1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/GetElementPtrTypeIterator.h"
49 #include "llvm/Support/InstVisitor.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/PatternMatch.h"
52 #include "llvm/Support/Compiler.h"
53 #include "llvm/ADT/DenseMap.h"
54 #include "llvm/ADT/SmallVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/ADT/STLExtras.h"
61 using namespace llvm::PatternMatch;
63 STATISTIC(NumCombined , "Number of insts combined");
64 STATISTIC(NumConstProp, "Number of constant folds");
65 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
66 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
67 STATISTIC(NumSunkInst , "Number of instructions sunk");
70 class VISIBILITY_HIDDEN InstCombiner
71 : public FunctionPass,
72 public InstVisitor<InstCombiner, Instruction*> {
73 // Worklist of all of the instructions that need to be simplified.
74 std::vector<Instruction*> Worklist;
75 DenseMap<Instruction*, unsigned> WorklistMap;
77 bool MustPreserveLCSSA;
79 /// AddToWorkList - Add the specified instruction to the worklist if it
80 /// isn't already in it.
81 void AddToWorkList(Instruction *I) {
82 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
83 Worklist.push_back(I);
86 // RemoveFromWorkList - remove I from the worklist if it exists.
87 void RemoveFromWorkList(Instruction *I) {
88 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
89 if (It == WorklistMap.end()) return; // Not in worklist.
91 // Don't bother moving everything down, just null out the slot.
92 Worklist[It->second] = 0;
94 WorklistMap.erase(It);
97 Instruction *RemoveOneFromWorkList() {
98 Instruction *I = Worklist.back();
100 WorklistMap.erase(I);
105 /// AddUsersToWorkList - When an instruction is simplified, add all users of
106 /// the instruction to the work lists because they might get more simplified
109 void AddUsersToWorkList(Value &I) {
110 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
112 AddToWorkList(cast<Instruction>(*UI));
115 /// AddUsesToWorkList - When an instruction is simplified, add operands to
116 /// the work lists because they might get more simplified now.
118 void AddUsesToWorkList(Instruction &I) {
119 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
120 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
124 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
125 /// dead. Add all of its operands to the worklist, turning them into
126 /// undef's to reduce the number of uses of those instructions.
128 /// Return the specified operand before it is turned into an undef.
130 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
131 Value *R = I.getOperand(op);
133 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
134 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
136 // Set the operand to undef to drop the use.
137 I.setOperand(i, UndefValue::get(Op->getType()));
144 virtual bool runOnFunction(Function &F);
146 bool DoOneIteration(Function &F, unsigned ItNum);
148 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
149 AU.addRequired<TargetData>();
150 AU.addPreservedID(LCSSAID);
151 AU.setPreservesCFG();
154 TargetData &getTargetData() const { return *TD; }
156 // Visitation implementation - Implement instruction combining for different
157 // instruction types. The semantics are as follows:
159 // null - No change was made
160 // I - Change was made, I is still valid, I may be dead though
161 // otherwise - Change was made, replace I with returned instruction
163 Instruction *visitAdd(BinaryOperator &I);
164 Instruction *visitSub(BinaryOperator &I);
165 Instruction *visitMul(BinaryOperator &I);
166 Instruction *visitURem(BinaryOperator &I);
167 Instruction *visitSRem(BinaryOperator &I);
168 Instruction *visitFRem(BinaryOperator &I);
169 Instruction *commonRemTransforms(BinaryOperator &I);
170 Instruction *commonIRemTransforms(BinaryOperator &I);
171 Instruction *commonDivTransforms(BinaryOperator &I);
172 Instruction *commonIDivTransforms(BinaryOperator &I);
173 Instruction *visitUDiv(BinaryOperator &I);
174 Instruction *visitSDiv(BinaryOperator &I);
175 Instruction *visitFDiv(BinaryOperator &I);
176 Instruction *visitAnd(BinaryOperator &I);
177 Instruction *visitOr (BinaryOperator &I);
178 Instruction *visitXor(BinaryOperator &I);
179 Instruction *visitShl(BinaryOperator &I);
180 Instruction *visitAShr(BinaryOperator &I);
181 Instruction *visitLShr(BinaryOperator &I);
182 Instruction *commonShiftTransforms(BinaryOperator &I);
183 Instruction *visitFCmpInst(FCmpInst &I);
184 Instruction *visitICmpInst(ICmpInst &I);
185 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
187 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
188 ICmpInst::Predicate Cond, Instruction &I);
189 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
191 Instruction *commonCastTransforms(CastInst &CI);
192 Instruction *commonIntCastTransforms(CastInst &CI);
193 Instruction *visitTrunc(CastInst &CI);
194 Instruction *visitZExt(CastInst &CI);
195 Instruction *visitSExt(CastInst &CI);
196 Instruction *visitFPTrunc(CastInst &CI);
197 Instruction *visitFPExt(CastInst &CI);
198 Instruction *visitFPToUI(CastInst &CI);
199 Instruction *visitFPToSI(CastInst &CI);
200 Instruction *visitUIToFP(CastInst &CI);
201 Instruction *visitSIToFP(CastInst &CI);
202 Instruction *visitPtrToInt(CastInst &CI);
203 Instruction *visitIntToPtr(CastInst &CI);
204 Instruction *visitBitCast(CastInst &CI);
205 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
207 Instruction *visitSelectInst(SelectInst &CI);
208 Instruction *visitCallInst(CallInst &CI);
209 Instruction *visitInvokeInst(InvokeInst &II);
210 Instruction *visitPHINode(PHINode &PN);
211 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
212 Instruction *visitAllocationInst(AllocationInst &AI);
213 Instruction *visitFreeInst(FreeInst &FI);
214 Instruction *visitLoadInst(LoadInst &LI);
215 Instruction *visitStoreInst(StoreInst &SI);
216 Instruction *visitBranchInst(BranchInst &BI);
217 Instruction *visitSwitchInst(SwitchInst &SI);
218 Instruction *visitInsertElementInst(InsertElementInst &IE);
219 Instruction *visitExtractElementInst(ExtractElementInst &EI);
220 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
222 // visitInstruction - Specify what to return for unhandled instructions...
223 Instruction *visitInstruction(Instruction &I) { return 0; }
226 Instruction *visitCallSite(CallSite CS);
227 bool transformConstExprCastCall(CallSite CS);
230 // InsertNewInstBefore - insert an instruction New before instruction Old
231 // in the program. Add the new instruction to the worklist.
233 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
234 assert(New && New->getParent() == 0 &&
235 "New instruction already inserted into a basic block!");
236 BasicBlock *BB = Old.getParent();
237 BB->getInstList().insert(&Old, New); // Insert inst
242 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
243 /// This also adds the cast to the worklist. Finally, this returns the
245 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
247 if (V->getType() == Ty) return V;
249 if (Constant *CV = dyn_cast<Constant>(V))
250 return ConstantExpr::getCast(opc, CV, Ty);
252 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
257 // ReplaceInstUsesWith - This method is to be used when an instruction is
258 // found to be dead, replacable with another preexisting expression. Here
259 // we add all uses of I to the worklist, replace all uses of I with the new
260 // value, then return I, so that the inst combiner will know that I was
263 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
264 AddUsersToWorkList(I); // Add all modified instrs to worklist
266 I.replaceAllUsesWith(V);
269 // If we are replacing the instruction with itself, this must be in a
270 // segment of unreachable code, so just clobber the instruction.
271 I.replaceAllUsesWith(UndefValue::get(I.getType()));
276 // UpdateValueUsesWith - This method is to be used when an value is
277 // found to be replacable with another preexisting expression or was
278 // updated. Here we add all uses of I to the worklist, replace all uses of
279 // I with the new value (unless the instruction was just updated), then
280 // return true, so that the inst combiner will know that I was modified.
282 bool UpdateValueUsesWith(Value *Old, Value *New) {
283 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
285 Old->replaceAllUsesWith(New);
286 if (Instruction *I = dyn_cast<Instruction>(Old))
288 if (Instruction *I = dyn_cast<Instruction>(New))
293 // EraseInstFromFunction - When dealing with an instruction that has side
294 // effects or produces a void value, we can't rely on DCE to delete the
295 // instruction. Instead, visit methods should return the value returned by
297 Instruction *EraseInstFromFunction(Instruction &I) {
298 assert(I.use_empty() && "Cannot erase instruction that is used!");
299 AddUsesToWorkList(I);
300 RemoveFromWorkList(&I);
302 return 0; // Don't do anything with FI
306 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
307 /// InsertBefore instruction. This is specialized a bit to avoid inserting
308 /// casts that are known to not do anything...
310 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
311 Value *V, const Type *DestTy,
312 Instruction *InsertBefore);
314 /// SimplifyCommutative - This performs a few simplifications for
315 /// commutative operators.
316 bool SimplifyCommutative(BinaryOperator &I);
318 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
319 /// most-complex to least-complex order.
320 bool SimplifyCompare(CmpInst &I);
322 bool SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
323 uint64_t &KnownZero, uint64_t &KnownOne,
326 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
327 APInt& KnownZero, APInt& KnownOne,
330 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
331 uint64_t &UndefElts, unsigned Depth = 0);
333 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
334 // PHI node as operand #0, see if we can fold the instruction into the PHI
335 // (which is only possible if all operands to the PHI are constants).
336 Instruction *FoldOpIntoPhi(Instruction &I);
338 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
339 // operator and they all are only used by the PHI, PHI together their
340 // inputs, and do the operation once, to the result of the PHI.
341 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
342 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
345 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
346 ConstantInt *AndRHS, BinaryOperator &TheAnd);
348 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
349 bool isSub, Instruction &I);
350 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
351 bool isSigned, bool Inside, Instruction &IB);
352 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
353 Instruction *MatchBSwap(BinaryOperator &I);
355 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
358 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
361 // getComplexity: Assign a complexity or rank value to LLVM Values...
362 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
363 static unsigned getComplexity(Value *V) {
364 if (isa<Instruction>(V)) {
365 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
369 if (isa<Argument>(V)) return 3;
370 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
373 // isOnlyUse - Return true if this instruction will be deleted if we stop using
375 static bool isOnlyUse(Value *V) {
376 return V->hasOneUse() || isa<Constant>(V);
379 // getPromotedType - Return the specified type promoted as it would be to pass
380 // though a va_arg area...
381 static const Type *getPromotedType(const Type *Ty) {
382 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
383 if (ITy->getBitWidth() < 32)
384 return Type::Int32Ty;
385 } else if (Ty == Type::FloatTy)
386 return Type::DoubleTy;
390 /// getBitCastOperand - If the specified operand is a CastInst or a constant
391 /// expression bitcast, return the operand value, otherwise return null.
392 static Value *getBitCastOperand(Value *V) {
393 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
394 return I->getOperand(0);
395 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
396 if (CE->getOpcode() == Instruction::BitCast)
397 return CE->getOperand(0);
401 /// This function is a wrapper around CastInst::isEliminableCastPair. It
402 /// simply extracts arguments and returns what that function returns.
403 static Instruction::CastOps
404 isEliminableCastPair(
405 const CastInst *CI, ///< The first cast instruction
406 unsigned opcode, ///< The opcode of the second cast instruction
407 const Type *DstTy, ///< The target type for the second cast instruction
408 TargetData *TD ///< The target data for pointer size
411 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
412 const Type *MidTy = CI->getType(); // B from above
414 // Get the opcodes of the two Cast instructions
415 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
416 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
418 return Instruction::CastOps(
419 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
420 DstTy, TD->getIntPtrType()));
423 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
424 /// in any code being generated. It does not require codegen if V is simple
425 /// enough or if the cast can be folded into other casts.
426 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
427 const Type *Ty, TargetData *TD) {
428 if (V->getType() == Ty || isa<Constant>(V)) return false;
430 // If this is another cast that can be eliminated, it isn't codegen either.
431 if (const CastInst *CI = dyn_cast<CastInst>(V))
432 if (isEliminableCastPair(CI, opcode, Ty, TD))
437 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
438 /// InsertBefore instruction. This is specialized a bit to avoid inserting
439 /// casts that are known to not do anything...
441 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
442 Value *V, const Type *DestTy,
443 Instruction *InsertBefore) {
444 if (V->getType() == DestTy) return V;
445 if (Constant *C = dyn_cast<Constant>(V))
446 return ConstantExpr::getCast(opcode, C, DestTy);
448 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
451 // SimplifyCommutative - This performs a few simplifications for commutative
454 // 1. Order operands such that they are listed from right (least complex) to
455 // left (most complex). This puts constants before unary operators before
458 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
459 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
461 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
462 bool Changed = false;
463 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
464 Changed = !I.swapOperands();
466 if (!I.isAssociative()) return Changed;
467 Instruction::BinaryOps Opcode = I.getOpcode();
468 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
469 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
470 if (isa<Constant>(I.getOperand(1))) {
471 Constant *Folded = ConstantExpr::get(I.getOpcode(),
472 cast<Constant>(I.getOperand(1)),
473 cast<Constant>(Op->getOperand(1)));
474 I.setOperand(0, Op->getOperand(0));
475 I.setOperand(1, Folded);
477 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
478 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
479 isOnlyUse(Op) && isOnlyUse(Op1)) {
480 Constant *C1 = cast<Constant>(Op->getOperand(1));
481 Constant *C2 = cast<Constant>(Op1->getOperand(1));
483 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
484 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
485 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
489 I.setOperand(0, New);
490 I.setOperand(1, Folded);
497 /// SimplifyCompare - For a CmpInst this function just orders the operands
498 /// so that theyare listed from right (least complex) to left (most complex).
499 /// This puts constants before unary operators before binary operators.
500 bool InstCombiner::SimplifyCompare(CmpInst &I) {
501 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
504 // Compare instructions are not associative so there's nothing else we can do.
508 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
509 // if the LHS is a constant zero (which is the 'negate' form).
511 static inline Value *dyn_castNegVal(Value *V) {
512 if (BinaryOperator::isNeg(V))
513 return BinaryOperator::getNegArgument(V);
515 // Constants can be considered to be negated values if they can be folded.
516 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
517 return ConstantExpr::getNeg(C);
521 static inline Value *dyn_castNotVal(Value *V) {
522 if (BinaryOperator::isNot(V))
523 return BinaryOperator::getNotArgument(V);
525 // Constants can be considered to be not'ed values...
526 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
527 return ConstantExpr::getNot(C);
531 // dyn_castFoldableMul - If this value is a multiply that can be folded into
532 // other computations (because it has a constant operand), return the
533 // non-constant operand of the multiply, and set CST to point to the multiplier.
534 // Otherwise, return null.
536 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
537 if (V->hasOneUse() && V->getType()->isInteger())
538 if (Instruction *I = dyn_cast<Instruction>(V)) {
539 if (I->getOpcode() == Instruction::Mul)
540 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
541 return I->getOperand(0);
542 if (I->getOpcode() == Instruction::Shl)
543 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
544 // The multiplier is really 1 << CST.
545 Constant *One = ConstantInt::get(V->getType(), 1);
546 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
547 return I->getOperand(0);
553 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
554 /// expression, return it.
555 static User *dyn_castGetElementPtr(Value *V) {
556 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
557 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
558 if (CE->getOpcode() == Instruction::GetElementPtr)
559 return cast<User>(V);
563 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
564 static ConstantInt *AddOne(ConstantInt *C) {
565 return cast<ConstantInt>(ConstantExpr::getAdd(C,
566 ConstantInt::get(C->getType(), 1)));
568 static ConstantInt *SubOne(ConstantInt *C) {
569 return cast<ConstantInt>(ConstantExpr::getSub(C,
570 ConstantInt::get(C->getType(), 1)));
573 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
574 /// known to be either zero or one and return them in the KnownZero/KnownOne
575 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
577 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
578 /// we cannot optimize based on the assumption that it is zero without changing
579 /// it to be an explicit zero. If we don't change it to zero, other code could
580 /// optimized based on the contradictory assumption that it is non-zero.
581 /// Because instcombine aggressively folds operations with undef args anyway,
582 /// this won't lose us code quality.
583 static void ComputeMaskedBits(Value *V, APInt Mask, APInt& KnownZero,
584 APInt& KnownOne, unsigned Depth = 0) {
585 assert(V && "No Value?");
586 assert(Depth <= 6 && "Limit Search Depth");
587 uint32_t BitWidth = Mask.getBitWidth();
588 const IntegerType *VTy = cast<IntegerType>(V->getType());
589 assert(VTy->getBitWidth() == BitWidth &&
590 KnownZero.getBitWidth() == BitWidth &&
591 KnownOne.getBitWidth() == BitWidth &&
592 "VTy, Mask, KnownOne and KnownZero should have same BitWidth");
593 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
594 // We know all of the bits for a constant!
595 KnownOne = CI->getValue() & Mask;
596 KnownZero = ~KnownOne & Mask;
600 if (Depth == 6 || Mask == 0)
601 return; // Limit search depth.
603 Instruction *I = dyn_cast<Instruction>(V);
606 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
607 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
608 Mask &= APInt::getAllOnesValue(BitWidth);
610 switch (I->getOpcode()) {
611 case Instruction::And:
612 // If either the LHS or the RHS are Zero, the result is zero.
613 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
615 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
616 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
617 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
619 // Output known-1 bits are only known if set in both the LHS & RHS.
620 KnownOne &= KnownOne2;
621 // Output known-0 are known to be clear if zero in either the LHS | RHS.
622 KnownZero |= KnownZero2;
624 case Instruction::Or:
625 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
627 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
628 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
629 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
631 // Output known-0 bits are only known if clear in both the LHS & RHS.
632 KnownZero &= KnownZero2;
633 // Output known-1 are known to be set if set in either the LHS | RHS.
634 KnownOne |= KnownOne2;
636 case Instruction::Xor: {
637 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
638 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
639 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
640 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
642 // Output known-0 bits are known if clear or set in both the LHS & RHS.
643 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
644 // Output known-1 are known to be set if set in only one of the LHS, RHS.
645 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
646 KnownZero = KnownZeroOut;
649 case Instruction::Select:
650 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
651 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
652 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
653 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
655 // Only known if known in both the LHS and RHS.
656 KnownOne &= KnownOne2;
657 KnownZero &= KnownZero2;
659 case Instruction::FPTrunc:
660 case Instruction::FPExt:
661 case Instruction::FPToUI:
662 case Instruction::FPToSI:
663 case Instruction::SIToFP:
664 case Instruction::PtrToInt:
665 case Instruction::UIToFP:
666 case Instruction::IntToPtr:
667 return; // Can't work with floating point or pointers
668 case Instruction::Trunc: {
669 // All these have integer operands
670 uint32_t SrcBitWidth =
671 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
672 ComputeMaskedBits(I->getOperand(0), Mask.zext(SrcBitWidth),
673 KnownZero.zext(SrcBitWidth), KnownOne.zext(SrcBitWidth), Depth+1);
674 KnownZero.trunc(BitWidth);
675 KnownOne.trunc(BitWidth);
678 case Instruction::BitCast: {
679 const Type *SrcTy = I->getOperand(0)->getType();
680 if (SrcTy->isInteger()) {
681 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
686 case Instruction::ZExt: {
687 // Compute the bits in the result that are not present in the input.
688 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
689 APInt NewBits(APInt::getAllOnesValue(BitWidth).shl(SrcTy->getBitWidth()));
691 uint32_t SrcBitWidth = SrcTy->getBitWidth();
692 ComputeMaskedBits(I->getOperand(0), Mask.trunc(SrcBitWidth),
693 KnownZero.trunc(SrcBitWidth), KnownOne.trunc(SrcBitWidth), Depth+1);
694 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
695 // The top bits are known to be zero.
696 KnownZero.zext(BitWidth);
697 KnownOne.zext(BitWidth);
698 KnownZero |= NewBits;
701 case Instruction::SExt: {
702 // Compute the bits in the result that are not present in the input.
703 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
704 APInt NewBits(APInt::getAllOnesValue(BitWidth).shl(SrcTy->getBitWidth()));
706 uint32_t SrcBitWidth = SrcTy->getBitWidth();
707 ComputeMaskedBits(I->getOperand(0), Mask.trunc(SrcBitWidth),
708 KnownZero.trunc(SrcBitWidth), KnownOne.trunc(SrcBitWidth), Depth+1);
709 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
710 KnownZero.zext(BitWidth);
711 KnownOne.zext(BitWidth);
713 // If the sign bit of the input is known set or clear, then we know the
714 // top bits of the result.
715 APInt InSignBit(APInt::getSignBit(SrcTy->getBitWidth()));
716 InSignBit.zext(BitWidth);
717 if ((KnownZero & InSignBit) != 0) { // Input sign bit known zero
718 KnownZero |= NewBits;
719 KnownOne &= ~NewBits;
720 } else if ((KnownOne & InSignBit) != 0) { // Input sign bit known set
722 KnownZero &= ~NewBits;
723 } else { // Input sign bit unknown
724 KnownZero &= ~NewBits;
725 KnownOne &= ~NewBits;
729 case Instruction::Shl:
730 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
731 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
732 uint64_t ShiftAmt = SA->getZExtValue();
733 Mask = APIntOps::lshr(Mask, ShiftAmt);
734 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
735 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
736 KnownZero <<= ShiftAmt;
737 KnownOne <<= ShiftAmt;
738 KnownZero |= APInt(BitWidth, 1ULL).shl(ShiftAmt)-1; // low bits known zero.
742 case Instruction::LShr:
743 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
744 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
745 // Compute the new bits that are at the top now.
746 uint64_t ShiftAmt = SA->getZExtValue();
747 APInt HighBits(APInt::getAllOnesValue(BitWidth).shl(BitWidth-ShiftAmt));
749 // Unsigned shift right.
751 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
752 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
753 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
754 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
755 KnownZero |= HighBits; // high bits known zero.
759 case Instruction::AShr:
760 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
761 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
762 // Compute the new bits that are at the top now.
763 uint64_t ShiftAmt = SA->getZExtValue();
764 APInt HighBits(APInt::getAllOnesValue(BitWidth).shl(BitWidth-ShiftAmt));
766 // Signed shift right.
768 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
769 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
770 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
771 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
773 // Handle the sign bits and adjust to where it is now in the mask.
774 APInt SignBit(APInt::getSignBit(BitWidth).lshr(ShiftAmt));
776 if ((KnownZero & SignBit) != 0) { // New bits are known zero.
777 KnownZero |= HighBits;
778 } else if ((KnownOne & SignBit) != 0) { // New bits are known one.
779 KnownOne |= HighBits;
787 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
788 /// known to be either zero or one and return them in the KnownZero/KnownOne
789 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
791 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
792 uint64_t &KnownOne, unsigned Depth = 0) {
793 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
794 // we cannot optimize based on the assumption that it is zero without changing
795 // it to be an explicit zero. If we don't change it to zero, other code could
796 // optimized based on the contradictory assumption that it is non-zero.
797 // Because instcombine aggressively folds operations with undef args anyway,
798 // this won't lose us code quality.
799 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
800 // We know all of the bits for a constant!
801 KnownOne = CI->getZExtValue() & Mask;
802 KnownZero = ~KnownOne & Mask;
806 KnownZero = KnownOne = 0; // Don't know anything.
807 if (Depth == 6 || Mask == 0)
808 return; // Limit search depth.
810 uint64_t KnownZero2, KnownOne2;
811 Instruction *I = dyn_cast<Instruction>(V);
814 Mask &= cast<IntegerType>(V->getType())->getBitMask();
816 switch (I->getOpcode()) {
817 case Instruction::And:
818 // If either the LHS or the RHS are Zero, the result is zero.
819 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
821 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
822 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
823 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
825 // Output known-1 bits are only known if set in both the LHS & RHS.
826 KnownOne &= KnownOne2;
827 // Output known-0 are known to be clear if zero in either the LHS | RHS.
828 KnownZero |= KnownZero2;
830 case Instruction::Or:
831 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
833 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
834 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
835 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
837 // Output known-0 bits are only known if clear in both the LHS & RHS.
838 KnownZero &= KnownZero2;
839 // Output known-1 are known to be set if set in either the LHS | RHS.
840 KnownOne |= KnownOne2;
842 case Instruction::Xor: {
843 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
844 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
845 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
846 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
848 // Output known-0 bits are known if clear or set in both the LHS & RHS.
849 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
850 // Output known-1 are known to be set if set in only one of the LHS, RHS.
851 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
852 KnownZero = KnownZeroOut;
855 case Instruction::Select:
856 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
857 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
858 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
859 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
861 // Only known if known in both the LHS and RHS.
862 KnownOne &= KnownOne2;
863 KnownZero &= KnownZero2;
865 case Instruction::FPTrunc:
866 case Instruction::FPExt:
867 case Instruction::FPToUI:
868 case Instruction::FPToSI:
869 case Instruction::SIToFP:
870 case Instruction::PtrToInt:
871 case Instruction::UIToFP:
872 case Instruction::IntToPtr:
873 return; // Can't work with floating point or pointers
874 case Instruction::Trunc:
875 // All these have integer operands
876 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
878 case Instruction::BitCast: {
879 const Type *SrcTy = I->getOperand(0)->getType();
880 if (SrcTy->isInteger()) {
881 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
886 case Instruction::ZExt: {
887 // Compute the bits in the result that are not present in the input.
888 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
889 uint64_t NotIn = ~SrcTy->getBitMask();
890 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
892 Mask &= SrcTy->getBitMask();
893 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
894 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
895 // The top bits are known to be zero.
896 KnownZero |= NewBits;
899 case Instruction::SExt: {
900 // Compute the bits in the result that are not present in the input.
901 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
902 uint64_t NotIn = ~SrcTy->getBitMask();
903 uint64_t NewBits = cast<IntegerType>(I->getType())->getBitMask() & NotIn;
905 Mask &= SrcTy->getBitMask();
906 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
907 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
909 // If the sign bit of the input is known set or clear, then we know the
910 // top bits of the result.
911 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
912 if (KnownZero & InSignBit) { // Input sign bit known zero
913 KnownZero |= NewBits;
914 KnownOne &= ~NewBits;
915 } else if (KnownOne & InSignBit) { // Input sign bit known set
917 KnownZero &= ~NewBits;
918 } else { // Input sign bit unknown
919 KnownZero &= ~NewBits;
920 KnownOne &= ~NewBits;
924 case Instruction::Shl:
925 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
926 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
927 uint64_t ShiftAmt = SA->getZExtValue();
929 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
930 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
931 KnownZero <<= ShiftAmt;
932 KnownOne <<= ShiftAmt;
933 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
937 case Instruction::LShr:
938 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
939 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
940 // Compute the new bits that are at the top now.
941 uint64_t ShiftAmt = SA->getZExtValue();
942 uint64_t HighBits = (1ULL << ShiftAmt)-1;
943 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
945 // Unsigned shift right.
947 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
948 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
949 KnownZero >>= ShiftAmt;
950 KnownOne >>= ShiftAmt;
951 KnownZero |= HighBits; // high bits known zero.
955 case Instruction::AShr:
956 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
957 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
958 // Compute the new bits that are at the top now.
959 uint64_t ShiftAmt = SA->getZExtValue();
960 uint64_t HighBits = (1ULL << ShiftAmt)-1;
961 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
963 // Signed shift right.
965 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
966 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
967 KnownZero >>= ShiftAmt;
968 KnownOne >>= ShiftAmt;
970 // Handle the sign bits.
971 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
972 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
974 if (KnownZero & SignBit) { // New bits are known zero.
975 KnownZero |= HighBits;
976 } else if (KnownOne & SignBit) { // New bits are known one.
977 KnownOne |= HighBits;
985 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
986 /// this predicate to simplify operations downstream. Mask is known to be zero
987 /// for bits that V cannot have.
988 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
989 uint64_t KnownZero, KnownOne;
990 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
991 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
992 return (KnownZero & Mask) == Mask;
995 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
996 /// this predicate to simplify operations downstream. Mask is known to be zero
997 /// for bits that V cannot have.
998 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
999 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1000 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
1001 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1002 return (KnownZero & Mask) == Mask;
1005 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1006 /// specified instruction is a constant integer. If so, check to see if there
1007 /// are any bits set in the constant that are not demanded. If so, shrink the
1008 /// constant and return true.
1009 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
1010 uint64_t Demanded) {
1011 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
1012 if (!OpC) return false;
1014 // If there are no bits set that aren't demanded, nothing to do.
1015 if ((~Demanded & OpC->getZExtValue()) == 0)
1018 // This is producing any bits that are not needed, shrink the RHS.
1019 uint64_t Val = Demanded & OpC->getZExtValue();
1020 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Val));
1024 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1025 /// specified instruction is a constant integer. If so, check to see if there
1026 /// are any bits set in the constant that are not demanded. If so, shrink the
1027 /// constant and return true.
1028 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
1030 assert(I && "No instruction?");
1031 assert(OpNo < I->getNumOperands() && "Operand index too large");
1033 // If the operand is not a constant integer, nothing to do.
1034 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
1035 if (!OpC) return false;
1037 // If there are no bits set that aren't demanded, nothing to do.
1038 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
1039 if ((~Demanded & OpC->getValue()) == 0)
1042 // This instruction is producing bits that are not demanded. Shrink the RHS.
1043 Demanded &= OpC->getValue();
1044 I->setOperand(OpNo, ConstantInt::get(Demanded));
1048 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
1049 // set of known zero and one bits, compute the maximum and minimum values that
1050 // could have the specified known zero and known one bits, returning them in
1052 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
1053 const APInt& KnownZero,
1054 const APInt& KnownOne,
1055 APInt& Min, APInt& Max) {
1056 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
1057 assert(KnownZero.getBitWidth() == BitWidth &&
1058 KnownOne.getBitWidth() == BitWidth &&
1059 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
1060 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1061 APInt TypeBits(APInt::getAllOnesValue(BitWidth));
1062 APInt UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
1064 APInt SignBit(APInt::getSignBit(BitWidth));
1066 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
1067 // bit if it is unknown.
1069 Max = KnownOne|UnknownBits;
1071 if ((SignBit & UnknownBits) != 0) { // Sign bit is unknown
1077 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
1078 // a set of known zero and one bits, compute the maximum and minimum values that
1079 // could have the specified known zero and known one bits, returning them in
1081 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
1082 const APInt& KnownZero,
1083 const APInt& KnownOne,
1086 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
1087 assert(KnownZero.getBitWidth() == BitWidth &&
1088 KnownOne.getBitWidth() == BitWidth &&
1089 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
1090 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1091 APInt TypeBits(APInt::getAllOnesValue(BitWidth));
1092 APInt UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
1094 // The minimum value is when the unknown bits are all zeros.
1096 // The maximum value is when the unknown bits are all ones.
1097 Max = KnownOne|UnknownBits;
1100 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
1101 /// DemandedMask bits of the result of V are ever used downstream. If we can
1102 /// use this information to simplify V, do so and return true. Otherwise,
1103 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
1104 /// the expression (used to simplify the caller). The KnownZero/One bits may
1105 /// only be accurate for those bits in the DemandedMask.
1106 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
1107 uint64_t &KnownZero, uint64_t &KnownOne,
1109 const IntegerType *VTy = cast<IntegerType>(V->getType());
1110 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1111 // We know all of the bits for a constant!
1112 KnownOne = CI->getZExtValue() & DemandedMask;
1113 KnownZero = ~KnownOne & DemandedMask;
1117 KnownZero = KnownOne = 0;
1118 if (!V->hasOneUse()) { // Other users may use these bits.
1119 if (Depth != 0) { // Not at the root.
1120 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
1121 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
1124 // If this is the root being simplified, allow it to have multiple uses,
1125 // just set the DemandedMask to all bits.
1126 DemandedMask = VTy->getBitMask();
1127 } else if (DemandedMask == 0) { // Not demanding any bits from V.
1128 if (V != UndefValue::get(VTy))
1129 return UpdateValueUsesWith(V, UndefValue::get(VTy));
1131 } else if (Depth == 6) { // Limit search depth.
1135 Instruction *I = dyn_cast<Instruction>(V);
1136 if (!I) return false; // Only analyze instructions.
1138 DemandedMask &= VTy->getBitMask();
1140 uint64_t KnownZero2 = 0, KnownOne2 = 0;
1141 switch (I->getOpcode()) {
1143 case Instruction::And:
1144 // If either the LHS or the RHS are Zero, the result is zero.
1145 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1146 KnownZero, KnownOne, Depth+1))
1148 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1150 // If something is known zero on the RHS, the bits aren't demanded on the
1152 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
1153 KnownZero2, KnownOne2, Depth+1))
1155 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1157 // If all of the demanded bits are known 1 on one side, return the other.
1158 // These bits cannot contribute to the result of the 'and'.
1159 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
1160 return UpdateValueUsesWith(I, I->getOperand(0));
1161 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
1162 return UpdateValueUsesWith(I, I->getOperand(1));
1164 // If all of the demanded bits in the inputs are known zeros, return zero.
1165 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
1166 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1168 // If the RHS is a constant, see if we can simplify it.
1169 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
1170 return UpdateValueUsesWith(I, I);
1172 // Output known-1 bits are only known if set in both the LHS & RHS.
1173 KnownOne &= KnownOne2;
1174 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1175 KnownZero |= KnownZero2;
1177 case Instruction::Or:
1178 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1179 KnownZero, KnownOne, Depth+1))
1181 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1182 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
1183 KnownZero2, KnownOne2, Depth+1))
1185 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1187 // If all of the demanded bits are known zero on one side, return the other.
1188 // These bits cannot contribute to the result of the 'or'.
1189 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
1190 return UpdateValueUsesWith(I, I->getOperand(0));
1191 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
1192 return UpdateValueUsesWith(I, I->getOperand(1));
1194 // If all of the potentially set bits on one side are known to be set on
1195 // the other side, just use the 'other' side.
1196 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
1197 (DemandedMask & (~KnownZero)))
1198 return UpdateValueUsesWith(I, I->getOperand(0));
1199 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
1200 (DemandedMask & (~KnownZero2)))
1201 return UpdateValueUsesWith(I, I->getOperand(1));
1203 // If the RHS is a constant, see if we can simplify it.
1204 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1205 return UpdateValueUsesWith(I, I);
1207 // Output known-0 bits are only known if clear in both the LHS & RHS.
1208 KnownZero &= KnownZero2;
1209 // Output known-1 are known to be set if set in either the LHS | RHS.
1210 KnownOne |= KnownOne2;
1212 case Instruction::Xor: {
1213 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1214 KnownZero, KnownOne, Depth+1))
1216 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1217 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1218 KnownZero2, KnownOne2, Depth+1))
1220 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1222 // If all of the demanded bits are known zero on one side, return the other.
1223 // These bits cannot contribute to the result of the 'xor'.
1224 if ((DemandedMask & KnownZero) == DemandedMask)
1225 return UpdateValueUsesWith(I, I->getOperand(0));
1226 if ((DemandedMask & KnownZero2) == DemandedMask)
1227 return UpdateValueUsesWith(I, I->getOperand(1));
1229 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1230 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1231 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1232 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1234 // If all of the demanded bits are known to be zero on one side or the
1235 // other, turn this into an *inclusive* or.
1236 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1237 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
1239 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1241 InsertNewInstBefore(Or, *I);
1242 return UpdateValueUsesWith(I, Or);
1245 // If all of the demanded bits on one side are known, and all of the set
1246 // bits on that side are also known to be set on the other side, turn this
1247 // into an AND, as we know the bits will be cleared.
1248 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1249 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
1250 if ((KnownOne & KnownOne2) == KnownOne) {
1251 Constant *AndC = ConstantInt::get(VTy, ~KnownOne & DemandedMask);
1253 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1254 InsertNewInstBefore(And, *I);
1255 return UpdateValueUsesWith(I, And);
1259 // If the RHS is a constant, see if we can simplify it.
1260 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1261 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1262 return UpdateValueUsesWith(I, I);
1264 KnownZero = KnownZeroOut;
1265 KnownOne = KnownOneOut;
1268 case Instruction::Select:
1269 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1270 KnownZero, KnownOne, Depth+1))
1272 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1273 KnownZero2, KnownOne2, Depth+1))
1275 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1276 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1278 // If the operands are constants, see if we can simplify them.
1279 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1280 return UpdateValueUsesWith(I, I);
1281 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1282 return UpdateValueUsesWith(I, I);
1284 // Only known if known in both the LHS and RHS.
1285 KnownOne &= KnownOne2;
1286 KnownZero &= KnownZero2;
1288 case Instruction::Trunc:
1289 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1290 KnownZero, KnownOne, Depth+1))
1292 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1294 case Instruction::BitCast:
1295 if (!I->getOperand(0)->getType()->isInteger())
1298 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1299 KnownZero, KnownOne, Depth+1))
1301 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1303 case Instruction::ZExt: {
1304 // Compute the bits in the result that are not present in the input.
1305 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1306 uint64_t NotIn = ~SrcTy->getBitMask();
1307 uint64_t NewBits = VTy->getBitMask() & NotIn;
1309 DemandedMask &= SrcTy->getBitMask();
1310 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1311 KnownZero, KnownOne, Depth+1))
1313 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1314 // The top bits are known to be zero.
1315 KnownZero |= NewBits;
1318 case Instruction::SExt: {
1319 // Compute the bits in the result that are not present in the input.
1320 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1321 uint64_t NotIn = ~SrcTy->getBitMask();
1322 uint64_t NewBits = VTy->getBitMask() & NotIn;
1324 // Get the sign bit for the source type
1325 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1326 int64_t InputDemandedBits = DemandedMask & SrcTy->getBitMask();
1328 // If any of the sign extended bits are demanded, we know that the sign
1330 if (NewBits & DemandedMask)
1331 InputDemandedBits |= InSignBit;
1333 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1334 KnownZero, KnownOne, Depth+1))
1336 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1338 // If the sign bit of the input is known set or clear, then we know the
1339 // top bits of the result.
1341 // If the input sign bit is known zero, or if the NewBits are not demanded
1342 // convert this into a zero extension.
1343 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1344 // Convert to ZExt cast
1345 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1346 return UpdateValueUsesWith(I, NewCast);
1347 } else if (KnownOne & InSignBit) { // Input sign bit known set
1348 KnownOne |= NewBits;
1349 KnownZero &= ~NewBits;
1350 } else { // Input sign bit unknown
1351 KnownZero &= ~NewBits;
1352 KnownOne &= ~NewBits;
1356 case Instruction::Add:
1357 // If there is a constant on the RHS, there are a variety of xformations
1359 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1360 // If null, this should be simplified elsewhere. Some of the xforms here
1361 // won't work if the RHS is zero.
1362 if (RHS->isNullValue())
1365 // Figure out what the input bits are. If the top bits of the and result
1366 // are not demanded, then the add doesn't demand them from its input
1369 // Shift the demanded mask up so that it's at the top of the uint64_t.
1370 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
1371 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1373 // If the top bit of the output is demanded, demand everything from the
1374 // input. Otherwise, we demand all the input bits except NLZ top bits.
1375 uint64_t InDemandedBits = ~0ULL >> (64-BitWidth+NLZ);
1377 // Find information about known zero/one bits in the input.
1378 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1379 KnownZero2, KnownOne2, Depth+1))
1382 // If the RHS of the add has bits set that can't affect the input, reduce
1384 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1385 return UpdateValueUsesWith(I, I);
1387 // Avoid excess work.
1388 if (KnownZero2 == 0 && KnownOne2 == 0)
1391 // Turn it into OR if input bits are zero.
1392 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1394 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1396 InsertNewInstBefore(Or, *I);
1397 return UpdateValueUsesWith(I, Or);
1400 // We can say something about the output known-zero and known-one bits,
1401 // depending on potential carries from the input constant and the
1402 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1403 // bits set and the RHS constant is 0x01001, then we know we have a known
1404 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1406 // To compute this, we first compute the potential carry bits. These are
1407 // the bits which may be modified. I'm not aware of a better way to do
1409 uint64_t RHSVal = RHS->getZExtValue();
1411 bool CarryIn = false;
1412 uint64_t CarryBits = 0;
1413 uint64_t CurBit = 1;
1414 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1415 // Record the current carry in.
1416 if (CarryIn) CarryBits |= CurBit;
1420 // This bit has a carry out unless it is "zero + zero" or
1421 // "zero + anything" with no carry in.
1422 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1423 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1424 } else if (!CarryIn &&
1425 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1426 CarryOut = false; // 0 + anything has no carry out if no carry in.
1428 // Otherwise, we have to assume we have a carry out.
1432 // This stage's carry out becomes the next stage's carry-in.
1436 // Now that we know which bits have carries, compute the known-1/0 sets.
1438 // Bits are known one if they are known zero in one operand and one in the
1439 // other, and there is no input carry.
1440 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1442 // Bits are known zero if they are known zero in both operands and there
1443 // is no input carry.
1444 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1446 // If the high-bits of this ADD are not demanded, then it does not demand
1447 // the high bits of its LHS or RHS.
1448 if ((DemandedMask & VTy->getSignBit()) == 0) {
1449 // Right fill the mask of bits for this ADD to demand the most
1450 // significant bit and all those below it.
1451 unsigned NLZ = CountLeadingZeros_64(DemandedMask);
1452 uint64_t DemandedFromOps = ~0ULL >> NLZ;
1453 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1454 KnownZero2, KnownOne2, Depth+1))
1456 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1457 KnownZero2, KnownOne2, Depth+1))
1462 case Instruction::Sub:
1463 // If the high-bits of this SUB are not demanded, then it does not demand
1464 // the high bits of its LHS or RHS.
1465 if ((DemandedMask & VTy->getSignBit()) == 0) {
1466 // Right fill the mask of bits for this SUB to demand the most
1467 // significant bit and all those below it.
1468 unsigned NLZ = CountLeadingZeros_64(DemandedMask);
1469 uint64_t DemandedFromOps = ~0ULL >> NLZ;
1470 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1471 KnownZero2, KnownOne2, Depth+1))
1473 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1474 KnownZero2, KnownOne2, Depth+1))
1478 case Instruction::Shl:
1479 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1480 uint64_t ShiftAmt = SA->getZExtValue();
1481 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1482 KnownZero, KnownOne, Depth+1))
1484 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1485 KnownZero <<= ShiftAmt;
1486 KnownOne <<= ShiftAmt;
1487 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1490 case Instruction::LShr:
1491 // For a logical shift right
1492 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1493 unsigned ShiftAmt = SA->getZExtValue();
1495 // Compute the new bits that are at the top now.
1496 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1497 HighBits <<= VTy->getBitWidth() - ShiftAmt;
1498 uint64_t TypeMask = VTy->getBitMask();
1499 // Unsigned shift right.
1500 if (SimplifyDemandedBits(I->getOperand(0),
1501 (DemandedMask << ShiftAmt) & TypeMask,
1502 KnownZero, KnownOne, Depth+1))
1504 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1505 KnownZero &= TypeMask;
1506 KnownOne &= TypeMask;
1507 KnownZero >>= ShiftAmt;
1508 KnownOne >>= ShiftAmt;
1509 KnownZero |= HighBits; // high bits known zero.
1512 case Instruction::AShr:
1513 // If this is an arithmetic shift right and only the low-bit is set, we can
1514 // always convert this into a logical shr, even if the shift amount is
1515 // variable. The low bit of the shift cannot be an input sign bit unless
1516 // the shift amount is >= the size of the datatype, which is undefined.
1517 if (DemandedMask == 1) {
1518 // Perform the logical shift right.
1519 Value *NewVal = BinaryOperator::createLShr(
1520 I->getOperand(0), I->getOperand(1), I->getName());
1521 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1522 return UpdateValueUsesWith(I, NewVal);
1525 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1526 unsigned ShiftAmt = SA->getZExtValue();
1528 // Compute the new bits that are at the top now.
1529 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1530 HighBits <<= VTy->getBitWidth() - ShiftAmt;
1531 uint64_t TypeMask = VTy->getBitMask();
1532 // Signed shift right.
1533 if (SimplifyDemandedBits(I->getOperand(0),
1534 (DemandedMask << ShiftAmt) & TypeMask,
1535 KnownZero, KnownOne, Depth+1))
1537 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1538 KnownZero &= TypeMask;
1539 KnownOne &= TypeMask;
1540 KnownZero >>= ShiftAmt;
1541 KnownOne >>= ShiftAmt;
1543 // Handle the sign bits.
1544 uint64_t SignBit = 1ULL << (VTy->getBitWidth()-1);
1545 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1547 // If the input sign bit is known to be zero, or if none of the top bits
1548 // are demanded, turn this into an unsigned shift right.
1549 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1550 // Perform the logical shift right.
1551 Value *NewVal = BinaryOperator::createLShr(
1552 I->getOperand(0), SA, I->getName());
1553 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1554 return UpdateValueUsesWith(I, NewVal);
1555 } else if (KnownOne & SignBit) { // New bits are known one.
1556 KnownOne |= HighBits;
1562 // If the client is only demanding bits that we know, return the known
1564 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1565 return UpdateValueUsesWith(I, ConstantInt::get(VTy, KnownOne));
1569 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
1570 /// value based on the demanded bits. When this function is called, it is known
1571 /// that only the bits set in DemandedMask of the result of V are ever used
1572 /// downstream. Consequently, depending on the mask and V, it may be possible
1573 /// to replace V with a constant or one of its operands. In such cases, this
1574 /// function does the replacement and returns true. In all other cases, it
1575 /// returns false after analyzing the expression and setting KnownOne and known
1576 /// to be one in the expression. KnownZero contains all the bits that are known
1577 /// to be zero in the expression. These are provided to potentially allow the
1578 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
1579 /// the expression. KnownOne and KnownZero always follow the invariant that
1580 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
1581 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
1582 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
1583 /// and KnownOne must all be the same.
1584 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
1585 APInt& KnownZero, APInt& KnownOne,
1587 assert(V != 0 && "Null pointer of Value???");
1588 assert(Depth <= 6 && "Limit Search Depth");
1589 uint32_t BitWidth = DemandedMask.getBitWidth();
1590 const IntegerType *VTy = cast<IntegerType>(V->getType());
1591 assert(VTy->getBitWidth() == BitWidth &&
1592 KnownZero.getBitWidth() == BitWidth &&
1593 KnownOne.getBitWidth() == BitWidth &&
1594 "Value *V, DemandedMask, KnownZero and KnownOne \
1595 must have same BitWidth");
1596 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1597 // We know all of the bits for a constant!
1598 KnownOne = CI->getValue() & DemandedMask;
1599 KnownZero = ~KnownOne & DemandedMask;
1605 if (!V->hasOneUse()) { // Other users may use these bits.
1606 if (Depth != 0) { // Not at the root.
1607 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
1608 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
1611 // If this is the root being simplified, allow it to have multiple uses,
1612 // just set the DemandedMask to all bits.
1613 DemandedMask = APInt::getAllOnesValue(BitWidth);
1614 } else if (DemandedMask == 0) { // Not demanding any bits from V.
1615 if (V != UndefValue::get(VTy))
1616 return UpdateValueUsesWith(V, UndefValue::get(VTy));
1618 } else if (Depth == 6) { // Limit search depth.
1622 Instruction *I = dyn_cast<Instruction>(V);
1623 if (!I) return false; // Only analyze instructions.
1625 DemandedMask &= APInt::getAllOnesValue(BitWidth);
1627 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1628 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
1629 switch (I->getOpcode()) {
1631 case Instruction::And:
1632 // If either the LHS or the RHS are Zero, the result is zero.
1633 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1634 RHSKnownZero, RHSKnownOne, Depth+1))
1636 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1637 "Bits known to be one AND zero?");
1639 // If something is known zero on the RHS, the bits aren't demanded on the
1641 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
1642 LHSKnownZero, LHSKnownOne, Depth+1))
1644 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1645 "Bits known to be one AND zero?");
1647 // If all of the demanded bits are known 1 on one side, return the other.
1648 // These bits cannot contribute to the result of the 'and'.
1649 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1650 (DemandedMask & ~LHSKnownZero))
1651 return UpdateValueUsesWith(I, I->getOperand(0));
1652 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1653 (DemandedMask & ~RHSKnownZero))
1654 return UpdateValueUsesWith(I, I->getOperand(1));
1656 // If all of the demanded bits in the inputs are known zeros, return zero.
1657 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1658 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1660 // If the RHS is a constant, see if we can simplify it.
1661 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1662 return UpdateValueUsesWith(I, I);
1664 // Output known-1 bits are only known if set in both the LHS & RHS.
1665 RHSKnownOne &= LHSKnownOne;
1666 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1667 RHSKnownZero |= LHSKnownZero;
1669 case Instruction::Or:
1670 // If either the LHS or the RHS are One, the result is One.
1671 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1672 RHSKnownZero, RHSKnownOne, Depth+1))
1674 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1675 "Bits known to be one AND zero?");
1676 // If something is known one on the RHS, the bits aren't demanded on the
1678 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1679 LHSKnownZero, LHSKnownOne, Depth+1))
1681 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1682 "Bits known to be one AND zero?");
1684 // If all of the demanded bits are known zero on one side, return the other.
1685 // These bits cannot contribute to the result of the 'or'.
1686 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1687 (DemandedMask & ~LHSKnownOne))
1688 return UpdateValueUsesWith(I, I->getOperand(0));
1689 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1690 (DemandedMask & ~RHSKnownOne))
1691 return UpdateValueUsesWith(I, I->getOperand(1));
1693 // If all of the potentially set bits on one side are known to be set on
1694 // the other side, just use the 'other' side.
1695 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1696 (DemandedMask & (~RHSKnownZero)))
1697 return UpdateValueUsesWith(I, I->getOperand(0));
1698 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1699 (DemandedMask & (~LHSKnownZero)))
1700 return UpdateValueUsesWith(I, I->getOperand(1));
1702 // If the RHS is a constant, see if we can simplify it.
1703 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1704 return UpdateValueUsesWith(I, I);
1706 // Output known-0 bits are only known if clear in both the LHS & RHS.
1707 RHSKnownZero &= LHSKnownZero;
1708 // Output known-1 are known to be set if set in either the LHS | RHS.
1709 RHSKnownOne |= LHSKnownOne;
1711 case Instruction::Xor: {
1712 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1713 RHSKnownZero, RHSKnownOne, Depth+1))
1715 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1716 "Bits known to be one AND zero?");
1717 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1718 LHSKnownZero, LHSKnownOne, Depth+1))
1720 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1721 "Bits known to be one AND zero?");
1723 // If all of the demanded bits are known zero on one side, return the other.
1724 // These bits cannot contribute to the result of the 'xor'.
1725 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1726 return UpdateValueUsesWith(I, I->getOperand(0));
1727 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1728 return UpdateValueUsesWith(I, I->getOperand(1));
1730 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1731 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1732 (RHSKnownOne & LHSKnownOne);
1733 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1734 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1735 (RHSKnownOne & LHSKnownZero);
1737 // If all of the demanded bits are known to be zero on one side or the
1738 // other, turn this into an *inclusive* or.
1739 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1740 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1742 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1744 InsertNewInstBefore(Or, *I);
1745 return UpdateValueUsesWith(I, Or);
1748 // If all of the demanded bits on one side are known, and all of the set
1749 // bits on that side are also known to be set on the other side, turn this
1750 // into an AND, as we know the bits will be cleared.
1751 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1752 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1754 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1755 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1757 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1758 InsertNewInstBefore(And, *I);
1759 return UpdateValueUsesWith(I, And);
1763 // If the RHS is a constant, see if we can simplify it.
1764 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1765 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1766 return UpdateValueUsesWith(I, I);
1768 RHSKnownZero = KnownZeroOut;
1769 RHSKnownOne = KnownOneOut;
1772 case Instruction::Select:
1773 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1774 RHSKnownZero, RHSKnownOne, Depth+1))
1776 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1777 LHSKnownZero, LHSKnownOne, Depth+1))
1779 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1780 "Bits known to be one AND zero?");
1781 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1782 "Bits known to be one AND zero?");
1784 // If the operands are constants, see if we can simplify them.
1785 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1786 return UpdateValueUsesWith(I, I);
1787 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1788 return UpdateValueUsesWith(I, I);
1790 // Only known if known in both the LHS and RHS.
1791 RHSKnownOne &= LHSKnownOne;
1792 RHSKnownZero &= LHSKnownZero;
1794 case Instruction::Trunc: {
1796 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1797 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask.zext(truncBf),
1798 RHSKnownZero.zext(truncBf), RHSKnownOne.zext(truncBf), Depth+1))
1800 DemandedMask.trunc(BitWidth);
1801 RHSKnownZero.trunc(BitWidth);
1802 RHSKnownOne.trunc(BitWidth);
1803 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1804 "Bits known to be one AND zero?");
1807 case Instruction::BitCast:
1808 if (!I->getOperand(0)->getType()->isInteger())
1811 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1812 RHSKnownZero, RHSKnownOne, Depth+1))
1814 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1815 "Bits known to be one AND zero?");
1817 case Instruction::ZExt: {
1818 // Compute the bits in the result that are not present in the input.
1819 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1820 APInt NewBits(APInt::getAllOnesValue(BitWidth).shl(SrcTy->getBitWidth()));
1822 DemandedMask &= SrcTy->getMask().zext(BitWidth);
1823 uint32_t zextBf = SrcTy->getBitWidth();
1824 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask.trunc(zextBf),
1825 RHSKnownZero.trunc(zextBf), RHSKnownOne.trunc(zextBf), Depth+1))
1827 DemandedMask.zext(BitWidth);
1828 RHSKnownZero.zext(BitWidth);
1829 RHSKnownOne.zext(BitWidth);
1830 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1831 "Bits known to be one AND zero?");
1832 // The top bits are known to be zero.
1833 RHSKnownZero |= NewBits;
1836 case Instruction::SExt: {
1837 // Compute the bits in the result that are not present in the input.
1838 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1839 APInt NewBits(APInt::getAllOnesValue(BitWidth).shl(SrcTy->getBitWidth()));
1841 // Get the sign bit for the source type
1842 APInt InSignBit(APInt::getSignBit(SrcTy->getPrimitiveSizeInBits()));
1843 InSignBit.zext(BitWidth);
1844 APInt InputDemandedBits = DemandedMask &
1845 SrcTy->getMask().zext(BitWidth);
1847 // If any of the sign extended bits are demanded, we know that the sign
1849 if ((NewBits & DemandedMask) != 0)
1850 InputDemandedBits |= InSignBit;
1852 uint32_t sextBf = SrcTy->getBitWidth();
1853 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits.trunc(sextBf),
1854 RHSKnownZero.trunc(sextBf), RHSKnownOne.trunc(sextBf), Depth+1))
1856 InputDemandedBits.zext(BitWidth);
1857 RHSKnownZero.zext(BitWidth);
1858 RHSKnownOne.zext(BitWidth);
1859 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1860 "Bits known to be one AND zero?");
1862 // If the sign bit of the input is known set or clear, then we know the
1863 // top bits of the result.
1865 // If the input sign bit is known zero, or if the NewBits are not demanded
1866 // convert this into a zero extension.
1867 if ((RHSKnownZero & InSignBit) != 0 || (NewBits & ~DemandedMask) == NewBits)
1869 // Convert to ZExt cast
1870 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1871 return UpdateValueUsesWith(I, NewCast);
1872 } else if ((RHSKnownOne & InSignBit) != 0) { // Input sign bit known set
1873 RHSKnownOne |= NewBits;
1874 RHSKnownZero &= ~NewBits;
1875 } else { // Input sign bit unknown
1876 RHSKnownZero &= ~NewBits;
1877 RHSKnownOne &= ~NewBits;
1881 case Instruction::Add: {
1882 // Figure out what the input bits are. If the top bits of the and result
1883 // are not demanded, then the add doesn't demand them from its input
1885 unsigned NLZ = DemandedMask.countLeadingZeros();
1887 // If there is a constant on the RHS, there are a variety of xformations
1889 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1890 // If null, this should be simplified elsewhere. Some of the xforms here
1891 // won't work if the RHS is zero.
1895 // If the top bit of the output is demanded, demand everything from the
1896 // input. Otherwise, we demand all the input bits except NLZ top bits.
1897 APInt InDemandedBits(APInt::getAllOnesValue(BitWidth).lshr(NLZ));
1899 // Find information about known zero/one bits in the input.
1900 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1901 LHSKnownZero, LHSKnownOne, Depth+1))
1904 // If the RHS of the add has bits set that can't affect the input, reduce
1906 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1907 return UpdateValueUsesWith(I, I);
1909 // Avoid excess work.
1910 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1913 // Turn it into OR if input bits are zero.
1914 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1916 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1918 InsertNewInstBefore(Or, *I);
1919 return UpdateValueUsesWith(I, Or);
1922 // We can say something about the output known-zero and known-one bits,
1923 // depending on potential carries from the input constant and the
1924 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1925 // bits set and the RHS constant is 0x01001, then we know we have a known
1926 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1928 // To compute this, we first compute the potential carry bits. These are
1929 // the bits which may be modified. I'm not aware of a better way to do
1931 APInt RHSVal(RHS->getValue());
1933 bool CarryIn = false;
1934 APInt CarryBits(BitWidth, 0);
1935 const uint64_t *LHSKnownZeroRawVal = LHSKnownZero.getRawData(),
1936 *RHSRawVal = RHSVal.getRawData();
1937 for (uint32_t i = 0; i != RHSVal.getNumWords(); ++i) {
1938 uint64_t AddVal = ~LHSKnownZeroRawVal[i] + RHSRawVal[i],
1939 XorVal = ~LHSKnownZeroRawVal[i] ^ RHSRawVal[i];
1940 uint64_t WordCarryBits = AddVal ^ XorVal + CarryIn;
1941 if (AddVal < RHSRawVal[i])
1945 CarryBits.setWordToValue(i, WordCarryBits);
1948 // Now that we know which bits have carries, compute the known-1/0 sets.
1950 // Bits are known one if they are known zero in one operand and one in the
1951 // other, and there is no input carry.
1952 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1953 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1955 // Bits are known zero if they are known zero in both operands and there
1956 // is no input carry.
1957 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1959 // If the high-bits of this ADD are not demanded, then it does not demand
1960 // the high bits of its LHS or RHS.
1961 if ((DemandedMask & APInt::getSignBit(BitWidth)) == 0) {
1962 // Right fill the mask of bits for this ADD to demand the most
1963 // significant bit and all those below it.
1964 APInt DemandedFromOps = APInt::getAllOnesValue(BitWidth).lshr(NLZ);
1965 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1966 LHSKnownZero, LHSKnownOne, Depth+1))
1968 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1969 LHSKnownZero, LHSKnownOne, Depth+1))
1975 case Instruction::Sub:
1976 // If the high-bits of this SUB are not demanded, then it does not demand
1977 // the high bits of its LHS or RHS.
1978 if ((DemandedMask & APInt::getSignBit(BitWidth)) == 0) {
1979 // Right fill the mask of bits for this SUB to demand the most
1980 // significant bit and all those below it.
1981 unsigned NLZ = DemandedMask.countLeadingZeros();
1982 APInt DemandedFromOps(APInt::getAllOnesValue(BitWidth).lshr(NLZ));
1983 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1984 LHSKnownZero, LHSKnownOne, Depth+1))
1986 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1987 LHSKnownZero, LHSKnownOne, Depth+1))
1991 case Instruction::Shl:
1992 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1993 uint64_t ShiftAmt = SA->getZExtValue();
1994 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask.lshr(ShiftAmt),
1995 RHSKnownZero, RHSKnownOne, Depth+1))
1997 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1998 "Bits known to be one AND zero?");
1999 RHSKnownZero <<= ShiftAmt;
2000 RHSKnownOne <<= ShiftAmt;
2001 // low bits known zero.
2003 RHSKnownZero |= APInt::getAllOnesValue(ShiftAmt).zextOrCopy(BitWidth);
2006 case Instruction::LShr:
2007 // For a logical shift right
2008 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
2009 unsigned ShiftAmt = SA->getZExtValue();
2011 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
2012 // Unsigned shift right.
2013 if (SimplifyDemandedBits(I->getOperand(0),
2014 (DemandedMask.shl(ShiftAmt)) & TypeMask,
2015 RHSKnownZero, RHSKnownOne, Depth+1))
2017 assert((RHSKnownZero & RHSKnownOne) == 0 &&
2018 "Bits known to be one AND zero?");
2019 RHSKnownZero &= TypeMask;
2020 RHSKnownOne &= TypeMask;
2021 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
2022 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
2024 // Compute the new bits that are at the top now.
2025 APInt HighBits(APInt::getAllOnesValue(BitWidth).shl(
2026 BitWidth - ShiftAmt));
2027 RHSKnownZero |= HighBits; // high bits known zero.
2031 case Instruction::AShr:
2032 // If this is an arithmetic shift right and only the low-bit is set, we can
2033 // always convert this into a logical shr, even if the shift amount is
2034 // variable. The low bit of the shift cannot be an input sign bit unless
2035 // the shift amount is >= the size of the datatype, which is undefined.
2036 if (DemandedMask == 1) {
2037 // Perform the logical shift right.
2038 Value *NewVal = BinaryOperator::createLShr(
2039 I->getOperand(0), I->getOperand(1), I->getName());
2040 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
2041 return UpdateValueUsesWith(I, NewVal);
2044 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
2045 unsigned ShiftAmt = SA->getZExtValue();
2047 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
2048 // Signed shift right.
2049 if (SimplifyDemandedBits(I->getOperand(0),
2050 (DemandedMask.shl(ShiftAmt)) & TypeMask,
2051 RHSKnownZero, RHSKnownOne, Depth+1))
2053 assert((RHSKnownZero & RHSKnownOne) == 0 &&
2054 "Bits known to be one AND zero?");
2055 // Compute the new bits that are at the top now.
2056 APInt HighBits(APInt::getAllOnesValue(BitWidth).shl(BitWidth - ShiftAmt));
2057 RHSKnownZero &= TypeMask;
2058 RHSKnownOne &= TypeMask;
2059 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
2060 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
2062 // Handle the sign bits.
2063 APInt SignBit(APInt::getSignBit(BitWidth));
2064 // Adjust to where it is now in the mask.
2065 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
2067 // If the input sign bit is known to be zero, or if none of the top bits
2068 // are demanded, turn this into an unsigned shift right.
2069 if ((RHSKnownZero & SignBit) != 0 ||
2070 (HighBits & ~DemandedMask) == HighBits) {
2071 // Perform the logical shift right.
2072 Value *NewVal = BinaryOperator::createLShr(
2073 I->getOperand(0), SA, I->getName());
2074 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
2075 return UpdateValueUsesWith(I, NewVal);
2076 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
2077 RHSKnownOne |= HighBits;
2083 // If the client is only demanding bits that we know, return the known
2085 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
2086 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
2091 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
2092 /// 64 or fewer elements. DemandedElts contains the set of elements that are
2093 /// actually used by the caller. This method analyzes which elements of the
2094 /// operand are undef and returns that information in UndefElts.
2096 /// If the information about demanded elements can be used to simplify the
2097 /// operation, the operation is simplified, then the resultant value is
2098 /// returned. This returns null if no change was made.
2099 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
2100 uint64_t &UndefElts,
2102 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
2103 assert(VWidth <= 64 && "Vector too wide to analyze!");
2104 uint64_t EltMask = ~0ULL >> (64-VWidth);
2105 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
2106 "Invalid DemandedElts!");
2108 if (isa<UndefValue>(V)) {
2109 // If the entire vector is undefined, just return this info.
2110 UndefElts = EltMask;
2112 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
2113 UndefElts = EltMask;
2114 return UndefValue::get(V->getType());
2118 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
2119 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
2120 Constant *Undef = UndefValue::get(EltTy);
2122 std::vector<Constant*> Elts;
2123 for (unsigned i = 0; i != VWidth; ++i)
2124 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
2125 Elts.push_back(Undef);
2126 UndefElts |= (1ULL << i);
2127 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
2128 Elts.push_back(Undef);
2129 UndefElts |= (1ULL << i);
2130 } else { // Otherwise, defined.
2131 Elts.push_back(CP->getOperand(i));
2134 // If we changed the constant, return it.
2135 Constant *NewCP = ConstantVector::get(Elts);
2136 return NewCP != CP ? NewCP : 0;
2137 } else if (isa<ConstantAggregateZero>(V)) {
2138 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
2140 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
2141 Constant *Zero = Constant::getNullValue(EltTy);
2142 Constant *Undef = UndefValue::get(EltTy);
2143 std::vector<Constant*> Elts;
2144 for (unsigned i = 0; i != VWidth; ++i)
2145 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
2146 UndefElts = DemandedElts ^ EltMask;
2147 return ConstantVector::get(Elts);
2150 if (!V->hasOneUse()) { // Other users may use these bits.
2151 if (Depth != 0) { // Not at the root.
2152 // TODO: Just compute the UndefElts information recursively.
2156 } else if (Depth == 10) { // Limit search depth.
2160 Instruction *I = dyn_cast<Instruction>(V);
2161 if (!I) return false; // Only analyze instructions.
2163 bool MadeChange = false;
2164 uint64_t UndefElts2;
2166 switch (I->getOpcode()) {
2169 case Instruction::InsertElement: {
2170 // If this is a variable index, we don't know which element it overwrites.
2171 // demand exactly the same input as we produce.
2172 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
2174 // Note that we can't propagate undef elt info, because we don't know
2175 // which elt is getting updated.
2176 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
2177 UndefElts2, Depth+1);
2178 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
2182 // If this is inserting an element that isn't demanded, remove this
2184 unsigned IdxNo = Idx->getZExtValue();
2185 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
2186 return AddSoonDeadInstToWorklist(*I, 0);
2188 // Otherwise, the element inserted overwrites whatever was there, so the
2189 // input demanded set is simpler than the output set.
2190 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
2191 DemandedElts & ~(1ULL << IdxNo),
2192 UndefElts, Depth+1);
2193 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
2195 // The inserted element is defined.
2196 UndefElts |= 1ULL << IdxNo;
2200 case Instruction::And:
2201 case Instruction::Or:
2202 case Instruction::Xor:
2203 case Instruction::Add:
2204 case Instruction::Sub:
2205 case Instruction::Mul:
2206 // div/rem demand all inputs, because they don't want divide by zero.
2207 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
2208 UndefElts, Depth+1);
2209 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
2210 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
2211 UndefElts2, Depth+1);
2212 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
2214 // Output elements are undefined if both are undefined. Consider things
2215 // like undef&0. The result is known zero, not undef.
2216 UndefElts &= UndefElts2;
2219 case Instruction::Call: {
2220 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
2222 switch (II->getIntrinsicID()) {
2225 // Binary vector operations that work column-wise. A dest element is a
2226 // function of the corresponding input elements from the two inputs.
2227 case Intrinsic::x86_sse_sub_ss:
2228 case Intrinsic::x86_sse_mul_ss:
2229 case Intrinsic::x86_sse_min_ss:
2230 case Intrinsic::x86_sse_max_ss:
2231 case Intrinsic::x86_sse2_sub_sd:
2232 case Intrinsic::x86_sse2_mul_sd:
2233 case Intrinsic::x86_sse2_min_sd:
2234 case Intrinsic::x86_sse2_max_sd:
2235 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
2236 UndefElts, Depth+1);
2237 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
2238 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
2239 UndefElts2, Depth+1);
2240 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
2242 // If only the low elt is demanded and this is a scalarizable intrinsic,
2243 // scalarize it now.
2244 if (DemandedElts == 1) {
2245 switch (II->getIntrinsicID()) {
2247 case Intrinsic::x86_sse_sub_ss:
2248 case Intrinsic::x86_sse_mul_ss:
2249 case Intrinsic::x86_sse2_sub_sd:
2250 case Intrinsic::x86_sse2_mul_sd:
2251 // TODO: Lower MIN/MAX/ABS/etc
2252 Value *LHS = II->getOperand(1);
2253 Value *RHS = II->getOperand(2);
2254 // Extract the element as scalars.
2255 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
2256 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
2258 switch (II->getIntrinsicID()) {
2259 default: assert(0 && "Case stmts out of sync!");
2260 case Intrinsic::x86_sse_sub_ss:
2261 case Intrinsic::x86_sse2_sub_sd:
2262 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
2263 II->getName()), *II);
2265 case Intrinsic::x86_sse_mul_ss:
2266 case Intrinsic::x86_sse2_mul_sd:
2267 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
2268 II->getName()), *II);
2273 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
2275 InsertNewInstBefore(New, *II);
2276 AddSoonDeadInstToWorklist(*II, 0);
2281 // Output elements are undefined if both are undefined. Consider things
2282 // like undef&0. The result is known zero, not undef.
2283 UndefElts &= UndefElts2;
2289 return MadeChange ? I : 0;
2292 /// @returns true if the specified compare instruction is
2293 /// true when both operands are equal...
2294 /// @brief Determine if the ICmpInst returns true if both operands are equal
2295 static bool isTrueWhenEqual(ICmpInst &ICI) {
2296 ICmpInst::Predicate pred = ICI.getPredicate();
2297 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
2298 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
2299 pred == ICmpInst::ICMP_SLE;
2302 /// AssociativeOpt - Perform an optimization on an associative operator. This
2303 /// function is designed to check a chain of associative operators for a
2304 /// potential to apply a certain optimization. Since the optimization may be
2305 /// applicable if the expression was reassociated, this checks the chain, then
2306 /// reassociates the expression as necessary to expose the optimization
2307 /// opportunity. This makes use of a special Functor, which must define
2308 /// 'shouldApply' and 'apply' methods.
2310 template<typename Functor>
2311 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
2312 unsigned Opcode = Root.getOpcode();
2313 Value *LHS = Root.getOperand(0);
2315 // Quick check, see if the immediate LHS matches...
2316 if (F.shouldApply(LHS))
2317 return F.apply(Root);
2319 // Otherwise, if the LHS is not of the same opcode as the root, return.
2320 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2321 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
2322 // Should we apply this transform to the RHS?
2323 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
2325 // If not to the RHS, check to see if we should apply to the LHS...
2326 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
2327 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
2331 // If the functor wants to apply the optimization to the RHS of LHSI,
2332 // reassociate the expression from ((? op A) op B) to (? op (A op B))
2334 BasicBlock *BB = Root.getParent();
2336 // Now all of the instructions are in the current basic block, go ahead
2337 // and perform the reassociation.
2338 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
2340 // First move the selected RHS to the LHS of the root...
2341 Root.setOperand(0, LHSI->getOperand(1));
2343 // Make what used to be the LHS of the root be the user of the root...
2344 Value *ExtraOperand = TmpLHSI->getOperand(1);
2345 if (&Root == TmpLHSI) {
2346 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
2349 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
2350 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
2351 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
2352 BasicBlock::iterator ARI = &Root; ++ARI;
2353 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
2356 // Now propagate the ExtraOperand down the chain of instructions until we
2358 while (TmpLHSI != LHSI) {
2359 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
2360 // Move the instruction to immediately before the chain we are
2361 // constructing to avoid breaking dominance properties.
2362 NextLHSI->getParent()->getInstList().remove(NextLHSI);
2363 BB->getInstList().insert(ARI, NextLHSI);
2366 Value *NextOp = NextLHSI->getOperand(1);
2367 NextLHSI->setOperand(1, ExtraOperand);
2369 ExtraOperand = NextOp;
2372 // Now that the instructions are reassociated, have the functor perform
2373 // the transformation...
2374 return F.apply(Root);
2377 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
2383 // AddRHS - Implements: X + X --> X << 1
2386 AddRHS(Value *rhs) : RHS(rhs) {}
2387 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2388 Instruction *apply(BinaryOperator &Add) const {
2389 return BinaryOperator::createShl(Add.getOperand(0),
2390 ConstantInt::get(Add.getType(), 1));
2394 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
2396 struct AddMaskingAnd {
2398 AddMaskingAnd(Constant *c) : C2(c) {}
2399 bool shouldApply(Value *LHS) const {
2401 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
2402 ConstantExpr::getAnd(C1, C2)->isNullValue();
2404 Instruction *apply(BinaryOperator &Add) const {
2405 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
2409 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
2411 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
2412 if (Constant *SOC = dyn_cast<Constant>(SO))
2413 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
2415 return IC->InsertNewInstBefore(CastInst::create(
2416 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
2419 // Figure out if the constant is the left or the right argument.
2420 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
2421 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
2423 if (Constant *SOC = dyn_cast<Constant>(SO)) {
2425 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
2426 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
2429 Value *Op0 = SO, *Op1 = ConstOperand;
2431 std::swap(Op0, Op1);
2433 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2434 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
2435 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2436 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
2437 SO->getName()+".cmp");
2439 assert(0 && "Unknown binary instruction type!");
2442 return IC->InsertNewInstBefore(New, I);
2445 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
2446 // constant as the other operand, try to fold the binary operator into the
2447 // select arguments. This also works for Cast instructions, which obviously do
2448 // not have a second operand.
2449 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
2451 // Don't modify shared select instructions
2452 if (!SI->hasOneUse()) return 0;
2453 Value *TV = SI->getOperand(1);
2454 Value *FV = SI->getOperand(2);
2456 if (isa<Constant>(TV) || isa<Constant>(FV)) {
2457 // Bool selects with constant operands can be folded to logical ops.
2458 if (SI->getType() == Type::Int1Ty) return 0;
2460 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
2461 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
2463 return new SelectInst(SI->getCondition(), SelectTrueVal,
2470 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
2471 /// node as operand #0, see if we can fold the instruction into the PHI (which
2472 /// is only possible if all operands to the PHI are constants).
2473 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
2474 PHINode *PN = cast<PHINode>(I.getOperand(0));
2475 unsigned NumPHIValues = PN->getNumIncomingValues();
2476 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
2478 // Check to see if all of the operands of the PHI are constants. If there is
2479 // one non-constant value, remember the BB it is. If there is more than one
2480 // or if *it* is a PHI, bail out.
2481 BasicBlock *NonConstBB = 0;
2482 for (unsigned i = 0; i != NumPHIValues; ++i)
2483 if (!isa<Constant>(PN->getIncomingValue(i))) {
2484 if (NonConstBB) return 0; // More than one non-const value.
2485 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2486 NonConstBB = PN->getIncomingBlock(i);
2488 // If the incoming non-constant value is in I's block, we have an infinite
2490 if (NonConstBB == I.getParent())
2494 // If there is exactly one non-constant value, we can insert a copy of the
2495 // operation in that block. However, if this is a critical edge, we would be
2496 // inserting the computation one some other paths (e.g. inside a loop). Only
2497 // do this if the pred block is unconditionally branching into the phi block.
2499 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2500 if (!BI || !BI->isUnconditional()) return 0;
2503 // Okay, we can do the transformation: create the new PHI node.
2504 PHINode *NewPN = new PHINode(I.getType(), "");
2505 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2506 InsertNewInstBefore(NewPN, *PN);
2507 NewPN->takeName(PN);
2509 // Next, add all of the operands to the PHI.
2510 if (I.getNumOperands() == 2) {
2511 Constant *C = cast<Constant>(I.getOperand(1));
2512 for (unsigned i = 0; i != NumPHIValues; ++i) {
2514 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2515 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2516 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2518 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2520 assert(PN->getIncomingBlock(i) == NonConstBB);
2521 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2522 InV = BinaryOperator::create(BO->getOpcode(),
2523 PN->getIncomingValue(i), C, "phitmp",
2524 NonConstBB->getTerminator());
2525 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2526 InV = CmpInst::create(CI->getOpcode(),
2528 PN->getIncomingValue(i), C, "phitmp",
2529 NonConstBB->getTerminator());
2531 assert(0 && "Unknown binop!");
2533 AddToWorkList(cast<Instruction>(InV));
2535 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2538 CastInst *CI = cast<CastInst>(&I);
2539 const Type *RetTy = CI->getType();
2540 for (unsigned i = 0; i != NumPHIValues; ++i) {
2542 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2543 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2545 assert(PN->getIncomingBlock(i) == NonConstBB);
2546 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
2547 I.getType(), "phitmp",
2548 NonConstBB->getTerminator());
2549 AddToWorkList(cast<Instruction>(InV));
2551 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2554 return ReplaceInstUsesWith(I, NewPN);
2557 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2558 bool Changed = SimplifyCommutative(I);
2559 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2561 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2562 // X + undef -> undef
2563 if (isa<UndefValue>(RHS))
2564 return ReplaceInstUsesWith(I, RHS);
2567 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2568 if (RHSC->isNullValue())
2569 return ReplaceInstUsesWith(I, LHS);
2570 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2571 if (CFP->isExactlyValue(-0.0))
2572 return ReplaceInstUsesWith(I, LHS);
2575 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2576 // X + (signbit) --> X ^ signbit
2577 uint64_t Val = CI->getZExtValue();
2578 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
2579 return BinaryOperator::createXor(LHS, RHS);
2581 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2582 // (X & 254)+1 -> (X&254)|1
2583 uint64_t KnownZero, KnownOne;
2584 if (!isa<VectorType>(I.getType()) &&
2585 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
2586 KnownZero, KnownOne))
2590 if (isa<PHINode>(LHS))
2591 if (Instruction *NV = FoldOpIntoPhi(I))
2594 ConstantInt *XorRHS = 0;
2596 if (isa<ConstantInt>(RHSC) &&
2597 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2598 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
2599 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
2600 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
2602 uint64_t C0080Val = 1ULL << 31;
2603 int64_t CFF80Val = -C0080Val;
2606 if (TySizeBits > Size) {
2608 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2609 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2610 if (RHSSExt == CFF80Val) {
2611 if (XorRHS->getZExtValue() == C0080Val)
2613 } else if (RHSZExt == C0080Val) {
2614 if (XorRHS->getSExtValue() == CFF80Val)
2618 // This is a sign extend if the top bits are known zero.
2619 uint64_t Mask = ~0ULL;
2620 Mask <<= 64-(TySizeBits-Size);
2621 Mask &= cast<IntegerType>(XorLHS->getType())->getBitMask();
2622 if (!MaskedValueIsZero(XorLHS, Mask))
2623 Size = 0; // Not a sign ext, but can't be any others either.
2630 } while (Size >= 8);
2633 const Type *MiddleType = 0;
2636 case 32: MiddleType = Type::Int32Ty; break;
2637 case 16: MiddleType = Type::Int16Ty; break;
2638 case 8: MiddleType = Type::Int8Ty; break;
2641 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2642 InsertNewInstBefore(NewTrunc, I);
2643 return new SExtInst(NewTrunc, I.getType());
2649 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2650 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2652 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2653 if (RHSI->getOpcode() == Instruction::Sub)
2654 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2655 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2657 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2658 if (LHSI->getOpcode() == Instruction::Sub)
2659 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2660 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2665 if (Value *V = dyn_castNegVal(LHS))
2666 return BinaryOperator::createSub(RHS, V);
2669 if (!isa<Constant>(RHS))
2670 if (Value *V = dyn_castNegVal(RHS))
2671 return BinaryOperator::createSub(LHS, V);
2675 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2676 if (X == RHS) // X*C + X --> X * (C+1)
2677 return BinaryOperator::createMul(RHS, AddOne(C2));
2679 // X*C1 + X*C2 --> X * (C1+C2)
2681 if (X == dyn_castFoldableMul(RHS, C1))
2682 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
2685 // X + X*C --> X * (C+1)
2686 if (dyn_castFoldableMul(RHS, C2) == LHS)
2687 return BinaryOperator::createMul(LHS, AddOne(C2));
2689 // X + ~X --> -1 since ~X = -X-1
2690 if (dyn_castNotVal(LHS) == RHS ||
2691 dyn_castNotVal(RHS) == LHS)
2692 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
2695 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2696 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2697 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2700 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2702 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
2703 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
2704 return BinaryOperator::createSub(C, X);
2707 // (X & FF00) + xx00 -> (X+xx00) & FF00
2708 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2709 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2710 if (Anded == CRHS) {
2711 // See if all bits from the first bit set in the Add RHS up are included
2712 // in the mask. First, get the rightmost bit.
2713 uint64_t AddRHSV = CRHS->getZExtValue();
2715 // Form a mask of all bits from the lowest bit added through the top.
2716 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
2717 AddRHSHighBits &= C2->getType()->getBitMask();
2719 // See if the and mask includes all of these bits.
2720 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
2722 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2723 // Okay, the xform is safe. Insert the new add pronto.
2724 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2725 LHS->getName()), I);
2726 return BinaryOperator::createAnd(NewAdd, C2);
2731 // Try to fold constant add into select arguments.
2732 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2733 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2737 // add (cast *A to intptrtype) B ->
2738 // cast (GEP (cast *A to sbyte*) B) ->
2741 CastInst *CI = dyn_cast<CastInst>(LHS);
2744 CI = dyn_cast<CastInst>(RHS);
2747 if (CI && CI->getType()->isSized() &&
2748 (CI->getType()->getPrimitiveSizeInBits() ==
2749 TD->getIntPtrType()->getPrimitiveSizeInBits())
2750 && isa<PointerType>(CI->getOperand(0)->getType())) {
2751 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2752 PointerType::get(Type::Int8Ty), I);
2753 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2754 return new PtrToIntInst(I2, CI->getType());
2758 return Changed ? &I : 0;
2761 // isSignBit - Return true if the value represented by the constant only has the
2762 // highest order bit set.
2763 static bool isSignBit(ConstantInt *CI) {
2764 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
2765 return CI->getValue() == APInt::getSignBit(NumBits);
2768 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2769 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2771 if (Op0 == Op1) // sub X, X -> 0
2772 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2774 // If this is a 'B = x-(-A)', change to B = x+A...
2775 if (Value *V = dyn_castNegVal(Op1))
2776 return BinaryOperator::createAdd(Op0, V);
2778 if (isa<UndefValue>(Op0))
2779 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2780 if (isa<UndefValue>(Op1))
2781 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2783 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2784 // Replace (-1 - A) with (~A)...
2785 if (C->isAllOnesValue())
2786 return BinaryOperator::createNot(Op1);
2788 // C - ~X == X + (1+C)
2790 if (match(Op1, m_Not(m_Value(X))))
2791 return BinaryOperator::createAdd(X,
2792 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
2793 // -(X >>u 31) -> (X >>s 31)
2794 // -(X >>s 31) -> (X >>u 31)
2795 if (C->isNullValue()) {
2796 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2797 if (SI->getOpcode() == Instruction::LShr) {
2798 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2799 // Check to see if we are shifting out everything but the sign bit.
2800 if (CU->getZExtValue() ==
2801 SI->getType()->getPrimitiveSizeInBits()-1) {
2802 // Ok, the transformation is safe. Insert AShr.
2803 return BinaryOperator::create(Instruction::AShr,
2804 SI->getOperand(0), CU, SI->getName());
2808 else if (SI->getOpcode() == Instruction::AShr) {
2809 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2810 // Check to see if we are shifting out everything but the sign bit.
2811 if (CU->getZExtValue() ==
2812 SI->getType()->getPrimitiveSizeInBits()-1) {
2813 // Ok, the transformation is safe. Insert LShr.
2814 return BinaryOperator::createLShr(
2815 SI->getOperand(0), CU, SI->getName());
2821 // Try to fold constant sub into select arguments.
2822 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2823 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2826 if (isa<PHINode>(Op0))
2827 if (Instruction *NV = FoldOpIntoPhi(I))
2831 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2832 if (Op1I->getOpcode() == Instruction::Add &&
2833 !Op0->getType()->isFPOrFPVector()) {
2834 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2835 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2836 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2837 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2838 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2839 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2840 // C1-(X+C2) --> (C1-C2)-X
2841 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2842 Op1I->getOperand(0));
2846 if (Op1I->hasOneUse()) {
2847 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2848 // is not used by anyone else...
2850 if (Op1I->getOpcode() == Instruction::Sub &&
2851 !Op1I->getType()->isFPOrFPVector()) {
2852 // Swap the two operands of the subexpr...
2853 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2854 Op1I->setOperand(0, IIOp1);
2855 Op1I->setOperand(1, IIOp0);
2857 // Create the new top level add instruction...
2858 return BinaryOperator::createAdd(Op0, Op1);
2861 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2863 if (Op1I->getOpcode() == Instruction::And &&
2864 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2865 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2868 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2869 return BinaryOperator::createAnd(Op0, NewNot);
2872 // 0 - (X sdiv C) -> (X sdiv -C)
2873 if (Op1I->getOpcode() == Instruction::SDiv)
2874 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2875 if (CSI->isNullValue())
2876 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2877 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2878 ConstantExpr::getNeg(DivRHS));
2880 // X - X*C --> X * (1-C)
2881 ConstantInt *C2 = 0;
2882 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2884 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2885 return BinaryOperator::createMul(Op0, CP1);
2890 if (!Op0->getType()->isFPOrFPVector())
2891 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2892 if (Op0I->getOpcode() == Instruction::Add) {
2893 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2894 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2895 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2896 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2897 } else if (Op0I->getOpcode() == Instruction::Sub) {
2898 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2899 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2903 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2904 if (X == Op1) { // X*C - X --> X * (C-1)
2905 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2906 return BinaryOperator::createMul(Op1, CP1);
2909 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2910 if (X == dyn_castFoldableMul(Op1, C2))
2911 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2916 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2917 /// really just returns true if the most significant (sign) bit is set.
2918 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2920 case ICmpInst::ICMP_SLT:
2921 // True if LHS s< RHS and RHS == 0
2922 return RHS->isNullValue();
2923 case ICmpInst::ICMP_SLE:
2924 // True if LHS s<= RHS and RHS == -1
2925 return RHS->isAllOnesValue();
2926 case ICmpInst::ICMP_UGE:
2927 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2928 return RHS->getZExtValue() == (1ULL <<
2929 (RHS->getType()->getPrimitiveSizeInBits()-1));
2930 case ICmpInst::ICMP_UGT:
2931 // True if LHS u> RHS and RHS == high-bit-mask - 1
2932 return RHS->getZExtValue() ==
2933 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2939 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2940 bool Changed = SimplifyCommutative(I);
2941 Value *Op0 = I.getOperand(0);
2943 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2944 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2946 // Simplify mul instructions with a constant RHS...
2947 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2948 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2950 // ((X << C1)*C2) == (X * (C2 << C1))
2951 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2952 if (SI->getOpcode() == Instruction::Shl)
2953 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2954 return BinaryOperator::createMul(SI->getOperand(0),
2955 ConstantExpr::getShl(CI, ShOp));
2957 if (CI->isNullValue())
2958 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2959 if (CI->equalsInt(1)) // X * 1 == X
2960 return ReplaceInstUsesWith(I, Op0);
2961 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2962 return BinaryOperator::createNeg(Op0, I.getName());
2964 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2965 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2966 uint64_t C = Log2_64(Val);
2967 return BinaryOperator::createShl(Op0,
2968 ConstantInt::get(Op0->getType(), C));
2970 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2971 if (Op1F->isNullValue())
2972 return ReplaceInstUsesWith(I, Op1);
2974 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2975 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2976 if (Op1F->getValue() == 1.0)
2977 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2980 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2981 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2982 isa<ConstantInt>(Op0I->getOperand(1))) {
2983 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2984 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2986 InsertNewInstBefore(Add, I);
2987 Value *C1C2 = ConstantExpr::getMul(Op1,
2988 cast<Constant>(Op0I->getOperand(1)));
2989 return BinaryOperator::createAdd(Add, C1C2);
2993 // Try to fold constant mul into select arguments.
2994 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2995 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2998 if (isa<PHINode>(Op0))
2999 if (Instruction *NV = FoldOpIntoPhi(I))
3003 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
3004 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
3005 return BinaryOperator::createMul(Op0v, Op1v);
3007 // If one of the operands of the multiply is a cast from a boolean value, then
3008 // we know the bool is either zero or one, so this is a 'masking' multiply.
3009 // See if we can simplify things based on how the boolean was originally
3011 CastInst *BoolCast = 0;
3012 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
3013 if (CI->getOperand(0)->getType() == Type::Int1Ty)
3016 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
3017 if (CI->getOperand(0)->getType() == Type::Int1Ty)
3020 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
3021 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
3022 const Type *SCOpTy = SCIOp0->getType();
3024 // If the icmp is true iff the sign bit of X is set, then convert this
3025 // multiply into a shift/and combination.
3026 if (isa<ConstantInt>(SCIOp1) &&
3027 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
3028 // Shift the X value right to turn it into "all signbits".
3029 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
3030 SCOpTy->getPrimitiveSizeInBits()-1);
3032 InsertNewInstBefore(
3033 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
3034 BoolCast->getOperand(0)->getName()+
3037 // If the multiply type is not the same as the source type, sign extend
3038 // or truncate to the multiply type.
3039 if (I.getType() != V->getType()) {
3040 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
3041 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
3042 Instruction::CastOps opcode =
3043 (SrcBits == DstBits ? Instruction::BitCast :
3044 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
3045 V = InsertCastBefore(opcode, V, I.getType(), I);
3048 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
3049 return BinaryOperator::createAnd(V, OtherOp);
3054 return Changed ? &I : 0;
3057 /// This function implements the transforms on div instructions that work
3058 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
3059 /// used by the visitors to those instructions.
3060 /// @brief Transforms common to all three div instructions
3061 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
3062 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3065 if (isa<UndefValue>(Op0))
3066 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3068 // X / undef -> undef
3069 if (isa<UndefValue>(Op1))
3070 return ReplaceInstUsesWith(I, Op1);
3072 // Handle cases involving: div X, (select Cond, Y, Z)
3073 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3074 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
3075 // same basic block, then we replace the select with Y, and the condition
3076 // of the select with false (if the cond value is in the same BB). If the
3077 // select has uses other than the div, this allows them to be simplified
3078 // also. Note that div X, Y is just as good as div X, 0 (undef)
3079 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3080 if (ST->isNullValue()) {
3081 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3082 if (CondI && CondI->getParent() == I.getParent())
3083 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3084 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3085 I.setOperand(1, SI->getOperand(2));
3087 UpdateValueUsesWith(SI, SI->getOperand(2));
3091 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
3092 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3093 if (ST->isNullValue()) {
3094 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3095 if (CondI && CondI->getParent() == I.getParent())
3096 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3097 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3098 I.setOperand(1, SI->getOperand(1));
3100 UpdateValueUsesWith(SI, SI->getOperand(1));
3108 /// This function implements the transforms common to both integer division
3109 /// instructions (udiv and sdiv). It is called by the visitors to those integer
3110 /// division instructions.
3111 /// @brief Common integer divide transforms
3112 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
3113 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3115 if (Instruction *Common = commonDivTransforms(I))
3118 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3120 if (RHS->equalsInt(1))
3121 return ReplaceInstUsesWith(I, Op0);
3123 // (X / C1) / C2 -> X / (C1*C2)
3124 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3125 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3126 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3127 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
3128 ConstantExpr::getMul(RHS, LHSRHS));
3131 if (!RHS->isNullValue()) { // avoid X udiv 0
3132 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3133 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3135 if (isa<PHINode>(Op0))
3136 if (Instruction *NV = FoldOpIntoPhi(I))
3141 // 0 / X == 0, we don't need to preserve faults!
3142 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3143 if (LHS->equalsInt(0))
3144 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3149 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3150 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3152 // Handle the integer div common cases
3153 if (Instruction *Common = commonIDivTransforms(I))
3156 // X udiv C^2 -> X >> C
3157 // Check to see if this is an unsigned division with an exact power of 2,
3158 // if so, convert to a right shift.
3159 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3160 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
3161 if (isPowerOf2_64(Val)) {
3162 uint64_t ShiftAmt = Log2_64(Val);
3163 return BinaryOperator::createLShr(Op0,
3164 ConstantInt::get(Op0->getType(), ShiftAmt));
3168 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3169 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3170 if (RHSI->getOpcode() == Instruction::Shl &&
3171 isa<ConstantInt>(RHSI->getOperand(0))) {
3172 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
3173 if (isPowerOf2_64(C1)) {
3174 Value *N = RHSI->getOperand(1);
3175 const Type *NTy = N->getType();
3176 if (uint64_t C2 = Log2_64(C1)) {
3177 Constant *C2V = ConstantInt::get(NTy, C2);
3178 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
3180 return BinaryOperator::createLShr(Op0, N);
3185 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3186 // where C1&C2 are powers of two.
3187 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3188 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3189 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3190 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
3191 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
3192 // Compute the shift amounts
3193 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
3194 // Construct the "on true" case of the select
3195 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3196 Instruction *TSI = BinaryOperator::createLShr(
3197 Op0, TC, SI->getName()+".t");
3198 TSI = InsertNewInstBefore(TSI, I);
3200 // Construct the "on false" case of the select
3201 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3202 Instruction *FSI = BinaryOperator::createLShr(
3203 Op0, FC, SI->getName()+".f");
3204 FSI = InsertNewInstBefore(FSI, I);
3206 // construct the select instruction and return it.
3207 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
3213 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3214 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3216 // Handle the integer div common cases
3217 if (Instruction *Common = commonIDivTransforms(I))
3220 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3222 if (RHS->isAllOnesValue())
3223 return BinaryOperator::createNeg(Op0);
3226 if (Value *LHSNeg = dyn_castNegVal(Op0))
3227 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
3230 // If the sign bits of both operands are zero (i.e. we can prove they are
3231 // unsigned inputs), turn this into a udiv.
3232 if (I.getType()->isInteger()) {
3233 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
3234 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3235 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
3242 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3243 return commonDivTransforms(I);
3246 /// GetFactor - If we can prove that the specified value is at least a multiple
3247 /// of some factor, return that factor.
3248 static Constant *GetFactor(Value *V) {
3249 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3252 // Unless we can be tricky, we know this is a multiple of 1.
3253 Constant *Result = ConstantInt::get(V->getType(), 1);
3255 Instruction *I = dyn_cast<Instruction>(V);
3256 if (!I) return Result;
3258 if (I->getOpcode() == Instruction::Mul) {
3259 // Handle multiplies by a constant, etc.
3260 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
3261 GetFactor(I->getOperand(1)));
3262 } else if (I->getOpcode() == Instruction::Shl) {
3263 // (X<<C) -> X * (1 << C)
3264 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
3265 ShRHS = ConstantExpr::getShl(Result, ShRHS);
3266 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
3268 } else if (I->getOpcode() == Instruction::And) {
3269 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
3270 // X & 0xFFF0 is known to be a multiple of 16.
3271 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
3272 if (Zeros != V->getType()->getPrimitiveSizeInBits())
3273 return ConstantExpr::getShl(Result,
3274 ConstantInt::get(Result->getType(), Zeros));
3276 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
3277 // Only handle int->int casts.
3278 if (!CI->isIntegerCast())
3280 Value *Op = CI->getOperand(0);
3281 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
3286 /// This function implements the transforms on rem instructions that work
3287 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3288 /// is used by the visitors to those instructions.
3289 /// @brief Transforms common to all three rem instructions
3290 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3291 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3293 // 0 % X == 0, we don't need to preserve faults!
3294 if (Constant *LHS = dyn_cast<Constant>(Op0))
3295 if (LHS->isNullValue())
3296 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3298 if (isa<UndefValue>(Op0)) // undef % X -> 0
3299 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3300 if (isa<UndefValue>(Op1))
3301 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3303 // Handle cases involving: rem X, (select Cond, Y, Z)
3304 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3305 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
3306 // the same basic block, then we replace the select with Y, and the
3307 // condition of the select with false (if the cond value is in the same
3308 // BB). If the select has uses other than the div, this allows them to be
3310 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3311 if (ST->isNullValue()) {
3312 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3313 if (CondI && CondI->getParent() == I.getParent())
3314 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3315 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3316 I.setOperand(1, SI->getOperand(2));
3318 UpdateValueUsesWith(SI, SI->getOperand(2));
3321 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
3322 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3323 if (ST->isNullValue()) {
3324 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3325 if (CondI && CondI->getParent() == I.getParent())
3326 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3327 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3328 I.setOperand(1, SI->getOperand(1));
3330 UpdateValueUsesWith(SI, SI->getOperand(1));
3338 /// This function implements the transforms common to both integer remainder
3339 /// instructions (urem and srem). It is called by the visitors to those integer
3340 /// remainder instructions.
3341 /// @brief Common integer remainder transforms
3342 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3343 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3345 if (Instruction *common = commonRemTransforms(I))
3348 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3349 // X % 0 == undef, we don't need to preserve faults!
3350 if (RHS->equalsInt(0))
3351 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3353 if (RHS->equalsInt(1)) // X % 1 == 0
3354 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3356 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3357 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3358 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3360 } else if (isa<PHINode>(Op0I)) {
3361 if (Instruction *NV = FoldOpIntoPhi(I))
3364 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
3365 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
3366 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3373 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3374 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3376 if (Instruction *common = commonIRemTransforms(I))
3379 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3380 // X urem C^2 -> X and C
3381 // Check to see if this is an unsigned remainder with an exact power of 2,
3382 // if so, convert to a bitwise and.
3383 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3384 if (isPowerOf2_64(C->getZExtValue()))
3385 return BinaryOperator::createAnd(Op0, SubOne(C));
3388 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3389 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3390 if (RHSI->getOpcode() == Instruction::Shl &&
3391 isa<ConstantInt>(RHSI->getOperand(0))) {
3392 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
3393 if (isPowerOf2_64(C1)) {
3394 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3395 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
3397 return BinaryOperator::createAnd(Op0, Add);
3402 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3403 // where C1&C2 are powers of two.
3404 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3405 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3406 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3407 // STO == 0 and SFO == 0 handled above.
3408 if (isPowerOf2_64(STO->getZExtValue()) &&
3409 isPowerOf2_64(SFO->getZExtValue())) {
3410 Value *TrueAnd = InsertNewInstBefore(
3411 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3412 Value *FalseAnd = InsertNewInstBefore(
3413 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3414 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
3422 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3423 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3425 if (Instruction *common = commonIRemTransforms(I))
3428 if (Value *RHSNeg = dyn_castNegVal(Op1))
3429 if (!isa<ConstantInt>(RHSNeg) ||
3430 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
3432 AddUsesToWorkList(I);
3433 I.setOperand(1, RHSNeg);
3437 // If the top bits of both operands are zero (i.e. we can prove they are
3438 // unsigned inputs), turn this into a urem.
3439 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
3440 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3441 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3442 return BinaryOperator::createURem(Op0, Op1, I.getName());
3448 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3449 return commonRemTransforms(I);
3452 // isMaxValueMinusOne - return true if this is Max-1
3453 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
3454 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3456 // Calculate 0111111111..11111
3457 APInt Val(APInt::getSignedMaxValue(TypeBits));
3458 return C->getValue() == Val-1;
3460 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
3463 // isMinValuePlusOne - return true if this is Min+1
3464 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
3466 // Calculate 1111111111000000000000
3467 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3468 APInt Val(APInt::getSignedMinValue(TypeBits));
3469 return C->getValue() == Val+1;
3471 return C->getValue() == 1; // unsigned
3474 // isOneBitSet - Return true if there is exactly one bit set in the specified
3476 static bool isOneBitSet(const ConstantInt *CI) {
3477 return CI->getValue().isPowerOf2();
3480 #if 0 // Currently unused
3481 // isLowOnes - Return true if the constant is of the form 0+1+.
3482 static bool isLowOnes(const ConstantInt *CI) {
3483 uint64_t V = CI->getZExtValue();
3485 // There won't be bits set in parts that the type doesn't contain.
3486 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
3488 uint64_t U = V+1; // If it is low ones, this should be a power of two.
3489 return U && V && (U & V) == 0;
3493 // isHighOnes - Return true if the constant is of the form 1+0+.
3494 // This is the same as lowones(~X).
3495 static bool isHighOnes(const ConstantInt *CI) {
3496 return (~CI->getValue() + 1).isPowerOf2();
3499 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3500 /// are carefully arranged to allow folding of expressions such as:
3502 /// (A < B) | (A > B) --> (A != B)
3504 /// Note that this is only valid if the first and second predicates have the
3505 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3507 /// Three bits are used to represent the condition, as follows:
3512 /// <=> Value Definition
3513 /// 000 0 Always false
3520 /// 111 7 Always true
3522 static unsigned getICmpCode(const ICmpInst *ICI) {
3523 switch (ICI->getPredicate()) {
3525 case ICmpInst::ICMP_UGT: return 1; // 001
3526 case ICmpInst::ICMP_SGT: return 1; // 001
3527 case ICmpInst::ICMP_EQ: return 2; // 010
3528 case ICmpInst::ICMP_UGE: return 3; // 011
3529 case ICmpInst::ICMP_SGE: return 3; // 011
3530 case ICmpInst::ICMP_ULT: return 4; // 100
3531 case ICmpInst::ICMP_SLT: return 4; // 100
3532 case ICmpInst::ICMP_NE: return 5; // 101
3533 case ICmpInst::ICMP_ULE: return 6; // 110
3534 case ICmpInst::ICMP_SLE: return 6; // 110
3537 assert(0 && "Invalid ICmp predicate!");
3542 /// getICmpValue - This is the complement of getICmpCode, which turns an
3543 /// opcode and two operands into either a constant true or false, or a brand
3544 /// new /// ICmp instruction. The sign is passed in to determine which kind
3545 /// of predicate to use in new icmp instructions.
3546 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3548 default: assert(0 && "Illegal ICmp code!");
3549 case 0: return ConstantInt::getFalse();
3552 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3554 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3555 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3558 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3560 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3563 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3565 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3566 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3569 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3571 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3572 case 7: return ConstantInt::getTrue();
3576 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3577 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3578 (ICmpInst::isSignedPredicate(p1) &&
3579 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3580 (ICmpInst::isSignedPredicate(p2) &&
3581 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3585 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3586 struct FoldICmpLogical {
3589 ICmpInst::Predicate pred;
3590 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3591 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3592 pred(ICI->getPredicate()) {}
3593 bool shouldApply(Value *V) const {
3594 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3595 if (PredicatesFoldable(pred, ICI->getPredicate()))
3596 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
3597 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
3600 Instruction *apply(Instruction &Log) const {
3601 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3602 if (ICI->getOperand(0) != LHS) {
3603 assert(ICI->getOperand(1) == LHS);
3604 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3607 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3608 unsigned LHSCode = getICmpCode(ICI);
3609 unsigned RHSCode = getICmpCode(RHSICI);
3611 switch (Log.getOpcode()) {
3612 case Instruction::And: Code = LHSCode & RHSCode; break;
3613 case Instruction::Or: Code = LHSCode | RHSCode; break;
3614 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3615 default: assert(0 && "Illegal logical opcode!"); return 0;
3618 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3619 ICmpInst::isSignedPredicate(ICI->getPredicate());
3621 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3622 if (Instruction *I = dyn_cast<Instruction>(RV))
3624 // Otherwise, it's a constant boolean value...
3625 return IC.ReplaceInstUsesWith(Log, RV);
3628 } // end anonymous namespace
3630 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3631 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3632 // guaranteed to be a binary operator.
3633 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3635 ConstantInt *AndRHS,
3636 BinaryOperator &TheAnd) {
3637 Value *X = Op->getOperand(0);
3638 Constant *Together = 0;
3640 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3642 switch (Op->getOpcode()) {
3643 case Instruction::Xor:
3644 if (Op->hasOneUse()) {
3645 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3646 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3647 InsertNewInstBefore(And, TheAnd);
3649 return BinaryOperator::createXor(And, Together);
3652 case Instruction::Or:
3653 if (Together == AndRHS) // (X | C) & C --> C
3654 return ReplaceInstUsesWith(TheAnd, AndRHS);
3656 if (Op->hasOneUse() && Together != OpRHS) {
3657 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3658 Instruction *Or = BinaryOperator::createOr(X, Together);
3659 InsertNewInstBefore(Or, TheAnd);
3661 return BinaryOperator::createAnd(Or, AndRHS);
3664 case Instruction::Add:
3665 if (Op->hasOneUse()) {
3666 // Adding a one to a single bit bit-field should be turned into an XOR
3667 // of the bit. First thing to check is to see if this AND is with a
3668 // single bit constant.
3669 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
3671 // Clear bits that are not part of the constant.
3672 AndRHSV &= AndRHS->getType()->getBitMask();
3674 // If there is only one bit set...
3675 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3676 // Ok, at this point, we know that we are masking the result of the
3677 // ADD down to exactly one bit. If the constant we are adding has
3678 // no bits set below this bit, then we can eliminate the ADD.
3679 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
3681 // Check to see if any bits below the one bit set in AndRHSV are set.
3682 if ((AddRHS & (AndRHSV-1)) == 0) {
3683 // If not, the only thing that can effect the output of the AND is
3684 // the bit specified by AndRHSV. If that bit is set, the effect of
3685 // the XOR is to toggle the bit. If it is clear, then the ADD has
3687 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3688 TheAnd.setOperand(0, X);
3691 // Pull the XOR out of the AND.
3692 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3693 InsertNewInstBefore(NewAnd, TheAnd);
3694 NewAnd->takeName(Op);
3695 return BinaryOperator::createXor(NewAnd, AndRHS);
3702 case Instruction::Shl: {
3703 // We know that the AND will not produce any of the bits shifted in, so if
3704 // the anded constant includes them, clear them now!
3706 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3707 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
3708 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
3710 if (CI == ShlMask) { // Masking out bits that the shift already masks
3711 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3712 } else if (CI != AndRHS) { // Reducing bits set in and.
3713 TheAnd.setOperand(1, CI);
3718 case Instruction::LShr:
3720 // We know that the AND will not produce any of the bits shifted in, so if
3721 // the anded constant includes them, clear them now! This only applies to
3722 // unsigned shifts, because a signed shr may bring in set bits!
3724 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3725 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
3726 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
3728 if (CI == ShrMask) { // Masking out bits that the shift already masks.
3729 return ReplaceInstUsesWith(TheAnd, Op);
3730 } else if (CI != AndRHS) {
3731 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3736 case Instruction::AShr:
3738 // See if this is shifting in some sign extension, then masking it out
3740 if (Op->hasOneUse()) {
3741 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3742 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
3743 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
3744 if (C == AndRHS) { // Masking out bits shifted in.
3745 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3746 // Make the argument unsigned.
3747 Value *ShVal = Op->getOperand(0);
3748 ShVal = InsertNewInstBefore(
3749 BinaryOperator::createLShr(ShVal, OpRHS,
3750 Op->getName()), TheAnd);
3751 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3760 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3761 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3762 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3763 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3764 /// insert new instructions.
3765 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3766 bool isSigned, bool Inside,
3768 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3769 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3770 "Lo is not <= Hi in range emission code!");
3773 if (Lo == Hi) // Trivially false.
3774 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3776 // V >= Min && V < Hi --> V < Hi
3777 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3778 ICmpInst::Predicate pred = (isSigned ?
3779 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3780 return new ICmpInst(pred, V, Hi);
3783 // Emit V-Lo <u Hi-Lo
3784 Constant *NegLo = ConstantExpr::getNeg(Lo);
3785 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3786 InsertNewInstBefore(Add, IB);
3787 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3788 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3791 if (Lo == Hi) // Trivially true.
3792 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3794 // V < Min || V >= Hi -> V > Hi-1
3795 Hi = SubOne(cast<ConstantInt>(Hi));
3796 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3797 ICmpInst::Predicate pred = (isSigned ?
3798 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3799 return new ICmpInst(pred, V, Hi);
3802 // Emit V-Lo >u Hi-1-Lo
3803 // Note that Hi has already had one subtracted from it, above.
3804 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3805 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3806 InsertNewInstBefore(Add, IB);
3807 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3808 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3811 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3812 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3813 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3814 // not, since all 1s are not contiguous.
3815 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
3816 uint64_t V = Val->getZExtValue();
3817 if (!isShiftedMask_64(V)) return false;
3819 // look for the first zero bit after the run of ones
3820 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
3821 // look for the first non-zero bit
3822 ME = 64-CountLeadingZeros_64(V);
3828 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3829 /// where isSub determines whether the operator is a sub. If we can fold one of
3830 /// the following xforms:
3832 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3833 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3834 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3836 /// return (A +/- B).
3838 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3839 ConstantInt *Mask, bool isSub,
3841 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3842 if (!LHSI || LHSI->getNumOperands() != 2 ||
3843 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3845 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3847 switch (LHSI->getOpcode()) {
3849 case Instruction::And:
3850 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3851 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3852 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3855 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3856 // part, we don't need any explicit masks to take them out of A. If that
3857 // is all N is, ignore it.
3859 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3860 uint64_t Mask = cast<IntegerType>(RHS->getType())->getBitMask();
3862 if (MaskedValueIsZero(RHS, Mask))
3867 case Instruction::Or:
3868 case Instruction::Xor:
3869 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3870 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3871 ConstantExpr::getAnd(N, Mask)->isNullValue())
3878 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3880 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3881 return InsertNewInstBefore(New, I);
3884 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3885 bool Changed = SimplifyCommutative(I);
3886 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3888 if (isa<UndefValue>(Op1)) // X & undef -> 0
3889 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3893 return ReplaceInstUsesWith(I, Op1);
3895 // See if we can simplify any instructions used by the instruction whose sole
3896 // purpose is to compute bits we don't care about.
3897 uint64_t KnownZero, KnownOne;
3898 if (!isa<VectorType>(I.getType())) {
3899 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3900 KnownZero, KnownOne))
3903 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3904 if (CP->isAllOnesValue())
3905 return ReplaceInstUsesWith(I, I.getOperand(0));
3909 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3910 uint64_t AndRHSMask = AndRHS->getZExtValue();
3911 uint64_t TypeMask = cast<IntegerType>(Op0->getType())->getBitMask();
3912 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3914 // Optimize a variety of ((val OP C1) & C2) combinations...
3915 if (isa<BinaryOperator>(Op0)) {
3916 Instruction *Op0I = cast<Instruction>(Op0);
3917 Value *Op0LHS = Op0I->getOperand(0);
3918 Value *Op0RHS = Op0I->getOperand(1);
3919 switch (Op0I->getOpcode()) {
3920 case Instruction::Xor:
3921 case Instruction::Or:
3922 // If the mask is only needed on one incoming arm, push it up.
3923 if (Op0I->hasOneUse()) {
3924 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3925 // Not masking anything out for the LHS, move to RHS.
3926 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3927 Op0RHS->getName()+".masked");
3928 InsertNewInstBefore(NewRHS, I);
3929 return BinaryOperator::create(
3930 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3932 if (!isa<Constant>(Op0RHS) &&
3933 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3934 // Not masking anything out for the RHS, move to LHS.
3935 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3936 Op0LHS->getName()+".masked");
3937 InsertNewInstBefore(NewLHS, I);
3938 return BinaryOperator::create(
3939 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3944 case Instruction::Add:
3945 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3946 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3947 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3948 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3949 return BinaryOperator::createAnd(V, AndRHS);
3950 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3951 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3954 case Instruction::Sub:
3955 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3956 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3957 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3958 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3959 return BinaryOperator::createAnd(V, AndRHS);
3963 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3964 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3966 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3967 // If this is an integer truncation or change from signed-to-unsigned, and
3968 // if the source is an and/or with immediate, transform it. This
3969 // frequently occurs for bitfield accesses.
3970 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3971 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3972 CastOp->getNumOperands() == 2)
3973 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3974 if (CastOp->getOpcode() == Instruction::And) {
3975 // Change: and (cast (and X, C1) to T), C2
3976 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3977 // This will fold the two constants together, which may allow
3978 // other simplifications.
3979 Instruction *NewCast = CastInst::createTruncOrBitCast(
3980 CastOp->getOperand(0), I.getType(),
3981 CastOp->getName()+".shrunk");
3982 NewCast = InsertNewInstBefore(NewCast, I);
3983 // trunc_or_bitcast(C1)&C2
3984 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3985 C3 = ConstantExpr::getAnd(C3, AndRHS);
3986 return BinaryOperator::createAnd(NewCast, C3);
3987 } else if (CastOp->getOpcode() == Instruction::Or) {
3988 // Change: and (cast (or X, C1) to T), C2
3989 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3990 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3991 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3992 return ReplaceInstUsesWith(I, AndRHS);
3997 // Try to fold constant and into select arguments.
3998 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3999 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4001 if (isa<PHINode>(Op0))
4002 if (Instruction *NV = FoldOpIntoPhi(I))
4006 Value *Op0NotVal = dyn_castNotVal(Op0);
4007 Value *Op1NotVal = dyn_castNotVal(Op1);
4009 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4010 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4012 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4013 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4014 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
4015 I.getName()+".demorgan");
4016 InsertNewInstBefore(Or, I);
4017 return BinaryOperator::createNot(Or);
4021 Value *A = 0, *B = 0;
4022 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
4023 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4024 return ReplaceInstUsesWith(I, Op1);
4025 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
4026 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4027 return ReplaceInstUsesWith(I, Op0);
4029 if (Op0->hasOneUse() &&
4030 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4031 if (A == Op1) { // (A^B)&A -> A&(A^B)
4032 I.swapOperands(); // Simplify below
4033 std::swap(Op0, Op1);
4034 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4035 cast<BinaryOperator>(Op0)->swapOperands();
4036 I.swapOperands(); // Simplify below
4037 std::swap(Op0, Op1);
4040 if (Op1->hasOneUse() &&
4041 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4042 if (B == Op0) { // B&(A^B) -> B&(B^A)
4043 cast<BinaryOperator>(Op1)->swapOperands();
4046 if (A == Op0) { // A&(A^B) -> A & ~B
4047 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
4048 InsertNewInstBefore(NotB, I);
4049 return BinaryOperator::createAnd(A, NotB);
4054 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4055 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4056 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4059 Value *LHSVal, *RHSVal;
4060 ConstantInt *LHSCst, *RHSCst;
4061 ICmpInst::Predicate LHSCC, RHSCC;
4062 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4063 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4064 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
4065 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
4066 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4067 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4068 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4069 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
4070 // Ensure that the larger constant is on the RHS.
4071 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
4072 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
4073 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
4074 ICmpInst *LHS = cast<ICmpInst>(Op0);
4075 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
4076 std::swap(LHS, RHS);
4077 std::swap(LHSCst, RHSCst);
4078 std::swap(LHSCC, RHSCC);
4081 // At this point, we know we have have two icmp instructions
4082 // comparing a value against two constants and and'ing the result
4083 // together. Because of the above check, we know that we only have
4084 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
4085 // (from the FoldICmpLogical check above), that the two constants
4086 // are not equal and that the larger constant is on the RHS
4087 assert(LHSCst != RHSCst && "Compares not folded above?");
4090 default: assert(0 && "Unknown integer condition code!");
4091 case ICmpInst::ICMP_EQ:
4093 default: assert(0 && "Unknown integer condition code!");
4094 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
4095 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
4096 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
4097 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4098 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
4099 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
4100 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
4101 return ReplaceInstUsesWith(I, LHS);
4103 case ICmpInst::ICMP_NE:
4105 default: assert(0 && "Unknown integer condition code!");
4106 case ICmpInst::ICMP_ULT:
4107 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
4108 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
4109 break; // (X != 13 & X u< 15) -> no change
4110 case ICmpInst::ICMP_SLT:
4111 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
4112 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
4113 break; // (X != 13 & X s< 15) -> no change
4114 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
4115 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
4116 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
4117 return ReplaceInstUsesWith(I, RHS);
4118 case ICmpInst::ICMP_NE:
4119 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
4120 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4121 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4122 LHSVal->getName()+".off");
4123 InsertNewInstBefore(Add, I);
4124 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
4125 ConstantInt::get(Add->getType(), 1));
4127 break; // (X != 13 & X != 15) -> no change
4130 case ICmpInst::ICMP_ULT:
4132 default: assert(0 && "Unknown integer condition code!");
4133 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
4134 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
4135 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4136 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
4138 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
4139 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
4140 return ReplaceInstUsesWith(I, LHS);
4141 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
4145 case ICmpInst::ICMP_SLT:
4147 default: assert(0 && "Unknown integer condition code!");
4148 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
4149 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
4150 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4151 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
4153 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
4154 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
4155 return ReplaceInstUsesWith(I, LHS);
4156 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
4160 case ICmpInst::ICMP_UGT:
4162 default: assert(0 && "Unknown integer condition code!");
4163 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
4164 return ReplaceInstUsesWith(I, LHS);
4165 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
4166 return ReplaceInstUsesWith(I, RHS);
4167 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
4169 case ICmpInst::ICMP_NE:
4170 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
4171 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4172 break; // (X u> 13 & X != 15) -> no change
4173 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
4174 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
4176 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
4180 case ICmpInst::ICMP_SGT:
4182 default: assert(0 && "Unknown integer condition code!");
4183 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
4184 return ReplaceInstUsesWith(I, LHS);
4185 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
4186 return ReplaceInstUsesWith(I, RHS);
4187 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
4189 case ICmpInst::ICMP_NE:
4190 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
4191 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4192 break; // (X s> 13 & X != 15) -> no change
4193 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
4194 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
4196 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
4204 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4205 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4206 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4207 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4208 const Type *SrcTy = Op0C->getOperand(0)->getType();
4209 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4210 // Only do this if the casts both really cause code to be generated.
4211 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4213 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4215 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
4216 Op1C->getOperand(0),
4218 InsertNewInstBefore(NewOp, I);
4219 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4223 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4224 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4225 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4226 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4227 SI0->getOperand(1) == SI1->getOperand(1) &&
4228 (SI0->hasOneUse() || SI1->hasOneUse())) {
4229 Instruction *NewOp =
4230 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
4232 SI0->getName()), I);
4233 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4234 SI1->getOperand(1));
4238 return Changed ? &I : 0;
4241 /// CollectBSwapParts - Look to see if the specified value defines a single byte
4242 /// in the result. If it does, and if the specified byte hasn't been filled in
4243 /// yet, fill it in and return false.
4244 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
4245 Instruction *I = dyn_cast<Instruction>(V);
4246 if (I == 0) return true;
4248 // If this is an or instruction, it is an inner node of the bswap.
4249 if (I->getOpcode() == Instruction::Or)
4250 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
4251 CollectBSwapParts(I->getOperand(1), ByteValues);
4253 // If this is a shift by a constant int, and it is "24", then its operand
4254 // defines a byte. We only handle unsigned types here.
4255 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
4256 // Not shifting the entire input by N-1 bytes?
4257 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
4258 8*(ByteValues.size()-1))
4262 if (I->getOpcode() == Instruction::Shl) {
4263 // X << 24 defines the top byte with the lowest of the input bytes.
4264 DestNo = ByteValues.size()-1;
4266 // X >>u 24 defines the low byte with the highest of the input bytes.
4270 // If the destination byte value is already defined, the values are or'd
4271 // together, which isn't a bswap (unless it's an or of the same bits).
4272 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
4274 ByteValues[DestNo] = I->getOperand(0);
4278 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
4280 Value *Shift = 0, *ShiftLHS = 0;
4281 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
4282 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
4283 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
4285 Instruction *SI = cast<Instruction>(Shift);
4287 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
4288 if (ShiftAmt->getZExtValue() & 7 ||
4289 ShiftAmt->getZExtValue() > 8*ByteValues.size())
4292 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
4294 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
4295 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
4297 // Unknown mask for bswap.
4298 if (DestByte == ByteValues.size()) return true;
4300 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
4302 if (SI->getOpcode() == Instruction::Shl)
4303 SrcByte = DestByte - ShiftBytes;
4305 SrcByte = DestByte + ShiftBytes;
4307 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
4308 if (SrcByte != ByteValues.size()-DestByte-1)
4311 // If the destination byte value is already defined, the values are or'd
4312 // together, which isn't a bswap (unless it's an or of the same bits).
4313 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
4315 ByteValues[DestByte] = SI->getOperand(0);
4319 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4320 /// If so, insert the new bswap intrinsic and return it.
4321 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4322 // We cannot bswap one byte.
4323 if (I.getType() == Type::Int8Ty)
4326 /// ByteValues - For each byte of the result, we keep track of which value
4327 /// defines each byte.
4328 SmallVector<Value*, 8> ByteValues;
4329 ByteValues.resize(TD->getTypeSize(I.getType()));
4331 // Try to find all the pieces corresponding to the bswap.
4332 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
4333 CollectBSwapParts(I.getOperand(1), ByteValues))
4336 // Check to see if all of the bytes come from the same value.
4337 Value *V = ByteValues[0];
4338 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4340 // Check to make sure that all of the bytes come from the same value.
4341 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4342 if (ByteValues[i] != V)
4345 // If they do then *success* we can turn this into a bswap. Figure out what
4346 // bswap to make it into.
4347 Module *M = I.getParent()->getParent()->getParent();
4348 const char *FnName = 0;
4349 if (I.getType() == Type::Int16Ty)
4350 FnName = "llvm.bswap.i16";
4351 else if (I.getType() == Type::Int32Ty)
4352 FnName = "llvm.bswap.i32";
4353 else if (I.getType() == Type::Int64Ty)
4354 FnName = "llvm.bswap.i64";
4356 assert(0 && "Unknown integer type!");
4357 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
4358 return new CallInst(F, V);
4362 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4363 bool Changed = SimplifyCommutative(I);
4364 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4366 if (isa<UndefValue>(Op1))
4367 return ReplaceInstUsesWith(I, // X | undef -> -1
4368 ConstantInt::getAllOnesValue(I.getType()));
4372 return ReplaceInstUsesWith(I, Op0);
4374 // See if we can simplify any instructions used by the instruction whose sole
4375 // purpose is to compute bits we don't care about.
4376 uint64_t KnownZero, KnownOne;
4377 if (!isa<VectorType>(I.getType()) &&
4378 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
4379 KnownZero, KnownOne))
4383 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4384 ConstantInt *C1 = 0; Value *X = 0;
4385 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4386 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4387 Instruction *Or = BinaryOperator::createOr(X, RHS);
4388 InsertNewInstBefore(Or, I);
4390 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
4393 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4394 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4395 Instruction *Or = BinaryOperator::createOr(X, RHS);
4396 InsertNewInstBefore(Or, I);
4398 return BinaryOperator::createXor(Or,
4399 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
4402 // Try to fold constant and into select arguments.
4403 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4404 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4406 if (isa<PHINode>(Op0))
4407 if (Instruction *NV = FoldOpIntoPhi(I))
4411 Value *A = 0, *B = 0;
4412 ConstantInt *C1 = 0, *C2 = 0;
4414 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4415 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4416 return ReplaceInstUsesWith(I, Op1);
4417 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4418 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4419 return ReplaceInstUsesWith(I, Op0);
4421 // (A | B) | C and A | (B | C) -> bswap if possible.
4422 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4423 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4424 match(Op1, m_Or(m_Value(), m_Value())) ||
4425 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4426 match(Op1, m_Shift(m_Value(), m_Value())))) {
4427 if (Instruction *BSwap = MatchBSwap(I))
4431 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4432 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4433 MaskedValueIsZero(Op1, C1->getZExtValue())) {
4434 Instruction *NOr = BinaryOperator::createOr(A, Op1);
4435 InsertNewInstBefore(NOr, I);
4437 return BinaryOperator::createXor(NOr, C1);
4440 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4441 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4442 MaskedValueIsZero(Op0, C1->getZExtValue())) {
4443 Instruction *NOr = BinaryOperator::createOr(A, Op0);
4444 InsertNewInstBefore(NOr, I);
4446 return BinaryOperator::createXor(NOr, C1);
4449 // (A & C1)|(B & C2)
4450 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
4451 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
4453 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
4454 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
4457 // If we have: ((V + N) & C1) | (V & C2)
4458 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4459 // replace with V+N.
4460 if (C1 == ConstantExpr::getNot(C2)) {
4461 Value *V1 = 0, *V2 = 0;
4462 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
4463 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4464 // Add commutes, try both ways.
4465 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
4466 return ReplaceInstUsesWith(I, A);
4467 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
4468 return ReplaceInstUsesWith(I, A);
4470 // Or commutes, try both ways.
4471 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
4472 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4473 // Add commutes, try both ways.
4474 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
4475 return ReplaceInstUsesWith(I, B);
4476 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
4477 return ReplaceInstUsesWith(I, B);
4482 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4483 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4484 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4485 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4486 SI0->getOperand(1) == SI1->getOperand(1) &&
4487 (SI0->hasOneUse() || SI1->hasOneUse())) {
4488 Instruction *NewOp =
4489 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4491 SI0->getName()), I);
4492 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4493 SI1->getOperand(1));
4497 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4498 if (A == Op1) // ~A | A == -1
4499 return ReplaceInstUsesWith(I,
4500 ConstantInt::getAllOnesValue(I.getType()));
4504 // Note, A is still live here!
4505 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4507 return ReplaceInstUsesWith(I,
4508 ConstantInt::getAllOnesValue(I.getType()));
4510 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4511 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4512 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4513 I.getName()+".demorgan"), I);
4514 return BinaryOperator::createNot(And);
4518 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4519 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4520 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4523 Value *LHSVal, *RHSVal;
4524 ConstantInt *LHSCst, *RHSCst;
4525 ICmpInst::Predicate LHSCC, RHSCC;
4526 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4527 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4528 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4529 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4530 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4531 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4532 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4533 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
4534 // Ensure that the larger constant is on the RHS.
4535 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
4536 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
4537 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
4538 ICmpInst *LHS = cast<ICmpInst>(Op0);
4539 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
4540 std::swap(LHS, RHS);
4541 std::swap(LHSCst, RHSCst);
4542 std::swap(LHSCC, RHSCC);
4545 // At this point, we know we have have two icmp instructions
4546 // comparing a value against two constants and or'ing the result
4547 // together. Because of the above check, we know that we only have
4548 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4549 // FoldICmpLogical check above), that the two constants are not
4551 assert(LHSCst != RHSCst && "Compares not folded above?");
4554 default: assert(0 && "Unknown integer condition code!");
4555 case ICmpInst::ICMP_EQ:
4557 default: assert(0 && "Unknown integer condition code!");
4558 case ICmpInst::ICMP_EQ:
4559 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4560 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4561 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4562 LHSVal->getName()+".off");
4563 InsertNewInstBefore(Add, I);
4564 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4565 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4567 break; // (X == 13 | X == 15) -> no change
4568 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4569 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4571 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4572 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4573 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4574 return ReplaceInstUsesWith(I, RHS);
4577 case ICmpInst::ICMP_NE:
4579 default: assert(0 && "Unknown integer condition code!");
4580 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4581 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4582 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4583 return ReplaceInstUsesWith(I, LHS);
4584 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4585 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4586 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4587 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4590 case ICmpInst::ICMP_ULT:
4592 default: assert(0 && "Unknown integer condition code!");
4593 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4595 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4596 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4598 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4600 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4601 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4602 return ReplaceInstUsesWith(I, RHS);
4603 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4607 case ICmpInst::ICMP_SLT:
4609 default: assert(0 && "Unknown integer condition code!");
4610 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4612 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4613 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4615 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4617 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4618 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4619 return ReplaceInstUsesWith(I, RHS);
4620 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4624 case ICmpInst::ICMP_UGT:
4626 default: assert(0 && "Unknown integer condition code!");
4627 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4628 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4629 return ReplaceInstUsesWith(I, LHS);
4630 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4632 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4633 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4634 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4635 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4639 case ICmpInst::ICMP_SGT:
4641 default: assert(0 && "Unknown integer condition code!");
4642 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4643 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4644 return ReplaceInstUsesWith(I, LHS);
4645 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4647 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4648 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4649 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4650 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4658 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4659 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4660 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4661 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4662 const Type *SrcTy = Op0C->getOperand(0)->getType();
4663 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4664 // Only do this if the casts both really cause code to be generated.
4665 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4667 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4669 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4670 Op1C->getOperand(0),
4672 InsertNewInstBefore(NewOp, I);
4673 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4678 return Changed ? &I : 0;
4681 // XorSelf - Implements: X ^ X --> 0
4684 XorSelf(Value *rhs) : RHS(rhs) {}
4685 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4686 Instruction *apply(BinaryOperator &Xor) const {
4692 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4693 bool Changed = SimplifyCommutative(I);
4694 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4696 if (isa<UndefValue>(Op1))
4697 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4699 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4700 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4701 assert(Result == &I && "AssociativeOpt didn't work?");
4702 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4705 // See if we can simplify any instructions used by the instruction whose sole
4706 // purpose is to compute bits we don't care about.
4707 uint64_t KnownZero, KnownOne;
4708 if (!isa<VectorType>(I.getType()) &&
4709 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
4710 KnownZero, KnownOne))
4713 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4714 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
4715 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4716 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
4717 return new ICmpInst(ICI->getInversePredicate(),
4718 ICI->getOperand(0), ICI->getOperand(1));
4720 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4721 // ~(c-X) == X-c-1 == X+(-c-1)
4722 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4723 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4724 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4725 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4726 ConstantInt::get(I.getType(), 1));
4727 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4730 // ~(~X & Y) --> (X | ~Y)
4731 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
4732 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4733 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4735 BinaryOperator::createNot(Op0I->getOperand(1),
4736 Op0I->getOperand(1)->getName()+".not");
4737 InsertNewInstBefore(NotY, I);
4738 return BinaryOperator::createOr(Op0NotVal, NotY);
4742 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4743 if (Op0I->getOpcode() == Instruction::Add) {
4744 // ~(X-c) --> (-c-1)-X
4745 if (RHS->isAllOnesValue()) {
4746 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4747 return BinaryOperator::createSub(
4748 ConstantExpr::getSub(NegOp0CI,
4749 ConstantInt::get(I.getType(), 1)),
4750 Op0I->getOperand(0));
4752 } else if (Op0I->getOpcode() == Instruction::Or) {
4753 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4754 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
4755 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4756 // Anything in both C1 and C2 is known to be zero, remove it from
4758 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
4759 NewRHS = ConstantExpr::getAnd(NewRHS,
4760 ConstantExpr::getNot(CommonBits));
4761 AddToWorkList(Op0I);
4762 I.setOperand(0, Op0I->getOperand(0));
4763 I.setOperand(1, NewRHS);
4769 // Try to fold constant and into select arguments.
4770 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4771 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4773 if (isa<PHINode>(Op0))
4774 if (Instruction *NV = FoldOpIntoPhi(I))
4778 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4780 return ReplaceInstUsesWith(I,
4781 ConstantInt::getAllOnesValue(I.getType()));
4783 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4785 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
4788 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4791 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4792 if (A == Op0) { // B^(B|A) == (A|B)^B
4793 Op1I->swapOperands();
4795 std::swap(Op0, Op1);
4796 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4797 I.swapOperands(); // Simplified below.
4798 std::swap(Op0, Op1);
4800 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4801 if (Op0 == A) // A^(A^B) == B
4802 return ReplaceInstUsesWith(I, B);
4803 else if (Op0 == B) // A^(B^A) == B
4804 return ReplaceInstUsesWith(I, A);
4805 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4806 if (A == Op0) // A^(A&B) -> A^(B&A)
4807 Op1I->swapOperands();
4808 if (B == Op0) { // A^(B&A) -> (B&A)^A
4809 I.swapOperands(); // Simplified below.
4810 std::swap(Op0, Op1);
4815 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4818 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4819 if (A == Op1) // (B|A)^B == (A|B)^B
4821 if (B == Op1) { // (A|B)^B == A & ~B
4823 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4824 return BinaryOperator::createAnd(A, NotB);
4826 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4827 if (Op1 == A) // (A^B)^A == B
4828 return ReplaceInstUsesWith(I, B);
4829 else if (Op1 == B) // (B^A)^A == B
4830 return ReplaceInstUsesWith(I, A);
4831 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4832 if (A == Op1) // (A&B)^A -> (B&A)^A
4834 if (B == Op1 && // (B&A)^A == ~B & A
4835 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4837 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4838 return BinaryOperator::createAnd(N, Op1);
4843 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4844 if (Op0I && Op1I && Op0I->isShift() &&
4845 Op0I->getOpcode() == Op1I->getOpcode() &&
4846 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4847 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4848 Instruction *NewOp =
4849 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4850 Op1I->getOperand(0),
4851 Op0I->getName()), I);
4852 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4853 Op1I->getOperand(1));
4857 Value *A, *B, *C, *D;
4858 // (A & B)^(A | B) -> A ^ B
4859 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4860 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4861 if ((A == C && B == D) || (A == D && B == C))
4862 return BinaryOperator::createXor(A, B);
4864 // (A | B)^(A & B) -> A ^ B
4865 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4866 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4867 if ((A == C && B == D) || (A == D && B == C))
4868 return BinaryOperator::createXor(A, B);
4872 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4873 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4874 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4875 // (X & Y)^(X & Y) -> (Y^Z) & X
4876 Value *X = 0, *Y = 0, *Z = 0;
4878 X = A, Y = B, Z = D;
4880 X = A, Y = B, Z = C;
4882 X = B, Y = A, Z = D;
4884 X = B, Y = A, Z = C;
4887 Instruction *NewOp =
4888 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4889 return BinaryOperator::createAnd(NewOp, X);
4894 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4895 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4896 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4899 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4900 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4901 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4902 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4903 const Type *SrcTy = Op0C->getOperand(0)->getType();
4904 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4905 // Only do this if the casts both really cause code to be generated.
4906 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4908 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4910 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4911 Op1C->getOperand(0),
4913 InsertNewInstBefore(NewOp, I);
4914 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4918 return Changed ? &I : 0;
4921 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4922 /// overflowed for this type.
4923 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4924 ConstantInt *In2, bool IsSigned = false) {
4925 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4928 if (In2->getValue().isNegative())
4929 return Result->getValue().sgt(In1->getValue());
4931 return Result->getValue().slt(In1->getValue());
4933 return Result->getValue().ult(In1->getValue());
4936 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4937 /// code necessary to compute the offset from the base pointer (without adding
4938 /// in the base pointer). Return the result as a signed integer of intptr size.
4939 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4940 TargetData &TD = IC.getTargetData();
4941 gep_type_iterator GTI = gep_type_begin(GEP);
4942 const Type *IntPtrTy = TD.getIntPtrType();
4943 Value *Result = Constant::getNullValue(IntPtrTy);
4945 // Build a mask for high order bits.
4946 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4948 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4949 Value *Op = GEP->getOperand(i);
4950 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4951 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4952 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4953 if (!OpC->isNullValue()) {
4954 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4955 Scale = ConstantExpr::getMul(OpC, Scale);
4956 if (Constant *RC = dyn_cast<Constant>(Result))
4957 Result = ConstantExpr::getAdd(RC, Scale);
4959 // Emit an add instruction.
4960 Result = IC.InsertNewInstBefore(
4961 BinaryOperator::createAdd(Result, Scale,
4962 GEP->getName()+".offs"), I);
4966 // Convert to correct type.
4967 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4968 Op->getName()+".c"), I);
4970 // We'll let instcombine(mul) convert this to a shl if possible.
4971 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4972 GEP->getName()+".idx"), I);
4974 // Emit an add instruction.
4975 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4976 GEP->getName()+".offs"), I);
4982 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4983 /// else. At this point we know that the GEP is on the LHS of the comparison.
4984 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4985 ICmpInst::Predicate Cond,
4987 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4989 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4990 if (isa<PointerType>(CI->getOperand(0)->getType()))
4991 RHS = CI->getOperand(0);
4993 Value *PtrBase = GEPLHS->getOperand(0);
4994 if (PtrBase == RHS) {
4995 // As an optimization, we don't actually have to compute the actual value of
4996 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4997 // each index is zero or not.
4998 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4999 Instruction *InVal = 0;
5000 gep_type_iterator GTI = gep_type_begin(GEPLHS);
5001 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
5003 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
5004 if (isa<UndefValue>(C)) // undef index -> undef.
5005 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
5006 if (C->isNullValue())
5008 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
5009 EmitIt = false; // This is indexing into a zero sized array?
5010 } else if (isa<ConstantInt>(C))
5011 return ReplaceInstUsesWith(I, // No comparison is needed here.
5012 ConstantInt::get(Type::Int1Ty,
5013 Cond == ICmpInst::ICMP_NE));
5018 new ICmpInst(Cond, GEPLHS->getOperand(i),
5019 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
5023 InVal = InsertNewInstBefore(InVal, I);
5024 InsertNewInstBefore(Comp, I);
5025 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
5026 InVal = BinaryOperator::createOr(InVal, Comp);
5027 else // True if all are equal
5028 InVal = BinaryOperator::createAnd(InVal, Comp);
5036 // No comparison is needed here, all indexes = 0
5037 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5038 Cond == ICmpInst::ICMP_EQ));
5041 // Only lower this if the icmp is the only user of the GEP or if we expect
5042 // the result to fold to a constant!
5043 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
5044 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5045 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
5046 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5047 Constant::getNullValue(Offset->getType()));
5049 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5050 // If the base pointers are different, but the indices are the same, just
5051 // compare the base pointer.
5052 if (PtrBase != GEPRHS->getOperand(0)) {
5053 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5054 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5055 GEPRHS->getOperand(0)->getType();
5057 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5058 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5059 IndicesTheSame = false;
5063 // If all indices are the same, just compare the base pointers.
5065 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5066 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5068 // Otherwise, the base pointers are different and the indices are
5069 // different, bail out.
5073 // If one of the GEPs has all zero indices, recurse.
5074 bool AllZeros = true;
5075 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5076 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5077 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5082 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5083 ICmpInst::getSwappedPredicate(Cond), I);
5085 // If the other GEP has all zero indices, recurse.
5087 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5088 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5089 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5094 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5096 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5097 // If the GEPs only differ by one index, compare it.
5098 unsigned NumDifferences = 0; // Keep track of # differences.
5099 unsigned DiffOperand = 0; // The operand that differs.
5100 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5101 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5102 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5103 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5104 // Irreconcilable differences.
5108 if (NumDifferences++) break;
5113 if (NumDifferences == 0) // SAME GEP?
5114 return ReplaceInstUsesWith(I, // No comparison is needed here.
5115 ConstantInt::get(Type::Int1Ty,
5116 Cond == ICmpInst::ICMP_EQ));
5117 else if (NumDifferences == 1) {
5118 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5119 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5120 // Make sure we do a signed comparison here.
5121 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5125 // Only lower this if the icmp is the only user of the GEP or if we expect
5126 // the result to fold to a constant!
5127 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5128 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5129 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5130 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5131 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5132 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5138 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5139 bool Changed = SimplifyCompare(I);
5140 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5142 // Fold trivial predicates.
5143 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5144 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5145 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5146 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5148 // Simplify 'fcmp pred X, X'
5150 switch (I.getPredicate()) {
5151 default: assert(0 && "Unknown predicate!");
5152 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5153 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5154 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5155 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5156 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5157 case FCmpInst::FCMP_OLT: // True if ordered and less than
5158 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5159 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5161 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5162 case FCmpInst::FCMP_ULT: // True if unordered or less than
5163 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5164 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5165 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5166 I.setPredicate(FCmpInst::FCMP_UNO);
5167 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5170 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5171 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5172 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5173 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5174 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5175 I.setPredicate(FCmpInst::FCMP_ORD);
5176 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5181 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5182 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5184 // Handle fcmp with constant RHS
5185 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5186 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5187 switch (LHSI->getOpcode()) {
5188 case Instruction::PHI:
5189 if (Instruction *NV = FoldOpIntoPhi(I))
5192 case Instruction::Select:
5193 // If either operand of the select is a constant, we can fold the
5194 // comparison into the select arms, which will cause one to be
5195 // constant folded and the select turned into a bitwise or.
5196 Value *Op1 = 0, *Op2 = 0;
5197 if (LHSI->hasOneUse()) {
5198 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5199 // Fold the known value into the constant operand.
5200 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5201 // Insert a new FCmp of the other select operand.
5202 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5203 LHSI->getOperand(2), RHSC,
5205 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5206 // Fold the known value into the constant operand.
5207 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5208 // Insert a new FCmp of the other select operand.
5209 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5210 LHSI->getOperand(1), RHSC,
5216 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5221 return Changed ? &I : 0;
5224 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5225 bool Changed = SimplifyCompare(I);
5226 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5227 const Type *Ty = Op0->getType();
5231 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5232 isTrueWhenEqual(I)));
5234 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5235 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5237 // icmp of GlobalValues can never equal each other as long as they aren't
5238 // external weak linkage type.
5239 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
5240 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
5241 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
5242 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5243 !isTrueWhenEqual(I)));
5245 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5246 // addresses never equal each other! We already know that Op0 != Op1.
5247 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5248 isa<ConstantPointerNull>(Op0)) &&
5249 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5250 isa<ConstantPointerNull>(Op1)))
5251 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5252 !isTrueWhenEqual(I)));
5254 // icmp's with boolean values can always be turned into bitwise operations
5255 if (Ty == Type::Int1Ty) {
5256 switch (I.getPredicate()) {
5257 default: assert(0 && "Invalid icmp instruction!");
5258 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5259 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
5260 InsertNewInstBefore(Xor, I);
5261 return BinaryOperator::createNot(Xor);
5263 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5264 return BinaryOperator::createXor(Op0, Op1);
5266 case ICmpInst::ICMP_UGT:
5267 case ICmpInst::ICMP_SGT:
5268 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5270 case ICmpInst::ICMP_ULT:
5271 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5272 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
5273 InsertNewInstBefore(Not, I);
5274 return BinaryOperator::createAnd(Not, Op1);
5276 case ICmpInst::ICMP_UGE:
5277 case ICmpInst::ICMP_SGE:
5278 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5280 case ICmpInst::ICMP_ULE:
5281 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5282 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
5283 InsertNewInstBefore(Not, I);
5284 return BinaryOperator::createOr(Not, Op1);
5289 // See if we are doing a comparison between a constant and an instruction that
5290 // can be folded into the comparison.
5291 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5292 switch (I.getPredicate()) {
5294 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5295 if (CI->isMinValue(false))
5296 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5297 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5298 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5299 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5300 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5303 case ICmpInst::ICMP_SLT:
5304 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5305 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5306 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5307 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5308 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5309 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5312 case ICmpInst::ICMP_UGT:
5313 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5314 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5315 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5316 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5317 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5318 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5321 case ICmpInst::ICMP_SGT:
5322 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5323 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5324 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5325 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5326 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5327 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5330 case ICmpInst::ICMP_ULE:
5331 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5332 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5333 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5334 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5335 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5336 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5339 case ICmpInst::ICMP_SLE:
5340 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5341 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5342 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5343 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5344 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5345 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5348 case ICmpInst::ICMP_UGE:
5349 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5350 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5351 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5352 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5353 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5354 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5357 case ICmpInst::ICMP_SGE:
5358 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5359 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5360 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5361 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5362 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5363 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5367 // If we still have a icmp le or icmp ge instruction, turn it into the
5368 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5369 // already been handled above, this requires little checking.
5371 if (I.getPredicate() == ICmpInst::ICMP_ULE)
5372 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5373 if (I.getPredicate() == ICmpInst::ICMP_SLE)
5374 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5375 if (I.getPredicate() == ICmpInst::ICMP_UGE)
5376 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5377 if (I.getPredicate() == ICmpInst::ICMP_SGE)
5378 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5380 // See if we can fold the comparison based on bits known to be zero or one
5382 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5383 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5384 if (SimplifyDemandedBits(Op0, APInt::getAllOnesValue(BitWidth),
5385 KnownZero, KnownOne, 0))
5388 // Given the known and unknown bits, compute a range that the LHS could be
5390 if ((KnownOne | KnownZero) != 0) {
5391 // Compute the Min, Max and RHS values based on the known bits. For the
5392 // EQ and NE we use unsigned values.
5393 APInt Min(BitWidth, 0), Max(BitWidth, 0), RHSVal(CI->getValue());
5394 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5395 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5398 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5401 switch (I.getPredicate()) { // LE/GE have been folded already.
5402 default: assert(0 && "Unknown icmp opcode!");
5403 case ICmpInst::ICMP_EQ:
5404 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5405 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5407 case ICmpInst::ICMP_NE:
5408 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5409 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5411 case ICmpInst::ICMP_ULT:
5412 if (Max.ult(RHSVal))
5413 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5414 if (Min.ugt(RHSVal))
5415 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5417 case ICmpInst::ICMP_UGT:
5418 if (Min.ugt(RHSVal))
5419 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5420 if (Max.ult(RHSVal))
5421 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5423 case ICmpInst::ICMP_SLT:
5424 if (Max.slt(RHSVal))
5425 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5426 if (Min.sgt(RHSVal))
5427 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5429 case ICmpInst::ICMP_SGT:
5430 if (Min.sgt(RHSVal))
5431 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5432 if (Max.slt(RHSVal))
5433 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5438 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5439 // instruction, see if that instruction also has constants so that the
5440 // instruction can be folded into the icmp
5441 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5442 switch (LHSI->getOpcode()) {
5443 case Instruction::And:
5444 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5445 LHSI->getOperand(0)->hasOneUse()) {
5446 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5448 // If the LHS is an AND of a truncating cast, we can widen the
5449 // and/compare to be the input width without changing the value
5450 // produced, eliminating a cast.
5451 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
5452 // We can do this transformation if either the AND constant does not
5453 // have its sign bit set or if it is an equality comparison.
5454 // Extending a relational comparison when we're checking the sign
5455 // bit would not work.
5456 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
5457 (I.isEquality() || AndCST->getValue().isPositive() &&
5458 CI->getValue().isPositive())) {
5459 ConstantInt *NewCST;
5461 APInt NewCSTVal(AndCST->getValue()), NewCIVal(CI->getValue());
5462 uint32_t BitWidth = cast<IntegerType>(
5463 Cast->getOperand(0)->getType())->getBitWidth();
5464 NewCST = ConstantInt::get(NewCSTVal.zext(BitWidth));
5465 NewCI = ConstantInt::get(NewCIVal.zext(BitWidth));
5466 Instruction *NewAnd =
5467 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
5469 InsertNewInstBefore(NewAnd, I);
5470 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
5474 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5475 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5476 // happens a LOT in code produced by the C front-end, for bitfield
5478 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5479 if (Shift && !Shift->isShift())
5483 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5484 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5485 const Type *AndTy = AndCST->getType(); // Type of the and.
5487 // We can fold this as long as we can't shift unknown bits
5488 // into the mask. This can only happen with signed shift
5489 // rights, as they sign-extend.
5491 bool CanFold = Shift->isLogicalShift();
5493 // To test for the bad case of the signed shr, see if any
5494 // of the bits shifted in could be tested after the mask.
5495 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
5496 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
5498 Constant *OShAmt = ConstantInt::get(AndTy, ShAmtVal);
5500 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
5502 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
5508 if (Shift->getOpcode() == Instruction::Shl)
5509 NewCst = ConstantExpr::getLShr(CI, ShAmt);
5511 NewCst = ConstantExpr::getShl(CI, ShAmt);
5513 // Check to see if we are shifting out any of the bits being
5515 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
5516 // If we shifted bits out, the fold is not going to work out.
5517 // As a special case, check to see if this means that the
5518 // result is always true or false now.
5519 if (I.getPredicate() == ICmpInst::ICMP_EQ)
5520 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5521 if (I.getPredicate() == ICmpInst::ICMP_NE)
5522 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5524 I.setOperand(1, NewCst);
5525 Constant *NewAndCST;
5526 if (Shift->getOpcode() == Instruction::Shl)
5527 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5529 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5530 LHSI->setOperand(1, NewAndCST);
5531 LHSI->setOperand(0, Shift->getOperand(0));
5532 AddToWorkList(Shift); // Shift is dead.
5533 AddUsesToWorkList(I);
5539 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5540 // preferable because it allows the C<<Y expression to be hoisted out
5541 // of a loop if Y is invariant and X is not.
5542 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
5543 I.isEquality() && !Shift->isArithmeticShift() &&
5544 isa<Instruction>(Shift->getOperand(0))) {
5547 if (Shift->getOpcode() == Instruction::LShr) {
5548 NS = BinaryOperator::createShl(AndCST,
5549 Shift->getOperand(1), "tmp");
5551 // Insert a logical shift.
5552 NS = BinaryOperator::createLShr(AndCST,
5553 Shift->getOperand(1), "tmp");
5555 InsertNewInstBefore(cast<Instruction>(NS), I);
5557 // Compute X & (C << Y).
5558 Instruction *NewAnd = BinaryOperator::createAnd(
5559 Shift->getOperand(0), NS, LHSI->getName());
5560 InsertNewInstBefore(NewAnd, I);
5562 I.setOperand(0, NewAnd);
5568 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
5569 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5570 if (I.isEquality()) {
5571 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
5573 // Check that the shift amount is in range. If not, don't perform
5574 // undefined shifts. When the shift is visited it will be
5576 if (ShAmt->getZExtValue() >= TypeBits)
5579 // If we are comparing against bits always shifted out, the
5580 // comparison cannot succeed.
5582 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
5583 if (Comp != CI) {// Comparing against a bit that we know is zero.
5584 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
5585 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5586 return ReplaceInstUsesWith(I, Cst);
5589 if (LHSI->hasOneUse()) {
5590 // Otherwise strength reduce the shift into an and.
5591 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
5592 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
5593 Constant *Mask = ConstantInt::get(CI->getType(), Val);
5596 BinaryOperator::createAnd(LHSI->getOperand(0),
5597 Mask, LHSI->getName()+".mask");
5598 Value *And = InsertNewInstBefore(AndI, I);
5599 return new ICmpInst(I.getPredicate(), And,
5600 ConstantExpr::getLShr(CI, ShAmt));
5606 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5607 case Instruction::AShr:
5608 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5609 if (I.isEquality()) {
5610 // Check that the shift amount is in range. If not, don't perform
5611 // undefined shifts. When the shift is visited it will be
5613 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
5614 if (ShAmt->getZExtValue() >= TypeBits)
5617 // If we are comparing against bits always shifted out, the
5618 // comparison cannot succeed.
5620 if (LHSI->getOpcode() == Instruction::LShr)
5621 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
5624 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
5627 if (Comp != CI) {// Comparing against a bit that we know is zero.
5628 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
5629 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5630 return ReplaceInstUsesWith(I, Cst);
5633 if (LHSI->hasOneUse() || CI->isNullValue()) {
5634 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
5636 // Otherwise strength reduce the shift into an and.
5637 APInt Val(APInt::getAllOnesValue(TypeBits).shl(ShAmtVal));
5638 Constant *Mask = ConstantInt::get(Val);
5641 BinaryOperator::createAnd(LHSI->getOperand(0),
5642 Mask, LHSI->getName()+".mask");
5643 Value *And = InsertNewInstBefore(AndI, I);
5644 return new ICmpInst(I.getPredicate(), And,
5645 ConstantExpr::getShl(CI, ShAmt));
5651 case Instruction::SDiv:
5652 case Instruction::UDiv:
5653 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5654 // Fold this div into the comparison, producing a range check.
5655 // Determine, based on the divide type, what the range is being
5656 // checked. If there is an overflow on the low or high side, remember
5657 // it, otherwise compute the range [low, hi) bounding the new value.
5658 // See: InsertRangeTest above for the kinds of replacements possible.
5659 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5660 // FIXME: If the operand types don't match the type of the divide
5661 // then don't attempt this transform. The code below doesn't have the
5662 // logic to deal with a signed divide and an unsigned compare (and
5663 // vice versa). This is because (x /s C1) <s C2 produces different
5664 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5665 // (x /u C1) <u C2. Simply casting the operands and result won't
5666 // work. :( The if statement below tests that condition and bails
5668 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
5669 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
5671 if (DivRHS->isZero())
5672 break; // Don't hack on div by zero
5674 // Initialize the variables that will indicate the nature of the
5676 bool LoOverflow = false, HiOverflow = false;
5677 ConstantInt *LoBound = 0, *HiBound = 0;
5679 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5680 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5681 // C2 (CI). By solving for X we can turn this into a range check
5682 // instead of computing a divide.
5684 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
5686 // Determine if the product overflows by seeing if the product is
5687 // not equal to the divide. Make sure we do the same kind of divide
5688 // as in the LHS instruction that we're folding.
5689 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5690 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
5692 // Get the ICmp opcode
5693 ICmpInst::Predicate predicate = I.getPredicate();
5695 if (!DivIsSigned) { // udiv
5697 LoOverflow = ProdOV;
5698 HiOverflow = ProdOV ||
5699 AddWithOverflow(HiBound, LoBound, DivRHS, false);
5700 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5701 if (CI->isNullValue()) { // (X / pos) op 0
5703 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5705 } else if (CI->getValue().isPositive()) { // (X / pos) op pos
5707 LoOverflow = ProdOV;
5708 HiOverflow = ProdOV ||
5709 AddWithOverflow(HiBound, Prod, DivRHS, true);
5710 } else { // (X / pos) op neg
5711 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5712 LoOverflow = AddWithOverflow(LoBound, Prod,
5713 cast<ConstantInt>(DivRHSH), true);
5714 HiBound = AddOne(Prod);
5715 HiOverflow = ProdOV;
5717 } else { // Divisor is < 0.
5718 if (CI->isNullValue()) { // (X / neg) op 0
5719 LoBound = AddOne(DivRHS);
5720 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5721 if (HiBound == DivRHS)
5722 LoBound = 0; // - INTMIN = INTMIN
5723 } else if (CI->getValue().isPositive()) { // (X / neg) op pos
5724 HiOverflow = LoOverflow = ProdOV;
5726 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS),
5728 HiBound = AddOne(Prod);
5729 } else { // (X / neg) op neg
5731 LoOverflow = HiOverflow = ProdOV;
5732 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
5735 // Dividing by a negate swaps the condition.
5736 predicate = ICmpInst::getSwappedPredicate(predicate);
5740 Value *X = LHSI->getOperand(0);
5741 switch (predicate) {
5742 default: assert(0 && "Unhandled icmp opcode!");
5743 case ICmpInst::ICMP_EQ:
5744 if (LoOverflow && HiOverflow)
5745 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5746 else if (HiOverflow)
5747 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5748 ICmpInst::ICMP_UGE, X, LoBound);
5749 else if (LoOverflow)
5750 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5751 ICmpInst::ICMP_ULT, X, HiBound);
5753 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5755 case ICmpInst::ICMP_NE:
5756 if (LoOverflow && HiOverflow)
5757 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5758 else if (HiOverflow)
5759 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5760 ICmpInst::ICMP_ULT, X, LoBound);
5761 else if (LoOverflow)
5762 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5763 ICmpInst::ICMP_UGE, X, HiBound);
5765 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5767 case ICmpInst::ICMP_ULT:
5768 case ICmpInst::ICMP_SLT:
5770 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5771 return new ICmpInst(predicate, X, LoBound);
5772 case ICmpInst::ICMP_UGT:
5773 case ICmpInst::ICMP_SGT:
5775 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5776 if (predicate == ICmpInst::ICMP_UGT)
5777 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5779 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5786 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5787 if (I.isEquality()) {
5788 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
5790 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5791 // the second operand is a constant, simplify a bit.
5792 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
5793 switch (BO->getOpcode()) {
5794 case Instruction::SRem:
5795 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5796 if (CI->isZero() && isa<ConstantInt>(BO->getOperand(1)) &&
5798 APInt V(cast<ConstantInt>(BO->getOperand(1))->getValue());
5799 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5800 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
5801 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
5802 return new ICmpInst(I.getPredicate(), NewRem,
5803 Constant::getNullValue(BO->getType()));
5807 case Instruction::Add:
5808 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5809 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5810 if (BO->hasOneUse())
5811 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5812 ConstantExpr::getSub(CI, BOp1C));
5813 } else if (CI->isNullValue()) {
5814 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5815 // efficiently invertible, or if the add has just this one use.
5816 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5818 if (Value *NegVal = dyn_castNegVal(BOp1))
5819 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
5820 else if (Value *NegVal = dyn_castNegVal(BOp0))
5821 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
5822 else if (BO->hasOneUse()) {
5823 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5824 InsertNewInstBefore(Neg, I);
5826 return new ICmpInst(I.getPredicate(), BOp0, Neg);
5830 case Instruction::Xor:
5831 // For the xor case, we can xor two constants together, eliminating
5832 // the explicit xor.
5833 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5834 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5835 ConstantExpr::getXor(CI, BOC));
5838 case Instruction::Sub:
5839 // Replace (([sub|xor] A, B) != 0) with (A != B)
5841 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5845 case Instruction::Or:
5846 // If bits are being or'd in that are not present in the constant we
5847 // are comparing against, then the comparison could never succeed!
5848 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5849 Constant *NotCI = ConstantExpr::getNot(CI);
5850 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5851 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5856 case Instruction::And:
5857 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5858 // If bits are being compared against that are and'd out, then the
5859 // comparison can never succeed!
5860 if (!ConstantExpr::getAnd(CI,
5861 ConstantExpr::getNot(BOC))->isNullValue())
5862 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5865 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5866 if (CI == BOC && isOneBitSet(CI))
5867 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5868 ICmpInst::ICMP_NE, Op0,
5869 Constant::getNullValue(CI->getType()));
5871 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5872 if (isSignBit(BOC)) {
5873 Value *X = BO->getOperand(0);
5874 Constant *Zero = Constant::getNullValue(X->getType());
5875 ICmpInst::Predicate pred = isICMP_NE ?
5876 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5877 return new ICmpInst(pred, X, Zero);
5880 // ((X & ~7) == 0) --> X < 8
5881 if (CI->isNullValue() && isHighOnes(BOC)) {
5882 Value *X = BO->getOperand(0);
5883 Constant *NegX = ConstantExpr::getNeg(BOC);
5884 ICmpInst::Predicate pred = isICMP_NE ?
5885 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5886 return new ICmpInst(pred, X, NegX);
5892 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5893 // Handle set{eq|ne} <intrinsic>, intcst.
5894 switch (II->getIntrinsicID()) {
5896 case Intrinsic::bswap_i16:
5897 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5898 AddToWorkList(II); // Dead?
5899 I.setOperand(0, II->getOperand(1));
5900 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5901 ByteSwap_16(CI->getZExtValue())));
5903 case Intrinsic::bswap_i32:
5904 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5905 AddToWorkList(II); // Dead?
5906 I.setOperand(0, II->getOperand(1));
5907 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5908 ByteSwap_32(CI->getZExtValue())));
5910 case Intrinsic::bswap_i64:
5911 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5912 AddToWorkList(II); // Dead?
5913 I.setOperand(0, II->getOperand(1));
5914 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5915 ByteSwap_64(CI->getZExtValue())));
5919 } else { // Not a ICMP_EQ/ICMP_NE
5920 // If the LHS is a cast from an integral value of the same size, then
5921 // since we know the RHS is a constant, try to simlify.
5922 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5923 Value *CastOp = Cast->getOperand(0);
5924 const Type *SrcTy = CastOp->getType();
5925 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5926 if (SrcTy->isInteger() &&
5927 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5928 // If this is an unsigned comparison, try to make the comparison use
5929 // smaller constant values.
5930 switch (I.getPredicate()) {
5932 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5933 ConstantInt *CUI = cast<ConstantInt>(CI);
5934 if (CUI->getValue() == APInt::getSignBit(SrcTySize))
5935 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5936 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5939 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5940 ConstantInt *CUI = cast<ConstantInt>(CI);
5941 if (CUI->getValue() == APInt::getSignedMaxValue(SrcTySize))
5942 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5943 Constant::getNullValue(SrcTy));
5953 // Handle icmp with constant RHS
5954 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5955 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5956 switch (LHSI->getOpcode()) {
5957 case Instruction::GetElementPtr:
5958 if (RHSC->isNullValue()) {
5959 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5960 bool isAllZeros = true;
5961 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5962 if (!isa<Constant>(LHSI->getOperand(i)) ||
5963 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5968 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5969 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5973 case Instruction::PHI:
5974 if (Instruction *NV = FoldOpIntoPhi(I))
5977 case Instruction::Select:
5978 // If either operand of the select is a constant, we can fold the
5979 // comparison into the select arms, which will cause one to be
5980 // constant folded and the select turned into a bitwise or.
5981 Value *Op1 = 0, *Op2 = 0;
5982 if (LHSI->hasOneUse()) {
5983 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5984 // Fold the known value into the constant operand.
5985 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5986 // Insert a new ICmp of the other select operand.
5987 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5988 LHSI->getOperand(2), RHSC,
5990 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5991 // Fold the known value into the constant operand.
5992 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5993 // Insert a new ICmp of the other select operand.
5994 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5995 LHSI->getOperand(1), RHSC,
6001 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
6006 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6007 if (User *GEP = dyn_castGetElementPtr(Op0))
6008 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6010 if (User *GEP = dyn_castGetElementPtr(Op1))
6011 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6012 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6015 // Test to see if the operands of the icmp are casted versions of other
6016 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6018 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6019 if (isa<PointerType>(Op0->getType()) &&
6020 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6021 // We keep moving the cast from the left operand over to the right
6022 // operand, where it can often be eliminated completely.
6023 Op0 = CI->getOperand(0);
6025 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6026 // so eliminate it as well.
6027 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6028 Op1 = CI2->getOperand(0);
6030 // If Op1 is a constant, we can fold the cast into the constant.
6031 if (Op0->getType() != Op1->getType())
6032 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6033 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6035 // Otherwise, cast the RHS right before the icmp
6036 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
6038 return new ICmpInst(I.getPredicate(), Op0, Op1);
6042 if (isa<CastInst>(Op0)) {
6043 // Handle the special case of: icmp (cast bool to X), <cst>
6044 // This comes up when you have code like
6047 // For generality, we handle any zero-extension of any operand comparison
6048 // with a constant or another cast from the same type.
6049 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6050 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6054 if (I.isEquality()) {
6055 Value *A, *B, *C, *D;
6056 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6057 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6058 Value *OtherVal = A == Op1 ? B : A;
6059 return new ICmpInst(I.getPredicate(), OtherVal,
6060 Constant::getNullValue(A->getType()));
6063 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6064 // A^c1 == C^c2 --> A == C^(c1^c2)
6065 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
6066 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
6067 if (Op1->hasOneUse()) {
6068 Constant *NC = ConstantExpr::getXor(C1, C2);
6069 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
6070 return new ICmpInst(I.getPredicate(), A,
6071 InsertNewInstBefore(Xor, I));
6074 // A^B == A^D -> B == D
6075 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6076 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6077 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6078 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6082 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6083 (A == Op0 || B == Op0)) {
6084 // A == (A^B) -> B == 0
6085 Value *OtherVal = A == Op0 ? B : A;
6086 return new ICmpInst(I.getPredicate(), OtherVal,
6087 Constant::getNullValue(A->getType()));
6089 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
6090 // (A-B) == A -> B == 0
6091 return new ICmpInst(I.getPredicate(), B,
6092 Constant::getNullValue(B->getType()));
6094 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
6095 // A == (A-B) -> B == 0
6096 return new ICmpInst(I.getPredicate(), B,
6097 Constant::getNullValue(B->getType()));
6100 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6101 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6102 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6103 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6104 Value *X = 0, *Y = 0, *Z = 0;
6107 X = B; Y = D; Z = A;
6108 } else if (A == D) {
6109 X = B; Y = C; Z = A;
6110 } else if (B == C) {
6111 X = A; Y = D; Z = B;
6112 } else if (B == D) {
6113 X = A; Y = C; Z = B;
6116 if (X) { // Build (X^Y) & Z
6117 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
6118 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
6119 I.setOperand(0, Op1);
6120 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6125 return Changed ? &I : 0;
6128 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6129 // We only handle extending casts so far.
6131 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6132 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6133 Value *LHSCIOp = LHSCI->getOperand(0);
6134 const Type *SrcTy = LHSCIOp->getType();
6135 const Type *DestTy = LHSCI->getType();
6138 // We only handle extension cast instructions, so far. Enforce this.
6139 if (LHSCI->getOpcode() != Instruction::ZExt &&
6140 LHSCI->getOpcode() != Instruction::SExt)
6143 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6144 bool isSignedCmp = ICI.isSignedPredicate();
6146 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6147 // Not an extension from the same type?
6148 RHSCIOp = CI->getOperand(0);
6149 if (RHSCIOp->getType() != LHSCIOp->getType())
6152 // If the signedness of the two compares doesn't agree (i.e. one is a sext
6153 // and the other is a zext), then we can't handle this.
6154 if (CI->getOpcode() != LHSCI->getOpcode())
6157 // Likewise, if the signedness of the [sz]exts and the compare don't match,
6158 // then we can't handle this.
6159 if (isSignedExt != isSignedCmp && !ICI.isEquality())
6162 // Okay, just insert a compare of the reduced operands now!
6163 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6166 // If we aren't dealing with a constant on the RHS, exit early
6167 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6171 // Compute the constant that would happen if we truncated to SrcTy then
6172 // reextended to DestTy.
6173 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6174 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6176 // If the re-extended constant didn't change...
6178 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6179 // For example, we might have:
6180 // %A = sext short %X to uint
6181 // %B = icmp ugt uint %A, 1330
6182 // It is incorrect to transform this into
6183 // %B = icmp ugt short %X, 1330
6184 // because %A may have negative value.
6186 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6187 // OR operation is EQ/NE.
6188 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6189 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6194 // The re-extended constant changed so the constant cannot be represented
6195 // in the shorter type. Consequently, we cannot emit a simple comparison.
6197 // First, handle some easy cases. We know the result cannot be equal at this
6198 // point so handle the ICI.isEquality() cases
6199 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6200 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6201 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6202 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6204 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6205 // should have been folded away previously and not enter in here.
6208 // We're performing a signed comparison.
6209 if (cast<ConstantInt>(CI)->getValue().isNegative())
6210 Result = ConstantInt::getFalse(); // X < (small) --> false
6212 Result = ConstantInt::getTrue(); // X < (large) --> true
6214 // We're performing an unsigned comparison.
6216 // We're performing an unsigned comp with a sign extended value.
6217 // This is true if the input is >= 0. [aka >s -1]
6218 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6219 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6220 NegOne, ICI.getName()), ICI);
6222 // Unsigned extend & unsigned compare -> always true.
6223 Result = ConstantInt::getTrue();
6227 // Finally, return the value computed.
6228 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6229 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6230 return ReplaceInstUsesWith(ICI, Result);
6232 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6233 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6234 "ICmp should be folded!");
6235 if (Constant *CI = dyn_cast<Constant>(Result))
6236 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6238 return BinaryOperator::createNot(Result);
6242 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6243 return commonShiftTransforms(I);
6246 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6247 return commonShiftTransforms(I);
6250 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6251 return commonShiftTransforms(I);
6254 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6255 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6256 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6258 // shl X, 0 == X and shr X, 0 == X
6259 // shl 0, X == 0 and shr 0, X == 0
6260 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6261 Op0 == Constant::getNullValue(Op0->getType()))
6262 return ReplaceInstUsesWith(I, Op0);
6264 if (isa<UndefValue>(Op0)) {
6265 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6266 return ReplaceInstUsesWith(I, Op0);
6267 else // undef << X -> 0, undef >>u X -> 0
6268 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6270 if (isa<UndefValue>(Op1)) {
6271 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6272 return ReplaceInstUsesWith(I, Op0);
6273 else // X << undef, X >>u undef -> 0
6274 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6277 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6278 if (I.getOpcode() == Instruction::AShr)
6279 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6280 if (CSI->isAllOnesValue())
6281 return ReplaceInstUsesWith(I, CSI);
6283 // Try to fold constant and into select arguments.
6284 if (isa<Constant>(Op0))
6285 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6286 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6289 // See if we can turn a signed shr into an unsigned shr.
6290 if (I.isArithmeticShift()) {
6291 if (MaskedValueIsZero(Op0,
6292 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
6293 return BinaryOperator::createLShr(Op0, Op1, I.getName());
6297 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6298 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6303 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6304 BinaryOperator &I) {
6305 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6307 // See if we can simplify any instructions used by the instruction whose sole
6308 // purpose is to compute bits we don't care about.
6309 uint64_t KnownZero, KnownOne;
6310 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
6311 KnownZero, KnownOne))
6314 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6315 // of a signed value.
6317 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6318 if (Op1->getZExtValue() >= TypeBits) {
6319 if (I.getOpcode() != Instruction::AShr)
6320 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6322 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6327 // ((X*C1) << C2) == (X * (C1 << C2))
6328 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6329 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6330 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6331 return BinaryOperator::createMul(BO->getOperand(0),
6332 ConstantExpr::getShl(BOOp, Op1));
6334 // Try to fold constant and into select arguments.
6335 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6336 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6338 if (isa<PHINode>(Op0))
6339 if (Instruction *NV = FoldOpIntoPhi(I))
6342 if (Op0->hasOneUse()) {
6343 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6344 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6347 switch (Op0BO->getOpcode()) {
6349 case Instruction::Add:
6350 case Instruction::And:
6351 case Instruction::Or:
6352 case Instruction::Xor: {
6353 // These operators commute.
6354 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6355 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6356 match(Op0BO->getOperand(1),
6357 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6358 Instruction *YS = BinaryOperator::createShl(
6359 Op0BO->getOperand(0), Op1,
6361 InsertNewInstBefore(YS, I); // (Y << C)
6363 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6364 Op0BO->getOperand(1)->getName());
6365 InsertNewInstBefore(X, I); // (X + (Y << C))
6366 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
6367 C2 = ConstantExpr::getShl(C2, Op1);
6368 return BinaryOperator::createAnd(X, C2);
6371 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6372 Value *Op0BOOp1 = Op0BO->getOperand(1);
6373 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6375 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6376 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6378 Instruction *YS = BinaryOperator::createShl(
6379 Op0BO->getOperand(0), Op1,
6381 InsertNewInstBefore(YS, I); // (Y << C)
6383 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6384 V1->getName()+".mask");
6385 InsertNewInstBefore(XM, I); // X & (CC << C)
6387 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6392 case Instruction::Sub: {
6393 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6394 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6395 match(Op0BO->getOperand(0),
6396 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6397 Instruction *YS = BinaryOperator::createShl(
6398 Op0BO->getOperand(1), Op1,
6400 InsertNewInstBefore(YS, I); // (Y << C)
6402 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6403 Op0BO->getOperand(0)->getName());
6404 InsertNewInstBefore(X, I); // (X + (Y << C))
6405 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
6406 C2 = ConstantExpr::getShl(C2, Op1);
6407 return BinaryOperator::createAnd(X, C2);
6410 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6411 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6412 match(Op0BO->getOperand(0),
6413 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6414 m_ConstantInt(CC))) && V2 == Op1 &&
6415 cast<BinaryOperator>(Op0BO->getOperand(0))
6416 ->getOperand(0)->hasOneUse()) {
6417 Instruction *YS = BinaryOperator::createShl(
6418 Op0BO->getOperand(1), Op1,
6420 InsertNewInstBefore(YS, I); // (Y << C)
6422 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6423 V1->getName()+".mask");
6424 InsertNewInstBefore(XM, I); // X & (CC << C)
6426 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6434 // If the operand is an bitwise operator with a constant RHS, and the
6435 // shift is the only use, we can pull it out of the shift.
6436 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6437 bool isValid = true; // Valid only for And, Or, Xor
6438 bool highBitSet = false; // Transform if high bit of constant set?
6440 switch (Op0BO->getOpcode()) {
6441 default: isValid = false; break; // Do not perform transform!
6442 case Instruction::Add:
6443 isValid = isLeftShift;
6445 case Instruction::Or:
6446 case Instruction::Xor:
6449 case Instruction::And:
6454 // If this is a signed shift right, and the high bit is modified
6455 // by the logical operation, do not perform the transformation.
6456 // The highBitSet boolean indicates the value of the high bit of
6457 // the constant which would cause it to be modified for this
6460 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
6461 uint64_t Val = Op0C->getZExtValue();
6462 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
6466 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6468 Instruction *NewShift =
6469 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6470 InsertNewInstBefore(NewShift, I);
6471 NewShift->takeName(Op0BO);
6473 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6480 // Find out if this is a shift of a shift by a constant.
6481 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6482 if (ShiftOp && !ShiftOp->isShift())
6485 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6486 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6487 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
6488 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
6489 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6490 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6491 Value *X = ShiftOp->getOperand(0);
6493 unsigned AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6494 if (AmtSum > I.getType()->getPrimitiveSizeInBits())
6495 AmtSum = I.getType()->getPrimitiveSizeInBits();
6497 const IntegerType *Ty = cast<IntegerType>(I.getType());
6499 // Check for (X << c1) << c2 and (X >> c1) >> c2
6500 if (I.getOpcode() == ShiftOp->getOpcode()) {
6501 return BinaryOperator::create(I.getOpcode(), X,
6502 ConstantInt::get(Ty, AmtSum));
6503 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6504 I.getOpcode() == Instruction::AShr) {
6505 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6506 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6507 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6508 I.getOpcode() == Instruction::LShr) {
6509 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6510 Instruction *Shift =
6511 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6512 InsertNewInstBefore(Shift, I);
6514 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
6515 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6518 // Okay, if we get here, one shift must be left, and the other shift must be
6519 // right. See if the amounts are equal.
6520 if (ShiftAmt1 == ShiftAmt2) {
6521 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6522 if (I.getOpcode() == Instruction::Shl) {
6523 uint64_t Mask = Ty->getBitMask() << ShiftAmt1;
6524 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
6526 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6527 if (I.getOpcode() == Instruction::LShr) {
6528 uint64_t Mask = Ty->getBitMask() >> ShiftAmt1;
6529 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
6531 // We can simplify ((X << C) >>s C) into a trunc + sext.
6532 // NOTE: we could do this for any C, but that would make 'unusual' integer
6533 // types. For now, just stick to ones well-supported by the code
6535 const Type *SExtType = 0;
6536 switch (Ty->getBitWidth() - ShiftAmt1) {
6537 case 8 : SExtType = Type::Int8Ty; break;
6538 case 16: SExtType = Type::Int16Ty; break;
6539 case 32: SExtType = Type::Int32Ty; break;
6543 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6544 InsertNewInstBefore(NewTrunc, I);
6545 return new SExtInst(NewTrunc, Ty);
6547 // Otherwise, we can't handle it yet.
6548 } else if (ShiftAmt1 < ShiftAmt2) {
6549 unsigned ShiftDiff = ShiftAmt2-ShiftAmt1;
6551 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6552 if (I.getOpcode() == Instruction::Shl) {
6553 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6554 ShiftOp->getOpcode() == Instruction::AShr);
6555 Instruction *Shift =
6556 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6557 InsertNewInstBefore(Shift, I);
6559 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
6560 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6563 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6564 if (I.getOpcode() == Instruction::LShr) {
6565 assert(ShiftOp->getOpcode() == Instruction::Shl);
6566 Instruction *Shift =
6567 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6568 InsertNewInstBefore(Shift, I);
6570 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
6571 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6574 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6576 assert(ShiftAmt2 < ShiftAmt1);
6577 unsigned ShiftDiff = ShiftAmt1-ShiftAmt2;
6579 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6580 if (I.getOpcode() == Instruction::Shl) {
6581 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6582 ShiftOp->getOpcode() == Instruction::AShr);
6583 Instruction *Shift =
6584 BinaryOperator::create(ShiftOp->getOpcode(), X,
6585 ConstantInt::get(Ty, ShiftDiff));
6586 InsertNewInstBefore(Shift, I);
6588 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
6589 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6592 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6593 if (I.getOpcode() == Instruction::LShr) {
6594 assert(ShiftOp->getOpcode() == Instruction::Shl);
6595 Instruction *Shift =
6596 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6597 InsertNewInstBefore(Shift, I);
6599 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
6600 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6603 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6610 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6611 /// expression. If so, decompose it, returning some value X, such that Val is
6614 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6616 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6617 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6618 Offset = CI->getZExtValue();
6620 return ConstantInt::get(Type::Int32Ty, 0);
6621 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
6622 if (I->getNumOperands() == 2) {
6623 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6624 if (I->getOpcode() == Instruction::Shl) {
6625 // This is a value scaled by '1 << the shift amt'.
6626 Scale = 1U << CUI->getZExtValue();
6628 return I->getOperand(0);
6629 } else if (I->getOpcode() == Instruction::Mul) {
6630 // This value is scaled by 'CUI'.
6631 Scale = CUI->getZExtValue();
6633 return I->getOperand(0);
6634 } else if (I->getOpcode() == Instruction::Add) {
6635 // We have X+C. Check to see if we really have (X*C2)+C1,
6636 // where C1 is divisible by C2.
6639 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6640 Offset += CUI->getZExtValue();
6641 if (SubScale > 1 && (Offset % SubScale == 0)) {
6650 // Otherwise, we can't look past this.
6657 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6658 /// try to eliminate the cast by moving the type information into the alloc.
6659 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
6660 AllocationInst &AI) {
6661 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
6662 if (!PTy) return 0; // Not casting the allocation to a pointer type.
6664 // Remove any uses of AI that are dead.
6665 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6667 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6668 Instruction *User = cast<Instruction>(*UI++);
6669 if (isInstructionTriviallyDead(User)) {
6670 while (UI != E && *UI == User)
6671 ++UI; // If this instruction uses AI more than once, don't break UI.
6674 DOUT << "IC: DCE: " << *User;
6675 EraseInstFromFunction(*User);
6679 // Get the type really allocated and the type casted to.
6680 const Type *AllocElTy = AI.getAllocatedType();
6681 const Type *CastElTy = PTy->getElementType();
6682 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6684 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6685 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6686 if (CastElTyAlign < AllocElTyAlign) return 0;
6688 // If the allocation has multiple uses, only promote it if we are strictly
6689 // increasing the alignment of the resultant allocation. If we keep it the
6690 // same, we open the door to infinite loops of various kinds.
6691 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6693 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
6694 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
6695 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6697 // See if we can satisfy the modulus by pulling a scale out of the array
6699 unsigned ArraySizeScale, ArrayOffset;
6700 Value *NumElements = // See if the array size is a decomposable linear expr.
6701 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6703 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6705 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6706 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6708 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6713 // If the allocation size is constant, form a constant mul expression
6714 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6715 if (isa<ConstantInt>(NumElements))
6716 Amt = ConstantExpr::getMul(
6717 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6718 // otherwise multiply the amount and the number of elements
6719 else if (Scale != 1) {
6720 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6721 Amt = InsertNewInstBefore(Tmp, AI);
6725 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6726 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
6727 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6728 Amt = InsertNewInstBefore(Tmp, AI);
6731 AllocationInst *New;
6732 if (isa<MallocInst>(AI))
6733 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6735 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6736 InsertNewInstBefore(New, AI);
6739 // If the allocation has multiple uses, insert a cast and change all things
6740 // that used it to use the new cast. This will also hack on CI, but it will
6742 if (!AI.hasOneUse()) {
6743 AddUsesToWorkList(AI);
6744 // New is the allocation instruction, pointer typed. AI is the original
6745 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6746 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6747 InsertNewInstBefore(NewCast, AI);
6748 AI.replaceAllUsesWith(NewCast);
6750 return ReplaceInstUsesWith(CI, New);
6753 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6754 /// and return it as type Ty without inserting any new casts and without
6755 /// changing the computed value. This is used by code that tries to decide
6756 /// whether promoting or shrinking integer operations to wider or smaller types
6757 /// will allow us to eliminate a truncate or extend.
6759 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6760 /// extension operation if Ty is larger.
6761 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6762 int &NumCastsRemoved) {
6763 // We can always evaluate constants in another type.
6764 if (isa<ConstantInt>(V))
6767 Instruction *I = dyn_cast<Instruction>(V);
6768 if (!I) return false;
6770 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6772 switch (I->getOpcode()) {
6773 case Instruction::Add:
6774 case Instruction::Sub:
6775 case Instruction::And:
6776 case Instruction::Or:
6777 case Instruction::Xor:
6778 if (!I->hasOneUse()) return false;
6779 // These operators can all arbitrarily be extended or truncated.
6780 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
6781 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
6783 case Instruction::Shl:
6784 if (!I->hasOneUse()) return false;
6785 // If we are truncating the result of this SHL, and if it's a shift of a
6786 // constant amount, we can always perform a SHL in a smaller type.
6787 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6788 if (Ty->getBitWidth() < OrigTy->getBitWidth() &&
6789 CI->getZExtValue() < Ty->getBitWidth())
6790 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6793 case Instruction::LShr:
6794 if (!I->hasOneUse()) return false;
6795 // If this is a truncate of a logical shr, we can truncate it to a smaller
6796 // lshr iff we know that the bits we would otherwise be shifting in are
6798 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6799 uint32_t BitWidth = OrigTy->getBitWidth();
6800 if (Ty->getBitWidth() < OrigTy->getBitWidth() &&
6801 MaskedValueIsZero(I->getOperand(0),
6802 APInt::getAllOnesValue(BitWidth) &
6803 APInt::getAllOnesValue(Ty->getBitWidth()).zextOrTrunc(BitWidth).flip())
6804 && CI->getZExtValue() < Ty->getBitWidth()) {
6805 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
6809 case Instruction::Trunc:
6810 case Instruction::ZExt:
6811 case Instruction::SExt:
6812 // If this is a cast from the destination type, we can trivially eliminate
6813 // it, and this will remove a cast overall.
6814 if (I->getOperand(0)->getType() == Ty) {
6815 // If the first operand is itself a cast, and is eliminable, do not count
6816 // this as an eliminable cast. We would prefer to eliminate those two
6818 if (isa<CastInst>(I->getOperand(0)))
6826 // TODO: Can handle more cases here.
6833 /// EvaluateInDifferentType - Given an expression that
6834 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6835 /// evaluate the expression.
6836 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6838 if (Constant *C = dyn_cast<Constant>(V))
6839 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6841 // Otherwise, it must be an instruction.
6842 Instruction *I = cast<Instruction>(V);
6843 Instruction *Res = 0;
6844 switch (I->getOpcode()) {
6845 case Instruction::Add:
6846 case Instruction::Sub:
6847 case Instruction::And:
6848 case Instruction::Or:
6849 case Instruction::Xor:
6850 case Instruction::AShr:
6851 case Instruction::LShr:
6852 case Instruction::Shl: {
6853 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6854 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6855 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6856 LHS, RHS, I->getName());
6859 case Instruction::Trunc:
6860 case Instruction::ZExt:
6861 case Instruction::SExt:
6862 case Instruction::BitCast:
6863 // If the source type of the cast is the type we're trying for then we can
6864 // just return the source. There's no need to insert it because its not new.
6865 if (I->getOperand(0)->getType() == Ty)
6866 return I->getOperand(0);
6868 // Some other kind of cast, which shouldn't happen, so just ..
6871 // TODO: Can handle more cases here.
6872 assert(0 && "Unreachable!");
6876 return InsertNewInstBefore(Res, *I);
6879 /// @brief Implement the transforms common to all CastInst visitors.
6880 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6881 Value *Src = CI.getOperand(0);
6883 // Casting undef to anything results in undef so might as just replace it and
6884 // get rid of the cast.
6885 if (isa<UndefValue>(Src)) // cast undef -> undef
6886 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
6888 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
6889 // eliminate it now.
6890 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6891 if (Instruction::CastOps opc =
6892 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6893 // The first cast (CSrc) is eliminable so we need to fix up or replace
6894 // the second cast (CI). CSrc will then have a good chance of being dead.
6895 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6899 // If casting the result of a getelementptr instruction with no offset, turn
6900 // this into a cast of the original pointer!
6902 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6903 bool AllZeroOperands = true;
6904 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
6905 if (!isa<Constant>(GEP->getOperand(i)) ||
6906 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
6907 AllZeroOperands = false;
6910 if (AllZeroOperands) {
6911 // Changing the cast operand is usually not a good idea but it is safe
6912 // here because the pointer operand is being replaced with another
6913 // pointer operand so the opcode doesn't need to change.
6914 CI.setOperand(0, GEP->getOperand(0));
6919 // If we are casting a malloc or alloca to a pointer to a type of the same
6920 // size, rewrite the allocation instruction to allocate the "right" type.
6921 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6922 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6925 // If we are casting a select then fold the cast into the select
6926 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6927 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6930 // If we are casting a PHI then fold the cast into the PHI
6931 if (isa<PHINode>(Src))
6932 if (Instruction *NV = FoldOpIntoPhi(CI))
6938 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6939 /// integer types. This function implements the common transforms for all those
6941 /// @brief Implement the transforms common to CastInst with integer operands
6942 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6943 if (Instruction *Result = commonCastTransforms(CI))
6946 Value *Src = CI.getOperand(0);
6947 const Type *SrcTy = Src->getType();
6948 const Type *DestTy = CI.getType();
6949 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6950 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6952 // See if we can simplify any instructions used by the LHS whose sole
6953 // purpose is to compute bits we don't care about.
6954 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6955 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6956 KnownZero, KnownOne))
6959 // If the source isn't an instruction or has more than one use then we
6960 // can't do anything more.
6961 Instruction *SrcI = dyn_cast<Instruction>(Src);
6962 if (!SrcI || !Src->hasOneUse())
6965 // Attempt to propagate the cast into the instruction for int->int casts.
6966 int NumCastsRemoved = 0;
6967 if (!isa<BitCastInst>(CI) &&
6968 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6970 // If this cast is a truncate, evaluting in a different type always
6971 // eliminates the cast, so it is always a win. If this is a noop-cast
6972 // this just removes a noop cast which isn't pointful, but simplifies
6973 // the code. If this is a zero-extension, we need to do an AND to
6974 // maintain the clear top-part of the computation, so we require that
6975 // the input have eliminated at least one cast. If this is a sign
6976 // extension, we insert two new casts (to do the extension) so we
6977 // require that two casts have been eliminated.
6979 switch (CI.getOpcode()) {
6981 // All the others use floating point so we shouldn't actually
6982 // get here because of the check above.
6983 assert(0 && "Unknown cast type");
6984 case Instruction::Trunc:
6987 case Instruction::ZExt:
6988 DoXForm = NumCastsRemoved >= 1;
6990 case Instruction::SExt:
6991 DoXForm = NumCastsRemoved >= 2;
6993 case Instruction::BitCast:
6999 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7000 CI.getOpcode() == Instruction::SExt);
7001 assert(Res->getType() == DestTy);
7002 switch (CI.getOpcode()) {
7003 default: assert(0 && "Unknown cast type!");
7004 case Instruction::Trunc:
7005 case Instruction::BitCast:
7006 // Just replace this cast with the result.
7007 return ReplaceInstUsesWith(CI, Res);
7008 case Instruction::ZExt: {
7009 // We need to emit an AND to clear the high bits.
7010 assert(SrcBitSize < DestBitSize && "Not a zext?");
7011 Constant *C = ConstantInt::get(APInt::getAllOnesValue(SrcBitSize));
7012 C = ConstantExpr::getZExt(C, DestTy);
7013 return BinaryOperator::createAnd(Res, C);
7015 case Instruction::SExt:
7016 // We need to emit a cast to truncate, then a cast to sext.
7017 return CastInst::create(Instruction::SExt,
7018 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7024 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7025 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7027 switch (SrcI->getOpcode()) {
7028 case Instruction::Add:
7029 case Instruction::Mul:
7030 case Instruction::And:
7031 case Instruction::Or:
7032 case Instruction::Xor:
7033 // If we are discarding information, or just changing the sign,
7035 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7036 // Don't insert two casts if they cannot be eliminated. We allow
7037 // two casts to be inserted if the sizes are the same. This could
7038 // only be converting signedness, which is a noop.
7039 if (DestBitSize == SrcBitSize ||
7040 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7041 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7042 Instruction::CastOps opcode = CI.getOpcode();
7043 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7044 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7045 return BinaryOperator::create(
7046 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7050 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7051 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7052 SrcI->getOpcode() == Instruction::Xor &&
7053 Op1 == ConstantInt::getTrue() &&
7054 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7055 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7056 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
7059 case Instruction::SDiv:
7060 case Instruction::UDiv:
7061 case Instruction::SRem:
7062 case Instruction::URem:
7063 // If we are just changing the sign, rewrite.
7064 if (DestBitSize == SrcBitSize) {
7065 // Don't insert two casts if they cannot be eliminated. We allow
7066 // two casts to be inserted if the sizes are the same. This could
7067 // only be converting signedness, which is a noop.
7068 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7069 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7070 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7072 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7074 return BinaryOperator::create(
7075 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7080 case Instruction::Shl:
7081 // Allow changing the sign of the source operand. Do not allow
7082 // changing the size of the shift, UNLESS the shift amount is a
7083 // constant. We must not change variable sized shifts to a smaller
7084 // size, because it is undefined to shift more bits out than exist
7086 if (DestBitSize == SrcBitSize ||
7087 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7088 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7089 Instruction::BitCast : Instruction::Trunc);
7090 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7091 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7092 return BinaryOperator::createShl(Op0c, Op1c);
7095 case Instruction::AShr:
7096 // If this is a signed shr, and if all bits shifted in are about to be
7097 // truncated off, turn it into an unsigned shr to allow greater
7099 if (DestBitSize < SrcBitSize &&
7100 isa<ConstantInt>(Op1)) {
7101 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
7102 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7103 // Insert the new logical shift right.
7104 return BinaryOperator::createLShr(Op0, Op1);
7109 case Instruction::ICmp:
7110 // If we are just checking for a icmp eq of a single bit and casting it
7111 // to an integer, then shift the bit to the appropriate place and then
7112 // cast to integer to avoid the comparison.
7113 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
7114 APInt Op1CV(Op1C->getValue());
7115 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
7116 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
7117 // cast (X == 1) to int --> X iff X has only the low bit set.
7118 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
7119 // cast (X != 0) to int --> X iff X has only the low bit set.
7120 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
7121 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
7122 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
7123 if (Op1CV == 0 || Op1CV.isPowerOf2()) {
7124 // If Op1C some other power of two, convert:
7125 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7126 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7127 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7128 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
7130 // This only works for EQ and NE
7131 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
7132 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
7135 if ((KnownZero^TypeMask).isPowerOf2()) { // Exactly 1 possible 1?
7136 bool isNE = pred == ICmpInst::ICMP_NE;
7137 if (Op1CV != 0 && (Op1CV != (KnownZero^TypeMask))) {
7138 // (X&4) == 2 --> false
7139 // (X&4) != 2 --> true
7140 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7141 Res = ConstantExpr::getZExt(Res, CI.getType());
7142 return ReplaceInstUsesWith(CI, Res);
7145 unsigned ShiftAmt = (KnownZero^TypeMask).logBase2();
7148 // Perform a logical shr by shiftamt.
7149 // Insert the shift to put the result in the low bit.
7150 In = InsertNewInstBefore(
7151 BinaryOperator::createLShr(In,
7152 ConstantInt::get(In->getType(), ShiftAmt),
7153 In->getName()+".lobit"), CI);
7156 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7157 Constant *One = ConstantInt::get(In->getType(), 1);
7158 In = BinaryOperator::createXor(In, One, "tmp");
7159 InsertNewInstBefore(cast<Instruction>(In), CI);
7162 if (CI.getType() == In->getType())
7163 return ReplaceInstUsesWith(CI, In);
7165 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7174 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
7175 if (Instruction *Result = commonIntCastTransforms(CI))
7178 Value *Src = CI.getOperand(0);
7179 const Type *Ty = CI.getType();
7180 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
7181 unsigned SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7183 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7184 switch (SrcI->getOpcode()) {
7186 case Instruction::LShr:
7187 // We can shrink lshr to something smaller if we know the bits shifted in
7188 // are already zeros.
7189 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7190 unsigned ShAmt = ShAmtV->getZExtValue();
7192 // Get a mask for the bits shifting in.
7193 APInt Mask(APInt::getAllOnesValue(SrcBitWidth).lshr(
7194 SrcBitWidth-ShAmt).shl(DestBitWidth));
7195 Value* SrcIOp0 = SrcI->getOperand(0);
7196 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7197 if (ShAmt >= DestBitWidth) // All zeros.
7198 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7200 // Okay, we can shrink this. Truncate the input, then return a new
7202 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7203 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7205 return BinaryOperator::createLShr(V1, V2);
7207 } else { // This is a variable shr.
7209 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7210 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7211 // loop-invariant and CSE'd.
7212 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7213 Value *One = ConstantInt::get(SrcI->getType(), 1);
7215 Value *V = InsertNewInstBefore(
7216 BinaryOperator::createShl(One, SrcI->getOperand(1),
7218 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7219 SrcI->getOperand(0),
7221 Value *Zero = Constant::getNullValue(V->getType());
7222 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7232 Instruction *InstCombiner::visitZExt(CastInst &CI) {
7233 // If one of the common conversion will work ..
7234 if (Instruction *Result = commonIntCastTransforms(CI))
7237 Value *Src = CI.getOperand(0);
7239 // If this is a cast of a cast
7240 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7241 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7242 // types and if the sizes are just right we can convert this into a logical
7243 // 'and' which will be much cheaper than the pair of casts.
7244 if (isa<TruncInst>(CSrc)) {
7245 // Get the sizes of the types involved
7246 Value *A = CSrc->getOperand(0);
7247 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
7248 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7249 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7250 // If we're actually extending zero bits and the trunc is a no-op
7251 if (MidSize < DstSize && SrcSize == DstSize) {
7252 // Replace both of the casts with an And of the type mask.
7253 APInt AndValue(APInt::getAllOnesValue(MidSize).zext(SrcSize));
7254 Constant *AndConst = ConstantInt::get(AndValue);
7256 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7257 // Unfortunately, if the type changed, we need to cast it back.
7258 if (And->getType() != CI.getType()) {
7259 And->setName(CSrc->getName()+".mask");
7260 InsertNewInstBefore(And, CI);
7261 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7271 Instruction *InstCombiner::visitSExt(CastInst &CI) {
7272 return commonIntCastTransforms(CI);
7275 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
7276 return commonCastTransforms(CI);
7279 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7280 return commonCastTransforms(CI);
7283 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7284 return commonCastTransforms(CI);
7287 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7288 return commonCastTransforms(CI);
7291 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7292 return commonCastTransforms(CI);
7295 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7296 return commonCastTransforms(CI);
7299 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7300 return commonCastTransforms(CI);
7303 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
7304 return commonCastTransforms(CI);
7307 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
7309 // If the operands are integer typed then apply the integer transforms,
7310 // otherwise just apply the common ones.
7311 Value *Src = CI.getOperand(0);
7312 const Type *SrcTy = Src->getType();
7313 const Type *DestTy = CI.getType();
7315 if (SrcTy->isInteger() && DestTy->isInteger()) {
7316 if (Instruction *Result = commonIntCastTransforms(CI))
7319 if (Instruction *Result = commonCastTransforms(CI))
7324 // Get rid of casts from one type to the same type. These are useless and can
7325 // be replaced by the operand.
7326 if (DestTy == Src->getType())
7327 return ReplaceInstUsesWith(CI, Src);
7329 // If the source and destination are pointers, and this cast is equivalent to
7330 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
7331 // This can enhance SROA and other transforms that want type-safe pointers.
7332 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7333 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
7334 const Type *DstElTy = DstPTy->getElementType();
7335 const Type *SrcElTy = SrcPTy->getElementType();
7337 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7338 unsigned NumZeros = 0;
7339 while (SrcElTy != DstElTy &&
7340 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7341 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7342 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7346 // If we found a path from the src to dest, create the getelementptr now.
7347 if (SrcElTy == DstElTy) {
7348 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7349 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
7354 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7355 if (SVI->hasOneUse()) {
7356 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7357 // a bitconvert to a vector with the same # elts.
7358 if (isa<VectorType>(DestTy) &&
7359 cast<VectorType>(DestTy)->getNumElements() ==
7360 SVI->getType()->getNumElements()) {
7362 // If either of the operands is a cast from CI.getType(), then
7363 // evaluating the shuffle in the casted destination's type will allow
7364 // us to eliminate at least one cast.
7365 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7366 Tmp->getOperand(0)->getType() == DestTy) ||
7367 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7368 Tmp->getOperand(0)->getType() == DestTy)) {
7369 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7370 SVI->getOperand(0), DestTy, &CI);
7371 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7372 SVI->getOperand(1), DestTy, &CI);
7373 // Return a new shuffle vector. Use the same element ID's, as we
7374 // know the vector types match #elts.
7375 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7383 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7385 /// %D = select %cond, %C, %A
7387 /// %C = select %cond, %B, 0
7390 /// Assuming that the specified instruction is an operand to the select, return
7391 /// a bitmask indicating which operands of this instruction are foldable if they
7392 /// equal the other incoming value of the select.
7394 static unsigned GetSelectFoldableOperands(Instruction *I) {
7395 switch (I->getOpcode()) {
7396 case Instruction::Add:
7397 case Instruction::Mul:
7398 case Instruction::And:
7399 case Instruction::Or:
7400 case Instruction::Xor:
7401 return 3; // Can fold through either operand.
7402 case Instruction::Sub: // Can only fold on the amount subtracted.
7403 case Instruction::Shl: // Can only fold on the shift amount.
7404 case Instruction::LShr:
7405 case Instruction::AShr:
7408 return 0; // Cannot fold
7412 /// GetSelectFoldableConstant - For the same transformation as the previous
7413 /// function, return the identity constant that goes into the select.
7414 static Constant *GetSelectFoldableConstant(Instruction *I) {
7415 switch (I->getOpcode()) {
7416 default: assert(0 && "This cannot happen!"); abort();
7417 case Instruction::Add:
7418 case Instruction::Sub:
7419 case Instruction::Or:
7420 case Instruction::Xor:
7421 case Instruction::Shl:
7422 case Instruction::LShr:
7423 case Instruction::AShr:
7424 return Constant::getNullValue(I->getType());
7425 case Instruction::And:
7426 return ConstantInt::getAllOnesValue(I->getType());
7427 case Instruction::Mul:
7428 return ConstantInt::get(I->getType(), 1);
7432 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7433 /// have the same opcode and only one use each. Try to simplify this.
7434 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7436 if (TI->getNumOperands() == 1) {
7437 // If this is a non-volatile load or a cast from the same type,
7440 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7443 return 0; // unknown unary op.
7446 // Fold this by inserting a select from the input values.
7447 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7448 FI->getOperand(0), SI.getName()+".v");
7449 InsertNewInstBefore(NewSI, SI);
7450 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7454 // Only handle binary operators here.
7455 if (!isa<BinaryOperator>(TI))
7458 // Figure out if the operations have any operands in common.
7459 Value *MatchOp, *OtherOpT, *OtherOpF;
7461 if (TI->getOperand(0) == FI->getOperand(0)) {
7462 MatchOp = TI->getOperand(0);
7463 OtherOpT = TI->getOperand(1);
7464 OtherOpF = FI->getOperand(1);
7465 MatchIsOpZero = true;
7466 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7467 MatchOp = TI->getOperand(1);
7468 OtherOpT = TI->getOperand(0);
7469 OtherOpF = FI->getOperand(0);
7470 MatchIsOpZero = false;
7471 } else if (!TI->isCommutative()) {
7473 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7474 MatchOp = TI->getOperand(0);
7475 OtherOpT = TI->getOperand(1);
7476 OtherOpF = FI->getOperand(0);
7477 MatchIsOpZero = true;
7478 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7479 MatchOp = TI->getOperand(1);
7480 OtherOpT = TI->getOperand(0);
7481 OtherOpF = FI->getOperand(1);
7482 MatchIsOpZero = true;
7487 // If we reach here, they do have operations in common.
7488 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7489 OtherOpF, SI.getName()+".v");
7490 InsertNewInstBefore(NewSI, SI);
7492 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7494 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7496 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7498 assert(0 && "Shouldn't get here");
7502 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7503 Value *CondVal = SI.getCondition();
7504 Value *TrueVal = SI.getTrueValue();
7505 Value *FalseVal = SI.getFalseValue();
7507 // select true, X, Y -> X
7508 // select false, X, Y -> Y
7509 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7510 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7512 // select C, X, X -> X
7513 if (TrueVal == FalseVal)
7514 return ReplaceInstUsesWith(SI, TrueVal);
7516 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7517 return ReplaceInstUsesWith(SI, FalseVal);
7518 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7519 return ReplaceInstUsesWith(SI, TrueVal);
7520 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7521 if (isa<Constant>(TrueVal))
7522 return ReplaceInstUsesWith(SI, TrueVal);
7524 return ReplaceInstUsesWith(SI, FalseVal);
7527 if (SI.getType() == Type::Int1Ty) {
7528 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7529 if (C->getZExtValue()) {
7530 // Change: A = select B, true, C --> A = or B, C
7531 return BinaryOperator::createOr(CondVal, FalseVal);
7533 // Change: A = select B, false, C --> A = and !B, C
7535 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7536 "not."+CondVal->getName()), SI);
7537 return BinaryOperator::createAnd(NotCond, FalseVal);
7539 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7540 if (C->getZExtValue() == false) {
7541 // Change: A = select B, C, false --> A = and B, C
7542 return BinaryOperator::createAnd(CondVal, TrueVal);
7544 // Change: A = select B, C, true --> A = or !B, C
7546 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7547 "not."+CondVal->getName()), SI);
7548 return BinaryOperator::createOr(NotCond, TrueVal);
7553 // Selecting between two integer constants?
7554 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7555 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7556 // select C, 1, 0 -> cast C to int
7557 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
7558 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7559 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
7560 // select C, 0, 1 -> cast !C to int
7562 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7563 "not."+CondVal->getName()), SI);
7564 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7567 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7569 // (x <s 0) ? -1 : 0 -> ashr x, 31
7570 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
7571 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
7572 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7573 bool CanXForm = false;
7574 if (IC->isSignedPredicate())
7575 CanXForm = CmpCst->isNullValue() &&
7576 IC->getPredicate() == ICmpInst::ICMP_SLT;
7578 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
7579 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
7580 IC->getPredicate() == ICmpInst::ICMP_UGT;
7584 // The comparison constant and the result are not neccessarily the
7585 // same width. Make an all-ones value by inserting a AShr.
7586 Value *X = IC->getOperand(0);
7587 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
7588 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7589 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7591 InsertNewInstBefore(SRA, SI);
7593 // Finally, convert to the type of the select RHS. We figure out
7594 // if this requires a SExt, Trunc or BitCast based on the sizes.
7595 Instruction::CastOps opc = Instruction::BitCast;
7596 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
7597 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
7598 if (SRASize < SISize)
7599 opc = Instruction::SExt;
7600 else if (SRASize > SISize)
7601 opc = Instruction::Trunc;
7602 return CastInst::create(opc, SRA, SI.getType());
7607 // If one of the constants is zero (we know they can't both be) and we
7608 // have a fcmp instruction with zero, and we have an 'and' with the
7609 // non-constant value, eliminate this whole mess. This corresponds to
7610 // cases like this: ((X & 27) ? 27 : 0)
7611 if (TrueValC->isNullValue() || FalseValC->isNullValue())
7612 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7613 cast<Constant>(IC->getOperand(1))->isNullValue())
7614 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7615 if (ICA->getOpcode() == Instruction::And &&
7616 isa<ConstantInt>(ICA->getOperand(1)) &&
7617 (ICA->getOperand(1) == TrueValC ||
7618 ICA->getOperand(1) == FalseValC) &&
7619 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7620 // Okay, now we know that everything is set up, we just don't
7621 // know whether we have a icmp_ne or icmp_eq and whether the
7622 // true or false val is the zero.
7623 bool ShouldNotVal = !TrueValC->isNullValue();
7624 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7627 V = InsertNewInstBefore(BinaryOperator::create(
7628 Instruction::Xor, V, ICA->getOperand(1)), SI);
7629 return ReplaceInstUsesWith(SI, V);
7634 // See if we are selecting two values based on a comparison of the two values.
7635 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7636 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7637 // Transform (X == Y) ? X : Y -> Y
7638 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7639 return ReplaceInstUsesWith(SI, FalseVal);
7640 // Transform (X != Y) ? X : Y -> X
7641 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7642 return ReplaceInstUsesWith(SI, TrueVal);
7643 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7645 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7646 // Transform (X == Y) ? Y : X -> X
7647 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7648 return ReplaceInstUsesWith(SI, FalseVal);
7649 // Transform (X != Y) ? Y : X -> Y
7650 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7651 return ReplaceInstUsesWith(SI, TrueVal);
7652 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7656 // See if we are selecting two values based on a comparison of the two values.
7657 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7658 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7659 // Transform (X == Y) ? X : Y -> Y
7660 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7661 return ReplaceInstUsesWith(SI, FalseVal);
7662 // Transform (X != Y) ? X : Y -> X
7663 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7664 return ReplaceInstUsesWith(SI, TrueVal);
7665 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7667 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7668 // Transform (X == Y) ? Y : X -> X
7669 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7670 return ReplaceInstUsesWith(SI, FalseVal);
7671 // Transform (X != Y) ? Y : X -> Y
7672 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7673 return ReplaceInstUsesWith(SI, TrueVal);
7674 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7678 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7679 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7680 if (TI->hasOneUse() && FI->hasOneUse()) {
7681 Instruction *AddOp = 0, *SubOp = 0;
7683 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7684 if (TI->getOpcode() == FI->getOpcode())
7685 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7688 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7689 // even legal for FP.
7690 if (TI->getOpcode() == Instruction::Sub &&
7691 FI->getOpcode() == Instruction::Add) {
7692 AddOp = FI; SubOp = TI;
7693 } else if (FI->getOpcode() == Instruction::Sub &&
7694 TI->getOpcode() == Instruction::Add) {
7695 AddOp = TI; SubOp = FI;
7699 Value *OtherAddOp = 0;
7700 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7701 OtherAddOp = AddOp->getOperand(1);
7702 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7703 OtherAddOp = AddOp->getOperand(0);
7707 // So at this point we know we have (Y -> OtherAddOp):
7708 // select C, (add X, Y), (sub X, Z)
7709 Value *NegVal; // Compute -Z
7710 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7711 NegVal = ConstantExpr::getNeg(C);
7713 NegVal = InsertNewInstBefore(
7714 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7717 Value *NewTrueOp = OtherAddOp;
7718 Value *NewFalseOp = NegVal;
7720 std::swap(NewTrueOp, NewFalseOp);
7721 Instruction *NewSel =
7722 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7724 NewSel = InsertNewInstBefore(NewSel, SI);
7725 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7730 // See if we can fold the select into one of our operands.
7731 if (SI.getType()->isInteger()) {
7732 // See the comment above GetSelectFoldableOperands for a description of the
7733 // transformation we are doing here.
7734 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7735 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7736 !isa<Constant>(FalseVal))
7737 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7738 unsigned OpToFold = 0;
7739 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7741 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7746 Constant *C = GetSelectFoldableConstant(TVI);
7747 Instruction *NewSel =
7748 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7749 InsertNewInstBefore(NewSel, SI);
7750 NewSel->takeName(TVI);
7751 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7752 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7754 assert(0 && "Unknown instruction!!");
7759 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7760 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7761 !isa<Constant>(TrueVal))
7762 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7763 unsigned OpToFold = 0;
7764 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7766 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7771 Constant *C = GetSelectFoldableConstant(FVI);
7772 Instruction *NewSel =
7773 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7774 InsertNewInstBefore(NewSel, SI);
7775 NewSel->takeName(FVI);
7776 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7777 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7779 assert(0 && "Unknown instruction!!");
7784 if (BinaryOperator::isNot(CondVal)) {
7785 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7786 SI.setOperand(1, FalseVal);
7787 SI.setOperand(2, TrueVal);
7794 /// GetKnownAlignment - If the specified pointer has an alignment that we can
7795 /// determine, return it, otherwise return 0.
7796 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
7797 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7798 unsigned Align = GV->getAlignment();
7799 if (Align == 0 && TD)
7800 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7802 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7803 unsigned Align = AI->getAlignment();
7804 if (Align == 0 && TD) {
7805 if (isa<AllocaInst>(AI))
7806 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7807 else if (isa<MallocInst>(AI)) {
7808 // Malloc returns maximally aligned memory.
7809 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7812 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7815 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7819 } else if (isa<BitCastInst>(V) ||
7820 (isa<ConstantExpr>(V) &&
7821 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7822 User *CI = cast<User>(V);
7823 if (isa<PointerType>(CI->getOperand(0)->getType()))
7824 return GetKnownAlignment(CI->getOperand(0), TD);
7826 } else if (isa<GetElementPtrInst>(V) ||
7827 (isa<ConstantExpr>(V) &&
7828 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
7829 User *GEPI = cast<User>(V);
7830 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
7831 if (BaseAlignment == 0) return 0;
7833 // If all indexes are zero, it is just the alignment of the base pointer.
7834 bool AllZeroOperands = true;
7835 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7836 if (!isa<Constant>(GEPI->getOperand(i)) ||
7837 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7838 AllZeroOperands = false;
7841 if (AllZeroOperands)
7842 return BaseAlignment;
7844 // Otherwise, if the base alignment is >= the alignment we expect for the
7845 // base pointer type, then we know that the resultant pointer is aligned at
7846 // least as much as its type requires.
7849 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7850 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7851 if (TD->getABITypeAlignment(PtrTy->getElementType())
7853 const Type *GEPTy = GEPI->getType();
7854 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7855 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
7863 /// visitCallInst - CallInst simplification. This mostly only handles folding
7864 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7865 /// the heavy lifting.
7867 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7868 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7869 if (!II) return visitCallSite(&CI);
7871 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7873 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7874 bool Changed = false;
7876 // memmove/cpy/set of zero bytes is a noop.
7877 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7878 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7880 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7881 if (CI->getZExtValue() == 1) {
7882 // Replace the instruction with just byte operations. We would
7883 // transform other cases to loads/stores, but we don't know if
7884 // alignment is sufficient.
7888 // If we have a memmove and the source operation is a constant global,
7889 // then the source and dest pointers can't alias, so we can change this
7890 // into a call to memcpy.
7891 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7892 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7893 if (GVSrc->isConstant()) {
7894 Module *M = CI.getParent()->getParent()->getParent();
7896 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7898 Name = "llvm.memcpy.i32";
7900 Name = "llvm.memcpy.i64";
7901 Constant *MemCpy = M->getOrInsertFunction(Name,
7902 CI.getCalledFunction()->getFunctionType());
7903 CI.setOperand(0, MemCpy);
7908 // If we can determine a pointer alignment that is bigger than currently
7909 // set, update the alignment.
7910 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7911 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7912 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7913 unsigned Align = std::min(Alignment1, Alignment2);
7914 if (MI->getAlignment()->getZExtValue() < Align) {
7915 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7918 } else if (isa<MemSetInst>(MI)) {
7919 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7920 if (MI->getAlignment()->getZExtValue() < Alignment) {
7921 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7926 if (Changed) return II;
7928 switch (II->getIntrinsicID()) {
7930 case Intrinsic::ppc_altivec_lvx:
7931 case Intrinsic::ppc_altivec_lvxl:
7932 case Intrinsic::x86_sse_loadu_ps:
7933 case Intrinsic::x86_sse2_loadu_pd:
7934 case Intrinsic::x86_sse2_loadu_dq:
7935 // Turn PPC lvx -> load if the pointer is known aligned.
7936 // Turn X86 loadups -> load if the pointer is known aligned.
7937 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7938 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7939 PointerType::get(II->getType()), CI);
7940 return new LoadInst(Ptr);
7943 case Intrinsic::ppc_altivec_stvx:
7944 case Intrinsic::ppc_altivec_stvxl:
7945 // Turn stvx -> store if the pointer is known aligned.
7946 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7947 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7948 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7950 return new StoreInst(II->getOperand(1), Ptr);
7953 case Intrinsic::x86_sse_storeu_ps:
7954 case Intrinsic::x86_sse2_storeu_pd:
7955 case Intrinsic::x86_sse2_storeu_dq:
7956 case Intrinsic::x86_sse2_storel_dq:
7957 // Turn X86 storeu -> store if the pointer is known aligned.
7958 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7959 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7960 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7962 return new StoreInst(II->getOperand(2), Ptr);
7966 case Intrinsic::x86_sse_cvttss2si: {
7967 // These intrinsics only demands the 0th element of its input vector. If
7968 // we can simplify the input based on that, do so now.
7970 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7972 II->setOperand(1, V);
7978 case Intrinsic::ppc_altivec_vperm:
7979 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7980 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7981 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7983 // Check that all of the elements are integer constants or undefs.
7984 bool AllEltsOk = true;
7985 for (unsigned i = 0; i != 16; ++i) {
7986 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7987 !isa<UndefValue>(Mask->getOperand(i))) {
7994 // Cast the input vectors to byte vectors.
7995 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7996 II->getOperand(1), Mask->getType(), CI);
7997 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7998 II->getOperand(2), Mask->getType(), CI);
7999 Value *Result = UndefValue::get(Op0->getType());
8001 // Only extract each element once.
8002 Value *ExtractedElts[32];
8003 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8005 for (unsigned i = 0; i != 16; ++i) {
8006 if (isa<UndefValue>(Mask->getOperand(i)))
8008 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8009 Idx &= 31; // Match the hardware behavior.
8011 if (ExtractedElts[Idx] == 0) {
8013 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8014 InsertNewInstBefore(Elt, CI);
8015 ExtractedElts[Idx] = Elt;
8018 // Insert this value into the result vector.
8019 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
8020 InsertNewInstBefore(cast<Instruction>(Result), CI);
8022 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8027 case Intrinsic::stackrestore: {
8028 // If the save is right next to the restore, remove the restore. This can
8029 // happen when variable allocas are DCE'd.
8030 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8031 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8032 BasicBlock::iterator BI = SS;
8034 return EraseInstFromFunction(CI);
8038 // If the stack restore is in a return/unwind block and if there are no
8039 // allocas or calls between the restore and the return, nuke the restore.
8040 TerminatorInst *TI = II->getParent()->getTerminator();
8041 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
8042 BasicBlock::iterator BI = II;
8043 bool CannotRemove = false;
8044 for (++BI; &*BI != TI; ++BI) {
8045 if (isa<AllocaInst>(BI) ||
8046 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
8047 CannotRemove = true;
8052 return EraseInstFromFunction(CI);
8059 return visitCallSite(II);
8062 // InvokeInst simplification
8064 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8065 return visitCallSite(&II);
8068 // visitCallSite - Improvements for call and invoke instructions.
8070 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8071 bool Changed = false;
8073 // If the callee is a constexpr cast of a function, attempt to move the cast
8074 // to the arguments of the call/invoke.
8075 if (transformConstExprCastCall(CS)) return 0;
8077 Value *Callee = CS.getCalledValue();
8079 if (Function *CalleeF = dyn_cast<Function>(Callee))
8080 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8081 Instruction *OldCall = CS.getInstruction();
8082 // If the call and callee calling conventions don't match, this call must
8083 // be unreachable, as the call is undefined.
8084 new StoreInst(ConstantInt::getTrue(),
8085 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
8086 if (!OldCall->use_empty())
8087 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8088 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8089 return EraseInstFromFunction(*OldCall);
8093 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8094 // This instruction is not reachable, just remove it. We insert a store to
8095 // undef so that we know that this code is not reachable, despite the fact
8096 // that we can't modify the CFG here.
8097 new StoreInst(ConstantInt::getTrue(),
8098 UndefValue::get(PointerType::get(Type::Int1Ty)),
8099 CS.getInstruction());
8101 if (!CS.getInstruction()->use_empty())
8102 CS.getInstruction()->
8103 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8105 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8106 // Don't break the CFG, insert a dummy cond branch.
8107 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8108 ConstantInt::getTrue(), II);
8110 return EraseInstFromFunction(*CS.getInstruction());
8113 const PointerType *PTy = cast<PointerType>(Callee->getType());
8114 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8115 if (FTy->isVarArg()) {
8116 // See if we can optimize any arguments passed through the varargs area of
8118 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8119 E = CS.arg_end(); I != E; ++I)
8120 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8121 // If this cast does not effect the value passed through the varargs
8122 // area, we can eliminate the use of the cast.
8123 Value *Op = CI->getOperand(0);
8124 if (CI->isLosslessCast()) {
8131 return Changed ? CS.getInstruction() : 0;
8134 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8135 // attempt to move the cast to the arguments of the call/invoke.
8137 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8138 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8139 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8140 if (CE->getOpcode() != Instruction::BitCast ||
8141 !isa<Function>(CE->getOperand(0)))
8143 Function *Callee = cast<Function>(CE->getOperand(0));
8144 Instruction *Caller = CS.getInstruction();
8146 // Okay, this is a cast from a function to a different type. Unless doing so
8147 // would cause a type conversion of one of our arguments, change this call to
8148 // be a direct call with arguments casted to the appropriate types.
8150 const FunctionType *FT = Callee->getFunctionType();
8151 const Type *OldRetTy = Caller->getType();
8153 // Check to see if we are changing the return type...
8154 if (OldRetTy != FT->getReturnType()) {
8155 if (Callee->isDeclaration() && !Caller->use_empty() &&
8156 // Conversion is ok if changing from pointer to int of same size.
8157 !(isa<PointerType>(FT->getReturnType()) &&
8158 TD->getIntPtrType() == OldRetTy))
8159 return false; // Cannot transform this return value.
8161 // If the callsite is an invoke instruction, and the return value is used by
8162 // a PHI node in a successor, we cannot change the return type of the call
8163 // because there is no place to put the cast instruction (without breaking
8164 // the critical edge). Bail out in this case.
8165 if (!Caller->use_empty())
8166 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8167 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8169 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8170 if (PN->getParent() == II->getNormalDest() ||
8171 PN->getParent() == II->getUnwindDest())
8175 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8176 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8178 CallSite::arg_iterator AI = CS.arg_begin();
8179 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8180 const Type *ParamTy = FT->getParamType(i);
8181 const Type *ActTy = (*AI)->getType();
8182 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8183 //Either we can cast directly, or we can upconvert the argument
8184 bool isConvertible = ActTy == ParamTy ||
8185 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8186 (ParamTy->isInteger() && ActTy->isInteger() &&
8187 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8188 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8189 && c->getSExtValue() > 0);
8190 if (Callee->isDeclaration() && !isConvertible) return false;
8193 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8194 Callee->isDeclaration())
8195 return false; // Do not delete arguments unless we have a function body...
8197 // Okay, we decided that this is a safe thing to do: go ahead and start
8198 // inserting cast instructions as necessary...
8199 std::vector<Value*> Args;
8200 Args.reserve(NumActualArgs);
8202 AI = CS.arg_begin();
8203 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8204 const Type *ParamTy = FT->getParamType(i);
8205 if ((*AI)->getType() == ParamTy) {
8206 Args.push_back(*AI);
8208 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8209 false, ParamTy, false);
8210 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8211 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8215 // If the function takes more arguments than the call was taking, add them
8217 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8218 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8220 // If we are removing arguments to the function, emit an obnoxious warning...
8221 if (FT->getNumParams() < NumActualArgs)
8222 if (!FT->isVarArg()) {
8223 cerr << "WARNING: While resolving call to function '"
8224 << Callee->getName() << "' arguments were dropped!\n";
8226 // Add all of the arguments in their promoted form to the arg list...
8227 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8228 const Type *PTy = getPromotedType((*AI)->getType());
8229 if (PTy != (*AI)->getType()) {
8230 // Must promote to pass through va_arg area!
8231 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8233 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8234 InsertNewInstBefore(Cast, *Caller);
8235 Args.push_back(Cast);
8237 Args.push_back(*AI);
8242 if (FT->getReturnType() == Type::VoidTy)
8243 Caller->setName(""); // Void type should not have a name.
8246 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8247 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8248 &Args[0], Args.size(), Caller->getName(), Caller);
8249 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
8251 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
8252 if (cast<CallInst>(Caller)->isTailCall())
8253 cast<CallInst>(NC)->setTailCall();
8254 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8257 // Insert a cast of the return type as necessary.
8259 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
8260 if (NV->getType() != Type::VoidTy) {
8261 const Type *CallerTy = Caller->getType();
8262 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8264 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
8266 // If this is an invoke instruction, we should insert it after the first
8267 // non-phi, instruction in the normal successor block.
8268 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8269 BasicBlock::iterator I = II->getNormalDest()->begin();
8270 while (isa<PHINode>(I)) ++I;
8271 InsertNewInstBefore(NC, *I);
8273 // Otherwise, it's a call, just insert cast right after the call instr
8274 InsertNewInstBefore(NC, *Caller);
8276 AddUsersToWorkList(*Caller);
8278 NV = UndefValue::get(Caller->getType());
8282 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8283 Caller->replaceAllUsesWith(NV);
8284 Caller->eraseFromParent();
8285 RemoveFromWorkList(Caller);
8289 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8290 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8291 /// and a single binop.
8292 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8293 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8294 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8295 isa<CmpInst>(FirstInst));
8296 unsigned Opc = FirstInst->getOpcode();
8297 Value *LHSVal = FirstInst->getOperand(0);
8298 Value *RHSVal = FirstInst->getOperand(1);
8300 const Type *LHSType = LHSVal->getType();
8301 const Type *RHSType = RHSVal->getType();
8303 // Scan to see if all operands are the same opcode, all have one use, and all
8304 // kill their operands (i.e. the operands have one use).
8305 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8306 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8307 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8308 // Verify type of the LHS matches so we don't fold cmp's of different
8309 // types or GEP's with different index types.
8310 I->getOperand(0)->getType() != LHSType ||
8311 I->getOperand(1)->getType() != RHSType)
8314 // If they are CmpInst instructions, check their predicates
8315 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8316 if (cast<CmpInst>(I)->getPredicate() !=
8317 cast<CmpInst>(FirstInst)->getPredicate())
8320 // Keep track of which operand needs a phi node.
8321 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8322 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8325 // Otherwise, this is safe to transform, determine if it is profitable.
8327 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8328 // Indexes are often folded into load/store instructions, so we don't want to
8329 // hide them behind a phi.
8330 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8333 Value *InLHS = FirstInst->getOperand(0);
8334 Value *InRHS = FirstInst->getOperand(1);
8335 PHINode *NewLHS = 0, *NewRHS = 0;
8337 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8338 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8339 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8340 InsertNewInstBefore(NewLHS, PN);
8345 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8346 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8347 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8348 InsertNewInstBefore(NewRHS, PN);
8352 // Add all operands to the new PHIs.
8353 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8355 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8356 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8359 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8360 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8364 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8365 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8366 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8367 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8370 assert(isa<GetElementPtrInst>(FirstInst));
8371 return new GetElementPtrInst(LHSVal, RHSVal);
8375 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8376 /// of the block that defines it. This means that it must be obvious the value
8377 /// of the load is not changed from the point of the load to the end of the
8380 /// Finally, it is safe, but not profitable, to sink a load targetting a
8381 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8383 static bool isSafeToSinkLoad(LoadInst *L) {
8384 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8386 for (++BBI; BBI != E; ++BBI)
8387 if (BBI->mayWriteToMemory())
8390 // Check for non-address taken alloca. If not address-taken already, it isn't
8391 // profitable to do this xform.
8392 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8393 bool isAddressTaken = false;
8394 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8396 if (isa<LoadInst>(UI)) continue;
8397 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8398 // If storing TO the alloca, then the address isn't taken.
8399 if (SI->getOperand(1) == AI) continue;
8401 isAddressTaken = true;
8405 if (!isAddressTaken)
8413 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8414 // operator and they all are only used by the PHI, PHI together their
8415 // inputs, and do the operation once, to the result of the PHI.
8416 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8417 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8419 // Scan the instruction, looking for input operations that can be folded away.
8420 // If all input operands to the phi are the same instruction (e.g. a cast from
8421 // the same type or "+42") we can pull the operation through the PHI, reducing
8422 // code size and simplifying code.
8423 Constant *ConstantOp = 0;
8424 const Type *CastSrcTy = 0;
8425 bool isVolatile = false;
8426 if (isa<CastInst>(FirstInst)) {
8427 CastSrcTy = FirstInst->getOperand(0)->getType();
8428 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8429 // Can fold binop, compare or shift here if the RHS is a constant,
8430 // otherwise call FoldPHIArgBinOpIntoPHI.
8431 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8432 if (ConstantOp == 0)
8433 return FoldPHIArgBinOpIntoPHI(PN);
8434 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8435 isVolatile = LI->isVolatile();
8436 // We can't sink the load if the loaded value could be modified between the
8437 // load and the PHI.
8438 if (LI->getParent() != PN.getIncomingBlock(0) ||
8439 !isSafeToSinkLoad(LI))
8441 } else if (isa<GetElementPtrInst>(FirstInst)) {
8442 if (FirstInst->getNumOperands() == 2)
8443 return FoldPHIArgBinOpIntoPHI(PN);
8444 // Can't handle general GEPs yet.
8447 return 0; // Cannot fold this operation.
8450 // Check to see if all arguments are the same operation.
8451 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8452 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8453 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8454 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8457 if (I->getOperand(0)->getType() != CastSrcTy)
8458 return 0; // Cast operation must match.
8459 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8460 // We can't sink the load if the loaded value could be modified between
8461 // the load and the PHI.
8462 if (LI->isVolatile() != isVolatile ||
8463 LI->getParent() != PN.getIncomingBlock(i) ||
8464 !isSafeToSinkLoad(LI))
8466 } else if (I->getOperand(1) != ConstantOp) {
8471 // Okay, they are all the same operation. Create a new PHI node of the
8472 // correct type, and PHI together all of the LHS's of the instructions.
8473 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8474 PN.getName()+".in");
8475 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8477 Value *InVal = FirstInst->getOperand(0);
8478 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8480 // Add all operands to the new PHI.
8481 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8482 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8483 if (NewInVal != InVal)
8485 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8490 // The new PHI unions all of the same values together. This is really
8491 // common, so we handle it intelligently here for compile-time speed.
8495 InsertNewInstBefore(NewPN, PN);
8499 // Insert and return the new operation.
8500 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8501 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8502 else if (isa<LoadInst>(FirstInst))
8503 return new LoadInst(PhiVal, "", isVolatile);
8504 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8505 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8506 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8507 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8508 PhiVal, ConstantOp);
8510 assert(0 && "Unknown operation");
8514 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8516 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
8517 if (PN->use_empty()) return true;
8518 if (!PN->hasOneUse()) return false;
8520 // Remember this node, and if we find the cycle, return.
8521 if (!PotentiallyDeadPHIs.insert(PN).second)
8524 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8525 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8530 // PHINode simplification
8532 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8533 // If LCSSA is around, don't mess with Phi nodes
8534 if (MustPreserveLCSSA) return 0;
8536 if (Value *V = PN.hasConstantValue())
8537 return ReplaceInstUsesWith(PN, V);
8539 // If all PHI operands are the same operation, pull them through the PHI,
8540 // reducing code size.
8541 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8542 PN.getIncomingValue(0)->hasOneUse())
8543 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8546 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8547 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8548 // PHI)... break the cycle.
8549 if (PN.hasOneUse()) {
8550 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8551 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8552 std::set<PHINode*> PotentiallyDeadPHIs;
8553 PotentiallyDeadPHIs.insert(&PN);
8554 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8555 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8558 // If this phi has a single use, and if that use just computes a value for
8559 // the next iteration of a loop, delete the phi. This occurs with unused
8560 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8561 // common case here is good because the only other things that catch this
8562 // are induction variable analysis (sometimes) and ADCE, which is only run
8564 if (PHIUser->hasOneUse() &&
8565 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8566 PHIUser->use_back() == &PN) {
8567 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8574 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8575 Instruction *InsertPoint,
8577 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8578 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8579 // We must cast correctly to the pointer type. Ensure that we
8580 // sign extend the integer value if it is smaller as this is
8581 // used for address computation.
8582 Instruction::CastOps opcode =
8583 (VTySize < PtrSize ? Instruction::SExt :
8584 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8585 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8589 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8590 Value *PtrOp = GEP.getOperand(0);
8591 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
8592 // If so, eliminate the noop.
8593 if (GEP.getNumOperands() == 1)
8594 return ReplaceInstUsesWith(GEP, PtrOp);
8596 if (isa<UndefValue>(GEP.getOperand(0)))
8597 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8599 bool HasZeroPointerIndex = false;
8600 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8601 HasZeroPointerIndex = C->isNullValue();
8603 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8604 return ReplaceInstUsesWith(GEP, PtrOp);
8606 // Eliminate unneeded casts for indices.
8607 bool MadeChange = false;
8608 gep_type_iterator GTI = gep_type_begin(GEP);
8609 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
8610 if (isa<SequentialType>(*GTI)) {
8611 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8612 if (CI->getOpcode() == Instruction::ZExt ||
8613 CI->getOpcode() == Instruction::SExt) {
8614 const Type *SrcTy = CI->getOperand(0)->getType();
8615 // We can eliminate a cast from i32 to i64 iff the target
8616 // is a 32-bit pointer target.
8617 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8619 GEP.setOperand(i, CI->getOperand(0));
8623 // If we are using a wider index than needed for this platform, shrink it
8624 // to what we need. If the incoming value needs a cast instruction,
8625 // insert it. This explicit cast can make subsequent optimizations more
8627 Value *Op = GEP.getOperand(i);
8628 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
8629 if (Constant *C = dyn_cast<Constant>(Op)) {
8630 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8633 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8635 GEP.setOperand(i, Op);
8639 if (MadeChange) return &GEP;
8641 // Combine Indices - If the source pointer to this getelementptr instruction
8642 // is a getelementptr instruction, combine the indices of the two
8643 // getelementptr instructions into a single instruction.
8645 SmallVector<Value*, 8> SrcGEPOperands;
8646 if (User *Src = dyn_castGetElementPtr(PtrOp))
8647 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8649 if (!SrcGEPOperands.empty()) {
8650 // Note that if our source is a gep chain itself that we wait for that
8651 // chain to be resolved before we perform this transformation. This
8652 // avoids us creating a TON of code in some cases.
8654 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8655 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8656 return 0; // Wait until our source is folded to completion.
8658 SmallVector<Value*, 8> Indices;
8660 // Find out whether the last index in the source GEP is a sequential idx.
8661 bool EndsWithSequential = false;
8662 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8663 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8664 EndsWithSequential = !isa<StructType>(*I);
8666 // Can we combine the two pointer arithmetics offsets?
8667 if (EndsWithSequential) {
8668 // Replace: gep (gep %P, long B), long A, ...
8669 // With: T = long A+B; gep %P, T, ...
8671 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8672 if (SO1 == Constant::getNullValue(SO1->getType())) {
8674 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8677 // If they aren't the same type, convert both to an integer of the
8678 // target's pointer size.
8679 if (SO1->getType() != GO1->getType()) {
8680 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8681 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8682 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8683 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8685 unsigned PS = TD->getPointerSize();
8686 if (TD->getTypeSize(SO1->getType()) == PS) {
8687 // Convert GO1 to SO1's type.
8688 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8690 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8691 // Convert SO1 to GO1's type.
8692 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8694 const Type *PT = TD->getIntPtrType();
8695 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8696 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8700 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8701 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8703 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8704 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8708 // Recycle the GEP we already have if possible.
8709 if (SrcGEPOperands.size() == 2) {
8710 GEP.setOperand(0, SrcGEPOperands[0]);
8711 GEP.setOperand(1, Sum);
8714 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8715 SrcGEPOperands.end()-1);
8716 Indices.push_back(Sum);
8717 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8719 } else if (isa<Constant>(*GEP.idx_begin()) &&
8720 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8721 SrcGEPOperands.size() != 1) {
8722 // Otherwise we can do the fold if the first index of the GEP is a zero
8723 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8724 SrcGEPOperands.end());
8725 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8728 if (!Indices.empty())
8729 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
8730 Indices.size(), GEP.getName());
8732 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8733 // GEP of global variable. If all of the indices for this GEP are
8734 // constants, we can promote this to a constexpr instead of an instruction.
8736 // Scan for nonconstants...
8737 SmallVector<Constant*, 8> Indices;
8738 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8739 for (; I != E && isa<Constant>(*I); ++I)
8740 Indices.push_back(cast<Constant>(*I));
8742 if (I == E) { // If they are all constants...
8743 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8744 &Indices[0],Indices.size());
8746 // Replace all uses of the GEP with the new constexpr...
8747 return ReplaceInstUsesWith(GEP, CE);
8749 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8750 if (!isa<PointerType>(X->getType())) {
8751 // Not interesting. Source pointer must be a cast from pointer.
8752 } else if (HasZeroPointerIndex) {
8753 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8754 // into : GEP [10 x ubyte]* X, long 0, ...
8756 // This occurs when the program declares an array extern like "int X[];"
8758 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8759 const PointerType *XTy = cast<PointerType>(X->getType());
8760 if (const ArrayType *XATy =
8761 dyn_cast<ArrayType>(XTy->getElementType()))
8762 if (const ArrayType *CATy =
8763 dyn_cast<ArrayType>(CPTy->getElementType()))
8764 if (CATy->getElementType() == XATy->getElementType()) {
8765 // At this point, we know that the cast source type is a pointer
8766 // to an array of the same type as the destination pointer
8767 // array. Because the array type is never stepped over (there
8768 // is a leading zero) we can fold the cast into this GEP.
8769 GEP.setOperand(0, X);
8772 } else if (GEP.getNumOperands() == 2) {
8773 // Transform things like:
8774 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8775 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8776 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8777 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8778 if (isa<ArrayType>(SrcElTy) &&
8779 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8780 TD->getTypeSize(ResElTy)) {
8781 Value *V = InsertNewInstBefore(
8782 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8783 GEP.getOperand(1), GEP.getName()), GEP);
8784 // V and GEP are both pointer types --> BitCast
8785 return new BitCastInst(V, GEP.getType());
8788 // Transform things like:
8789 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8790 // (where tmp = 8*tmp2) into:
8791 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8793 if (isa<ArrayType>(SrcElTy) &&
8794 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8795 uint64_t ArrayEltSize =
8796 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8798 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8799 // allow either a mul, shift, or constant here.
8801 ConstantInt *Scale = 0;
8802 if (ArrayEltSize == 1) {
8803 NewIdx = GEP.getOperand(1);
8804 Scale = ConstantInt::get(NewIdx->getType(), 1);
8805 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8806 NewIdx = ConstantInt::get(CI->getType(), 1);
8808 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8809 if (Inst->getOpcode() == Instruction::Shl &&
8810 isa<ConstantInt>(Inst->getOperand(1))) {
8812 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
8813 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
8814 NewIdx = Inst->getOperand(0);
8815 } else if (Inst->getOpcode() == Instruction::Mul &&
8816 isa<ConstantInt>(Inst->getOperand(1))) {
8817 Scale = cast<ConstantInt>(Inst->getOperand(1));
8818 NewIdx = Inst->getOperand(0);
8822 // If the index will be to exactly the right offset with the scale taken
8823 // out, perform the transformation.
8824 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8825 if (isa<ConstantInt>(Scale))
8826 Scale = ConstantInt::get(Scale->getType(),
8827 Scale->getZExtValue() / ArrayEltSize);
8828 if (Scale->getZExtValue() != 1) {
8829 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8831 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8832 NewIdx = InsertNewInstBefore(Sc, GEP);
8835 // Insert the new GEP instruction.
8836 Instruction *NewGEP =
8837 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8838 NewIdx, GEP.getName());
8839 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8840 // The NewGEP must be pointer typed, so must the old one -> BitCast
8841 return new BitCastInst(NewGEP, GEP.getType());
8850 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8851 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8852 if (AI.isArrayAllocation()) // Check C != 1
8853 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8855 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8856 AllocationInst *New = 0;
8858 // Create and insert the replacement instruction...
8859 if (isa<MallocInst>(AI))
8860 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8862 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8863 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8866 InsertNewInstBefore(New, AI);
8868 // Scan to the end of the allocation instructions, to skip over a block of
8869 // allocas if possible...
8871 BasicBlock::iterator It = New;
8872 while (isa<AllocationInst>(*It)) ++It;
8874 // Now that I is pointing to the first non-allocation-inst in the block,
8875 // insert our getelementptr instruction...
8877 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8878 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8879 New->getName()+".sub", It);
8881 // Now make everything use the getelementptr instead of the original
8883 return ReplaceInstUsesWith(AI, V);
8884 } else if (isa<UndefValue>(AI.getArraySize())) {
8885 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8888 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8889 // Note that we only do this for alloca's, because malloc should allocate and
8890 // return a unique pointer, even for a zero byte allocation.
8891 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8892 TD->getTypeSize(AI.getAllocatedType()) == 0)
8893 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8898 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8899 Value *Op = FI.getOperand(0);
8901 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8902 if (CastInst *CI = dyn_cast<CastInst>(Op))
8903 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8904 FI.setOperand(0, CI->getOperand(0));
8908 // free undef -> unreachable.
8909 if (isa<UndefValue>(Op)) {
8910 // Insert a new store to null because we cannot modify the CFG here.
8911 new StoreInst(ConstantInt::getTrue(),
8912 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8913 return EraseInstFromFunction(FI);
8916 // If we have 'free null' delete the instruction. This can happen in stl code
8917 // when lots of inlining happens.
8918 if (isa<ConstantPointerNull>(Op))
8919 return EraseInstFromFunction(FI);
8925 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8926 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8927 User *CI = cast<User>(LI.getOperand(0));
8928 Value *CastOp = CI->getOperand(0);
8930 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8931 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8932 const Type *SrcPTy = SrcTy->getElementType();
8934 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8935 isa<VectorType>(DestPTy)) {
8936 // If the source is an array, the code below will not succeed. Check to
8937 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8939 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8940 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8941 if (ASrcTy->getNumElements() != 0) {
8943 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8944 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8945 SrcTy = cast<PointerType>(CastOp->getType());
8946 SrcPTy = SrcTy->getElementType();
8949 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8950 isa<VectorType>(SrcPTy)) &&
8951 // Do not allow turning this into a load of an integer, which is then
8952 // casted to a pointer, this pessimizes pointer analysis a lot.
8953 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8954 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8955 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8957 // Okay, we are casting from one integer or pointer type to another of
8958 // the same size. Instead of casting the pointer before the load, cast
8959 // the result of the loaded value.
8960 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8962 LI.isVolatile()),LI);
8963 // Now cast the result of the load.
8964 return new BitCastInst(NewLoad, LI.getType());
8971 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8972 /// from this value cannot trap. If it is not obviously safe to load from the
8973 /// specified pointer, we do a quick local scan of the basic block containing
8974 /// ScanFrom, to determine if the address is already accessed.
8975 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8976 // If it is an alloca or global variable, it is always safe to load from.
8977 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8979 // Otherwise, be a little bit agressive by scanning the local block where we
8980 // want to check to see if the pointer is already being loaded or stored
8981 // from/to. If so, the previous load or store would have already trapped,
8982 // so there is no harm doing an extra load (also, CSE will later eliminate
8983 // the load entirely).
8984 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8989 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8990 if (LI->getOperand(0) == V) return true;
8991 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8992 if (SI->getOperand(1) == V) return true;
8998 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8999 Value *Op = LI.getOperand(0);
9001 // load (cast X) --> cast (load X) iff safe
9002 if (isa<CastInst>(Op))
9003 if (Instruction *Res = InstCombineLoadCast(*this, LI))
9006 // None of the following transforms are legal for volatile loads.
9007 if (LI.isVolatile()) return 0;
9009 if (&LI.getParent()->front() != &LI) {
9010 BasicBlock::iterator BBI = &LI; --BBI;
9011 // If the instruction immediately before this is a store to the same
9012 // address, do a simple form of store->load forwarding.
9013 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9014 if (SI->getOperand(1) == LI.getOperand(0))
9015 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9016 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9017 if (LIB->getOperand(0) == LI.getOperand(0))
9018 return ReplaceInstUsesWith(LI, LIB);
9021 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
9022 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
9023 isa<UndefValue>(GEPI->getOperand(0))) {
9024 // Insert a new store to null instruction before the load to indicate
9025 // that this code is not reachable. We do this instead of inserting
9026 // an unreachable instruction directly because we cannot modify the
9028 new StoreInst(UndefValue::get(LI.getType()),
9029 Constant::getNullValue(Op->getType()), &LI);
9030 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9033 if (Constant *C = dyn_cast<Constant>(Op)) {
9034 // load null/undef -> undef
9035 if ((C->isNullValue() || isa<UndefValue>(C))) {
9036 // Insert a new store to null instruction before the load to indicate that
9037 // this code is not reachable. We do this instead of inserting an
9038 // unreachable instruction directly because we cannot modify the CFG.
9039 new StoreInst(UndefValue::get(LI.getType()),
9040 Constant::getNullValue(Op->getType()), &LI);
9041 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9044 // Instcombine load (constant global) into the value loaded.
9045 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9046 if (GV->isConstant() && !GV->isDeclaration())
9047 return ReplaceInstUsesWith(LI, GV->getInitializer());
9049 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9050 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9051 if (CE->getOpcode() == Instruction::GetElementPtr) {
9052 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9053 if (GV->isConstant() && !GV->isDeclaration())
9055 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9056 return ReplaceInstUsesWith(LI, V);
9057 if (CE->getOperand(0)->isNullValue()) {
9058 // Insert a new store to null instruction before the load to indicate
9059 // that this code is not reachable. We do this instead of inserting
9060 // an unreachable instruction directly because we cannot modify the
9062 new StoreInst(UndefValue::get(LI.getType()),
9063 Constant::getNullValue(Op->getType()), &LI);
9064 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9067 } else if (CE->isCast()) {
9068 if (Instruction *Res = InstCombineLoadCast(*this, LI))
9073 if (Op->hasOneUse()) {
9074 // Change select and PHI nodes to select values instead of addresses: this
9075 // helps alias analysis out a lot, allows many others simplifications, and
9076 // exposes redundancy in the code.
9078 // Note that we cannot do the transformation unless we know that the
9079 // introduced loads cannot trap! Something like this is valid as long as
9080 // the condition is always false: load (select bool %C, int* null, int* %G),
9081 // but it would not be valid if we transformed it to load from null
9084 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9085 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9086 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9087 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9088 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9089 SI->getOperand(1)->getName()+".val"), LI);
9090 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9091 SI->getOperand(2)->getName()+".val"), LI);
9092 return new SelectInst(SI->getCondition(), V1, V2);
9095 // load (select (cond, null, P)) -> load P
9096 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9097 if (C->isNullValue()) {
9098 LI.setOperand(0, SI->getOperand(2));
9102 // load (select (cond, P, null)) -> load P
9103 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9104 if (C->isNullValue()) {
9105 LI.setOperand(0, SI->getOperand(1));
9113 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9115 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9116 User *CI = cast<User>(SI.getOperand(1));
9117 Value *CastOp = CI->getOperand(0);
9119 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9120 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9121 const Type *SrcPTy = SrcTy->getElementType();
9123 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9124 // If the source is an array, the code below will not succeed. Check to
9125 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9127 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9128 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9129 if (ASrcTy->getNumElements() != 0) {
9131 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9132 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9133 SrcTy = cast<PointerType>(CastOp->getType());
9134 SrcPTy = SrcTy->getElementType();
9137 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9138 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9139 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9141 // Okay, we are casting from one integer or pointer type to another of
9142 // the same size. Instead of casting the pointer before
9143 // the store, cast the value to be stored.
9145 Value *SIOp0 = SI.getOperand(0);
9146 Instruction::CastOps opcode = Instruction::BitCast;
9147 const Type* CastSrcTy = SIOp0->getType();
9148 const Type* CastDstTy = SrcPTy;
9149 if (isa<PointerType>(CastDstTy)) {
9150 if (CastSrcTy->isInteger())
9151 opcode = Instruction::IntToPtr;
9152 } else if (isa<IntegerType>(CastDstTy)) {
9153 if (isa<PointerType>(SIOp0->getType()))
9154 opcode = Instruction::PtrToInt;
9156 if (Constant *C = dyn_cast<Constant>(SIOp0))
9157 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9159 NewCast = IC.InsertNewInstBefore(
9160 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9162 return new StoreInst(NewCast, CastOp);
9169 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9170 Value *Val = SI.getOperand(0);
9171 Value *Ptr = SI.getOperand(1);
9173 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9174 EraseInstFromFunction(SI);
9179 // If the RHS is an alloca with a single use, zapify the store, making the
9181 if (Ptr->hasOneUse()) {
9182 if (isa<AllocaInst>(Ptr)) {
9183 EraseInstFromFunction(SI);
9188 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9189 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9190 GEP->getOperand(0)->hasOneUse()) {
9191 EraseInstFromFunction(SI);
9197 // Do really simple DSE, to catch cases where there are several consequtive
9198 // stores to the same location, separated by a few arithmetic operations. This
9199 // situation often occurs with bitfield accesses.
9200 BasicBlock::iterator BBI = &SI;
9201 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9205 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9206 // Prev store isn't volatile, and stores to the same location?
9207 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9210 EraseInstFromFunction(*PrevSI);
9216 // If this is a load, we have to stop. However, if the loaded value is from
9217 // the pointer we're loading and is producing the pointer we're storing,
9218 // then *this* store is dead (X = load P; store X -> P).
9219 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9220 if (LI == Val && LI->getOperand(0) == Ptr) {
9221 EraseInstFromFunction(SI);
9225 // Otherwise, this is a load from some other location. Stores before it
9230 // Don't skip over loads or things that can modify memory.
9231 if (BBI->mayWriteToMemory())
9236 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9238 // store X, null -> turns into 'unreachable' in SimplifyCFG
9239 if (isa<ConstantPointerNull>(Ptr)) {
9240 if (!isa<UndefValue>(Val)) {
9241 SI.setOperand(0, UndefValue::get(Val->getType()));
9242 if (Instruction *U = dyn_cast<Instruction>(Val))
9243 AddToWorkList(U); // Dropped a use.
9246 return 0; // Do not modify these!
9249 // store undef, Ptr -> noop
9250 if (isa<UndefValue>(Val)) {
9251 EraseInstFromFunction(SI);
9256 // If the pointer destination is a cast, see if we can fold the cast into the
9258 if (isa<CastInst>(Ptr))
9259 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9261 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9263 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9267 // If this store is the last instruction in the basic block, and if the block
9268 // ends with an unconditional branch, try to move it to the successor block.
9270 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9271 if (BI->isUnconditional()) {
9272 // Check to see if the successor block has exactly two incoming edges. If
9273 // so, see if the other predecessor contains a store to the same location.
9274 // if so, insert a PHI node (if needed) and move the stores down.
9275 BasicBlock *Dest = BI->getSuccessor(0);
9277 pred_iterator PI = pred_begin(Dest);
9278 BasicBlock *Other = 0;
9279 if (*PI != BI->getParent())
9282 if (PI != pred_end(Dest)) {
9283 if (*PI != BI->getParent())
9288 if (++PI != pred_end(Dest))
9291 if (Other) { // If only one other pred...
9292 BBI = Other->getTerminator();
9293 // Make sure this other block ends in an unconditional branch and that
9294 // there is an instruction before the branch.
9295 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
9296 BBI != Other->begin()) {
9298 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
9300 // If this instruction is a store to the same location.
9301 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
9302 // Okay, we know we can perform this transformation. Insert a PHI
9303 // node now if we need it.
9304 Value *MergedVal = OtherStore->getOperand(0);
9305 if (MergedVal != SI.getOperand(0)) {
9306 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9307 PN->reserveOperandSpace(2);
9308 PN->addIncoming(SI.getOperand(0), SI.getParent());
9309 PN->addIncoming(OtherStore->getOperand(0), Other);
9310 MergedVal = InsertNewInstBefore(PN, Dest->front());
9313 // Advance to a place where it is safe to insert the new store and
9315 BBI = Dest->begin();
9316 while (isa<PHINode>(BBI)) ++BBI;
9317 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9318 OtherStore->isVolatile()), *BBI);
9320 // Nuke the old stores.
9321 EraseInstFromFunction(SI);
9322 EraseInstFromFunction(*OtherStore);
9334 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9335 // Change br (not X), label True, label False to: br X, label False, True
9337 BasicBlock *TrueDest;
9338 BasicBlock *FalseDest;
9339 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9340 !isa<Constant>(X)) {
9341 // Swap Destinations and condition...
9343 BI.setSuccessor(0, FalseDest);
9344 BI.setSuccessor(1, TrueDest);
9348 // Cannonicalize fcmp_one -> fcmp_oeq
9349 FCmpInst::Predicate FPred; Value *Y;
9350 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9351 TrueDest, FalseDest)))
9352 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9353 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9354 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9355 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9356 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9357 NewSCC->takeName(I);
9358 // Swap Destinations and condition...
9359 BI.setCondition(NewSCC);
9360 BI.setSuccessor(0, FalseDest);
9361 BI.setSuccessor(1, TrueDest);
9362 RemoveFromWorkList(I);
9363 I->eraseFromParent();
9364 AddToWorkList(NewSCC);
9368 // Cannonicalize icmp_ne -> icmp_eq
9369 ICmpInst::Predicate IPred;
9370 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9371 TrueDest, FalseDest)))
9372 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9373 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9374 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9375 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9376 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9377 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9378 NewSCC->takeName(I);
9379 // Swap Destinations and condition...
9380 BI.setCondition(NewSCC);
9381 BI.setSuccessor(0, FalseDest);
9382 BI.setSuccessor(1, TrueDest);
9383 RemoveFromWorkList(I);
9384 I->eraseFromParent();;
9385 AddToWorkList(NewSCC);
9392 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9393 Value *Cond = SI.getCondition();
9394 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9395 if (I->getOpcode() == Instruction::Add)
9396 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9397 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9398 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9399 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9401 SI.setOperand(0, I->getOperand(0));
9409 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9410 /// is to leave as a vector operation.
9411 static bool CheapToScalarize(Value *V, bool isConstant) {
9412 if (isa<ConstantAggregateZero>(V))
9414 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9415 if (isConstant) return true;
9416 // If all elts are the same, we can extract.
9417 Constant *Op0 = C->getOperand(0);
9418 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9419 if (C->getOperand(i) != Op0)
9423 Instruction *I = dyn_cast<Instruction>(V);
9424 if (!I) return false;
9426 // Insert element gets simplified to the inserted element or is deleted if
9427 // this is constant idx extract element and its a constant idx insertelt.
9428 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9429 isa<ConstantInt>(I->getOperand(2)))
9431 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9433 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9434 if (BO->hasOneUse() &&
9435 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9436 CheapToScalarize(BO->getOperand(1), isConstant)))
9438 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9439 if (CI->hasOneUse() &&
9440 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9441 CheapToScalarize(CI->getOperand(1), isConstant)))
9447 /// Read and decode a shufflevector mask.
9449 /// It turns undef elements into values that are larger than the number of
9450 /// elements in the input.
9451 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9452 unsigned NElts = SVI->getType()->getNumElements();
9453 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9454 return std::vector<unsigned>(NElts, 0);
9455 if (isa<UndefValue>(SVI->getOperand(2)))
9456 return std::vector<unsigned>(NElts, 2*NElts);
9458 std::vector<unsigned> Result;
9459 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9460 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9461 if (isa<UndefValue>(CP->getOperand(i)))
9462 Result.push_back(NElts*2); // undef -> 8
9464 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9468 /// FindScalarElement - Given a vector and an element number, see if the scalar
9469 /// value is already around as a register, for example if it were inserted then
9470 /// extracted from the vector.
9471 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9472 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9473 const VectorType *PTy = cast<VectorType>(V->getType());
9474 unsigned Width = PTy->getNumElements();
9475 if (EltNo >= Width) // Out of range access.
9476 return UndefValue::get(PTy->getElementType());
9478 if (isa<UndefValue>(V))
9479 return UndefValue::get(PTy->getElementType());
9480 else if (isa<ConstantAggregateZero>(V))
9481 return Constant::getNullValue(PTy->getElementType());
9482 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9483 return CP->getOperand(EltNo);
9484 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9485 // If this is an insert to a variable element, we don't know what it is.
9486 if (!isa<ConstantInt>(III->getOperand(2)))
9488 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9490 // If this is an insert to the element we are looking for, return the
9493 return III->getOperand(1);
9495 // Otherwise, the insertelement doesn't modify the value, recurse on its
9497 return FindScalarElement(III->getOperand(0), EltNo);
9498 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9499 unsigned InEl = getShuffleMask(SVI)[EltNo];
9501 return FindScalarElement(SVI->getOperand(0), InEl);
9502 else if (InEl < Width*2)
9503 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9505 return UndefValue::get(PTy->getElementType());
9508 // Otherwise, we don't know.
9512 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9514 // If packed val is undef, replace extract with scalar undef.
9515 if (isa<UndefValue>(EI.getOperand(0)))
9516 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9518 // If packed val is constant 0, replace extract with scalar 0.
9519 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9520 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9522 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9523 // If packed val is constant with uniform operands, replace EI
9524 // with that operand
9525 Constant *op0 = C->getOperand(0);
9526 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9527 if (C->getOperand(i) != op0) {
9532 return ReplaceInstUsesWith(EI, op0);
9535 // If extracting a specified index from the vector, see if we can recursively
9536 // find a previously computed scalar that was inserted into the vector.
9537 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9538 // This instruction only demands the single element from the input vector.
9539 // If the input vector has a single use, simplify it based on this use
9541 uint64_t IndexVal = IdxC->getZExtValue();
9542 if (EI.getOperand(0)->hasOneUse()) {
9544 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9547 EI.setOperand(0, V);
9552 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9553 return ReplaceInstUsesWith(EI, Elt);
9556 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9557 if (I->hasOneUse()) {
9558 // Push extractelement into predecessor operation if legal and
9559 // profitable to do so
9560 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9561 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9562 if (CheapToScalarize(BO, isConstantElt)) {
9563 ExtractElementInst *newEI0 =
9564 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9565 EI.getName()+".lhs");
9566 ExtractElementInst *newEI1 =
9567 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9568 EI.getName()+".rhs");
9569 InsertNewInstBefore(newEI0, EI);
9570 InsertNewInstBefore(newEI1, EI);
9571 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9573 } else if (isa<LoadInst>(I)) {
9574 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9575 PointerType::get(EI.getType()), EI);
9576 GetElementPtrInst *GEP =
9577 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9578 InsertNewInstBefore(GEP, EI);
9579 return new LoadInst(GEP);
9582 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9583 // Extracting the inserted element?
9584 if (IE->getOperand(2) == EI.getOperand(1))
9585 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9586 // If the inserted and extracted elements are constants, they must not
9587 // be the same value, extract from the pre-inserted value instead.
9588 if (isa<Constant>(IE->getOperand(2)) &&
9589 isa<Constant>(EI.getOperand(1))) {
9590 AddUsesToWorkList(EI);
9591 EI.setOperand(0, IE->getOperand(0));
9594 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9595 // If this is extracting an element from a shufflevector, figure out where
9596 // it came from and extract from the appropriate input element instead.
9597 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9598 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9600 if (SrcIdx < SVI->getType()->getNumElements())
9601 Src = SVI->getOperand(0);
9602 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9603 SrcIdx -= SVI->getType()->getNumElements();
9604 Src = SVI->getOperand(1);
9606 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9608 return new ExtractElementInst(Src, SrcIdx);
9615 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9616 /// elements from either LHS or RHS, return the shuffle mask and true.
9617 /// Otherwise, return false.
9618 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9619 std::vector<Constant*> &Mask) {
9620 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9621 "Invalid CollectSingleShuffleElements");
9622 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9624 if (isa<UndefValue>(V)) {
9625 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9627 } else if (V == LHS) {
9628 for (unsigned i = 0; i != NumElts; ++i)
9629 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9631 } else if (V == RHS) {
9632 for (unsigned i = 0; i != NumElts; ++i)
9633 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9635 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9636 // If this is an insert of an extract from some other vector, include it.
9637 Value *VecOp = IEI->getOperand(0);
9638 Value *ScalarOp = IEI->getOperand(1);
9639 Value *IdxOp = IEI->getOperand(2);
9641 if (!isa<ConstantInt>(IdxOp))
9643 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9645 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9646 // Okay, we can handle this if the vector we are insertinting into is
9648 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9649 // If so, update the mask to reflect the inserted undef.
9650 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9653 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9654 if (isa<ConstantInt>(EI->getOperand(1)) &&
9655 EI->getOperand(0)->getType() == V->getType()) {
9656 unsigned ExtractedIdx =
9657 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9659 // This must be extracting from either LHS or RHS.
9660 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9661 // Okay, we can handle this if the vector we are insertinting into is
9663 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9664 // If so, update the mask to reflect the inserted value.
9665 if (EI->getOperand(0) == LHS) {
9666 Mask[InsertedIdx & (NumElts-1)] =
9667 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9669 assert(EI->getOperand(0) == RHS);
9670 Mask[InsertedIdx & (NumElts-1)] =
9671 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9680 // TODO: Handle shufflevector here!
9685 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9686 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9687 /// that computes V and the LHS value of the shuffle.
9688 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9690 assert(isa<VectorType>(V->getType()) &&
9691 (RHS == 0 || V->getType() == RHS->getType()) &&
9692 "Invalid shuffle!");
9693 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9695 if (isa<UndefValue>(V)) {
9696 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9698 } else if (isa<ConstantAggregateZero>(V)) {
9699 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9701 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9702 // If this is an insert of an extract from some other vector, include it.
9703 Value *VecOp = IEI->getOperand(0);
9704 Value *ScalarOp = IEI->getOperand(1);
9705 Value *IdxOp = IEI->getOperand(2);
9707 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9708 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9709 EI->getOperand(0)->getType() == V->getType()) {
9710 unsigned ExtractedIdx =
9711 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9712 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9714 // Either the extracted from or inserted into vector must be RHSVec,
9715 // otherwise we'd end up with a shuffle of three inputs.
9716 if (EI->getOperand(0) == RHS || RHS == 0) {
9717 RHS = EI->getOperand(0);
9718 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9719 Mask[InsertedIdx & (NumElts-1)] =
9720 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9725 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9726 // Everything but the extracted element is replaced with the RHS.
9727 for (unsigned i = 0; i != NumElts; ++i) {
9728 if (i != InsertedIdx)
9729 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9734 // If this insertelement is a chain that comes from exactly these two
9735 // vectors, return the vector and the effective shuffle.
9736 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9737 return EI->getOperand(0);
9742 // TODO: Handle shufflevector here!
9744 // Otherwise, can't do anything fancy. Return an identity vector.
9745 for (unsigned i = 0; i != NumElts; ++i)
9746 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9750 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9751 Value *VecOp = IE.getOperand(0);
9752 Value *ScalarOp = IE.getOperand(1);
9753 Value *IdxOp = IE.getOperand(2);
9755 // If the inserted element was extracted from some other vector, and if the
9756 // indexes are constant, try to turn this into a shufflevector operation.
9757 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9758 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9759 EI->getOperand(0)->getType() == IE.getType()) {
9760 unsigned NumVectorElts = IE.getType()->getNumElements();
9761 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9762 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9764 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9765 return ReplaceInstUsesWith(IE, VecOp);
9767 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9768 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9770 // If we are extracting a value from a vector, then inserting it right
9771 // back into the same place, just use the input vector.
9772 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9773 return ReplaceInstUsesWith(IE, VecOp);
9775 // We could theoretically do this for ANY input. However, doing so could
9776 // turn chains of insertelement instructions into a chain of shufflevector
9777 // instructions, and right now we do not merge shufflevectors. As such,
9778 // only do this in a situation where it is clear that there is benefit.
9779 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9780 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9781 // the values of VecOp, except then one read from EIOp0.
9782 // Build a new shuffle mask.
9783 std::vector<Constant*> Mask;
9784 if (isa<UndefValue>(VecOp))
9785 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9787 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9788 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9791 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9792 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9793 ConstantVector::get(Mask));
9796 // If this insertelement isn't used by some other insertelement, turn it
9797 // (and any insertelements it points to), into one big shuffle.
9798 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9799 std::vector<Constant*> Mask;
9801 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9802 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9803 // We now have a shuffle of LHS, RHS, Mask.
9804 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9813 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9814 Value *LHS = SVI.getOperand(0);
9815 Value *RHS = SVI.getOperand(1);
9816 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9818 bool MadeChange = false;
9820 // Undefined shuffle mask -> undefined value.
9821 if (isa<UndefValue>(SVI.getOperand(2)))
9822 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9824 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9825 // the undef, change them to undefs.
9826 if (isa<UndefValue>(SVI.getOperand(1))) {
9827 // Scan to see if there are any references to the RHS. If so, replace them
9828 // with undef element refs and set MadeChange to true.
9829 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9830 if (Mask[i] >= e && Mask[i] != 2*e) {
9837 // Remap any references to RHS to use LHS.
9838 std::vector<Constant*> Elts;
9839 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9841 Elts.push_back(UndefValue::get(Type::Int32Ty));
9843 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9845 SVI.setOperand(2, ConstantVector::get(Elts));
9849 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9850 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9851 if (LHS == RHS || isa<UndefValue>(LHS)) {
9852 if (isa<UndefValue>(LHS) && LHS == RHS) {
9853 // shuffle(undef,undef,mask) -> undef.
9854 return ReplaceInstUsesWith(SVI, LHS);
9857 // Remap any references to RHS to use LHS.
9858 std::vector<Constant*> Elts;
9859 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9861 Elts.push_back(UndefValue::get(Type::Int32Ty));
9863 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9864 (Mask[i] < e && isa<UndefValue>(LHS)))
9865 Mask[i] = 2*e; // Turn into undef.
9867 Mask[i] &= (e-1); // Force to LHS.
9868 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9871 SVI.setOperand(0, SVI.getOperand(1));
9872 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9873 SVI.setOperand(2, ConstantVector::get(Elts));
9874 LHS = SVI.getOperand(0);
9875 RHS = SVI.getOperand(1);
9879 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9880 bool isLHSID = true, isRHSID = true;
9882 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9883 if (Mask[i] >= e*2) continue; // Ignore undef values.
9884 // Is this an identity shuffle of the LHS value?
9885 isLHSID &= (Mask[i] == i);
9887 // Is this an identity shuffle of the RHS value?
9888 isRHSID &= (Mask[i]-e == i);
9891 // Eliminate identity shuffles.
9892 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9893 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9895 // If the LHS is a shufflevector itself, see if we can combine it with this
9896 // one without producing an unusual shuffle. Here we are really conservative:
9897 // we are absolutely afraid of producing a shuffle mask not in the input
9898 // program, because the code gen may not be smart enough to turn a merged
9899 // shuffle into two specific shuffles: it may produce worse code. As such,
9900 // we only merge two shuffles if the result is one of the two input shuffle
9901 // masks. In this case, merging the shuffles just removes one instruction,
9902 // which we know is safe. This is good for things like turning:
9903 // (splat(splat)) -> splat.
9904 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9905 if (isa<UndefValue>(RHS)) {
9906 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9908 std::vector<unsigned> NewMask;
9909 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9911 NewMask.push_back(2*e);
9913 NewMask.push_back(LHSMask[Mask[i]]);
9915 // If the result mask is equal to the src shuffle or this shuffle mask, do
9917 if (NewMask == LHSMask || NewMask == Mask) {
9918 std::vector<Constant*> Elts;
9919 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9920 if (NewMask[i] >= e*2) {
9921 Elts.push_back(UndefValue::get(Type::Int32Ty));
9923 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9926 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9927 LHSSVI->getOperand(1),
9928 ConstantVector::get(Elts));
9933 return MadeChange ? &SVI : 0;
9939 /// TryToSinkInstruction - Try to move the specified instruction from its
9940 /// current block into the beginning of DestBlock, which can only happen if it's
9941 /// safe to move the instruction past all of the instructions between it and the
9942 /// end of its block.
9943 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9944 assert(I->hasOneUse() && "Invariants didn't hold!");
9946 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9947 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9949 // Do not sink alloca instructions out of the entry block.
9950 if (isa<AllocaInst>(I) && I->getParent() ==
9951 &DestBlock->getParent()->getEntryBlock())
9954 // We can only sink load instructions if there is nothing between the load and
9955 // the end of block that could change the value.
9956 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9957 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9959 if (Scan->mayWriteToMemory())
9963 BasicBlock::iterator InsertPos = DestBlock->begin();
9964 while (isa<PHINode>(InsertPos)) ++InsertPos;
9966 I->moveBefore(InsertPos);
9972 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9973 /// all reachable code to the worklist.
9975 /// This has a couple of tricks to make the code faster and more powerful. In
9976 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9977 /// them to the worklist (this significantly speeds up instcombine on code where
9978 /// many instructions are dead or constant). Additionally, if we find a branch
9979 /// whose condition is a known constant, we only visit the reachable successors.
9981 static void AddReachableCodeToWorklist(BasicBlock *BB,
9982 SmallPtrSet<BasicBlock*, 64> &Visited,
9984 const TargetData *TD) {
9985 // We have now visited this block! If we've already been here, bail out.
9986 if (!Visited.insert(BB)) return;
9988 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9989 Instruction *Inst = BBI++;
9991 // DCE instruction if trivially dead.
9992 if (isInstructionTriviallyDead(Inst)) {
9994 DOUT << "IC: DCE: " << *Inst;
9995 Inst->eraseFromParent();
9999 // ConstantProp instruction if trivially constant.
10000 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10001 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10002 Inst->replaceAllUsesWith(C);
10004 Inst->eraseFromParent();
10008 IC.AddToWorkList(Inst);
10011 // Recursively visit successors. If this is a branch or switch on a constant,
10012 // only visit the reachable successor.
10013 TerminatorInst *TI = BB->getTerminator();
10014 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10015 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10016 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10017 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, IC, TD);
10020 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10021 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10022 // See if this is an explicit destination.
10023 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10024 if (SI->getCaseValue(i) == Cond) {
10025 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, IC, TD);
10029 // Otherwise it is the default destination.
10030 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, IC, TD);
10035 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10036 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, IC, TD);
10039 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10040 bool Changed = false;
10041 TD = &getAnalysis<TargetData>();
10043 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10044 << F.getNameStr() << "\n");
10047 // Do a depth-first traversal of the function, populate the worklist with
10048 // the reachable instructions. Ignore blocks that are not reachable. Keep
10049 // track of which blocks we visit.
10050 SmallPtrSet<BasicBlock*, 64> Visited;
10051 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10053 // Do a quick scan over the function. If we find any blocks that are
10054 // unreachable, remove any instructions inside of them. This prevents
10055 // the instcombine code from having to deal with some bad special cases.
10056 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10057 if (!Visited.count(BB)) {
10058 Instruction *Term = BB->getTerminator();
10059 while (Term != BB->begin()) { // Remove instrs bottom-up
10060 BasicBlock::iterator I = Term; --I;
10062 DOUT << "IC: DCE: " << *I;
10065 if (!I->use_empty())
10066 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10067 I->eraseFromParent();
10072 while (!Worklist.empty()) {
10073 Instruction *I = RemoveOneFromWorkList();
10074 if (I == 0) continue; // skip null values.
10076 // Check to see if we can DCE the instruction.
10077 if (isInstructionTriviallyDead(I)) {
10078 // Add operands to the worklist.
10079 if (I->getNumOperands() < 4)
10080 AddUsesToWorkList(*I);
10083 DOUT << "IC: DCE: " << *I;
10085 I->eraseFromParent();
10086 RemoveFromWorkList(I);
10090 // Instruction isn't dead, see if we can constant propagate it.
10091 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10092 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10094 // Add operands to the worklist.
10095 AddUsesToWorkList(*I);
10096 ReplaceInstUsesWith(*I, C);
10099 I->eraseFromParent();
10100 RemoveFromWorkList(I);
10104 // See if we can trivially sink this instruction to a successor basic block.
10105 if (I->hasOneUse()) {
10106 BasicBlock *BB = I->getParent();
10107 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10108 if (UserParent != BB) {
10109 bool UserIsSuccessor = false;
10110 // See if the user is one of our successors.
10111 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10112 if (*SI == UserParent) {
10113 UserIsSuccessor = true;
10117 // If the user is one of our immediate successors, and if that successor
10118 // only has us as a predecessors (we'd have to split the critical edge
10119 // otherwise), we can keep going.
10120 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10121 next(pred_begin(UserParent)) == pred_end(UserParent))
10122 // Okay, the CFG is simple enough, try to sink this instruction.
10123 Changed |= TryToSinkInstruction(I, UserParent);
10127 // Now that we have an instruction, try combining it to simplify it...
10128 if (Instruction *Result = visit(*I)) {
10130 // Should we replace the old instruction with a new one?
10132 DOUT << "IC: Old = " << *I
10133 << " New = " << *Result;
10135 // Everything uses the new instruction now.
10136 I->replaceAllUsesWith(Result);
10138 // Push the new instruction and any users onto the worklist.
10139 AddToWorkList(Result);
10140 AddUsersToWorkList(*Result);
10142 // Move the name to the new instruction first.
10143 Result->takeName(I);
10145 // Insert the new instruction into the basic block...
10146 BasicBlock *InstParent = I->getParent();
10147 BasicBlock::iterator InsertPos = I;
10149 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10150 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10153 InstParent->getInstList().insert(InsertPos, Result);
10155 // Make sure that we reprocess all operands now that we reduced their
10157 AddUsesToWorkList(*I);
10159 // Instructions can end up on the worklist more than once. Make sure
10160 // we do not process an instruction that has been deleted.
10161 RemoveFromWorkList(I);
10163 // Erase the old instruction.
10164 InstParent->getInstList().erase(I);
10166 DOUT << "IC: MOD = " << *I;
10168 // If the instruction was modified, it's possible that it is now dead.
10169 // if so, remove it.
10170 if (isInstructionTriviallyDead(I)) {
10171 // Make sure we process all operands now that we are reducing their
10173 AddUsesToWorkList(*I);
10175 // Instructions may end up in the worklist more than once. Erase all
10176 // occurrences of this instruction.
10177 RemoveFromWorkList(I);
10178 I->eraseFromParent();
10181 AddUsersToWorkList(*I);
10188 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10193 bool InstCombiner::runOnFunction(Function &F) {
10194 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10196 bool EverMadeChange = false;
10198 // Iterate while there is work to do.
10199 unsigned Iteration = 0;
10200 while (DoOneIteration(F, Iteration++))
10201 EverMadeChange = true;
10202 return EverMadeChange;
10205 FunctionPass *llvm::createInstructionCombiningPass() {
10206 return new InstCombiner();