instruction before the load to indicate
8166 // that this code is not reachable. We do this instead of inserting
8167 // an unreachable instruction directly because we cannot modify the
8169 new StoreInst(UndefValue::get(LI.getType()),
8170 Constant::getNullValue(Op->getType()), &LI);
8171 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8174 } else if (CE->isCast()) {
8175 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8180 if (Op->hasOneUse()) {
8181 // Change select and PHI nodes to select values instead of addresses: this
8182 // helps alias analysis out a lot, allows many others simplifications, and
8183 // exposes redundancy in the code.
8185 // Note that we cannot do the transformation unless we know that the
8186 // introduced loads cannot trap! Something like this is valid as long as
8187 // the condition is always false: load (select bool %C, int* null, int* %G),
8188 // but it would not be valid if we transformed it to load from null
8191 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8192 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8193 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8194 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8195 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8196 SI->getOperand(1)->getName()+".val"), LI);
8197 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8198 SI->getOperand(2)->getName()+".val"), LI);
8199 return new SelectInst(SI->getCondition(), V1, V2);
8202 // load (select (cond, null, P)) -> load P
8203 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8204 if (C->isNullValue()) {
8205 LI.setOperand(0, SI->getOperand(2));
8209 // load (select (cond, P, null)) -> load P
8210 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8211 if (C->isNullValue()) {
8212 LI.setOperand(0, SI->getOperand(1));
8220 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8222 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8223 User *CI = cast<User>(SI.getOperand(1));
8224 Value *CastOp = CI->getOperand(0);
8226 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8227 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8228 const Type *SrcPTy = SrcTy->getElementType();
8230 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8231 // If the source is an array, the code below will not succeed. Check to
8232 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8234 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8235 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8236 if (ASrcTy->getNumElements() != 0) {
8238 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8239 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8240 SrcTy = cast<PointerType>(CastOp->getType());
8241 SrcPTy = SrcTy->getElementType();
8244 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8245 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8246 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8248 // Okay, we are casting from one integer or pointer type to another of
8249 // the same size. Instead of casting the pointer before
8250 // the store, cast the value to be stored.
8252 Value *SIOp0 = SI.getOperand(0);
8253 Instruction::CastOps opcode = Instruction::BitCast;
8254 const Type* CastSrcTy = SIOp0->getType();
8255 const Type* CastDstTy = SrcPTy;
8256 if (isa<PointerType>(CastDstTy)) {
8257 if (CastSrcTy->isInteger())
8258 opcode = Instruction::IntToPtr;
8259 } else if (isa<IntegerType>(CastDstTy)) {
8260 if (isa<PointerType>(SIOp0->getType()))
8261 opcode = Instruction::PtrToInt;
8263 if (Constant *C = dyn_cast<Constant>(SIOp0))
8264 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8266 NewCast = IC.InsertNewInstBefore(
8267 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8269 return new StoreInst(NewCast, CastOp);
8276 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8277 Value *Val = SI.getOperand(0);
8278 Value *Ptr = SI.getOperand(1);
8280 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8281 EraseInstFromFunction(SI);
8286 // If the RHS is an alloca with a single use, zapify the store, making the
8288 if (Ptr->hasOneUse()) {
8289 if (isa<AllocaInst>(Ptr)) {
8290 EraseInstFromFunction(SI);
8295 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8296 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8297 GEP->getOperand(0)->hasOneUse()) {
8298 EraseInstFromFunction(SI);
8304 // Do really simple DSE, to catch cases where there are several consequtive
8305 // stores to the same location, separated by a few arithmetic operations. This
8306 // situation often occurs with bitfield accesses.
8307 BasicBlock::iterator BBI = &SI;
8308 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8312 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8313 // Prev store isn't volatile, and stores to the same location?
8314 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8317 EraseInstFromFunction(*PrevSI);
8323 // If this is a load, we have to stop. However, if the loaded value is from
8324 // the pointer we're loading and is producing the pointer we're storing,
8325 // then *this* store is dead (X = load P; store X -> P).
8326 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8327 if (LI == Val && LI->getOperand(0) == Ptr) {
8328 EraseInstFromFunction(SI);
8332 // Otherwise, this is a load from some other location. Stores before it
8337 // Don't skip over loads or things that can modify memory.
8338 if (BBI->mayWriteToMemory())
8343 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8345 // store X, null -> turns into 'unreachable' in SimplifyCFG
8346 if (isa<ConstantPointerNull>(Ptr)) {
8347 if (!isa<UndefValue>(Val)) {
8348 SI.setOperand(0, UndefValue::get(Val->getType()));
8349 if (Instruction *U = dyn_cast<Instruction>(Val))
8350 WorkList.push_back(U); // Dropped a use.
8353 return 0; // Do not modify these!
8356 // store undef, Ptr -> noop
8357 if (isa<UndefValue>(Val)) {
8358 EraseInstFromFunction(SI);
8363 // If the pointer destination is a cast, see if we can fold the cast into the
8365 if (isa<CastInst>(Ptr))
8366 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8368 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8370 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8374 // If this store is the last instruction in the basic block, and if the block
8375 // ends with an unconditional branch, try to move it to the successor block.
8377 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8378 if (BI->isUnconditional()) {
8379 // Check to see if the successor block has exactly two incoming edges. If
8380 // so, see if the other predecessor contains a store to the same location.
8381 // if so, insert a PHI node (if needed) and move the stores down.
8382 BasicBlock *Dest = BI->getSuccessor(0);
8384 pred_iterator PI = pred_begin(Dest);
8385 BasicBlock *Other = 0;
8386 if (*PI != BI->getParent())
8389 if (PI != pred_end(Dest)) {
8390 if (*PI != BI->getParent())
8395 if (++PI != pred_end(Dest))
8398 if (Other) { // If only one other pred...
8399 BBI = Other->getTerminator();
8400 // Make sure this other block ends in an unconditional branch and that
8401 // there is an instruction before the branch.
8402 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8403 BBI != Other->begin()) {
8405 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8407 // If this instruction is a store to the same location.
8408 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8409 // Okay, we know we can perform this transformation. Insert a PHI
8410 // node now if we need it.
8411 Value *MergedVal = OtherStore->getOperand(0);
8412 if (MergedVal != SI.getOperand(0)) {
8413 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8414 PN->reserveOperandSpace(2);
8415 PN->addIncoming(SI.getOperand(0), SI.getParent());
8416 PN->addIncoming(OtherStore->getOperand(0), Other);
8417 MergedVal = InsertNewInstBefore(PN, Dest->front());
8420 // Advance to a place where it is safe to insert the new store and
8422 BBI = Dest->begin();
8423 while (isa<PHINode>(BBI)) ++BBI;
8424 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8425 OtherStore->isVolatile()), *BBI);
8427 // Nuke the old stores.
8428 EraseInstFromFunction(SI);
8429 EraseInstFromFunction(*OtherStore);
8441 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8442 // Change br (not X), label True, label False to: br X, label False, True
8444 BasicBlock *TrueDest;
8445 BasicBlock *FalseDest;
8446 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8447 !isa<Constant>(X)) {
8448 // Swap Destinations and condition...
8450 BI.setSuccessor(0, FalseDest);
8451 BI.setSuccessor(1, TrueDest);
8455 // Cannonicalize fcmp_one -> fcmp_oeq
8456 FCmpInst::Predicate FPred; Value *Y;
8457 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8458 TrueDest, FalseDest)))
8459 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8460 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8461 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8462 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8463 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
8464 NewSCC->takeName(I);
8465 // Swap Destinations and condition...
8466 BI.setCondition(NewSCC);
8467 BI.setSuccessor(0, FalseDest);
8468 BI.setSuccessor(1, TrueDest);
8469 removeFromWorkList(I);
8470 I->eraseFromParent();
8471 WorkList.push_back(NewSCC);
8475 // Cannonicalize icmp_ne -> icmp_eq
8476 ICmpInst::Predicate IPred;
8477 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8478 TrueDest, FalseDest)))
8479 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8480 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8481 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8482 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8483 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8484 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
8485 NewSCC->takeName(I);
8486 // Swap Destinations and condition...
8487 BI.setCondition(NewSCC);
8488 BI.setSuccessor(0, FalseDest);
8489 BI.setSuccessor(1, TrueDest);
8490 removeFromWorkList(I);
8491 I->eraseFromParent();;
8492 WorkList.push_back(NewSCC);
8499 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8500 Value *Cond = SI.getCondition();
8501 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8502 if (I->getOpcode() == Instruction::Add)
8503 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8504 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8505 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8506 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8508 SI.setOperand(0, I->getOperand(0));
8509 WorkList.push_back(I);
8516 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8517 /// is to leave as a vector operation.
8518 static bool CheapToScalarize(Value *V, bool isConstant) {
8519 if (isa<ConstantAggregateZero>(V))
8521 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
8522 if (isConstant) return true;
8523 // If all elts are the same, we can extract.
8524 Constant *Op0 = C->getOperand(0);
8525 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8526 if (C->getOperand(i) != Op0)
8530 Instruction *I = dyn_cast<Instruction>(V);
8531 if (!I) return false;
8533 // Insert element gets simplified to the inserted element or is deleted if
8534 // this is constant idx extract element and its a constant idx insertelt.
8535 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8536 isa<ConstantInt>(I->getOperand(2)))
8538 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8540 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8541 if (BO->hasOneUse() &&
8542 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8543 CheapToScalarize(BO->getOperand(1), isConstant)))
8545 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8546 if (CI->hasOneUse() &&
8547 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8548 CheapToScalarize(CI->getOperand(1), isConstant)))
8554 /// Read and decode a shufflevector mask.
8556 /// It turns undef elements into values that are larger than the number of
8557 /// elements in the input.
8558 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8559 unsigned NElts = SVI->getType()->getNumElements();
8560 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8561 return std::vector<unsigned>(NElts, 0);
8562 if (isa<UndefValue>(SVI->getOperand(2)))
8563 return std::vector<unsigned>(NElts, 2*NElts);
8565 std::vector<unsigned> Result;
8566 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
8567 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8568 if (isa<UndefValue>(CP->getOperand(i)))
8569 Result.push_back(NElts*2); // undef -> 8
8571 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8575 /// FindScalarElement - Given a vector and an element number, see if the scalar
8576 /// value is already around as a register, for example if it were inserted then
8577 /// extracted from the vector.
8578 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8579 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
8580 const VectorType *PTy = cast<VectorType>(V->getType());
8581 unsigned Width = PTy->getNumElements();
8582 if (EltNo >= Width) // Out of range access.
8583 return UndefValue::get(PTy->getElementType());
8585 if (isa<UndefValue>(V))
8586 return UndefValue::get(PTy->getElementType());
8587 else if (isa<ConstantAggregateZero>(V))
8588 return Constant::getNullValue(PTy->getElementType());
8589 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
8590 return CP->getOperand(EltNo);
8591 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8592 // If this is an insert to a variable element, we don't know what it is.
8593 if (!isa<ConstantInt>(III->getOperand(2)))
8595 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8597 // If this is an insert to the element we are looking for, return the
8600 return III->getOperand(1);
8602 // Otherwise, the insertelement doesn't modify the value, recurse on its
8604 return FindScalarElement(III->getOperand(0), EltNo);
8605 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8606 unsigned InEl = getShuffleMask(SVI)[EltNo];
8608 return FindScalarElement(SVI->getOperand(0), InEl);
8609 else if (InEl < Width*2)
8610 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8612 return UndefValue::get(PTy->getElementType());
8615 // Otherwise, we don't know.
8619 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8621 // If packed val is undef, replace extract with scalar undef.
8622 if (isa<UndefValue>(EI.getOperand(0)))
8623 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8625 // If packed val is constant 0, replace extract with scalar 0.
8626 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8627 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8629 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
8630 // If packed val is constant with uniform operands, replace EI
8631 // with that operand
8632 Constant *op0 = C->getOperand(0);
8633 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8634 if (C->getOperand(i) != op0) {
8639 return ReplaceInstUsesWith(EI, op0);
8642 // If extracting a specified index from the vector, see if we can recursively
8643 // find a previously computed scalar that was inserted into the vector.
8644 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8645 // This instruction only demands the single element from the input vector.
8646 // If the input vector has a single use, simplify it based on this use
8648 uint64_t IndexVal = IdxC->getZExtValue();
8649 if (EI.getOperand(0)->hasOneUse()) {
8651 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8654 EI.setOperand(0, V);
8659 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8660 return ReplaceInstUsesWith(EI, Elt);
8663 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8664 if (I->hasOneUse()) {
8665 // Push extractelement into predecessor operation if legal and
8666 // profitable to do so
8667 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8668 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8669 if (CheapToScalarize(BO, isConstantElt)) {
8670 ExtractElementInst *newEI0 =
8671 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8672 EI.getName()+".lhs");
8673 ExtractElementInst *newEI1 =
8674 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8675 EI.getName()+".rhs");
8676 InsertNewInstBefore(newEI0, EI);
8677 InsertNewInstBefore(newEI1, EI);
8678 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8680 } else if (isa<LoadInst>(I)) {
8681 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8682 PointerType::get(EI.getType()), EI);
8683 GetElementPtrInst *GEP =
8684 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8685 InsertNewInstBefore(GEP, EI);
8686 return new LoadInst(GEP);
8689 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8690 // Extracting the inserted element?
8691 if (IE->getOperand(2) == EI.getOperand(1))
8692 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8693 // If the inserted and extracted elements are constants, they must not
8694 // be the same value, extract from the pre-inserted value instead.
8695 if (isa<Constant>(IE->getOperand(2)) &&
8696 isa<Constant>(EI.getOperand(1))) {
8697 AddUsesToWorkList(EI);
8698 EI.setOperand(0, IE->getOperand(0));
8701 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8702 // If this is extracting an element from a shufflevector, figure out where
8703 // it came from and extract from the appropriate input element instead.
8704 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8705 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8707 if (SrcIdx < SVI->getType()->getNumElements())
8708 Src = SVI->getOperand(0);
8709 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8710 SrcIdx -= SVI->getType()->getNumElements();
8711 Src = SVI->getOperand(1);
8713 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8715 return new ExtractElementInst(Src, SrcIdx);
8722 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8723 /// elements from either LHS or RHS, return the shuffle mask and true.
8724 /// Otherwise, return false.
8725 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8726 std::vector<Constant*> &Mask) {
8727 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8728 "Invalid CollectSingleShuffleElements");
8729 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8731 if (isa<UndefValue>(V)) {
8732 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8734 } else if (V == LHS) {
8735 for (unsigned i = 0; i != NumElts; ++i)
8736 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8738 } else if (V == RHS) {
8739 for (unsigned i = 0; i != NumElts; ++i)
8740 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8742 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8743 // If this is an insert of an extract from some other vector, include it.
8744 Value *VecOp = IEI->getOperand(0);
8745 Value *ScalarOp = IEI->getOperand(1);
8746 Value *IdxOp = IEI->getOperand(2);
8748 if (!isa<ConstantInt>(IdxOp))
8750 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8752 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8753 // Okay, we can handle this if the vector we are insertinting into is
8755 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8756 // If so, update the mask to reflect the inserted undef.
8757 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
8760 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8761 if (isa<ConstantInt>(EI->getOperand(1)) &&
8762 EI->getOperand(0)->getType() == V->getType()) {
8763 unsigned ExtractedIdx =
8764 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8766 // This must be extracting from either LHS or RHS.
8767 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8768 // Okay, we can handle this if the vector we are insertinting into is
8770 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8771 // If so, update the mask to reflect the inserted value.
8772 if (EI->getOperand(0) == LHS) {
8773 Mask[InsertedIdx & (NumElts-1)] =
8774 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8776 assert(EI->getOperand(0) == RHS);
8777 Mask[InsertedIdx & (NumElts-1)] =
8778 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
8787 // TODO: Handle shufflevector here!
8792 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8793 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8794 /// that computes V and the LHS value of the shuffle.
8795 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8797 assert(isa<VectorType>(V->getType()) &&
8798 (RHS == 0 || V->getType() == RHS->getType()) &&
8799 "Invalid shuffle!");
8800 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8802 if (isa<UndefValue>(V)) {
8803 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8805 } else if (isa<ConstantAggregateZero>(V)) {
8806 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
8808 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8809 // If this is an insert of an extract from some other vector, include it.
8810 Value *VecOp = IEI->getOperand(0);
8811 Value *ScalarOp = IEI->getOperand(1);
8812 Value *IdxOp = IEI->getOperand(2);
8814 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8815 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8816 EI->getOperand(0)->getType() == V->getType()) {
8817 unsigned ExtractedIdx =
8818 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8819 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8821 // Either the extracted from or inserted into vector must be RHSVec,
8822 // otherwise we'd end up with a shuffle of three inputs.
8823 if (EI->getOperand(0) == RHS || RHS == 0) {
8824 RHS = EI->getOperand(0);
8825 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8826 Mask[InsertedIdx & (NumElts-1)] =
8827 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
8832 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8833 // Everything but the extracted element is replaced with the RHS.
8834 for (unsigned i = 0; i != NumElts; ++i) {
8835 if (i != InsertedIdx)
8836 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
8841 // If this insertelement is a chain that comes from exactly these two
8842 // vectors, return the vector and the effective shuffle.
8843 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8844 return EI->getOperand(0);
8849 // TODO: Handle shufflevector here!
8851 // Otherwise, can't do anything fancy. Return an identity vector.
8852 for (unsigned i = 0; i != NumElts; ++i)
8853 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8857 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8858 Value *VecOp = IE.getOperand(0);
8859 Value *ScalarOp = IE.getOperand(1);
8860 Value *IdxOp = IE.getOperand(2);
8862 // If the inserted element was extracted from some other vector, and if the
8863 // indexes are constant, try to turn this into a shufflevector operation.
8864 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8865 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8866 EI->getOperand(0)->getType() == IE.getType()) {
8867 unsigned NumVectorElts = IE.getType()->getNumElements();
8868 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8869 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8871 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8872 return ReplaceInstUsesWith(IE, VecOp);
8874 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8875 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8877 // If we are extracting a value from a vector, then inserting it right
8878 // back into the same place, just use the input vector.
8879 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8880 return ReplaceInstUsesWith(IE, VecOp);
8882 // We could theoretically do this for ANY input. However, doing so could
8883 // turn chains of insertelement instructions into a chain of shufflevector
8884 // instructions, and right now we do not merge shufflevectors. As such,
8885 // only do this in a situation where it is clear that there is benefit.
8886 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8887 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8888 // the values of VecOp, except then one read from EIOp0.
8889 // Build a new shuffle mask.
8890 std::vector<Constant*> Mask;
8891 if (isa<UndefValue>(VecOp))
8892 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
8894 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8895 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
8898 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8899 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8900 ConstantVector::get(Mask));
8903 // If this insertelement isn't used by some other insertelement, turn it
8904 // (and any insertelements it points to), into one big shuffle.
8905 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8906 std::vector<Constant*> Mask;
8908 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8909 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8910 // We now have a shuffle of LHS, RHS, Mask.
8911 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
8920 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8921 Value *LHS = SVI.getOperand(0);
8922 Value *RHS = SVI.getOperand(1);
8923 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8925 bool MadeChange = false;
8927 // Undefined shuffle mask -> undefined value.
8928 if (isa<UndefValue>(SVI.getOperand(2)))
8929 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8931 // If we have shuffle(x, undef, mask) and any elements of mask refer to
8932 // the undef, change them to undefs.
8933 if (isa<UndefValue>(SVI.getOperand(1))) {
8934 // Scan to see if there are any references to the RHS. If so, replace them
8935 // with undef element refs and set MadeChange to true.
8936 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8937 if (Mask[i] >= e && Mask[i] != 2*e) {
8944 // Remap any references to RHS to use LHS.
8945 std::vector<Constant*> Elts;
8946 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8948 Elts.push_back(UndefValue::get(Type::Int32Ty));
8950 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8952 SVI.setOperand(2, ConstantVector::get(Elts));
8956 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8957 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8958 if (LHS == RHS || isa<UndefValue>(LHS)) {
8959 if (isa<UndefValue>(LHS) && LHS == RHS) {
8960 // shuffle(undef,undef,mask) -> undef.
8961 return ReplaceInstUsesWith(SVI, LHS);
8964 // Remap any references to RHS to use LHS.
8965 std::vector<Constant*> Elts;
8966 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8968 Elts.push_back(UndefValue::get(Type::Int32Ty));
8970 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8971 (Mask[i] < e && isa<UndefValue>(LHS)))
8972 Mask[i] = 2*e; // Turn into undef.
8974 Mask[i] &= (e-1); // Force to LHS.
8975 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8978 SVI.setOperand(0, SVI.getOperand(1));
8979 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8980 SVI.setOperand(2, ConstantVector::get(Elts));
8981 LHS = SVI.getOperand(0);
8982 RHS = SVI.getOperand(1);
8986 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8987 bool isLHSID = true, isRHSID = true;
8989 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8990 if (Mask[i] >= e*2) continue; // Ignore undef values.
8991 // Is this an identity shuffle of the LHS value?
8992 isLHSID &= (Mask[i] == i);
8994 // Is this an identity shuffle of the RHS value?
8995 isRHSID &= (Mask[i]-e == i);
8998 // Eliminate identity shuffles.
8999 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9000 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9002 // If the LHS is a shufflevector itself, see if we can combine it with this
9003 // one without producing an unusual shuffle. Here we are really conservative:
9004 // we are absolutely afraid of producing a shuffle mask not in the input
9005 // program, because the code gen may not be smart enough to turn a merged
9006 // shuffle into two specific shuffles: it may produce worse code. As such,
9007 // we only merge two shuffles if the result is one of the two input shuffle
9008 // masks. In this case, merging the shuffles just removes one instruction,
9009 // which we know is safe. This is good for things like turning:
9010 // (splat(splat)) -> splat.
9011 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9012 if (isa<UndefValue>(RHS)) {
9013 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9015 std::vector<unsigned> NewMask;
9016 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9018 NewMask.push_back(2*e);
9020 NewMask.push_back(LHSMask[Mask[i]]);
9022 // If the result mask is equal to the src shuffle or this shuffle mask, do
9024 if (NewMask == LHSMask || NewMask == Mask) {
9025 std::vector<Constant*> Elts;
9026 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9027 if (NewMask[i] >= e*2) {
9028 Elts.push_back(UndefValue::get(Type::Int32Ty));
9030 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9033 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9034 LHSSVI->getOperand(1),
9035 ConstantVector::get(Elts));
9040 return MadeChange ? &SVI : 0;
9045 void InstCombiner::removeFromWorkList(Instruction *I) {
9046 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
9051 /// TryToSinkInstruction - Try to move the specified instruction from its
9052 /// current block into the beginning of DestBlock, which can only happen if it's
9053 /// safe to move the instruction past all of the instructions between it and the
9054 /// end of its block.
9055 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9056 assert(I->hasOneUse() && "Invariants didn't hold!");
9058 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9059 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9061 // Do not sink alloca instructions out of the entry block.
9062 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
9065 // We can only sink load instructions if there is nothing between the load and
9066 // the end of block that could change the value.
9067 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9068 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9070 if (Scan->mayWriteToMemory())
9074 BasicBlock::iterator InsertPos = DestBlock->begin();
9075 while (isa<PHINode>(InsertPos)) ++InsertPos;
9077 I->moveBefore(InsertPos);
9083 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9084 /// all reachable code to the worklist.
9086 /// This has a couple of tricks to make the code faster and more powerful. In
9087 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9088 /// them to the worklist (this significantly speeds up instcombine on code where
9089 /// many instructions are dead or constant). Additionally, if we find a branch
9090 /// whose condition is a known constant, we only visit the reachable successors.
9092 static void AddReachableCodeToWorklist(BasicBlock *BB,
9093 SmallPtrSet<BasicBlock*, 64> &Visited,
9094 std::vector<Instruction*> &WorkList,
9095 const TargetData *TD) {
9096 // We have now visited this block! If we've already been here, bail out.
9097 if (!Visited.insert(BB)) return;
9099 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9100 Instruction *Inst = BBI++;
9102 // DCE instruction if trivially dead.
9103 if (isInstructionTriviallyDead(Inst)) {
9105 DOUT << "IC: DCE: " << *Inst;
9106 Inst->eraseFromParent();
9110 // ConstantProp instruction if trivially constant.
9111 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9112 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9113 Inst->replaceAllUsesWith(C);
9115 Inst->eraseFromParent();
9119 WorkList.push_back(Inst);
9122 // Recursively visit successors. If this is a branch or switch on a constant,
9123 // only visit the reachable successor.
9124 TerminatorInst *TI = BB->getTerminator();
9125 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9126 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9127 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9128 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
9132 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9133 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9134 // See if this is an explicit destination.
9135 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9136 if (SI->getCaseValue(i) == Cond) {
9137 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
9141 // Otherwise it is the default destination.
9142 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
9147 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9148 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
9151 bool InstCombiner::runOnFunction(Function &F) {
9152 bool Changed = false;
9153 TD = &getAnalysis<TargetData>();
9156 // Do a depth-first traversal of the function, populate the worklist with
9157 // the reachable instructions. Ignore blocks that are not reachable. Keep
9158 // track of which blocks we visit.
9159 SmallPtrSet<BasicBlock*, 64> Visited;
9160 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
9162 // Do a quick scan over the function. If we find any blocks that are
9163 // unreachable, remove any instructions inside of them. This prevents
9164 // the instcombine code from having to deal with some bad special cases.
9165 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9166 if (!Visited.count(BB)) {
9167 Instruction *Term = BB->getTerminator();
9168 while (Term != BB->begin()) { // Remove instrs bottom-up
9169 BasicBlock::iterator I = Term; --I;
9171 DOUT << "IC: DCE: " << *I;
9174 if (!I->use_empty())
9175 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9176 I->eraseFromParent();
9181 while (!WorkList.empty()) {
9182 Instruction *I = WorkList.back(); // Get an instruction from the worklist
9183 WorkList.pop_back();
9185 // Check to see if we can DCE the instruction.
9186 if (isInstructionTriviallyDead(I)) {
9187 // Add operands to the worklist.
9188 if (I->getNumOperands() < 4)
9189 AddUsesToWorkList(*I);
9192 DOUT << "IC: DCE: " << *I;
9194 I->eraseFromParent();
9195 removeFromWorkList(I);
9199 // Instruction isn't dead, see if we can constant propagate it.
9200 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9201 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9203 // Add operands to the worklist.
9204 AddUsesToWorkList(*I);
9205 ReplaceInstUsesWith(*I, C);
9208 I->eraseFromParent();
9209 removeFromWorkList(I);
9213 // See if we can trivially sink this instruction to a successor basic block.
9214 if (I->hasOneUse()) {
9215 BasicBlock *BB = I->getParent();
9216 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9217 if (UserParent != BB) {
9218 bool UserIsSuccessor = false;
9219 // See if the user is one of our successors.
9220 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9221 if (*SI == UserParent) {
9222 UserIsSuccessor = true;
9226 // If the user is one of our immediate successors, and if that successor
9227 // only has us as a predecessors (we'd have to split the critical edge
9228 // otherwise), we can keep going.
9229 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9230 next(pred_begin(UserParent)) == pred_end(UserParent))
9231 // Okay, the CFG is simple enough, try to sink this instruction.
9232 Changed |= TryToSinkInstruction(I, UserParent);
9236 // Now that we have an instruction, try combining it to simplify it...
9237 if (Instruction *Result = visit(*I)) {
9239 // Should we replace the old instruction with a new one?
9241 DOUT << "IC: Old = " << *I
9242 << " New = " << *Result;
9244 // Everything uses the new instruction now.
9245 I->replaceAllUsesWith(Result);
9247 // Push the new instruction and any users onto the worklist.
9248 WorkList.push_back(Result);
9249 AddUsersToWorkList(*Result);
9251 // Move the name to the new instruction first.
9252 Result->takeName(I);
9254 // Insert the new instruction into the basic block...
9255 BasicBlock *InstParent = I->getParent();
9256 BasicBlock::iterator InsertPos = I;
9258 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9259 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9262 InstParent->getInstList().insert(InsertPos, Result);
9264 // Make sure that we reprocess all operands now that we reduced their
9266 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9267 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9268 WorkList.push_back(OpI);
9270 // Instructions can end up on the worklist more than once. Make sure
9271 // we do not process an instruction that has been deleted.
9272 removeFromWorkList(I);
9274 // Erase the old instruction.
9275 InstParent->getInstList().erase(I);
9277 DOUT << "IC: MOD = " << *I;
9279 // If the instruction was modified, it's possible that it is now dead.
9280 // if so, remove it.
9281 if (isInstructionTriviallyDead(I)) {
9282 // Make sure we process all operands now that we are reducing their
9284 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9285 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9286 WorkList.push_back(OpI);
9288 // Instructions may end up in the worklist more than once. Erase all
9289 // occurrences of this instruction.
9290 removeFromWorkList(I);
9291 I->eraseFromParent();
9293 WorkList.push_back(Result);
9294 AddUsersToWorkList(*Result);
9304 FunctionPass *llvm::createInstructionCombiningPass() {
9305 return new InstCombiner();