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;
996 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
997 /// this predicate to simplify operations downstream. Mask is known to be zero
998 /// for bits that V cannot have.
999 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
1000 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1001 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
1002 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1003 return (KnownZero & Mask) == Mask;
1007 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1008 /// specified instruction is a constant integer. If so, check to see if there
1009 /// are any bits set in the constant that are not demanded. If so, shrink the
1010 /// constant and return true.
1011 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
1012 uint64_t Demanded) {
1013 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
1014 if (!OpC) return false;
1016 // If there are no bits set that aren't demanded, nothing to do.
1017 if ((~Demanded & OpC->getZExtValue()) == 0)
1020 // This is producing any bits that are not needed, shrink the RHS.
1021 uint64_t Val = Demanded & OpC->getZExtValue();
1022 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Val));
1026 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1027 /// specified instruction is a constant integer. If so, check to see if there
1028 /// are any bits set in the constant that are not demanded. If so, shrink the
1029 /// constant and return true.
1030 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
1032 assert(I && "No instruction?");
1033 assert(OpNo < I->getNumOperands() && "Operand index too large");
1035 // If the operand is not a constant integer, nothing to do.
1036 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
1037 if (!OpC) return false;
1039 // If there are no bits set that aren't demanded, nothing to do.
1040 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
1041 if ((~Demanded & OpC->getValue()) == 0)
1044 // This instruction is producing bits that are not demanded. Shrink the RHS.
1045 Demanded &= OpC->getValue();
1046 I->setOperand(OpNo, ConstantInt::get(Demanded));
1050 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
1051 // set of known zero and one bits, compute the maximum and minimum values that
1052 // could have the specified known zero and known one bits, returning them in
1054 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
1057 int64_t &Min, int64_t &Max) {
1058 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
1059 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
1061 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
1063 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
1064 // bit if it is unknown.
1066 Max = KnownOne|UnknownBits;
1068 if (SignBit & UnknownBits) { // Sign bit is unknown
1073 // Sign extend the min/max values.
1074 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
1075 Min = (Min << ShAmt) >> ShAmt;
1076 Max = (Max << ShAmt) >> ShAmt;
1079 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
1080 // a set of known zero and one bits, compute the maximum and minimum values that
1081 // could have the specified known zero and known one bits, returning them in
1083 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
1088 uint64_t TypeBits = cast<IntegerType>(Ty)->getBitMask();
1089 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
1091 // The minimum value is when the unknown bits are all zeros.
1093 // The maximum value is when the unknown bits are all ones.
1094 Max = KnownOne|UnknownBits;
1098 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
1099 /// DemandedMask bits of the result of V are ever used downstream. If we can
1100 /// use this information to simplify V, do so and return true. Otherwise,
1101 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
1102 /// the expression (used to simplify the caller). The KnownZero/One bits may
1103 /// only be accurate for those bits in the DemandedMask.
1104 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
1105 uint64_t &KnownZero, uint64_t &KnownOne,
1107 const IntegerType *VTy = cast<IntegerType>(V->getType());
1108 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1109 // We know all of the bits for a constant!
1110 KnownOne = CI->getZExtValue() & DemandedMask;
1111 KnownZero = ~KnownOne & DemandedMask;
1115 KnownZero = KnownOne = 0;
1116 if (!V->hasOneUse()) { // Other users may use these bits.
1117 if (Depth != 0) { // Not at the root.
1118 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
1119 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
1122 // If this is the root being simplified, allow it to have multiple uses,
1123 // just set the DemandedMask to all bits.
1124 DemandedMask = VTy->getBitMask();
1125 } else if (DemandedMask == 0) { // Not demanding any bits from V.
1126 if (V != UndefValue::get(VTy))
1127 return UpdateValueUsesWith(V, UndefValue::get(VTy));
1129 } else if (Depth == 6) { // Limit search depth.
1133 Instruction *I = dyn_cast<Instruction>(V);
1134 if (!I) return false; // Only analyze instructions.
1136 DemandedMask &= VTy->getBitMask();
1138 uint64_t KnownZero2 = 0, KnownOne2 = 0;
1139 switch (I->getOpcode()) {
1141 case Instruction::And:
1142 // If either the LHS or the RHS are Zero, the result is zero.
1143 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1144 KnownZero, KnownOne, Depth+1))
1146 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1148 // If something is known zero on the RHS, the bits aren't demanded on the
1150 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
1151 KnownZero2, KnownOne2, Depth+1))
1153 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1155 // If all of the demanded bits are known 1 on one side, return the other.
1156 // These bits cannot contribute to the result of the 'and'.
1157 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
1158 return UpdateValueUsesWith(I, I->getOperand(0));
1159 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
1160 return UpdateValueUsesWith(I, I->getOperand(1));
1162 // If all of the demanded bits in the inputs are known zeros, return zero.
1163 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
1164 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1166 // If the RHS is a constant, see if we can simplify it.
1167 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
1168 return UpdateValueUsesWith(I, I);
1170 // Output known-1 bits are only known if set in both the LHS & RHS.
1171 KnownOne &= KnownOne2;
1172 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1173 KnownZero |= KnownZero2;
1175 case Instruction::Or:
1176 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1177 KnownZero, KnownOne, Depth+1))
1179 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1180 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
1181 KnownZero2, KnownOne2, Depth+1))
1183 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1185 // If all of the demanded bits are known zero on one side, return the other.
1186 // These bits cannot contribute to the result of the 'or'.
1187 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
1188 return UpdateValueUsesWith(I, I->getOperand(0));
1189 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
1190 return UpdateValueUsesWith(I, I->getOperand(1));
1192 // If all of the potentially set bits on one side are known to be set on
1193 // the other side, just use the 'other' side.
1194 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
1195 (DemandedMask & (~KnownZero)))
1196 return UpdateValueUsesWith(I, I->getOperand(0));
1197 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
1198 (DemandedMask & (~KnownZero2)))
1199 return UpdateValueUsesWith(I, I->getOperand(1));
1201 // If the RHS is a constant, see if we can simplify it.
1202 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1203 return UpdateValueUsesWith(I, I);
1205 // Output known-0 bits are only known if clear in both the LHS & RHS.
1206 KnownZero &= KnownZero2;
1207 // Output known-1 are known to be set if set in either the LHS | RHS.
1208 KnownOne |= KnownOne2;
1210 case Instruction::Xor: {
1211 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1212 KnownZero, KnownOne, Depth+1))
1214 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1215 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1216 KnownZero2, KnownOne2, Depth+1))
1218 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1220 // If all of the demanded bits are known zero on one side, return the other.
1221 // These bits cannot contribute to the result of the 'xor'.
1222 if ((DemandedMask & KnownZero) == DemandedMask)
1223 return UpdateValueUsesWith(I, I->getOperand(0));
1224 if ((DemandedMask & KnownZero2) == DemandedMask)
1225 return UpdateValueUsesWith(I, I->getOperand(1));
1227 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1228 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
1229 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1230 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
1232 // If all of the demanded bits are known to be zero on one side or the
1233 // other, turn this into an *inclusive* or.
1234 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1235 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
1237 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1239 InsertNewInstBefore(Or, *I);
1240 return UpdateValueUsesWith(I, Or);
1243 // If all of the demanded bits on one side are known, and all of the set
1244 // bits on that side are also known to be set on the other side, turn this
1245 // into an AND, as we know the bits will be cleared.
1246 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1247 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
1248 if ((KnownOne & KnownOne2) == KnownOne) {
1249 Constant *AndC = ConstantInt::get(VTy, ~KnownOne & DemandedMask);
1251 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1252 InsertNewInstBefore(And, *I);
1253 return UpdateValueUsesWith(I, And);
1257 // If the RHS is a constant, see if we can simplify it.
1258 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1259 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1260 return UpdateValueUsesWith(I, I);
1262 KnownZero = KnownZeroOut;
1263 KnownOne = KnownOneOut;
1266 case Instruction::Select:
1267 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1268 KnownZero, KnownOne, Depth+1))
1270 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1271 KnownZero2, KnownOne2, Depth+1))
1273 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1274 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1276 // If the operands are constants, see if we can simplify them.
1277 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1278 return UpdateValueUsesWith(I, I);
1279 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1280 return UpdateValueUsesWith(I, I);
1282 // Only known if known in both the LHS and RHS.
1283 KnownOne &= KnownOne2;
1284 KnownZero &= KnownZero2;
1286 case Instruction::Trunc:
1287 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1288 KnownZero, KnownOne, Depth+1))
1290 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1292 case Instruction::BitCast:
1293 if (!I->getOperand(0)->getType()->isInteger())
1296 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1297 KnownZero, KnownOne, Depth+1))
1299 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1301 case Instruction::ZExt: {
1302 // Compute the bits in the result that are not present in the input.
1303 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1304 uint64_t NotIn = ~SrcTy->getBitMask();
1305 uint64_t NewBits = VTy->getBitMask() & NotIn;
1307 DemandedMask &= SrcTy->getBitMask();
1308 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1309 KnownZero, KnownOne, Depth+1))
1311 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1312 // The top bits are known to be zero.
1313 KnownZero |= NewBits;
1316 case Instruction::SExt: {
1317 // Compute the bits in the result that are not present in the input.
1318 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1319 uint64_t NotIn = ~SrcTy->getBitMask();
1320 uint64_t NewBits = VTy->getBitMask() & NotIn;
1322 // Get the sign bit for the source type
1323 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1324 int64_t InputDemandedBits = DemandedMask & SrcTy->getBitMask();
1326 // If any of the sign extended bits are demanded, we know that the sign
1328 if (NewBits & DemandedMask)
1329 InputDemandedBits |= InSignBit;
1331 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1332 KnownZero, KnownOne, Depth+1))
1334 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1336 // If the sign bit of the input is known set or clear, then we know the
1337 // top bits of the result.
1339 // If the input sign bit is known zero, or if the NewBits are not demanded
1340 // convert this into a zero extension.
1341 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1342 // Convert to ZExt cast
1343 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1344 return UpdateValueUsesWith(I, NewCast);
1345 } else if (KnownOne & InSignBit) { // Input sign bit known set
1346 KnownOne |= NewBits;
1347 KnownZero &= ~NewBits;
1348 } else { // Input sign bit unknown
1349 KnownZero &= ~NewBits;
1350 KnownOne &= ~NewBits;
1354 case Instruction::Add:
1355 // If there is a constant on the RHS, there are a variety of xformations
1357 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1358 // If null, this should be simplified elsewhere. Some of the xforms here
1359 // won't work if the RHS is zero.
1360 if (RHS->isNullValue())
1363 // Figure out what the input bits are. If the top bits of the and result
1364 // are not demanded, then the add doesn't demand them from its input
1367 // Shift the demanded mask up so that it's at the top of the uint64_t.
1368 unsigned BitWidth = VTy->getPrimitiveSizeInBits();
1369 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1371 // If the top bit of the output is demanded, demand everything from the
1372 // input. Otherwise, we demand all the input bits except NLZ top bits.
1373 uint64_t InDemandedBits = ~0ULL >> (64-BitWidth+NLZ);
1375 // Find information about known zero/one bits in the input.
1376 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1377 KnownZero2, KnownOne2, Depth+1))
1380 // If the RHS of the add has bits set that can't affect the input, reduce
1382 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1383 return UpdateValueUsesWith(I, I);
1385 // Avoid excess work.
1386 if (KnownZero2 == 0 && KnownOne2 == 0)
1389 // Turn it into OR if input bits are zero.
1390 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1392 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1394 InsertNewInstBefore(Or, *I);
1395 return UpdateValueUsesWith(I, Or);
1398 // We can say something about the output known-zero and known-one bits,
1399 // depending on potential carries from the input constant and the
1400 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1401 // bits set and the RHS constant is 0x01001, then we know we have a known
1402 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1404 // To compute this, we first compute the potential carry bits. These are
1405 // the bits which may be modified. I'm not aware of a better way to do
1407 uint64_t RHSVal = RHS->getZExtValue();
1409 bool CarryIn = false;
1410 uint64_t CarryBits = 0;
1411 uint64_t CurBit = 1;
1412 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1413 // Record the current carry in.
1414 if (CarryIn) CarryBits |= CurBit;
1418 // This bit has a carry out unless it is "zero + zero" or
1419 // "zero + anything" with no carry in.
1420 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1421 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1422 } else if (!CarryIn &&
1423 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1424 CarryOut = false; // 0 + anything has no carry out if no carry in.
1426 // Otherwise, we have to assume we have a carry out.
1430 // This stage's carry out becomes the next stage's carry-in.
1434 // Now that we know which bits have carries, compute the known-1/0 sets.
1436 // Bits are known one if they are known zero in one operand and one in the
1437 // other, and there is no input carry.
1438 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1440 // Bits are known zero if they are known zero in both operands and there
1441 // is no input carry.
1442 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1444 // If the high-bits of this ADD are not demanded, then it does not demand
1445 // the high bits of its LHS or RHS.
1446 if ((DemandedMask & VTy->getSignBit()) == 0) {
1447 // Right fill the mask of bits for this ADD to demand the most
1448 // significant bit and all those below it.
1449 unsigned NLZ = CountLeadingZeros_64(DemandedMask);
1450 uint64_t DemandedFromOps = ~0ULL >> NLZ;
1451 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1452 KnownZero2, KnownOne2, Depth+1))
1454 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1455 KnownZero2, KnownOne2, Depth+1))
1460 case Instruction::Sub:
1461 // If the high-bits of this SUB are not demanded, then it does not demand
1462 // the high bits of its LHS or RHS.
1463 if ((DemandedMask & VTy->getSignBit()) == 0) {
1464 // Right fill the mask of bits for this SUB to demand the most
1465 // significant bit and all those below it.
1466 unsigned NLZ = CountLeadingZeros_64(DemandedMask);
1467 uint64_t DemandedFromOps = ~0ULL >> NLZ;
1468 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1469 KnownZero2, KnownOne2, Depth+1))
1471 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1472 KnownZero2, KnownOne2, Depth+1))
1476 case Instruction::Shl:
1477 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1478 uint64_t ShiftAmt = SA->getZExtValue();
1479 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1480 KnownZero, KnownOne, Depth+1))
1482 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1483 KnownZero <<= ShiftAmt;
1484 KnownOne <<= ShiftAmt;
1485 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1488 case Instruction::LShr:
1489 // For a logical shift right
1490 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1491 unsigned ShiftAmt = SA->getZExtValue();
1493 // Compute the new bits that are at the top now.
1494 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1495 HighBits <<= VTy->getBitWidth() - ShiftAmt;
1496 uint64_t TypeMask = VTy->getBitMask();
1497 // Unsigned shift right.
1498 if (SimplifyDemandedBits(I->getOperand(0),
1499 (DemandedMask << ShiftAmt) & TypeMask,
1500 KnownZero, KnownOne, Depth+1))
1502 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1503 KnownZero &= TypeMask;
1504 KnownOne &= TypeMask;
1505 KnownZero >>= ShiftAmt;
1506 KnownOne >>= ShiftAmt;
1507 KnownZero |= HighBits; // high bits known zero.
1510 case Instruction::AShr:
1511 // If this is an arithmetic shift right and only the low-bit is set, we can
1512 // always convert this into a logical shr, even if the shift amount is
1513 // variable. The low bit of the shift cannot be an input sign bit unless
1514 // the shift amount is >= the size of the datatype, which is undefined.
1515 if (DemandedMask == 1) {
1516 // Perform the logical shift right.
1517 Value *NewVal = BinaryOperator::createLShr(
1518 I->getOperand(0), I->getOperand(1), I->getName());
1519 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1520 return UpdateValueUsesWith(I, NewVal);
1523 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1524 unsigned ShiftAmt = SA->getZExtValue();
1526 // Compute the new bits that are at the top now.
1527 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1528 HighBits <<= VTy->getBitWidth() - ShiftAmt;
1529 uint64_t TypeMask = VTy->getBitMask();
1530 // Signed shift right.
1531 if (SimplifyDemandedBits(I->getOperand(0),
1532 (DemandedMask << ShiftAmt) & TypeMask,
1533 KnownZero, KnownOne, Depth+1))
1535 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1536 KnownZero &= TypeMask;
1537 KnownOne &= TypeMask;
1538 KnownZero >>= ShiftAmt;
1539 KnownOne >>= ShiftAmt;
1541 // Handle the sign bits.
1542 uint64_t SignBit = 1ULL << (VTy->getBitWidth()-1);
1543 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1545 // If the input sign bit is known to be zero, or if none of the top bits
1546 // are demanded, turn this into an unsigned shift right.
1547 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1548 // Perform the logical shift right.
1549 Value *NewVal = BinaryOperator::createLShr(
1550 I->getOperand(0), SA, I->getName());
1551 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1552 return UpdateValueUsesWith(I, NewVal);
1553 } else if (KnownOne & SignBit) { // New bits are known one.
1554 KnownOne |= HighBits;
1560 // If the client is only demanding bits that we know, return the known
1562 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1563 return UpdateValueUsesWith(I, ConstantInt::get(VTy, KnownOne));
1567 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
1568 /// value based on the demanded bits. When this function is called, it is known
1569 /// that only the bits set in DemandedMask of the result of V are ever used
1570 /// downstream. Consequently, depending on the mask and V, it may be possible
1571 /// to replace V with a constant or one of its operands. In such cases, this
1572 /// function does the replacement and returns true. In all other cases, it
1573 /// returns false after analyzing the expression and setting KnownOne and known
1574 /// to be one in the expression. KnownZero contains all the bits that are known
1575 /// to be zero in the expression. These are provided to potentially allow the
1576 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
1577 /// the expression. KnownOne and KnownZero always follow the invariant that
1578 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
1579 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
1580 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
1581 /// and KnownOne must all be the same.
1582 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
1583 APInt& KnownZero, APInt& KnownOne,
1585 assert(V != 0 && "Null pointer of Value???");
1586 assert(Depth <= 6 && "Limit Search Depth");
1587 uint32_t BitWidth = DemandedMask.getBitWidth();
1588 const IntegerType *VTy = cast<IntegerType>(V->getType());
1589 assert(VTy->getBitWidth() == BitWidth &&
1590 KnownZero.getBitWidth() == BitWidth &&
1591 KnownOne.getBitWidth() == BitWidth &&
1592 "Value *V, DemandedMask, KnownZero and KnownOne \
1593 must have same BitWidth");
1594 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1595 // We know all of the bits for a constant!
1596 KnownOne = CI->getValue() & DemandedMask;
1597 KnownZero = ~KnownOne & DemandedMask;
1603 if (!V->hasOneUse()) { // Other users may use these bits.
1604 if (Depth != 0) { // Not at the root.
1605 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
1606 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
1609 // If this is the root being simplified, allow it to have multiple uses,
1610 // just set the DemandedMask to all bits.
1611 DemandedMask = APInt::getAllOnesValue(BitWidth);
1612 } else if (DemandedMask == 0) { // Not demanding any bits from V.
1613 if (V != UndefValue::get(VTy))
1614 return UpdateValueUsesWith(V, UndefValue::get(VTy));
1616 } else if (Depth == 6) { // Limit search depth.
1620 Instruction *I = dyn_cast<Instruction>(V);
1621 if (!I) return false; // Only analyze instructions.
1623 DemandedMask &= APInt::getAllOnesValue(BitWidth);
1625 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1626 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
1627 switch (I->getOpcode()) {
1629 case Instruction::And:
1630 // If either the LHS or the RHS are Zero, the result is zero.
1631 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1632 RHSKnownZero, RHSKnownOne, Depth+1))
1634 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1635 "Bits known to be one AND zero?");
1637 // If something is known zero on the RHS, the bits aren't demanded on the
1639 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
1640 LHSKnownZero, LHSKnownOne, Depth+1))
1642 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1643 "Bits known to be one AND zero?");
1645 // If all of the demanded bits are known 1 on one side, return the other.
1646 // These bits cannot contribute to the result of the 'and'.
1647 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1648 (DemandedMask & ~LHSKnownZero))
1649 return UpdateValueUsesWith(I, I->getOperand(0));
1650 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1651 (DemandedMask & ~RHSKnownZero))
1652 return UpdateValueUsesWith(I, I->getOperand(1));
1654 // If all of the demanded bits in the inputs are known zeros, return zero.
1655 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1656 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1658 // If the RHS is a constant, see if we can simplify it.
1659 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1660 return UpdateValueUsesWith(I, I);
1662 // Output known-1 bits are only known if set in both the LHS & RHS.
1663 RHSKnownOne &= LHSKnownOne;
1664 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1665 RHSKnownZero |= LHSKnownZero;
1667 case Instruction::Or:
1668 // If either the LHS or the RHS are One, the result is One.
1669 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1670 RHSKnownZero, RHSKnownOne, Depth+1))
1672 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1673 "Bits known to be one AND zero?");
1674 // If something is known one on the RHS, the bits aren't demanded on the
1676 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1677 LHSKnownZero, LHSKnownOne, Depth+1))
1679 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1680 "Bits known to be one AND zero?");
1682 // If all of the demanded bits are known zero on one side, return the other.
1683 // These bits cannot contribute to the result of the 'or'.
1684 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1685 (DemandedMask & ~LHSKnownOne))
1686 return UpdateValueUsesWith(I, I->getOperand(0));
1687 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1688 (DemandedMask & ~RHSKnownOne))
1689 return UpdateValueUsesWith(I, I->getOperand(1));
1691 // If all of the potentially set bits on one side are known to be set on
1692 // the other side, just use the 'other' side.
1693 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1694 (DemandedMask & (~RHSKnownZero)))
1695 return UpdateValueUsesWith(I, I->getOperand(0));
1696 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1697 (DemandedMask & (~LHSKnownZero)))
1698 return UpdateValueUsesWith(I, I->getOperand(1));
1700 // If the RHS is a constant, see if we can simplify it.
1701 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1702 return UpdateValueUsesWith(I, I);
1704 // Output known-0 bits are only known if clear in both the LHS & RHS.
1705 RHSKnownZero &= LHSKnownZero;
1706 // Output known-1 are known to be set if set in either the LHS | RHS.
1707 RHSKnownOne |= LHSKnownOne;
1709 case Instruction::Xor: {
1710 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1711 RHSKnownZero, RHSKnownOne, Depth+1))
1713 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1714 "Bits known to be one AND zero?");
1715 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1716 LHSKnownZero, LHSKnownOne, Depth+1))
1718 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1719 "Bits known to be one AND zero?");
1721 // If all of the demanded bits are known zero on one side, return the other.
1722 // These bits cannot contribute to the result of the 'xor'.
1723 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1724 return UpdateValueUsesWith(I, I->getOperand(0));
1725 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1726 return UpdateValueUsesWith(I, I->getOperand(1));
1728 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1729 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1730 (RHSKnownOne & LHSKnownOne);
1731 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1732 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1733 (RHSKnownOne & LHSKnownZero);
1735 // If all of the demanded bits are known to be zero on one side or the
1736 // other, turn this into an *inclusive* or.
1737 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1738 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1740 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1742 InsertNewInstBefore(Or, *I);
1743 return UpdateValueUsesWith(I, Or);
1746 // If all of the demanded bits on one side are known, and all of the set
1747 // bits on that side are also known to be set on the other side, turn this
1748 // into an AND, as we know the bits will be cleared.
1749 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1750 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1752 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1753 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1755 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1756 InsertNewInstBefore(And, *I);
1757 return UpdateValueUsesWith(I, And);
1761 // If the RHS is a constant, see if we can simplify it.
1762 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1763 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1764 return UpdateValueUsesWith(I, I);
1766 RHSKnownZero = KnownZeroOut;
1767 RHSKnownOne = KnownOneOut;
1770 case Instruction::Select:
1771 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1772 RHSKnownZero, RHSKnownOne, Depth+1))
1774 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1775 LHSKnownZero, LHSKnownOne, Depth+1))
1777 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1778 "Bits known to be one AND zero?");
1779 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1780 "Bits known to be one AND zero?");
1782 // If the operands are constants, see if we can simplify them.
1783 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1784 return UpdateValueUsesWith(I, I);
1785 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1786 return UpdateValueUsesWith(I, I);
1788 // Only known if known in both the LHS and RHS.
1789 RHSKnownOne &= LHSKnownOne;
1790 RHSKnownZero &= LHSKnownZero;
1792 case Instruction::Trunc: {
1794 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1795 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask.zext(truncBf),
1796 RHSKnownZero.zext(truncBf), RHSKnownOne.zext(truncBf), Depth+1))
1798 DemandedMask.trunc(BitWidth);
1799 RHSKnownZero.trunc(BitWidth);
1800 RHSKnownOne.trunc(BitWidth);
1801 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1802 "Bits known to be one AND zero?");
1805 case Instruction::BitCast:
1806 if (!I->getOperand(0)->getType()->isInteger())
1809 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1810 RHSKnownZero, RHSKnownOne, Depth+1))
1812 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1813 "Bits known to be one AND zero?");
1815 case Instruction::ZExt: {
1816 // Compute the bits in the result that are not present in the input.
1817 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1818 APInt NewBits(APInt::getAllOnesValue(BitWidth).shl(SrcTy->getBitWidth()));
1820 DemandedMask &= SrcTy->getMask().zext(BitWidth);
1821 uint32_t zextBf = SrcTy->getBitWidth();
1822 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask.trunc(zextBf),
1823 RHSKnownZero.trunc(zextBf), RHSKnownOne.trunc(zextBf), Depth+1))
1825 DemandedMask.zext(BitWidth);
1826 RHSKnownZero.zext(BitWidth);
1827 RHSKnownOne.zext(BitWidth);
1828 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1829 "Bits known to be one AND zero?");
1830 // The top bits are known to be zero.
1831 RHSKnownZero |= NewBits;
1834 case Instruction::SExt: {
1835 // Compute the bits in the result that are not present in the input.
1836 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1837 APInt NewBits(APInt::getAllOnesValue(BitWidth).shl(SrcTy->getBitWidth()));
1839 // Get the sign bit for the source type
1840 APInt InSignBit(APInt::getSignBit(SrcTy->getPrimitiveSizeInBits()));
1841 InSignBit.zext(BitWidth);
1842 APInt InputDemandedBits = DemandedMask &
1843 SrcTy->getMask().zext(BitWidth);
1845 // If any of the sign extended bits are demanded, we know that the sign
1847 if ((NewBits & DemandedMask) != 0)
1848 InputDemandedBits |= InSignBit;
1850 uint32_t sextBf = SrcTy->getBitWidth();
1851 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits.trunc(sextBf),
1852 RHSKnownZero.trunc(sextBf), RHSKnownOne.trunc(sextBf), Depth+1))
1854 InputDemandedBits.zext(BitWidth);
1855 RHSKnownZero.zext(BitWidth);
1856 RHSKnownOne.zext(BitWidth);
1857 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1858 "Bits known to be one AND zero?");
1860 // If the sign bit of the input is known set or clear, then we know the
1861 // top bits of the result.
1863 // If the input sign bit is known zero, or if the NewBits are not demanded
1864 // convert this into a zero extension.
1865 if ((RHSKnownZero & InSignBit) != 0 || (NewBits & ~DemandedMask) == NewBits)
1867 // Convert to ZExt cast
1868 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1869 return UpdateValueUsesWith(I, NewCast);
1870 } else if ((RHSKnownOne & InSignBit) != 0) { // Input sign bit known set
1871 RHSKnownOne |= NewBits;
1872 RHSKnownZero &= ~NewBits;
1873 } else { // Input sign bit unknown
1874 RHSKnownZero &= ~NewBits;
1875 RHSKnownOne &= ~NewBits;
1879 case Instruction::Add: {
1880 // Figure out what the input bits are. If the top bits of the and result
1881 // are not demanded, then the add doesn't demand them from its input
1883 unsigned NLZ = DemandedMask.countLeadingZeros();
1885 // If there is a constant on the RHS, there are a variety of xformations
1887 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1888 // If null, this should be simplified elsewhere. Some of the xforms here
1889 // won't work if the RHS is zero.
1893 // If the top bit of the output is demanded, demand everything from the
1894 // input. Otherwise, we demand all the input bits except NLZ top bits.
1895 APInt InDemandedBits(APInt::getAllOnesValue(BitWidth).lshr(NLZ));
1897 // Find information about known zero/one bits in the input.
1898 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1899 LHSKnownZero, LHSKnownOne, Depth+1))
1902 // If the RHS of the add has bits set that can't affect the input, reduce
1904 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1905 return UpdateValueUsesWith(I, I);
1907 // Avoid excess work.
1908 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1911 // Turn it into OR if input bits are zero.
1912 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1914 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1916 InsertNewInstBefore(Or, *I);
1917 return UpdateValueUsesWith(I, Or);
1920 // We can say something about the output known-zero and known-one bits,
1921 // depending on potential carries from the input constant and the
1922 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1923 // bits set and the RHS constant is 0x01001, then we know we have a known
1924 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1926 // To compute this, we first compute the potential carry bits. These are
1927 // the bits which may be modified. I'm not aware of a better way to do
1929 APInt RHSVal(RHS->getValue());
1931 bool CarryIn = false;
1932 APInt CarryBits(BitWidth, 0);
1933 const uint64_t *LHSKnownZeroRawVal = LHSKnownZero.getRawData(),
1934 *RHSRawVal = RHSVal.getRawData();
1935 for (uint32_t i = 0; i != RHSVal.getNumWords(); ++i) {
1936 uint64_t AddVal = ~LHSKnownZeroRawVal[i] + RHSRawVal[i],
1937 XorVal = ~LHSKnownZeroRawVal[i] ^ RHSRawVal[i];
1938 uint64_t WordCarryBits = AddVal ^ XorVal + CarryIn;
1939 if (AddVal < RHSRawVal[i])
1943 CarryBits.setWordToValue(i, WordCarryBits);
1946 // Now that we know which bits have carries, compute the known-1/0 sets.
1948 // Bits are known one if they are known zero in one operand and one in the
1949 // other, and there is no input carry.
1950 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1951 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1953 // Bits are known zero if they are known zero in both operands and there
1954 // is no input carry.
1955 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1957 // If the high-bits of this ADD are not demanded, then it does not demand
1958 // the high bits of its LHS or RHS.
1959 if ((DemandedMask & APInt::getSignBit(BitWidth)) == 0) {
1960 // Right fill the mask of bits for this ADD to demand the most
1961 // significant bit and all those below it.
1962 APInt DemandedFromOps = APInt::getAllOnesValue(BitWidth).lshr(NLZ);
1963 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1964 LHSKnownZero, LHSKnownOne, Depth+1))
1966 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1967 LHSKnownZero, LHSKnownOne, Depth+1))
1973 case Instruction::Sub:
1974 // If the high-bits of this SUB are not demanded, then it does not demand
1975 // the high bits of its LHS or RHS.
1976 if ((DemandedMask & APInt::getSignBit(BitWidth)) == 0) {
1977 // Right fill the mask of bits for this SUB to demand the most
1978 // significant bit and all those below it.
1979 unsigned NLZ = DemandedMask.countLeadingZeros();
1980 APInt DemandedFromOps(APInt::getAllOnesValue(BitWidth).lshr(NLZ));
1981 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1982 LHSKnownZero, LHSKnownOne, Depth+1))
1984 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1985 LHSKnownZero, LHSKnownOne, Depth+1))
1989 case Instruction::Shl:
1990 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1991 uint64_t ShiftAmt = SA->getZExtValue();
1992 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask.lshr(ShiftAmt),
1993 RHSKnownZero, RHSKnownOne, Depth+1))
1995 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1996 "Bits known to be one AND zero?");
1997 RHSKnownZero <<= ShiftAmt;
1998 RHSKnownOne <<= ShiftAmt;
1999 // low bits known zero.
2001 RHSKnownZero |= APInt::getAllOnesValue(ShiftAmt).zextOrCopy(BitWidth);
2004 case Instruction::LShr:
2005 // For a logical shift right
2006 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
2007 unsigned ShiftAmt = SA->getZExtValue();
2009 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
2010 // Unsigned shift right.
2011 if (SimplifyDemandedBits(I->getOperand(0),
2012 (DemandedMask.shl(ShiftAmt)) & TypeMask,
2013 RHSKnownZero, RHSKnownOne, Depth+1))
2015 assert((RHSKnownZero & RHSKnownOne) == 0 &&
2016 "Bits known to be one AND zero?");
2017 RHSKnownZero &= TypeMask;
2018 RHSKnownOne &= TypeMask;
2019 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
2020 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
2022 // Compute the new bits that are at the top now.
2023 APInt HighBits(APInt::getAllOnesValue(BitWidth).shl(
2024 BitWidth - ShiftAmt));
2025 RHSKnownZero |= HighBits; // high bits known zero.
2029 case Instruction::AShr:
2030 // If this is an arithmetic shift right and only the low-bit is set, we can
2031 // always convert this into a logical shr, even if the shift amount is
2032 // variable. The low bit of the shift cannot be an input sign bit unless
2033 // the shift amount is >= the size of the datatype, which is undefined.
2034 if (DemandedMask == 1) {
2035 // Perform the logical shift right.
2036 Value *NewVal = BinaryOperator::createLShr(
2037 I->getOperand(0), I->getOperand(1), I->getName());
2038 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
2039 return UpdateValueUsesWith(I, NewVal);
2042 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
2043 unsigned ShiftAmt = SA->getZExtValue();
2045 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
2046 // Signed shift right.
2047 if (SimplifyDemandedBits(I->getOperand(0),
2048 (DemandedMask.shl(ShiftAmt)) & TypeMask,
2049 RHSKnownZero, RHSKnownOne, Depth+1))
2051 assert((RHSKnownZero & RHSKnownOne) == 0 &&
2052 "Bits known to be one AND zero?");
2053 // Compute the new bits that are at the top now.
2054 APInt HighBits(APInt::getAllOnesValue(BitWidth).shl(BitWidth - ShiftAmt));
2055 RHSKnownZero &= TypeMask;
2056 RHSKnownOne &= TypeMask;
2057 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
2058 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
2060 // Handle the sign bits.
2061 APInt SignBit(APInt::getSignBit(BitWidth));
2062 // Adjust to where it is now in the mask.
2063 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
2065 // If the input sign bit is known to be zero, or if none of the top bits
2066 // are demanded, turn this into an unsigned shift right.
2067 if ((RHSKnownZero & SignBit) != 0 ||
2068 (HighBits & ~DemandedMask) == HighBits) {
2069 // Perform the logical shift right.
2070 Value *NewVal = BinaryOperator::createLShr(
2071 I->getOperand(0), SA, I->getName());
2072 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
2073 return UpdateValueUsesWith(I, NewVal);
2074 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
2075 RHSKnownOne |= HighBits;
2081 // If the client is only demanding bits that we know, return the known
2083 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
2084 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
2089 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
2090 /// 64 or fewer elements. DemandedElts contains the set of elements that are
2091 /// actually used by the caller. This method analyzes which elements of the
2092 /// operand are undef and returns that information in UndefElts.
2094 /// If the information about demanded elements can be used to simplify the
2095 /// operation, the operation is simplified, then the resultant value is
2096 /// returned. This returns null if no change was made.
2097 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
2098 uint64_t &UndefElts,
2100 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
2101 assert(VWidth <= 64 && "Vector too wide to analyze!");
2102 uint64_t EltMask = ~0ULL >> (64-VWidth);
2103 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
2104 "Invalid DemandedElts!");
2106 if (isa<UndefValue>(V)) {
2107 // If the entire vector is undefined, just return this info.
2108 UndefElts = EltMask;
2110 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
2111 UndefElts = EltMask;
2112 return UndefValue::get(V->getType());
2116 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
2117 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
2118 Constant *Undef = UndefValue::get(EltTy);
2120 std::vector<Constant*> Elts;
2121 for (unsigned i = 0; i != VWidth; ++i)
2122 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
2123 Elts.push_back(Undef);
2124 UndefElts |= (1ULL << i);
2125 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
2126 Elts.push_back(Undef);
2127 UndefElts |= (1ULL << i);
2128 } else { // Otherwise, defined.
2129 Elts.push_back(CP->getOperand(i));
2132 // If we changed the constant, return it.
2133 Constant *NewCP = ConstantVector::get(Elts);
2134 return NewCP != CP ? NewCP : 0;
2135 } else if (isa<ConstantAggregateZero>(V)) {
2136 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
2138 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
2139 Constant *Zero = Constant::getNullValue(EltTy);
2140 Constant *Undef = UndefValue::get(EltTy);
2141 std::vector<Constant*> Elts;
2142 for (unsigned i = 0; i != VWidth; ++i)
2143 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
2144 UndefElts = DemandedElts ^ EltMask;
2145 return ConstantVector::get(Elts);
2148 if (!V->hasOneUse()) { // Other users may use these bits.
2149 if (Depth != 0) { // Not at the root.
2150 // TODO: Just compute the UndefElts information recursively.
2154 } else if (Depth == 10) { // Limit search depth.
2158 Instruction *I = dyn_cast<Instruction>(V);
2159 if (!I) return false; // Only analyze instructions.
2161 bool MadeChange = false;
2162 uint64_t UndefElts2;
2164 switch (I->getOpcode()) {
2167 case Instruction::InsertElement: {
2168 // If this is a variable index, we don't know which element it overwrites.
2169 // demand exactly the same input as we produce.
2170 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
2172 // Note that we can't propagate undef elt info, because we don't know
2173 // which elt is getting updated.
2174 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
2175 UndefElts2, Depth+1);
2176 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
2180 // If this is inserting an element that isn't demanded, remove this
2182 unsigned IdxNo = Idx->getZExtValue();
2183 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
2184 return AddSoonDeadInstToWorklist(*I, 0);
2186 // Otherwise, the element inserted overwrites whatever was there, so the
2187 // input demanded set is simpler than the output set.
2188 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
2189 DemandedElts & ~(1ULL << IdxNo),
2190 UndefElts, Depth+1);
2191 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
2193 // The inserted element is defined.
2194 UndefElts |= 1ULL << IdxNo;
2198 case Instruction::And:
2199 case Instruction::Or:
2200 case Instruction::Xor:
2201 case Instruction::Add:
2202 case Instruction::Sub:
2203 case Instruction::Mul:
2204 // div/rem demand all inputs, because they don't want divide by zero.
2205 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
2206 UndefElts, Depth+1);
2207 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
2208 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
2209 UndefElts2, Depth+1);
2210 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
2212 // Output elements are undefined if both are undefined. Consider things
2213 // like undef&0. The result is known zero, not undef.
2214 UndefElts &= UndefElts2;
2217 case Instruction::Call: {
2218 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
2220 switch (II->getIntrinsicID()) {
2223 // Binary vector operations that work column-wise. A dest element is a
2224 // function of the corresponding input elements from the two inputs.
2225 case Intrinsic::x86_sse_sub_ss:
2226 case Intrinsic::x86_sse_mul_ss:
2227 case Intrinsic::x86_sse_min_ss:
2228 case Intrinsic::x86_sse_max_ss:
2229 case Intrinsic::x86_sse2_sub_sd:
2230 case Intrinsic::x86_sse2_mul_sd:
2231 case Intrinsic::x86_sse2_min_sd:
2232 case Intrinsic::x86_sse2_max_sd:
2233 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
2234 UndefElts, Depth+1);
2235 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
2236 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
2237 UndefElts2, Depth+1);
2238 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
2240 // If only the low elt is demanded and this is a scalarizable intrinsic,
2241 // scalarize it now.
2242 if (DemandedElts == 1) {
2243 switch (II->getIntrinsicID()) {
2245 case Intrinsic::x86_sse_sub_ss:
2246 case Intrinsic::x86_sse_mul_ss:
2247 case Intrinsic::x86_sse2_sub_sd:
2248 case Intrinsic::x86_sse2_mul_sd:
2249 // TODO: Lower MIN/MAX/ABS/etc
2250 Value *LHS = II->getOperand(1);
2251 Value *RHS = II->getOperand(2);
2252 // Extract the element as scalars.
2253 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
2254 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
2256 switch (II->getIntrinsicID()) {
2257 default: assert(0 && "Case stmts out of sync!");
2258 case Intrinsic::x86_sse_sub_ss:
2259 case Intrinsic::x86_sse2_sub_sd:
2260 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
2261 II->getName()), *II);
2263 case Intrinsic::x86_sse_mul_ss:
2264 case Intrinsic::x86_sse2_mul_sd:
2265 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
2266 II->getName()), *II);
2271 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
2273 InsertNewInstBefore(New, *II);
2274 AddSoonDeadInstToWorklist(*II, 0);
2279 // Output elements are undefined if both are undefined. Consider things
2280 // like undef&0. The result is known zero, not undef.
2281 UndefElts &= UndefElts2;
2287 return MadeChange ? I : 0;
2290 /// @returns true if the specified compare instruction is
2291 /// true when both operands are equal...
2292 /// @brief Determine if the ICmpInst returns true if both operands are equal
2293 static bool isTrueWhenEqual(ICmpInst &ICI) {
2294 ICmpInst::Predicate pred = ICI.getPredicate();
2295 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
2296 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
2297 pred == ICmpInst::ICMP_SLE;
2300 /// AssociativeOpt - Perform an optimization on an associative operator. This
2301 /// function is designed to check a chain of associative operators for a
2302 /// potential to apply a certain optimization. Since the optimization may be
2303 /// applicable if the expression was reassociated, this checks the chain, then
2304 /// reassociates the expression as necessary to expose the optimization
2305 /// opportunity. This makes use of a special Functor, which must define
2306 /// 'shouldApply' and 'apply' methods.
2308 template<typename Functor>
2309 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
2310 unsigned Opcode = Root.getOpcode();
2311 Value *LHS = Root.getOperand(0);
2313 // Quick check, see if the immediate LHS matches...
2314 if (F.shouldApply(LHS))
2315 return F.apply(Root);
2317 // Otherwise, if the LHS is not of the same opcode as the root, return.
2318 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2319 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
2320 // Should we apply this transform to the RHS?
2321 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
2323 // If not to the RHS, check to see if we should apply to the LHS...
2324 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
2325 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
2329 // If the functor wants to apply the optimization to the RHS of LHSI,
2330 // reassociate the expression from ((? op A) op B) to (? op (A op B))
2332 BasicBlock *BB = Root.getParent();
2334 // Now all of the instructions are in the current basic block, go ahead
2335 // and perform the reassociation.
2336 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
2338 // First move the selected RHS to the LHS of the root...
2339 Root.setOperand(0, LHSI->getOperand(1));
2341 // Make what used to be the LHS of the root be the user of the root...
2342 Value *ExtraOperand = TmpLHSI->getOperand(1);
2343 if (&Root == TmpLHSI) {
2344 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
2347 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
2348 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
2349 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
2350 BasicBlock::iterator ARI = &Root; ++ARI;
2351 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
2354 // Now propagate the ExtraOperand down the chain of instructions until we
2356 while (TmpLHSI != LHSI) {
2357 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
2358 // Move the instruction to immediately before the chain we are
2359 // constructing to avoid breaking dominance properties.
2360 NextLHSI->getParent()->getInstList().remove(NextLHSI);
2361 BB->getInstList().insert(ARI, NextLHSI);
2364 Value *NextOp = NextLHSI->getOperand(1);
2365 NextLHSI->setOperand(1, ExtraOperand);
2367 ExtraOperand = NextOp;
2370 // Now that the instructions are reassociated, have the functor perform
2371 // the transformation...
2372 return F.apply(Root);
2375 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
2381 // AddRHS - Implements: X + X --> X << 1
2384 AddRHS(Value *rhs) : RHS(rhs) {}
2385 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2386 Instruction *apply(BinaryOperator &Add) const {
2387 return BinaryOperator::createShl(Add.getOperand(0),
2388 ConstantInt::get(Add.getType(), 1));
2392 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
2394 struct AddMaskingAnd {
2396 AddMaskingAnd(Constant *c) : C2(c) {}
2397 bool shouldApply(Value *LHS) const {
2399 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
2400 ConstantExpr::getAnd(C1, C2)->isNullValue();
2402 Instruction *apply(BinaryOperator &Add) const {
2403 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
2407 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
2409 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
2410 if (Constant *SOC = dyn_cast<Constant>(SO))
2411 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
2413 return IC->InsertNewInstBefore(CastInst::create(
2414 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
2417 // Figure out if the constant is the left or the right argument.
2418 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
2419 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
2421 if (Constant *SOC = dyn_cast<Constant>(SO)) {
2423 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
2424 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
2427 Value *Op0 = SO, *Op1 = ConstOperand;
2429 std::swap(Op0, Op1);
2431 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2432 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
2433 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2434 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
2435 SO->getName()+".cmp");
2437 assert(0 && "Unknown binary instruction type!");
2440 return IC->InsertNewInstBefore(New, I);
2443 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
2444 // constant as the other operand, try to fold the binary operator into the
2445 // select arguments. This also works for Cast instructions, which obviously do
2446 // not have a second operand.
2447 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
2449 // Don't modify shared select instructions
2450 if (!SI->hasOneUse()) return 0;
2451 Value *TV = SI->getOperand(1);
2452 Value *FV = SI->getOperand(2);
2454 if (isa<Constant>(TV) || isa<Constant>(FV)) {
2455 // Bool selects with constant operands can be folded to logical ops.
2456 if (SI->getType() == Type::Int1Ty) return 0;
2458 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
2459 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
2461 return new SelectInst(SI->getCondition(), SelectTrueVal,
2468 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
2469 /// node as operand #0, see if we can fold the instruction into the PHI (which
2470 /// is only possible if all operands to the PHI are constants).
2471 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
2472 PHINode *PN = cast<PHINode>(I.getOperand(0));
2473 unsigned NumPHIValues = PN->getNumIncomingValues();
2474 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
2476 // Check to see if all of the operands of the PHI are constants. If there is
2477 // one non-constant value, remember the BB it is. If there is more than one
2478 // or if *it* is a PHI, bail out.
2479 BasicBlock *NonConstBB = 0;
2480 for (unsigned i = 0; i != NumPHIValues; ++i)
2481 if (!isa<Constant>(PN->getIncomingValue(i))) {
2482 if (NonConstBB) return 0; // More than one non-const value.
2483 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2484 NonConstBB = PN->getIncomingBlock(i);
2486 // If the incoming non-constant value is in I's block, we have an infinite
2488 if (NonConstBB == I.getParent())
2492 // If there is exactly one non-constant value, we can insert a copy of the
2493 // operation in that block. However, if this is a critical edge, we would be
2494 // inserting the computation one some other paths (e.g. inside a loop). Only
2495 // do this if the pred block is unconditionally branching into the phi block.
2497 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2498 if (!BI || !BI->isUnconditional()) return 0;
2501 // Okay, we can do the transformation: create the new PHI node.
2502 PHINode *NewPN = new PHINode(I.getType(), "");
2503 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2504 InsertNewInstBefore(NewPN, *PN);
2505 NewPN->takeName(PN);
2507 // Next, add all of the operands to the PHI.
2508 if (I.getNumOperands() == 2) {
2509 Constant *C = cast<Constant>(I.getOperand(1));
2510 for (unsigned i = 0; i != NumPHIValues; ++i) {
2512 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2513 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2514 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2516 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2518 assert(PN->getIncomingBlock(i) == NonConstBB);
2519 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2520 InV = BinaryOperator::create(BO->getOpcode(),
2521 PN->getIncomingValue(i), C, "phitmp",
2522 NonConstBB->getTerminator());
2523 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2524 InV = CmpInst::create(CI->getOpcode(),
2526 PN->getIncomingValue(i), C, "phitmp",
2527 NonConstBB->getTerminator());
2529 assert(0 && "Unknown binop!");
2531 AddToWorkList(cast<Instruction>(InV));
2533 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2536 CastInst *CI = cast<CastInst>(&I);
2537 const Type *RetTy = CI->getType();
2538 for (unsigned i = 0; i != NumPHIValues; ++i) {
2540 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2541 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2543 assert(PN->getIncomingBlock(i) == NonConstBB);
2544 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
2545 I.getType(), "phitmp",
2546 NonConstBB->getTerminator());
2547 AddToWorkList(cast<Instruction>(InV));
2549 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2552 return ReplaceInstUsesWith(I, NewPN);
2555 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2556 bool Changed = SimplifyCommutative(I);
2557 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2559 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2560 // X + undef -> undef
2561 if (isa<UndefValue>(RHS))
2562 return ReplaceInstUsesWith(I, RHS);
2565 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2566 if (RHSC->isNullValue())
2567 return ReplaceInstUsesWith(I, LHS);
2568 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2569 if (CFP->isExactlyValue(-0.0))
2570 return ReplaceInstUsesWith(I, LHS);
2573 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2574 // X + (signbit) --> X ^ signbit
2575 uint64_t Val = CI->getZExtValue();
2576 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
2577 return BinaryOperator::createXor(LHS, RHS);
2579 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2580 // (X & 254)+1 -> (X&254)|1
2581 uint64_t KnownZero, KnownOne;
2582 if (!isa<VectorType>(I.getType()) &&
2583 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
2584 KnownZero, KnownOne))
2588 if (isa<PHINode>(LHS))
2589 if (Instruction *NV = FoldOpIntoPhi(I))
2592 ConstantInt *XorRHS = 0;
2594 if (isa<ConstantInt>(RHSC) &&
2595 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2596 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
2597 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
2598 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
2600 uint64_t C0080Val = 1ULL << 31;
2601 int64_t CFF80Val = -C0080Val;
2604 if (TySizeBits > Size) {
2606 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2607 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2608 if (RHSSExt == CFF80Val) {
2609 if (XorRHS->getZExtValue() == C0080Val)
2611 } else if (RHSZExt == C0080Val) {
2612 if (XorRHS->getSExtValue() == CFF80Val)
2616 // This is a sign extend if the top bits are known zero.
2617 uint64_t Mask = ~0ULL;
2618 Mask <<= 64-(TySizeBits-Size);
2619 Mask &= cast<IntegerType>(XorLHS->getType())->getBitMask();
2620 if (!MaskedValueIsZero(XorLHS, Mask))
2621 Size = 0; // Not a sign ext, but can't be any others either.
2628 } while (Size >= 8);
2631 const Type *MiddleType = 0;
2634 case 32: MiddleType = Type::Int32Ty; break;
2635 case 16: MiddleType = Type::Int16Ty; break;
2636 case 8: MiddleType = Type::Int8Ty; break;
2639 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2640 InsertNewInstBefore(NewTrunc, I);
2641 return new SExtInst(NewTrunc, I.getType());
2647 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2648 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2650 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2651 if (RHSI->getOpcode() == Instruction::Sub)
2652 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2653 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2655 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2656 if (LHSI->getOpcode() == Instruction::Sub)
2657 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2658 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2663 if (Value *V = dyn_castNegVal(LHS))
2664 return BinaryOperator::createSub(RHS, V);
2667 if (!isa<Constant>(RHS))
2668 if (Value *V = dyn_castNegVal(RHS))
2669 return BinaryOperator::createSub(LHS, V);
2673 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2674 if (X == RHS) // X*C + X --> X * (C+1)
2675 return BinaryOperator::createMul(RHS, AddOne(C2));
2677 // X*C1 + X*C2 --> X * (C1+C2)
2679 if (X == dyn_castFoldableMul(RHS, C1))
2680 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
2683 // X + X*C --> X * (C+1)
2684 if (dyn_castFoldableMul(RHS, C2) == LHS)
2685 return BinaryOperator::createMul(LHS, AddOne(C2));
2687 // X + ~X --> -1 since ~X = -X-1
2688 if (dyn_castNotVal(LHS) == RHS ||
2689 dyn_castNotVal(RHS) == LHS)
2690 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
2693 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2694 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2695 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2698 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2700 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
2701 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
2702 return BinaryOperator::createSub(C, X);
2705 // (X & FF00) + xx00 -> (X+xx00) & FF00
2706 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2707 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2708 if (Anded == CRHS) {
2709 // See if all bits from the first bit set in the Add RHS up are included
2710 // in the mask. First, get the rightmost bit.
2711 uint64_t AddRHSV = CRHS->getZExtValue();
2713 // Form a mask of all bits from the lowest bit added through the top.
2714 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
2715 AddRHSHighBits &= C2->getType()->getBitMask();
2717 // See if the and mask includes all of these bits.
2718 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
2720 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2721 // Okay, the xform is safe. Insert the new add pronto.
2722 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2723 LHS->getName()), I);
2724 return BinaryOperator::createAnd(NewAdd, C2);
2729 // Try to fold constant add into select arguments.
2730 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2731 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2735 // add (cast *A to intptrtype) B ->
2736 // cast (GEP (cast *A to sbyte*) B) ->
2739 CastInst *CI = dyn_cast<CastInst>(LHS);
2742 CI = dyn_cast<CastInst>(RHS);
2745 if (CI && CI->getType()->isSized() &&
2746 (CI->getType()->getPrimitiveSizeInBits() ==
2747 TD->getIntPtrType()->getPrimitiveSizeInBits())
2748 && isa<PointerType>(CI->getOperand(0)->getType())) {
2749 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2750 PointerType::get(Type::Int8Ty), I);
2751 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2752 return new PtrToIntInst(I2, CI->getType());
2756 return Changed ? &I : 0;
2759 // isSignBit - Return true if the value represented by the constant only has the
2760 // highest order bit set.
2761 static bool isSignBit(ConstantInt *CI) {
2762 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
2763 return CI->getValue() == APInt::getSignBit(NumBits);
2766 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2767 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2769 if (Op0 == Op1) // sub X, X -> 0
2770 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2772 // If this is a 'B = x-(-A)', change to B = x+A...
2773 if (Value *V = dyn_castNegVal(Op1))
2774 return BinaryOperator::createAdd(Op0, V);
2776 if (isa<UndefValue>(Op0))
2777 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2778 if (isa<UndefValue>(Op1))
2779 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2781 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2782 // Replace (-1 - A) with (~A)...
2783 if (C->isAllOnesValue())
2784 return BinaryOperator::createNot(Op1);
2786 // C - ~X == X + (1+C)
2788 if (match(Op1, m_Not(m_Value(X))))
2789 return BinaryOperator::createAdd(X,
2790 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
2791 // -(X >>u 31) -> (X >>s 31)
2792 // -(X >>s 31) -> (X >>u 31)
2793 if (C->isNullValue()) {
2794 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2795 if (SI->getOpcode() == Instruction::LShr) {
2796 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2797 // Check to see if we are shifting out everything but the sign bit.
2798 if (CU->getZExtValue() ==
2799 SI->getType()->getPrimitiveSizeInBits()-1) {
2800 // Ok, the transformation is safe. Insert AShr.
2801 return BinaryOperator::create(Instruction::AShr,
2802 SI->getOperand(0), CU, SI->getName());
2806 else if (SI->getOpcode() == Instruction::AShr) {
2807 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2808 // Check to see if we are shifting out everything but the sign bit.
2809 if (CU->getZExtValue() ==
2810 SI->getType()->getPrimitiveSizeInBits()-1) {
2811 // Ok, the transformation is safe. Insert LShr.
2812 return BinaryOperator::createLShr(
2813 SI->getOperand(0), CU, SI->getName());
2819 // Try to fold constant sub into select arguments.
2820 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2821 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2824 if (isa<PHINode>(Op0))
2825 if (Instruction *NV = FoldOpIntoPhi(I))
2829 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2830 if (Op1I->getOpcode() == Instruction::Add &&
2831 !Op0->getType()->isFPOrFPVector()) {
2832 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2833 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2834 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2835 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2836 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2837 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2838 // C1-(X+C2) --> (C1-C2)-X
2839 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2840 Op1I->getOperand(0));
2844 if (Op1I->hasOneUse()) {
2845 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2846 // is not used by anyone else...
2848 if (Op1I->getOpcode() == Instruction::Sub &&
2849 !Op1I->getType()->isFPOrFPVector()) {
2850 // Swap the two operands of the subexpr...
2851 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2852 Op1I->setOperand(0, IIOp1);
2853 Op1I->setOperand(1, IIOp0);
2855 // Create the new top level add instruction...
2856 return BinaryOperator::createAdd(Op0, Op1);
2859 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2861 if (Op1I->getOpcode() == Instruction::And &&
2862 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2863 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2866 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2867 return BinaryOperator::createAnd(Op0, NewNot);
2870 // 0 - (X sdiv C) -> (X sdiv -C)
2871 if (Op1I->getOpcode() == Instruction::SDiv)
2872 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2873 if (CSI->isNullValue())
2874 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2875 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2876 ConstantExpr::getNeg(DivRHS));
2878 // X - X*C --> X * (1-C)
2879 ConstantInt *C2 = 0;
2880 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2882 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2883 return BinaryOperator::createMul(Op0, CP1);
2888 if (!Op0->getType()->isFPOrFPVector())
2889 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2890 if (Op0I->getOpcode() == Instruction::Add) {
2891 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2892 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2893 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2894 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2895 } else if (Op0I->getOpcode() == Instruction::Sub) {
2896 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2897 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2901 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2902 if (X == Op1) { // X*C - X --> X * (C-1)
2903 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2904 return BinaryOperator::createMul(Op1, CP1);
2907 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2908 if (X == dyn_castFoldableMul(Op1, C2))
2909 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2914 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2915 /// really just returns true if the most significant (sign) bit is set.
2916 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2918 case ICmpInst::ICMP_SLT:
2919 // True if LHS s< RHS and RHS == 0
2920 return RHS->isNullValue();
2921 case ICmpInst::ICMP_SLE:
2922 // True if LHS s<= RHS and RHS == -1
2923 return RHS->isAllOnesValue();
2924 case ICmpInst::ICMP_UGE:
2925 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2926 return RHS->getZExtValue() == (1ULL <<
2927 (RHS->getType()->getPrimitiveSizeInBits()-1));
2928 case ICmpInst::ICMP_UGT:
2929 // True if LHS u> RHS and RHS == high-bit-mask - 1
2930 return RHS->getZExtValue() ==
2931 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2937 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2938 bool Changed = SimplifyCommutative(I);
2939 Value *Op0 = I.getOperand(0);
2941 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2942 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2944 // Simplify mul instructions with a constant RHS...
2945 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2946 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2948 // ((X << C1)*C2) == (X * (C2 << C1))
2949 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2950 if (SI->getOpcode() == Instruction::Shl)
2951 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2952 return BinaryOperator::createMul(SI->getOperand(0),
2953 ConstantExpr::getShl(CI, ShOp));
2955 if (CI->isNullValue())
2956 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2957 if (CI->equalsInt(1)) // X * 1 == X
2958 return ReplaceInstUsesWith(I, Op0);
2959 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2960 return BinaryOperator::createNeg(Op0, I.getName());
2962 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2963 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2964 uint64_t C = Log2_64(Val);
2965 return BinaryOperator::createShl(Op0,
2966 ConstantInt::get(Op0->getType(), C));
2968 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2969 if (Op1F->isNullValue())
2970 return ReplaceInstUsesWith(I, Op1);
2972 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2973 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2974 if (Op1F->getValue() == 1.0)
2975 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2978 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2979 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2980 isa<ConstantInt>(Op0I->getOperand(1))) {
2981 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2982 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2984 InsertNewInstBefore(Add, I);
2985 Value *C1C2 = ConstantExpr::getMul(Op1,
2986 cast<Constant>(Op0I->getOperand(1)));
2987 return BinaryOperator::createAdd(Add, C1C2);
2991 // Try to fold constant mul into select arguments.
2992 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2993 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2996 if (isa<PHINode>(Op0))
2997 if (Instruction *NV = FoldOpIntoPhi(I))
3001 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
3002 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
3003 return BinaryOperator::createMul(Op0v, Op1v);
3005 // If one of the operands of the multiply is a cast from a boolean value, then
3006 // we know the bool is either zero or one, so this is a 'masking' multiply.
3007 // See if we can simplify things based on how the boolean was originally
3009 CastInst *BoolCast = 0;
3010 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
3011 if (CI->getOperand(0)->getType() == Type::Int1Ty)
3014 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
3015 if (CI->getOperand(0)->getType() == Type::Int1Ty)
3018 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
3019 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
3020 const Type *SCOpTy = SCIOp0->getType();
3022 // If the icmp is true iff the sign bit of X is set, then convert this
3023 // multiply into a shift/and combination.
3024 if (isa<ConstantInt>(SCIOp1) &&
3025 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
3026 // Shift the X value right to turn it into "all signbits".
3027 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
3028 SCOpTy->getPrimitiveSizeInBits()-1);
3030 InsertNewInstBefore(
3031 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
3032 BoolCast->getOperand(0)->getName()+
3035 // If the multiply type is not the same as the source type, sign extend
3036 // or truncate to the multiply type.
3037 if (I.getType() != V->getType()) {
3038 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
3039 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
3040 Instruction::CastOps opcode =
3041 (SrcBits == DstBits ? Instruction::BitCast :
3042 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
3043 V = InsertCastBefore(opcode, V, I.getType(), I);
3046 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
3047 return BinaryOperator::createAnd(V, OtherOp);
3052 return Changed ? &I : 0;
3055 /// This function implements the transforms on div instructions that work
3056 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
3057 /// used by the visitors to those instructions.
3058 /// @brief Transforms common to all three div instructions
3059 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
3060 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3063 if (isa<UndefValue>(Op0))
3064 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3066 // X / undef -> undef
3067 if (isa<UndefValue>(Op1))
3068 return ReplaceInstUsesWith(I, Op1);
3070 // Handle cases involving: div X, (select Cond, Y, Z)
3071 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3072 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
3073 // same basic block, then we replace the select with Y, and the condition
3074 // of the select with false (if the cond value is in the same BB). If the
3075 // select has uses other than the div, this allows them to be simplified
3076 // also. Note that div X, Y is just as good as div X, 0 (undef)
3077 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3078 if (ST->isNullValue()) {
3079 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3080 if (CondI && CondI->getParent() == I.getParent())
3081 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3082 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3083 I.setOperand(1, SI->getOperand(2));
3085 UpdateValueUsesWith(SI, SI->getOperand(2));
3089 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
3090 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3091 if (ST->isNullValue()) {
3092 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3093 if (CondI && CondI->getParent() == I.getParent())
3094 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3095 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3096 I.setOperand(1, SI->getOperand(1));
3098 UpdateValueUsesWith(SI, SI->getOperand(1));
3106 /// This function implements the transforms common to both integer division
3107 /// instructions (udiv and sdiv). It is called by the visitors to those integer
3108 /// division instructions.
3109 /// @brief Common integer divide transforms
3110 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
3111 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3113 if (Instruction *Common = commonDivTransforms(I))
3116 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3118 if (RHS->equalsInt(1))
3119 return ReplaceInstUsesWith(I, Op0);
3121 // (X / C1) / C2 -> X / (C1*C2)
3122 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3123 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3124 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3125 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
3126 ConstantExpr::getMul(RHS, LHSRHS));
3129 if (!RHS->isNullValue()) { // avoid X udiv 0
3130 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3131 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3133 if (isa<PHINode>(Op0))
3134 if (Instruction *NV = FoldOpIntoPhi(I))
3139 // 0 / X == 0, we don't need to preserve faults!
3140 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3141 if (LHS->equalsInt(0))
3142 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3147 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3148 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3150 // Handle the integer div common cases
3151 if (Instruction *Common = commonIDivTransforms(I))
3154 // X udiv C^2 -> X >> C
3155 // Check to see if this is an unsigned division with an exact power of 2,
3156 // if so, convert to a right shift.
3157 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3158 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
3159 if (isPowerOf2_64(Val)) {
3160 uint64_t ShiftAmt = Log2_64(Val);
3161 return BinaryOperator::createLShr(Op0,
3162 ConstantInt::get(Op0->getType(), ShiftAmt));
3166 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3167 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3168 if (RHSI->getOpcode() == Instruction::Shl &&
3169 isa<ConstantInt>(RHSI->getOperand(0))) {
3170 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
3171 if (isPowerOf2_64(C1)) {
3172 Value *N = RHSI->getOperand(1);
3173 const Type *NTy = N->getType();
3174 if (uint64_t C2 = Log2_64(C1)) {
3175 Constant *C2V = ConstantInt::get(NTy, C2);
3176 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
3178 return BinaryOperator::createLShr(Op0, N);
3183 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3184 // where C1&C2 are powers of two.
3185 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3186 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3187 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3188 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
3189 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
3190 // Compute the shift amounts
3191 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
3192 // Construct the "on true" case of the select
3193 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3194 Instruction *TSI = BinaryOperator::createLShr(
3195 Op0, TC, SI->getName()+".t");
3196 TSI = InsertNewInstBefore(TSI, I);
3198 // Construct the "on false" case of the select
3199 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3200 Instruction *FSI = BinaryOperator::createLShr(
3201 Op0, FC, SI->getName()+".f");
3202 FSI = InsertNewInstBefore(FSI, I);
3204 // construct the select instruction and return it.
3205 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
3211 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3212 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3214 // Handle the integer div common cases
3215 if (Instruction *Common = commonIDivTransforms(I))
3218 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3220 if (RHS->isAllOnesValue())
3221 return BinaryOperator::createNeg(Op0);
3224 if (Value *LHSNeg = dyn_castNegVal(Op0))
3225 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
3228 // If the sign bits of both operands are zero (i.e. we can prove they are
3229 // unsigned inputs), turn this into a udiv.
3230 if (I.getType()->isInteger()) {
3231 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
3232 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3233 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
3240 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3241 return commonDivTransforms(I);
3244 /// GetFactor - If we can prove that the specified value is at least a multiple
3245 /// of some factor, return that factor.
3246 static Constant *GetFactor(Value *V) {
3247 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3250 // Unless we can be tricky, we know this is a multiple of 1.
3251 Constant *Result = ConstantInt::get(V->getType(), 1);
3253 Instruction *I = dyn_cast<Instruction>(V);
3254 if (!I) return Result;
3256 if (I->getOpcode() == Instruction::Mul) {
3257 // Handle multiplies by a constant, etc.
3258 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
3259 GetFactor(I->getOperand(1)));
3260 } else if (I->getOpcode() == Instruction::Shl) {
3261 // (X<<C) -> X * (1 << C)
3262 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
3263 ShRHS = ConstantExpr::getShl(Result, ShRHS);
3264 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
3266 } else if (I->getOpcode() == Instruction::And) {
3267 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
3268 // X & 0xFFF0 is known to be a multiple of 16.
3269 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
3270 if (Zeros != V->getType()->getPrimitiveSizeInBits())
3271 return ConstantExpr::getShl(Result,
3272 ConstantInt::get(Result->getType(), Zeros));
3274 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
3275 // Only handle int->int casts.
3276 if (!CI->isIntegerCast())
3278 Value *Op = CI->getOperand(0);
3279 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
3284 /// This function implements the transforms on rem instructions that work
3285 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3286 /// is used by the visitors to those instructions.
3287 /// @brief Transforms common to all three rem instructions
3288 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3289 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3291 // 0 % X == 0, we don't need to preserve faults!
3292 if (Constant *LHS = dyn_cast<Constant>(Op0))
3293 if (LHS->isNullValue())
3294 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3296 if (isa<UndefValue>(Op0)) // undef % X -> 0
3297 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3298 if (isa<UndefValue>(Op1))
3299 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3301 // Handle cases involving: rem X, (select Cond, Y, Z)
3302 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3303 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
3304 // the same basic block, then we replace the select with Y, and the
3305 // condition of the select with false (if the cond value is in the same
3306 // BB). If the select has uses other than the div, this allows them to be
3308 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3309 if (ST->isNullValue()) {
3310 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3311 if (CondI && CondI->getParent() == I.getParent())
3312 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3313 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3314 I.setOperand(1, SI->getOperand(2));
3316 UpdateValueUsesWith(SI, SI->getOperand(2));
3319 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
3320 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3321 if (ST->isNullValue()) {
3322 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3323 if (CondI && CondI->getParent() == I.getParent())
3324 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3325 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3326 I.setOperand(1, SI->getOperand(1));
3328 UpdateValueUsesWith(SI, SI->getOperand(1));
3336 /// This function implements the transforms common to both integer remainder
3337 /// instructions (urem and srem). It is called by the visitors to those integer
3338 /// remainder instructions.
3339 /// @brief Common integer remainder transforms
3340 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3341 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3343 if (Instruction *common = commonRemTransforms(I))
3346 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3347 // X % 0 == undef, we don't need to preserve faults!
3348 if (RHS->equalsInt(0))
3349 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3351 if (RHS->equalsInt(1)) // X % 1 == 0
3352 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3354 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3355 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3356 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3358 } else if (isa<PHINode>(Op0I)) {
3359 if (Instruction *NV = FoldOpIntoPhi(I))
3362 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
3363 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
3364 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3371 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3372 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3374 if (Instruction *common = commonIRemTransforms(I))
3377 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3378 // X urem C^2 -> X and C
3379 // Check to see if this is an unsigned remainder with an exact power of 2,
3380 // if so, convert to a bitwise and.
3381 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3382 if (isPowerOf2_64(C->getZExtValue()))
3383 return BinaryOperator::createAnd(Op0, SubOne(C));
3386 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3387 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3388 if (RHSI->getOpcode() == Instruction::Shl &&
3389 isa<ConstantInt>(RHSI->getOperand(0))) {
3390 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
3391 if (isPowerOf2_64(C1)) {
3392 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3393 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
3395 return BinaryOperator::createAnd(Op0, Add);
3400 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3401 // where C1&C2 are powers of two.
3402 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3403 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3404 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3405 // STO == 0 and SFO == 0 handled above.
3406 if (isPowerOf2_64(STO->getZExtValue()) &&
3407 isPowerOf2_64(SFO->getZExtValue())) {
3408 Value *TrueAnd = InsertNewInstBefore(
3409 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3410 Value *FalseAnd = InsertNewInstBefore(
3411 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3412 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
3420 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3421 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3423 if (Instruction *common = commonIRemTransforms(I))
3426 if (Value *RHSNeg = dyn_castNegVal(Op1))
3427 if (!isa<ConstantInt>(RHSNeg) ||
3428 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
3430 AddUsesToWorkList(I);
3431 I.setOperand(1, RHSNeg);
3435 // If the top bits of both operands are zero (i.e. we can prove they are
3436 // unsigned inputs), turn this into a urem.
3437 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
3438 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3439 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3440 return BinaryOperator::createURem(Op0, Op1, I.getName());
3446 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3447 return commonRemTransforms(I);
3450 // isMaxValueMinusOne - return true if this is Max-1
3451 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
3452 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3454 // Calculate 0111111111..11111
3455 APInt Val(APInt::getSignedMaxValue(TypeBits));
3456 return C->getValue() == Val-1;
3458 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
3461 // isMinValuePlusOne - return true if this is Min+1
3462 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
3464 // Calculate 1111111111000000000000
3465 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3466 APInt Val(APInt::getSignedMinValue(TypeBits));
3467 return C->getValue() == Val+1;
3469 return C->getValue() == 1; // unsigned
3472 // isOneBitSet - Return true if there is exactly one bit set in the specified
3474 static bool isOneBitSet(const ConstantInt *CI) {
3475 return CI->getValue().isPowerOf2();
3478 #if 0 // Currently unused
3479 // isLowOnes - Return true if the constant is of the form 0+1+.
3480 static bool isLowOnes(const ConstantInt *CI) {
3481 uint64_t V = CI->getZExtValue();
3483 // There won't be bits set in parts that the type doesn't contain.
3484 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
3486 uint64_t U = V+1; // If it is low ones, this should be a power of two.
3487 return U && V && (U & V) == 0;
3491 // isHighOnes - Return true if the constant is of the form 1+0+.
3492 // This is the same as lowones(~X).
3493 static bool isHighOnes(const ConstantInt *CI) {
3494 return (~CI->getValue() + 1).isPowerOf2();
3497 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3498 /// are carefully arranged to allow folding of expressions such as:
3500 /// (A < B) | (A > B) --> (A != B)
3502 /// Note that this is only valid if the first and second predicates have the
3503 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3505 /// Three bits are used to represent the condition, as follows:
3510 /// <=> Value Definition
3511 /// 000 0 Always false
3518 /// 111 7 Always true
3520 static unsigned getICmpCode(const ICmpInst *ICI) {
3521 switch (ICI->getPredicate()) {
3523 case ICmpInst::ICMP_UGT: return 1; // 001
3524 case ICmpInst::ICMP_SGT: return 1; // 001
3525 case ICmpInst::ICMP_EQ: return 2; // 010
3526 case ICmpInst::ICMP_UGE: return 3; // 011
3527 case ICmpInst::ICMP_SGE: return 3; // 011
3528 case ICmpInst::ICMP_ULT: return 4; // 100
3529 case ICmpInst::ICMP_SLT: return 4; // 100
3530 case ICmpInst::ICMP_NE: return 5; // 101
3531 case ICmpInst::ICMP_ULE: return 6; // 110
3532 case ICmpInst::ICMP_SLE: return 6; // 110
3535 assert(0 && "Invalid ICmp predicate!");
3540 /// getICmpValue - This is the complement of getICmpCode, which turns an
3541 /// opcode and two operands into either a constant true or false, or a brand
3542 /// new /// ICmp instruction. The sign is passed in to determine which kind
3543 /// of predicate to use in new icmp instructions.
3544 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3546 default: assert(0 && "Illegal ICmp code!");
3547 case 0: return ConstantInt::getFalse();
3550 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3552 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3553 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3556 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3558 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3561 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3563 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3564 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3567 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3569 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3570 case 7: return ConstantInt::getTrue();
3574 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3575 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3576 (ICmpInst::isSignedPredicate(p1) &&
3577 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3578 (ICmpInst::isSignedPredicate(p2) &&
3579 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3583 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3584 struct FoldICmpLogical {
3587 ICmpInst::Predicate pred;
3588 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3589 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3590 pred(ICI->getPredicate()) {}
3591 bool shouldApply(Value *V) const {
3592 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3593 if (PredicatesFoldable(pred, ICI->getPredicate()))
3594 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
3595 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
3598 Instruction *apply(Instruction &Log) const {
3599 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3600 if (ICI->getOperand(0) != LHS) {
3601 assert(ICI->getOperand(1) == LHS);
3602 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3605 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3606 unsigned LHSCode = getICmpCode(ICI);
3607 unsigned RHSCode = getICmpCode(RHSICI);
3609 switch (Log.getOpcode()) {
3610 case Instruction::And: Code = LHSCode & RHSCode; break;
3611 case Instruction::Or: Code = LHSCode | RHSCode; break;
3612 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3613 default: assert(0 && "Illegal logical opcode!"); return 0;
3616 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3617 ICmpInst::isSignedPredicate(ICI->getPredicate());
3619 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3620 if (Instruction *I = dyn_cast<Instruction>(RV))
3622 // Otherwise, it's a constant boolean value...
3623 return IC.ReplaceInstUsesWith(Log, RV);
3626 } // end anonymous namespace
3628 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3629 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3630 // guaranteed to be a binary operator.
3631 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3633 ConstantInt *AndRHS,
3634 BinaryOperator &TheAnd) {
3635 Value *X = Op->getOperand(0);
3636 Constant *Together = 0;
3638 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3640 switch (Op->getOpcode()) {
3641 case Instruction::Xor:
3642 if (Op->hasOneUse()) {
3643 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3644 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3645 InsertNewInstBefore(And, TheAnd);
3647 return BinaryOperator::createXor(And, Together);
3650 case Instruction::Or:
3651 if (Together == AndRHS) // (X | C) & C --> C
3652 return ReplaceInstUsesWith(TheAnd, AndRHS);
3654 if (Op->hasOneUse() && Together != OpRHS) {
3655 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3656 Instruction *Or = BinaryOperator::createOr(X, Together);
3657 InsertNewInstBefore(Or, TheAnd);
3659 return BinaryOperator::createAnd(Or, AndRHS);
3662 case Instruction::Add:
3663 if (Op->hasOneUse()) {
3664 // Adding a one to a single bit bit-field should be turned into an XOR
3665 // of the bit. First thing to check is to see if this AND is with a
3666 // single bit constant.
3667 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
3669 // Clear bits that are not part of the constant.
3670 AndRHSV &= AndRHS->getType()->getBitMask();
3672 // If there is only one bit set...
3673 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3674 // Ok, at this point, we know that we are masking the result of the
3675 // ADD down to exactly one bit. If the constant we are adding has
3676 // no bits set below this bit, then we can eliminate the ADD.
3677 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
3679 // Check to see if any bits below the one bit set in AndRHSV are set.
3680 if ((AddRHS & (AndRHSV-1)) == 0) {
3681 // If not, the only thing that can effect the output of the AND is
3682 // the bit specified by AndRHSV. If that bit is set, the effect of
3683 // the XOR is to toggle the bit. If it is clear, then the ADD has
3685 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3686 TheAnd.setOperand(0, X);
3689 // Pull the XOR out of the AND.
3690 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3691 InsertNewInstBefore(NewAnd, TheAnd);
3692 NewAnd->takeName(Op);
3693 return BinaryOperator::createXor(NewAnd, AndRHS);
3700 case Instruction::Shl: {
3701 // We know that the AND will not produce any of the bits shifted in, so if
3702 // the anded constant includes them, clear them now!
3704 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3705 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
3706 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
3708 if (CI == ShlMask) { // Masking out bits that the shift already masks
3709 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3710 } else if (CI != AndRHS) { // Reducing bits set in and.
3711 TheAnd.setOperand(1, CI);
3716 case Instruction::LShr:
3718 // We know that the AND will not produce any of the bits shifted in, so if
3719 // the anded constant includes them, clear them now! This only applies to
3720 // unsigned shifts, because a signed shr may bring in set bits!
3722 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3723 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
3724 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
3726 if (CI == ShrMask) { // Masking out bits that the shift already masks.
3727 return ReplaceInstUsesWith(TheAnd, Op);
3728 } else if (CI != AndRHS) {
3729 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3734 case Instruction::AShr:
3736 // See if this is shifting in some sign extension, then masking it out
3738 if (Op->hasOneUse()) {
3739 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
3740 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
3741 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
3742 if (C == AndRHS) { // Masking out bits shifted in.
3743 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3744 // Make the argument unsigned.
3745 Value *ShVal = Op->getOperand(0);
3746 ShVal = InsertNewInstBefore(
3747 BinaryOperator::createLShr(ShVal, OpRHS,
3748 Op->getName()), TheAnd);
3749 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3758 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3759 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3760 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3761 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3762 /// insert new instructions.
3763 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3764 bool isSigned, bool Inside,
3766 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3767 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3768 "Lo is not <= Hi in range emission code!");
3771 if (Lo == Hi) // Trivially false.
3772 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3774 // V >= Min && V < Hi --> V < Hi
3775 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3776 ICmpInst::Predicate pred = (isSigned ?
3777 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3778 return new ICmpInst(pred, V, Hi);
3781 // Emit V-Lo <u Hi-Lo
3782 Constant *NegLo = ConstantExpr::getNeg(Lo);
3783 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3784 InsertNewInstBefore(Add, IB);
3785 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3786 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3789 if (Lo == Hi) // Trivially true.
3790 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3792 // V < Min || V >= Hi -> V > Hi-1
3793 Hi = SubOne(cast<ConstantInt>(Hi));
3794 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3795 ICmpInst::Predicate pred = (isSigned ?
3796 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3797 return new ICmpInst(pred, V, Hi);
3800 // Emit V-Lo >u Hi-1-Lo
3801 // Note that Hi has already had one subtracted from it, above.
3802 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3803 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3804 InsertNewInstBefore(Add, IB);
3805 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3806 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3809 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3810 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3811 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3812 // not, since all 1s are not contiguous.
3813 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
3814 uint64_t V = Val->getZExtValue();
3815 if (!isShiftedMask_64(V)) return false;
3817 // look for the first zero bit after the run of ones
3818 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
3819 // look for the first non-zero bit
3820 ME = 64-CountLeadingZeros_64(V);
3826 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3827 /// where isSub determines whether the operator is a sub. If we can fold one of
3828 /// the following xforms:
3830 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3831 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3832 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3834 /// return (A +/- B).
3836 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3837 ConstantInt *Mask, bool isSub,
3839 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3840 if (!LHSI || LHSI->getNumOperands() != 2 ||
3841 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3843 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3845 switch (LHSI->getOpcode()) {
3847 case Instruction::And:
3848 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3849 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3850 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3853 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3854 // part, we don't need any explicit masks to take them out of A. If that
3855 // is all N is, ignore it.
3857 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3858 uint64_t Mask = cast<IntegerType>(RHS->getType())->getBitMask();
3860 if (MaskedValueIsZero(RHS, Mask))
3865 case Instruction::Or:
3866 case Instruction::Xor:
3867 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3868 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3869 ConstantExpr::getAnd(N, Mask)->isNullValue())
3876 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3878 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3879 return InsertNewInstBefore(New, I);
3882 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3883 bool Changed = SimplifyCommutative(I);
3884 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3886 if (isa<UndefValue>(Op1)) // X & undef -> 0
3887 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3891 return ReplaceInstUsesWith(I, Op1);
3893 // See if we can simplify any instructions used by the instruction whose sole
3894 // purpose is to compute bits we don't care about.
3895 uint64_t KnownZero, KnownOne;
3896 if (!isa<VectorType>(I.getType())) {
3897 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
3898 KnownZero, KnownOne))
3901 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3902 if (CP->isAllOnesValue())
3903 return ReplaceInstUsesWith(I, I.getOperand(0));
3907 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3908 uint64_t AndRHSMask = AndRHS->getZExtValue();
3909 uint64_t TypeMask = cast<IntegerType>(Op0->getType())->getBitMask();
3910 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3912 // Optimize a variety of ((val OP C1) & C2) combinations...
3913 if (isa<BinaryOperator>(Op0)) {
3914 Instruction *Op0I = cast<Instruction>(Op0);
3915 Value *Op0LHS = Op0I->getOperand(0);
3916 Value *Op0RHS = Op0I->getOperand(1);
3917 switch (Op0I->getOpcode()) {
3918 case Instruction::Xor:
3919 case Instruction::Or:
3920 // If the mask is only needed on one incoming arm, push it up.
3921 if (Op0I->hasOneUse()) {
3922 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3923 // Not masking anything out for the LHS, move to RHS.
3924 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3925 Op0RHS->getName()+".masked");
3926 InsertNewInstBefore(NewRHS, I);
3927 return BinaryOperator::create(
3928 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3930 if (!isa<Constant>(Op0RHS) &&
3931 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3932 // Not masking anything out for the RHS, move to LHS.
3933 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3934 Op0LHS->getName()+".masked");
3935 InsertNewInstBefore(NewLHS, I);
3936 return BinaryOperator::create(
3937 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3942 case Instruction::Add:
3943 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3944 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3945 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3946 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3947 return BinaryOperator::createAnd(V, AndRHS);
3948 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3949 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3952 case Instruction::Sub:
3953 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3954 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3955 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3956 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3957 return BinaryOperator::createAnd(V, AndRHS);
3961 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3962 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3964 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3965 // If this is an integer truncation or change from signed-to-unsigned, and
3966 // if the source is an and/or with immediate, transform it. This
3967 // frequently occurs for bitfield accesses.
3968 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3969 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3970 CastOp->getNumOperands() == 2)
3971 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3972 if (CastOp->getOpcode() == Instruction::And) {
3973 // Change: and (cast (and X, C1) to T), C2
3974 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3975 // This will fold the two constants together, which may allow
3976 // other simplifications.
3977 Instruction *NewCast = CastInst::createTruncOrBitCast(
3978 CastOp->getOperand(0), I.getType(),
3979 CastOp->getName()+".shrunk");
3980 NewCast = InsertNewInstBefore(NewCast, I);
3981 // trunc_or_bitcast(C1)&C2
3982 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3983 C3 = ConstantExpr::getAnd(C3, AndRHS);
3984 return BinaryOperator::createAnd(NewCast, C3);
3985 } else if (CastOp->getOpcode() == Instruction::Or) {
3986 // Change: and (cast (or X, C1) to T), C2
3987 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3988 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3989 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3990 return ReplaceInstUsesWith(I, AndRHS);
3995 // Try to fold constant and into select arguments.
3996 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3997 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3999 if (isa<PHINode>(Op0))
4000 if (Instruction *NV = FoldOpIntoPhi(I))
4004 Value *Op0NotVal = dyn_castNotVal(Op0);
4005 Value *Op1NotVal = dyn_castNotVal(Op1);
4007 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4008 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4010 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4011 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4012 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
4013 I.getName()+".demorgan");
4014 InsertNewInstBefore(Or, I);
4015 return BinaryOperator::createNot(Or);
4019 Value *A = 0, *B = 0;
4020 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
4021 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4022 return ReplaceInstUsesWith(I, Op1);
4023 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
4024 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4025 return ReplaceInstUsesWith(I, Op0);
4027 if (Op0->hasOneUse() &&
4028 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4029 if (A == Op1) { // (A^B)&A -> A&(A^B)
4030 I.swapOperands(); // Simplify below
4031 std::swap(Op0, Op1);
4032 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4033 cast<BinaryOperator>(Op0)->swapOperands();
4034 I.swapOperands(); // Simplify below
4035 std::swap(Op0, Op1);
4038 if (Op1->hasOneUse() &&
4039 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4040 if (B == Op0) { // B&(A^B) -> B&(B^A)
4041 cast<BinaryOperator>(Op1)->swapOperands();
4044 if (A == Op0) { // A&(A^B) -> A & ~B
4045 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
4046 InsertNewInstBefore(NotB, I);
4047 return BinaryOperator::createAnd(A, NotB);
4052 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4053 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4054 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4057 Value *LHSVal, *RHSVal;
4058 ConstantInt *LHSCst, *RHSCst;
4059 ICmpInst::Predicate LHSCC, RHSCC;
4060 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4061 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4062 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
4063 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
4064 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4065 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4066 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4067 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
4068 // Ensure that the larger constant is on the RHS.
4069 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
4070 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
4071 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
4072 ICmpInst *LHS = cast<ICmpInst>(Op0);
4073 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
4074 std::swap(LHS, RHS);
4075 std::swap(LHSCst, RHSCst);
4076 std::swap(LHSCC, RHSCC);
4079 // At this point, we know we have have two icmp instructions
4080 // comparing a value against two constants and and'ing the result
4081 // together. Because of the above check, we know that we only have
4082 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
4083 // (from the FoldICmpLogical check above), that the two constants
4084 // are not equal and that the larger constant is on the RHS
4085 assert(LHSCst != RHSCst && "Compares not folded above?");
4088 default: assert(0 && "Unknown integer condition code!");
4089 case ICmpInst::ICMP_EQ:
4091 default: assert(0 && "Unknown integer condition code!");
4092 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
4093 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
4094 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
4095 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4096 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
4097 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
4098 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
4099 return ReplaceInstUsesWith(I, LHS);
4101 case ICmpInst::ICMP_NE:
4103 default: assert(0 && "Unknown integer condition code!");
4104 case ICmpInst::ICMP_ULT:
4105 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
4106 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
4107 break; // (X != 13 & X u< 15) -> no change
4108 case ICmpInst::ICMP_SLT:
4109 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
4110 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
4111 break; // (X != 13 & X s< 15) -> no change
4112 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
4113 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
4114 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
4115 return ReplaceInstUsesWith(I, RHS);
4116 case ICmpInst::ICMP_NE:
4117 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
4118 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4119 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4120 LHSVal->getName()+".off");
4121 InsertNewInstBefore(Add, I);
4122 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
4123 ConstantInt::get(Add->getType(), 1));
4125 break; // (X != 13 & X != 15) -> no change
4128 case ICmpInst::ICMP_ULT:
4130 default: assert(0 && "Unknown integer condition code!");
4131 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
4132 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
4133 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4134 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
4136 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
4137 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
4138 return ReplaceInstUsesWith(I, LHS);
4139 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
4143 case ICmpInst::ICMP_SLT:
4145 default: assert(0 && "Unknown integer condition code!");
4146 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
4147 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
4148 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4149 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
4151 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
4152 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
4153 return ReplaceInstUsesWith(I, LHS);
4154 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
4158 case ICmpInst::ICMP_UGT:
4160 default: assert(0 && "Unknown integer condition code!");
4161 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
4162 return ReplaceInstUsesWith(I, LHS);
4163 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
4164 return ReplaceInstUsesWith(I, RHS);
4165 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
4167 case ICmpInst::ICMP_NE:
4168 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
4169 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4170 break; // (X u> 13 & X != 15) -> no change
4171 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
4172 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
4174 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
4178 case ICmpInst::ICMP_SGT:
4180 default: assert(0 && "Unknown integer condition code!");
4181 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
4182 return ReplaceInstUsesWith(I, LHS);
4183 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
4184 return ReplaceInstUsesWith(I, RHS);
4185 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
4187 case ICmpInst::ICMP_NE:
4188 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
4189 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4190 break; // (X s> 13 & X != 15) -> no change
4191 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
4192 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
4194 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
4202 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4203 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4204 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4205 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4206 const Type *SrcTy = Op0C->getOperand(0)->getType();
4207 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4208 // Only do this if the casts both really cause code to be generated.
4209 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4211 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4213 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
4214 Op1C->getOperand(0),
4216 InsertNewInstBefore(NewOp, I);
4217 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4221 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4222 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4223 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4224 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4225 SI0->getOperand(1) == SI1->getOperand(1) &&
4226 (SI0->hasOneUse() || SI1->hasOneUse())) {
4227 Instruction *NewOp =
4228 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
4230 SI0->getName()), I);
4231 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4232 SI1->getOperand(1));
4236 return Changed ? &I : 0;
4239 /// CollectBSwapParts - Look to see if the specified value defines a single byte
4240 /// in the result. If it does, and if the specified byte hasn't been filled in
4241 /// yet, fill it in and return false.
4242 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
4243 Instruction *I = dyn_cast<Instruction>(V);
4244 if (I == 0) return true;
4246 // If this is an or instruction, it is an inner node of the bswap.
4247 if (I->getOpcode() == Instruction::Or)
4248 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
4249 CollectBSwapParts(I->getOperand(1), ByteValues);
4251 // If this is a shift by a constant int, and it is "24", then its operand
4252 // defines a byte. We only handle unsigned types here.
4253 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
4254 // Not shifting the entire input by N-1 bytes?
4255 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
4256 8*(ByteValues.size()-1))
4260 if (I->getOpcode() == Instruction::Shl) {
4261 // X << 24 defines the top byte with the lowest of the input bytes.
4262 DestNo = ByteValues.size()-1;
4264 // X >>u 24 defines the low byte with the highest of the input bytes.
4268 // If the destination byte value is already defined, the values are or'd
4269 // together, which isn't a bswap (unless it's an or of the same bits).
4270 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
4272 ByteValues[DestNo] = I->getOperand(0);
4276 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
4278 Value *Shift = 0, *ShiftLHS = 0;
4279 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
4280 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
4281 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
4283 Instruction *SI = cast<Instruction>(Shift);
4285 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
4286 if (ShiftAmt->getZExtValue() & 7 ||
4287 ShiftAmt->getZExtValue() > 8*ByteValues.size())
4290 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
4292 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
4293 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
4295 // Unknown mask for bswap.
4296 if (DestByte == ByteValues.size()) return true;
4298 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
4300 if (SI->getOpcode() == Instruction::Shl)
4301 SrcByte = DestByte - ShiftBytes;
4303 SrcByte = DestByte + ShiftBytes;
4305 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
4306 if (SrcByte != ByteValues.size()-DestByte-1)
4309 // If the destination byte value is already defined, the values are or'd
4310 // together, which isn't a bswap (unless it's an or of the same bits).
4311 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
4313 ByteValues[DestByte] = SI->getOperand(0);
4317 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4318 /// If so, insert the new bswap intrinsic and return it.
4319 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4320 // We cannot bswap one byte.
4321 if (I.getType() == Type::Int8Ty)
4324 /// ByteValues - For each byte of the result, we keep track of which value
4325 /// defines each byte.
4326 SmallVector<Value*, 8> ByteValues;
4327 ByteValues.resize(TD->getTypeSize(I.getType()));
4329 // Try to find all the pieces corresponding to the bswap.
4330 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
4331 CollectBSwapParts(I.getOperand(1), ByteValues))
4334 // Check to see if all of the bytes come from the same value.
4335 Value *V = ByteValues[0];
4336 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4338 // Check to make sure that all of the bytes come from the same value.
4339 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4340 if (ByteValues[i] != V)
4343 // If they do then *success* we can turn this into a bswap. Figure out what
4344 // bswap to make it into.
4345 Module *M = I.getParent()->getParent()->getParent();
4346 const char *FnName = 0;
4347 if (I.getType() == Type::Int16Ty)
4348 FnName = "llvm.bswap.i16";
4349 else if (I.getType() == Type::Int32Ty)
4350 FnName = "llvm.bswap.i32";
4351 else if (I.getType() == Type::Int64Ty)
4352 FnName = "llvm.bswap.i64";
4354 assert(0 && "Unknown integer type!");
4355 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
4356 return new CallInst(F, V);
4360 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4361 bool Changed = SimplifyCommutative(I);
4362 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4364 if (isa<UndefValue>(Op1))
4365 return ReplaceInstUsesWith(I, // X | undef -> -1
4366 ConstantInt::getAllOnesValue(I.getType()));
4370 return ReplaceInstUsesWith(I, Op0);
4372 // See if we can simplify any instructions used by the instruction whose sole
4373 // purpose is to compute bits we don't care about.
4374 uint64_t KnownZero, KnownOne;
4375 if (!isa<VectorType>(I.getType()) &&
4376 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
4377 KnownZero, KnownOne))
4381 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4382 ConstantInt *C1 = 0; Value *X = 0;
4383 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4384 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4385 Instruction *Or = BinaryOperator::createOr(X, RHS);
4386 InsertNewInstBefore(Or, I);
4388 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
4391 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4392 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4393 Instruction *Or = BinaryOperator::createOr(X, RHS);
4394 InsertNewInstBefore(Or, I);
4396 return BinaryOperator::createXor(Or,
4397 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
4400 // Try to fold constant and into select arguments.
4401 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4402 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4404 if (isa<PHINode>(Op0))
4405 if (Instruction *NV = FoldOpIntoPhi(I))
4409 Value *A = 0, *B = 0;
4410 ConstantInt *C1 = 0, *C2 = 0;
4412 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4413 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4414 return ReplaceInstUsesWith(I, Op1);
4415 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4416 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4417 return ReplaceInstUsesWith(I, Op0);
4419 // (A | B) | C and A | (B | C) -> bswap if possible.
4420 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4421 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4422 match(Op1, m_Or(m_Value(), m_Value())) ||
4423 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4424 match(Op1, m_Shift(m_Value(), m_Value())))) {
4425 if (Instruction *BSwap = MatchBSwap(I))
4429 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4430 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4431 MaskedValueIsZero(Op1, C1->getZExtValue())) {
4432 Instruction *NOr = BinaryOperator::createOr(A, Op1);
4433 InsertNewInstBefore(NOr, I);
4435 return BinaryOperator::createXor(NOr, C1);
4438 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4439 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4440 MaskedValueIsZero(Op0, C1->getZExtValue())) {
4441 Instruction *NOr = BinaryOperator::createOr(A, Op0);
4442 InsertNewInstBefore(NOr, I);
4444 return BinaryOperator::createXor(NOr, C1);
4447 // (A & C1)|(B & C2)
4448 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
4449 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
4451 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
4452 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
4455 // If we have: ((V + N) & C1) | (V & C2)
4456 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4457 // replace with V+N.
4458 if (C1 == ConstantExpr::getNot(C2)) {
4459 Value *V1 = 0, *V2 = 0;
4460 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
4461 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4462 // Add commutes, try both ways.
4463 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
4464 return ReplaceInstUsesWith(I, A);
4465 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
4466 return ReplaceInstUsesWith(I, A);
4468 // Or commutes, try both ways.
4469 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
4470 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4471 // Add commutes, try both ways.
4472 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
4473 return ReplaceInstUsesWith(I, B);
4474 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
4475 return ReplaceInstUsesWith(I, B);
4480 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4481 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4482 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4483 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4484 SI0->getOperand(1) == SI1->getOperand(1) &&
4485 (SI0->hasOneUse() || SI1->hasOneUse())) {
4486 Instruction *NewOp =
4487 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4489 SI0->getName()), I);
4490 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4491 SI1->getOperand(1));
4495 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4496 if (A == Op1) // ~A | A == -1
4497 return ReplaceInstUsesWith(I,
4498 ConstantInt::getAllOnesValue(I.getType()));
4502 // Note, A is still live here!
4503 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4505 return ReplaceInstUsesWith(I,
4506 ConstantInt::getAllOnesValue(I.getType()));
4508 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4509 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4510 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4511 I.getName()+".demorgan"), I);
4512 return BinaryOperator::createNot(And);
4516 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4517 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4518 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4521 Value *LHSVal, *RHSVal;
4522 ConstantInt *LHSCst, *RHSCst;
4523 ICmpInst::Predicate LHSCC, RHSCC;
4524 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4525 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4526 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4527 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4528 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4529 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4530 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4531 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
4532 // Ensure that the larger constant is on the RHS.
4533 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
4534 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
4535 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
4536 ICmpInst *LHS = cast<ICmpInst>(Op0);
4537 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
4538 std::swap(LHS, RHS);
4539 std::swap(LHSCst, RHSCst);
4540 std::swap(LHSCC, RHSCC);
4543 // At this point, we know we have have two icmp instructions
4544 // comparing a value against two constants and or'ing the result
4545 // together. Because of the above check, we know that we only have
4546 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4547 // FoldICmpLogical check above), that the two constants are not
4549 assert(LHSCst != RHSCst && "Compares not folded above?");
4552 default: assert(0 && "Unknown integer condition code!");
4553 case ICmpInst::ICMP_EQ:
4555 default: assert(0 && "Unknown integer condition code!");
4556 case ICmpInst::ICMP_EQ:
4557 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4558 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4559 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4560 LHSVal->getName()+".off");
4561 InsertNewInstBefore(Add, I);
4562 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4563 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4565 break; // (X == 13 | X == 15) -> no change
4566 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4567 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4569 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4570 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4571 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4572 return ReplaceInstUsesWith(I, RHS);
4575 case ICmpInst::ICMP_NE:
4577 default: assert(0 && "Unknown integer condition code!");
4578 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4579 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4580 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4581 return ReplaceInstUsesWith(I, LHS);
4582 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4583 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4584 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4585 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4588 case ICmpInst::ICMP_ULT:
4590 default: assert(0 && "Unknown integer condition code!");
4591 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4593 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4594 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4596 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4598 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4599 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4600 return ReplaceInstUsesWith(I, RHS);
4601 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4605 case ICmpInst::ICMP_SLT:
4607 default: assert(0 && "Unknown integer condition code!");
4608 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4610 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4611 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4613 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4615 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4616 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4617 return ReplaceInstUsesWith(I, RHS);
4618 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4622 case ICmpInst::ICMP_UGT:
4624 default: assert(0 && "Unknown integer condition code!");
4625 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4626 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4627 return ReplaceInstUsesWith(I, LHS);
4628 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4630 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4631 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4632 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4633 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4637 case ICmpInst::ICMP_SGT:
4639 default: assert(0 && "Unknown integer condition code!");
4640 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4641 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4642 return ReplaceInstUsesWith(I, LHS);
4643 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4645 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4646 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4647 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4648 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4656 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4657 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4658 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4659 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4660 const Type *SrcTy = Op0C->getOperand(0)->getType();
4661 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4662 // Only do this if the casts both really cause code to be generated.
4663 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4665 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4667 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4668 Op1C->getOperand(0),
4670 InsertNewInstBefore(NewOp, I);
4671 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4676 return Changed ? &I : 0;
4679 // XorSelf - Implements: X ^ X --> 0
4682 XorSelf(Value *rhs) : RHS(rhs) {}
4683 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4684 Instruction *apply(BinaryOperator &Xor) const {
4690 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4691 bool Changed = SimplifyCommutative(I);
4692 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4694 if (isa<UndefValue>(Op1))
4695 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4697 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4698 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4699 assert(Result == &I && "AssociativeOpt didn't work?");
4700 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4703 // See if we can simplify any instructions used by the instruction whose sole
4704 // purpose is to compute bits we don't care about.
4705 uint64_t KnownZero, KnownOne;
4706 if (!isa<VectorType>(I.getType()) &&
4707 SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
4708 KnownZero, KnownOne))
4711 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4712 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
4713 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4714 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
4715 return new ICmpInst(ICI->getInversePredicate(),
4716 ICI->getOperand(0), ICI->getOperand(1));
4718 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4719 // ~(c-X) == X-c-1 == X+(-c-1)
4720 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4721 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4722 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4723 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4724 ConstantInt::get(I.getType(), 1));
4725 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4728 // ~(~X & Y) --> (X | ~Y)
4729 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
4730 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4731 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4733 BinaryOperator::createNot(Op0I->getOperand(1),
4734 Op0I->getOperand(1)->getName()+".not");
4735 InsertNewInstBefore(NotY, I);
4736 return BinaryOperator::createOr(Op0NotVal, NotY);
4740 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4741 if (Op0I->getOpcode() == Instruction::Add) {
4742 // ~(X-c) --> (-c-1)-X
4743 if (RHS->isAllOnesValue()) {
4744 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4745 return BinaryOperator::createSub(
4746 ConstantExpr::getSub(NegOp0CI,
4747 ConstantInt::get(I.getType(), 1)),
4748 Op0I->getOperand(0));
4750 } else if (Op0I->getOpcode() == Instruction::Or) {
4751 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4752 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
4753 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4754 // Anything in both C1 and C2 is known to be zero, remove it from
4756 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
4757 NewRHS = ConstantExpr::getAnd(NewRHS,
4758 ConstantExpr::getNot(CommonBits));
4759 AddToWorkList(Op0I);
4760 I.setOperand(0, Op0I->getOperand(0));
4761 I.setOperand(1, NewRHS);
4767 // Try to fold constant and into select arguments.
4768 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4769 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4771 if (isa<PHINode>(Op0))
4772 if (Instruction *NV = FoldOpIntoPhi(I))
4776 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4778 return ReplaceInstUsesWith(I,
4779 ConstantInt::getAllOnesValue(I.getType()));
4781 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4783 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
4786 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4789 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4790 if (A == Op0) { // B^(B|A) == (A|B)^B
4791 Op1I->swapOperands();
4793 std::swap(Op0, Op1);
4794 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4795 I.swapOperands(); // Simplified below.
4796 std::swap(Op0, Op1);
4798 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4799 if (Op0 == A) // A^(A^B) == B
4800 return ReplaceInstUsesWith(I, B);
4801 else if (Op0 == B) // A^(B^A) == B
4802 return ReplaceInstUsesWith(I, A);
4803 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4804 if (A == Op0) // A^(A&B) -> A^(B&A)
4805 Op1I->swapOperands();
4806 if (B == Op0) { // A^(B&A) -> (B&A)^A
4807 I.swapOperands(); // Simplified below.
4808 std::swap(Op0, Op1);
4813 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4816 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4817 if (A == Op1) // (B|A)^B == (A|B)^B
4819 if (B == Op1) { // (A|B)^B == A & ~B
4821 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4822 return BinaryOperator::createAnd(A, NotB);
4824 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4825 if (Op1 == A) // (A^B)^A == B
4826 return ReplaceInstUsesWith(I, B);
4827 else if (Op1 == B) // (B^A)^A == B
4828 return ReplaceInstUsesWith(I, A);
4829 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4830 if (A == Op1) // (A&B)^A -> (B&A)^A
4832 if (B == Op1 && // (B&A)^A == ~B & A
4833 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4835 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4836 return BinaryOperator::createAnd(N, Op1);
4841 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4842 if (Op0I && Op1I && Op0I->isShift() &&
4843 Op0I->getOpcode() == Op1I->getOpcode() &&
4844 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4845 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4846 Instruction *NewOp =
4847 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4848 Op1I->getOperand(0),
4849 Op0I->getName()), I);
4850 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4851 Op1I->getOperand(1));
4855 Value *A, *B, *C, *D;
4856 // (A & B)^(A | B) -> A ^ B
4857 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4858 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4859 if ((A == C && B == D) || (A == D && B == C))
4860 return BinaryOperator::createXor(A, B);
4862 // (A | B)^(A & B) -> A ^ B
4863 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4864 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4865 if ((A == C && B == D) || (A == D && B == C))
4866 return BinaryOperator::createXor(A, B);
4870 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4871 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4872 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4873 // (X & Y)^(X & Y) -> (Y^Z) & X
4874 Value *X = 0, *Y = 0, *Z = 0;
4876 X = A, Y = B, Z = D;
4878 X = A, Y = B, Z = C;
4880 X = B, Y = A, Z = D;
4882 X = B, Y = A, Z = C;
4885 Instruction *NewOp =
4886 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4887 return BinaryOperator::createAnd(NewOp, X);
4892 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4893 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4894 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4897 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4898 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4899 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4900 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4901 const Type *SrcTy = Op0C->getOperand(0)->getType();
4902 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4903 // Only do this if the casts both really cause code to be generated.
4904 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4906 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4908 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4909 Op1C->getOperand(0),
4911 InsertNewInstBefore(NewOp, I);
4912 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4916 return Changed ? &I : 0;
4919 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4920 /// overflowed for this type.
4921 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4922 ConstantInt *In2, bool IsSigned = false) {
4923 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4926 if (In2->getValue().isNegative())
4927 return Result->getValue().sgt(In1->getValue());
4929 return Result->getValue().slt(In1->getValue());
4931 return Result->getValue().ult(In1->getValue());
4934 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4935 /// code necessary to compute the offset from the base pointer (without adding
4936 /// in the base pointer). Return the result as a signed integer of intptr size.
4937 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4938 TargetData &TD = IC.getTargetData();
4939 gep_type_iterator GTI = gep_type_begin(GEP);
4940 const Type *IntPtrTy = TD.getIntPtrType();
4941 Value *Result = Constant::getNullValue(IntPtrTy);
4943 // Build a mask for high order bits.
4944 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4946 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4947 Value *Op = GEP->getOperand(i);
4948 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4949 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4950 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4951 if (!OpC->isNullValue()) {
4952 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4953 Scale = ConstantExpr::getMul(OpC, Scale);
4954 if (Constant *RC = dyn_cast<Constant>(Result))
4955 Result = ConstantExpr::getAdd(RC, Scale);
4957 // Emit an add instruction.
4958 Result = IC.InsertNewInstBefore(
4959 BinaryOperator::createAdd(Result, Scale,
4960 GEP->getName()+".offs"), I);
4964 // Convert to correct type.
4965 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4966 Op->getName()+".c"), I);
4968 // We'll let instcombine(mul) convert this to a shl if possible.
4969 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4970 GEP->getName()+".idx"), I);
4972 // Emit an add instruction.
4973 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4974 GEP->getName()+".offs"), I);
4980 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4981 /// else. At this point we know that the GEP is on the LHS of the comparison.
4982 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4983 ICmpInst::Predicate Cond,
4985 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4987 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4988 if (isa<PointerType>(CI->getOperand(0)->getType()))
4989 RHS = CI->getOperand(0);
4991 Value *PtrBase = GEPLHS->getOperand(0);
4992 if (PtrBase == RHS) {
4993 // As an optimization, we don't actually have to compute the actual value of
4994 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4995 // each index is zero or not.
4996 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4997 Instruction *InVal = 0;
4998 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4999 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
5001 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
5002 if (isa<UndefValue>(C)) // undef index -> undef.
5003 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
5004 if (C->isNullValue())
5006 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
5007 EmitIt = false; // This is indexing into a zero sized array?
5008 } else if (isa<ConstantInt>(C))
5009 return ReplaceInstUsesWith(I, // No comparison is needed here.
5010 ConstantInt::get(Type::Int1Ty,
5011 Cond == ICmpInst::ICMP_NE));
5016 new ICmpInst(Cond, GEPLHS->getOperand(i),
5017 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
5021 InVal = InsertNewInstBefore(InVal, I);
5022 InsertNewInstBefore(Comp, I);
5023 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
5024 InVal = BinaryOperator::createOr(InVal, Comp);
5025 else // True if all are equal
5026 InVal = BinaryOperator::createAnd(InVal, Comp);
5034 // No comparison is needed here, all indexes = 0
5035 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5036 Cond == ICmpInst::ICMP_EQ));
5039 // Only lower this if the icmp is the only user of the GEP or if we expect
5040 // the result to fold to a constant!
5041 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
5042 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5043 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
5044 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5045 Constant::getNullValue(Offset->getType()));
5047 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5048 // If the base pointers are different, but the indices are the same, just
5049 // compare the base pointer.
5050 if (PtrBase != GEPRHS->getOperand(0)) {
5051 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5052 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5053 GEPRHS->getOperand(0)->getType();
5055 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5056 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5057 IndicesTheSame = false;
5061 // If all indices are the same, just compare the base pointers.
5063 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5064 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5066 // Otherwise, the base pointers are different and the indices are
5067 // different, bail out.
5071 // If one of the GEPs has all zero indices, recurse.
5072 bool AllZeros = true;
5073 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5074 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5075 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5080 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5081 ICmpInst::getSwappedPredicate(Cond), I);
5083 // If the other GEP has all zero indices, recurse.
5085 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5086 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5087 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5092 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5094 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5095 // If the GEPs only differ by one index, compare it.
5096 unsigned NumDifferences = 0; // Keep track of # differences.
5097 unsigned DiffOperand = 0; // The operand that differs.
5098 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5099 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5100 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5101 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5102 // Irreconcilable differences.
5106 if (NumDifferences++) break;
5111 if (NumDifferences == 0) // SAME GEP?
5112 return ReplaceInstUsesWith(I, // No comparison is needed here.
5113 ConstantInt::get(Type::Int1Ty,
5114 Cond == ICmpInst::ICMP_EQ));
5115 else if (NumDifferences == 1) {
5116 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5117 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5118 // Make sure we do a signed comparison here.
5119 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5123 // Only lower this if the icmp is the only user of the GEP or if we expect
5124 // the result to fold to a constant!
5125 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5126 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5127 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5128 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5129 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5130 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5136 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5137 bool Changed = SimplifyCompare(I);
5138 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5140 // Fold trivial predicates.
5141 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5142 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5143 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5144 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5146 // Simplify 'fcmp pred X, X'
5148 switch (I.getPredicate()) {
5149 default: assert(0 && "Unknown predicate!");
5150 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5151 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5152 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5153 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5154 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5155 case FCmpInst::FCMP_OLT: // True if ordered and less than
5156 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5157 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5159 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5160 case FCmpInst::FCMP_ULT: // True if unordered or less than
5161 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5162 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5163 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5164 I.setPredicate(FCmpInst::FCMP_UNO);
5165 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5168 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5169 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5170 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5171 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5172 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5173 I.setPredicate(FCmpInst::FCMP_ORD);
5174 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5179 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5180 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5182 // Handle fcmp with constant RHS
5183 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5184 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5185 switch (LHSI->getOpcode()) {
5186 case Instruction::PHI:
5187 if (Instruction *NV = FoldOpIntoPhi(I))
5190 case Instruction::Select:
5191 // If either operand of the select is a constant, we can fold the
5192 // comparison into the select arms, which will cause one to be
5193 // constant folded and the select turned into a bitwise or.
5194 Value *Op1 = 0, *Op2 = 0;
5195 if (LHSI->hasOneUse()) {
5196 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5197 // Fold the known value into the constant operand.
5198 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5199 // Insert a new FCmp of the other select operand.
5200 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5201 LHSI->getOperand(2), RHSC,
5203 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5204 // Fold the known value into the constant operand.
5205 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5206 // Insert a new FCmp of the other select operand.
5207 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5208 LHSI->getOperand(1), RHSC,
5214 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5219 return Changed ? &I : 0;
5222 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5223 bool Changed = SimplifyCompare(I);
5224 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5225 const Type *Ty = Op0->getType();
5229 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5230 isTrueWhenEqual(I)));
5232 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5233 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5235 // icmp of GlobalValues can never equal each other as long as they aren't
5236 // external weak linkage type.
5237 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
5238 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
5239 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
5240 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5241 !isTrueWhenEqual(I)));
5243 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5244 // addresses never equal each other! We already know that Op0 != Op1.
5245 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5246 isa<ConstantPointerNull>(Op0)) &&
5247 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5248 isa<ConstantPointerNull>(Op1)))
5249 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5250 !isTrueWhenEqual(I)));
5252 // icmp's with boolean values can always be turned into bitwise operations
5253 if (Ty == Type::Int1Ty) {
5254 switch (I.getPredicate()) {
5255 default: assert(0 && "Invalid icmp instruction!");
5256 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5257 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
5258 InsertNewInstBefore(Xor, I);
5259 return BinaryOperator::createNot(Xor);
5261 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5262 return BinaryOperator::createXor(Op0, Op1);
5264 case ICmpInst::ICMP_UGT:
5265 case ICmpInst::ICMP_SGT:
5266 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5268 case ICmpInst::ICMP_ULT:
5269 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5270 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
5271 InsertNewInstBefore(Not, I);
5272 return BinaryOperator::createAnd(Not, Op1);
5274 case ICmpInst::ICMP_UGE:
5275 case ICmpInst::ICMP_SGE:
5276 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5278 case ICmpInst::ICMP_ULE:
5279 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5280 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
5281 InsertNewInstBefore(Not, I);
5282 return BinaryOperator::createOr(Not, Op1);
5287 // See if we are doing a comparison between a constant and an instruction that
5288 // can be folded into the comparison.
5289 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5290 switch (I.getPredicate()) {
5292 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5293 if (CI->isMinValue(false))
5294 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5295 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5296 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5297 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5298 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5301 case ICmpInst::ICMP_SLT:
5302 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5303 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5304 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5305 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5306 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5307 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5310 case ICmpInst::ICMP_UGT:
5311 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5312 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5313 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5314 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5315 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5316 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5319 case ICmpInst::ICMP_SGT:
5320 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5321 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5322 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5323 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5324 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5325 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5328 case ICmpInst::ICMP_ULE:
5329 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5330 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5331 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5332 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5333 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5334 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5337 case ICmpInst::ICMP_SLE:
5338 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5339 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5340 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5341 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5342 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5343 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5346 case ICmpInst::ICMP_UGE:
5347 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5348 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5349 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5350 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5351 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5352 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5355 case ICmpInst::ICMP_SGE:
5356 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5357 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5358 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5359 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5360 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5361 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5365 // If we still have a icmp le or icmp ge instruction, turn it into the
5366 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5367 // already been handled above, this requires little checking.
5369 if (I.getPredicate() == ICmpInst::ICMP_ULE)
5370 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5371 if (I.getPredicate() == ICmpInst::ICMP_SLE)
5372 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5373 if (I.getPredicate() == ICmpInst::ICMP_UGE)
5374 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5375 if (I.getPredicate() == ICmpInst::ICMP_SGE)
5376 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5378 // See if we can fold the comparison based on bits known to be zero or one
5380 uint64_t KnownZero, KnownOne;
5381 if (SimplifyDemandedBits(Op0, cast<IntegerType>(Ty)->getBitMask(),
5382 KnownZero, KnownOne, 0))
5385 // Given the known and unknown bits, compute a range that the LHS could be
5387 if (KnownOne | KnownZero) {
5388 // Compute the Min, Max and RHS values based on the known bits. For the
5389 // EQ and NE we use unsigned values.
5390 uint64_t UMin = 0, UMax = 0, URHSVal = 0;
5391 int64_t SMin = 0, SMax = 0, SRHSVal = 0;
5392 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5393 SRHSVal = CI->getSExtValue();
5394 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
5397 URHSVal = CI->getZExtValue();
5398 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
5401 switch (I.getPredicate()) { // LE/GE have been folded already.
5402 default: assert(0 && "Unknown icmp opcode!");
5403 case ICmpInst::ICMP_EQ:
5404 if (UMax < URHSVal || UMin > URHSVal)
5405 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5407 case ICmpInst::ICMP_NE:
5408 if (UMax < URHSVal || UMin > URHSVal)
5409 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5411 case ICmpInst::ICMP_ULT:
5413 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5415 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5417 case ICmpInst::ICMP_UGT:
5419 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5421 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5423 case ICmpInst::ICMP_SLT:
5425 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5427 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5429 case ICmpInst::ICMP_SGT:
5431 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
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) &&
5458 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
5459 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
5460 ConstantInt *NewCST;
5462 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
5463 AndCST->getZExtValue());
5464 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
5465 CI->getZExtValue());
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 uint64_t Val = ~0ULL; // All ones.
5638 Val <<= ShAmtVal; // Shift over to the right spot.
5639 Val &= ~0ULL >> (64-TypeBits);
5640 Constant *Mask = ConstantInt::get(CI->getType(), Val);
5643 BinaryOperator::createAnd(LHSI->getOperand(0),
5644 Mask, LHSI->getName()+".mask");
5645 Value *And = InsertNewInstBefore(AndI, I);
5646 return new ICmpInst(I.getPredicate(), And,
5647 ConstantExpr::getShl(CI, ShAmt));
5653 case Instruction::SDiv:
5654 case Instruction::UDiv:
5655 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5656 // Fold this div into the comparison, producing a range check.
5657 // Determine, based on the divide type, what the range is being
5658 // checked. If there is an overflow on the low or high side, remember
5659 // it, otherwise compute the range [low, hi) bounding the new value.
5660 // See: InsertRangeTest above for the kinds of replacements possible.
5661 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5662 // FIXME: If the operand types don't match the type of the divide
5663 // then don't attempt this transform. The code below doesn't have the
5664 // logic to deal with a signed divide and an unsigned compare (and
5665 // vice versa). This is because (x /s C1) <s C2 produces different
5666 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5667 // (x /u C1) <u C2. Simply casting the operands and result won't
5668 // work. :( The if statement below tests that condition and bails
5670 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
5671 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
5673 if (DivRHS->isZero())
5674 break; // Don't hack on div by zero
5676 // Initialize the variables that will indicate the nature of the
5678 bool LoOverflow = false, HiOverflow = false;
5679 ConstantInt *LoBound = 0, *HiBound = 0;
5681 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5682 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5683 // C2 (CI). By solving for X we can turn this into a range check
5684 // instead of computing a divide.
5686 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
5688 // Determine if the product overflows by seeing if the product is
5689 // not equal to the divide. Make sure we do the same kind of divide
5690 // as in the LHS instruction that we're folding.
5691 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5692 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
5694 // Get the ICmp opcode
5695 ICmpInst::Predicate predicate = I.getPredicate();
5697 if (!DivIsSigned) { // udiv
5699 LoOverflow = ProdOV;
5700 HiOverflow = ProdOV ||
5701 AddWithOverflow(HiBound, LoBound, DivRHS, false);
5702 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5703 if (CI->isNullValue()) { // (X / pos) op 0
5705 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5707 } else if (CI->getValue().isPositive()) { // (X / pos) op pos
5709 LoOverflow = ProdOV;
5710 HiOverflow = ProdOV ||
5711 AddWithOverflow(HiBound, Prod, DivRHS, true);
5712 } else { // (X / pos) op neg
5713 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5714 LoOverflow = AddWithOverflow(LoBound, Prod,
5715 cast<ConstantInt>(DivRHSH), true);
5716 HiBound = AddOne(Prod);
5717 HiOverflow = ProdOV;
5719 } else { // Divisor is < 0.
5720 if (CI->isNullValue()) { // (X / neg) op 0
5721 LoBound = AddOne(DivRHS);
5722 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5723 if (HiBound == DivRHS)
5724 LoBound = 0; // - INTMIN = INTMIN
5725 } else if (CI->getValue().isPositive()) { // (X / neg) op pos
5726 HiOverflow = LoOverflow = ProdOV;
5728 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS),
5730 HiBound = AddOne(Prod);
5731 } else { // (X / neg) op neg
5733 LoOverflow = HiOverflow = ProdOV;
5734 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
5737 // Dividing by a negate swaps the condition.
5738 predicate = ICmpInst::getSwappedPredicate(predicate);
5742 Value *X = LHSI->getOperand(0);
5743 switch (predicate) {
5744 default: assert(0 && "Unhandled icmp opcode!");
5745 case ICmpInst::ICMP_EQ:
5746 if (LoOverflow && HiOverflow)
5747 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5748 else if (HiOverflow)
5749 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5750 ICmpInst::ICMP_UGE, X, LoBound);
5751 else if (LoOverflow)
5752 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5753 ICmpInst::ICMP_ULT, X, HiBound);
5755 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5757 case ICmpInst::ICMP_NE:
5758 if (LoOverflow && HiOverflow)
5759 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5760 else if (HiOverflow)
5761 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5762 ICmpInst::ICMP_ULT, X, LoBound);
5763 else if (LoOverflow)
5764 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5765 ICmpInst::ICMP_UGE, X, HiBound);
5767 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5769 case ICmpInst::ICMP_ULT:
5770 case ICmpInst::ICMP_SLT:
5772 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5773 return new ICmpInst(predicate, X, LoBound);
5774 case ICmpInst::ICMP_UGT:
5775 case ICmpInst::ICMP_SGT:
5777 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5778 if (predicate == ICmpInst::ICMP_UGT)
5779 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5781 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5788 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5789 if (I.isEquality()) {
5790 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
5792 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5793 // the second operand is a constant, simplify a bit.
5794 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
5795 switch (BO->getOpcode()) {
5796 case Instruction::SRem:
5797 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5798 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
5800 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
5801 if (V > 1 && isPowerOf2_64(V)) {
5802 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
5803 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
5804 return new ICmpInst(I.getPredicate(), NewRem,
5805 Constant::getNullValue(BO->getType()));
5809 case Instruction::Add:
5810 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5811 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5812 if (BO->hasOneUse())
5813 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5814 ConstantExpr::getSub(CI, BOp1C));
5815 } else if (CI->isNullValue()) {
5816 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5817 // efficiently invertible, or if the add has just this one use.
5818 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5820 if (Value *NegVal = dyn_castNegVal(BOp1))
5821 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
5822 else if (Value *NegVal = dyn_castNegVal(BOp0))
5823 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
5824 else if (BO->hasOneUse()) {
5825 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5826 InsertNewInstBefore(Neg, I);
5828 return new ICmpInst(I.getPredicate(), BOp0, Neg);
5832 case Instruction::Xor:
5833 // For the xor case, we can xor two constants together, eliminating
5834 // the explicit xor.
5835 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5836 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5837 ConstantExpr::getXor(CI, BOC));
5840 case Instruction::Sub:
5841 // Replace (([sub|xor] A, B) != 0) with (A != B)
5842 if (CI->isNullValue())
5843 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5847 case Instruction::Or:
5848 // If bits are being or'd in that are not present in the constant we
5849 // are comparing against, then the comparison could never succeed!
5850 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5851 Constant *NotCI = ConstantExpr::getNot(CI);
5852 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5853 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5858 case Instruction::And:
5859 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5860 // If bits are being compared against that are and'd out, then the
5861 // comparison can never succeed!
5862 if (!ConstantExpr::getAnd(CI,
5863 ConstantExpr::getNot(BOC))->isNullValue())
5864 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5867 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5868 if (CI == BOC && isOneBitSet(CI))
5869 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5870 ICmpInst::ICMP_NE, Op0,
5871 Constant::getNullValue(CI->getType()));
5873 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5874 if (isSignBit(BOC)) {
5875 Value *X = BO->getOperand(0);
5876 Constant *Zero = Constant::getNullValue(X->getType());
5877 ICmpInst::Predicate pred = isICMP_NE ?
5878 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5879 return new ICmpInst(pred, X, Zero);
5882 // ((X & ~7) == 0) --> X < 8
5883 if (CI->isNullValue() && isHighOnes(BOC)) {
5884 Value *X = BO->getOperand(0);
5885 Constant *NegX = ConstantExpr::getNeg(BOC);
5886 ICmpInst::Predicate pred = isICMP_NE ?
5887 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5888 return new ICmpInst(pred, X, NegX);
5894 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5895 // Handle set{eq|ne} <intrinsic>, intcst.
5896 switch (II->getIntrinsicID()) {
5898 case Intrinsic::bswap_i16:
5899 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5900 AddToWorkList(II); // Dead?
5901 I.setOperand(0, II->getOperand(1));
5902 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5903 ByteSwap_16(CI->getZExtValue())));
5905 case Intrinsic::bswap_i32:
5906 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5907 AddToWorkList(II); // Dead?
5908 I.setOperand(0, II->getOperand(1));
5909 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5910 ByteSwap_32(CI->getZExtValue())));
5912 case Intrinsic::bswap_i64:
5913 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5914 AddToWorkList(II); // Dead?
5915 I.setOperand(0, II->getOperand(1));
5916 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5917 ByteSwap_64(CI->getZExtValue())));
5921 } else { // Not a ICMP_EQ/ICMP_NE
5922 // If the LHS is a cast from an integral value of the same size, then
5923 // since we know the RHS is a constant, try to simlify.
5924 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5925 Value *CastOp = Cast->getOperand(0);
5926 const Type *SrcTy = CastOp->getType();
5927 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5928 if (SrcTy->isInteger() &&
5929 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5930 // If this is an unsigned comparison, try to make the comparison use
5931 // smaller constant values.
5932 switch (I.getPredicate()) {
5934 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5935 ConstantInt *CUI = cast<ConstantInt>(CI);
5936 if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
5937 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5938 ConstantInt::get(SrcTy, -1ULL));
5941 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5942 ConstantInt *CUI = cast<ConstantInt>(CI);
5943 if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
5944 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5945 Constant::getNullValue(SrcTy));
5955 // Handle icmp with constant RHS
5956 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5957 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5958 switch (LHSI->getOpcode()) {
5959 case Instruction::GetElementPtr:
5960 if (RHSC->isNullValue()) {
5961 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5962 bool isAllZeros = true;
5963 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5964 if (!isa<Constant>(LHSI->getOperand(i)) ||
5965 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5970 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5971 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5975 case Instruction::PHI:
5976 if (Instruction *NV = FoldOpIntoPhi(I))
5979 case Instruction::Select:
5980 // If either operand of the select is a constant, we can fold the
5981 // comparison into the select arms, which will cause one to be
5982 // constant folded and the select turned into a bitwise or.
5983 Value *Op1 = 0, *Op2 = 0;
5984 if (LHSI->hasOneUse()) {
5985 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5986 // Fold the known value into the constant operand.
5987 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5988 // Insert a new ICmp of the other select operand.
5989 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5990 LHSI->getOperand(2), RHSC,
5992 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5993 // Fold the known value into the constant operand.
5994 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5995 // Insert a new ICmp of the other select operand.
5996 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5997 LHSI->getOperand(1), RHSC,
6003 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
6008 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6009 if (User *GEP = dyn_castGetElementPtr(Op0))
6010 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6012 if (User *GEP = dyn_castGetElementPtr(Op1))
6013 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6014 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6017 // Test to see if the operands of the icmp are casted versions of other
6018 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6020 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6021 if (isa<PointerType>(Op0->getType()) &&
6022 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6023 // We keep moving the cast from the left operand over to the right
6024 // operand, where it can often be eliminated completely.
6025 Op0 = CI->getOperand(0);
6027 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6028 // so eliminate it as well.
6029 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6030 Op1 = CI2->getOperand(0);
6032 // If Op1 is a constant, we can fold the cast into the constant.
6033 if (Op0->getType() != Op1->getType())
6034 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6035 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6037 // Otherwise, cast the RHS right before the icmp
6038 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
6040 return new ICmpInst(I.getPredicate(), Op0, Op1);
6044 if (isa<CastInst>(Op0)) {
6045 // Handle the special case of: icmp (cast bool to X), <cst>
6046 // This comes up when you have code like
6049 // For generality, we handle any zero-extension of any operand comparison
6050 // with a constant or another cast from the same type.
6051 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6052 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6056 if (I.isEquality()) {
6057 Value *A, *B, *C, *D;
6058 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6059 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6060 Value *OtherVal = A == Op1 ? B : A;
6061 return new ICmpInst(I.getPredicate(), OtherVal,
6062 Constant::getNullValue(A->getType()));
6065 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6066 // A^c1 == C^c2 --> A == C^(c1^c2)
6067 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
6068 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
6069 if (Op1->hasOneUse()) {
6070 Constant *NC = ConstantExpr::getXor(C1, C2);
6071 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
6072 return new ICmpInst(I.getPredicate(), A,
6073 InsertNewInstBefore(Xor, I));
6076 // A^B == A^D -> B == D
6077 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6078 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6079 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6080 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6084 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6085 (A == Op0 || B == Op0)) {
6086 // A == (A^B) -> B == 0
6087 Value *OtherVal = A == Op0 ? B : A;
6088 return new ICmpInst(I.getPredicate(), OtherVal,
6089 Constant::getNullValue(A->getType()));
6091 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
6092 // (A-B) == A -> B == 0
6093 return new ICmpInst(I.getPredicate(), B,
6094 Constant::getNullValue(B->getType()));
6096 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
6097 // A == (A-B) -> B == 0
6098 return new ICmpInst(I.getPredicate(), B,
6099 Constant::getNullValue(B->getType()));
6102 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6103 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6104 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6105 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6106 Value *X = 0, *Y = 0, *Z = 0;
6109 X = B; Y = D; Z = A;
6110 } else if (A == D) {
6111 X = B; Y = C; Z = A;
6112 } else if (B == C) {
6113 X = A; Y = D; Z = B;
6114 } else if (B == D) {
6115 X = A; Y = C; Z = B;
6118 if (X) { // Build (X^Y) & Z
6119 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
6120 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
6121 I.setOperand(0, Op1);
6122 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6127 return Changed ? &I : 0;
6130 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6131 // We only handle extending casts so far.
6133 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6134 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6135 Value *LHSCIOp = LHSCI->getOperand(0);
6136 const Type *SrcTy = LHSCIOp->getType();
6137 const Type *DestTy = LHSCI->getType();
6140 // We only handle extension cast instructions, so far. Enforce this.
6141 if (LHSCI->getOpcode() != Instruction::ZExt &&
6142 LHSCI->getOpcode() != Instruction::SExt)
6145 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6146 bool isSignedCmp = ICI.isSignedPredicate();
6148 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6149 // Not an extension from the same type?
6150 RHSCIOp = CI->getOperand(0);
6151 if (RHSCIOp->getType() != LHSCIOp->getType())
6154 // If the signedness of the two compares doesn't agree (i.e. one is a sext
6155 // and the other is a zext), then we can't handle this.
6156 if (CI->getOpcode() != LHSCI->getOpcode())
6159 // Likewise, if the signedness of the [sz]exts and the compare don't match,
6160 // then we can't handle this.
6161 if (isSignedExt != isSignedCmp && !ICI.isEquality())
6164 // Okay, just insert a compare of the reduced operands now!
6165 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6168 // If we aren't dealing with a constant on the RHS, exit early
6169 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6173 // Compute the constant that would happen if we truncated to SrcTy then
6174 // reextended to DestTy.
6175 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6176 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6178 // If the re-extended constant didn't change...
6180 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6181 // For example, we might have:
6182 // %A = sext short %X to uint
6183 // %B = icmp ugt uint %A, 1330
6184 // It is incorrect to transform this into
6185 // %B = icmp ugt short %X, 1330
6186 // because %A may have negative value.
6188 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6189 // OR operation is EQ/NE.
6190 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6191 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6196 // The re-extended constant changed so the constant cannot be represented
6197 // in the shorter type. Consequently, we cannot emit a simple comparison.
6199 // First, handle some easy cases. We know the result cannot be equal at this
6200 // point so handle the ICI.isEquality() cases
6201 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6202 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6203 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6204 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6206 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6207 // should have been folded away previously and not enter in here.
6210 // We're performing a signed comparison.
6211 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
6212 Result = ConstantInt::getFalse(); // X < (small) --> false
6214 Result = ConstantInt::getTrue(); // X < (large) --> true
6216 // We're performing an unsigned comparison.
6218 // We're performing an unsigned comp with a sign extended value.
6219 // This is true if the input is >= 0. [aka >s -1]
6220 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6221 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6222 NegOne, ICI.getName()), ICI);
6224 // Unsigned extend & unsigned compare -> always true.
6225 Result = ConstantInt::getTrue();
6229 // Finally, return the value computed.
6230 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6231 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6232 return ReplaceInstUsesWith(ICI, Result);
6234 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6235 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6236 "ICmp should be folded!");
6237 if (Constant *CI = dyn_cast<Constant>(Result))
6238 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6240 return BinaryOperator::createNot(Result);
6244 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6245 return commonShiftTransforms(I);
6248 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6249 return commonShiftTransforms(I);
6252 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6253 return commonShiftTransforms(I);
6256 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6257 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6258 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6260 // shl X, 0 == X and shr X, 0 == X
6261 // shl 0, X == 0 and shr 0, X == 0
6262 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6263 Op0 == Constant::getNullValue(Op0->getType()))
6264 return ReplaceInstUsesWith(I, Op0);
6266 if (isa<UndefValue>(Op0)) {
6267 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6268 return ReplaceInstUsesWith(I, Op0);
6269 else // undef << X -> 0, undef >>u X -> 0
6270 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6272 if (isa<UndefValue>(Op1)) {
6273 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6274 return ReplaceInstUsesWith(I, Op0);
6275 else // X << undef, X >>u undef -> 0
6276 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6279 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6280 if (I.getOpcode() == Instruction::AShr)
6281 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6282 if (CSI->isAllOnesValue())
6283 return ReplaceInstUsesWith(I, CSI);
6285 // Try to fold constant and into select arguments.
6286 if (isa<Constant>(Op0))
6287 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6288 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6291 // See if we can turn a signed shr into an unsigned shr.
6292 if (I.isArithmeticShift()) {
6293 if (MaskedValueIsZero(Op0,
6294 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
6295 return BinaryOperator::createLShr(Op0, Op1, I.getName());
6299 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6300 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6305 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6306 BinaryOperator &I) {
6307 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6309 // See if we can simplify any instructions used by the instruction whose sole
6310 // purpose is to compute bits we don't care about.
6311 uint64_t KnownZero, KnownOne;
6312 if (SimplifyDemandedBits(&I, cast<IntegerType>(I.getType())->getBitMask(),
6313 KnownZero, KnownOne))
6316 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6317 // of a signed value.
6319 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6320 if (Op1->getZExtValue() >= TypeBits) {
6321 if (I.getOpcode() != Instruction::AShr)
6322 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6324 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6329 // ((X*C1) << C2) == (X * (C1 << C2))
6330 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6331 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6332 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6333 return BinaryOperator::createMul(BO->getOperand(0),
6334 ConstantExpr::getShl(BOOp, Op1));
6336 // Try to fold constant and into select arguments.
6337 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6338 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6340 if (isa<PHINode>(Op0))
6341 if (Instruction *NV = FoldOpIntoPhi(I))
6344 if (Op0->hasOneUse()) {
6345 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6346 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6349 switch (Op0BO->getOpcode()) {
6351 case Instruction::Add:
6352 case Instruction::And:
6353 case Instruction::Or:
6354 case Instruction::Xor: {
6355 // These operators commute.
6356 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6357 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6358 match(Op0BO->getOperand(1),
6359 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6360 Instruction *YS = BinaryOperator::createShl(
6361 Op0BO->getOperand(0), Op1,
6363 InsertNewInstBefore(YS, I); // (Y << C)
6365 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6366 Op0BO->getOperand(1)->getName());
6367 InsertNewInstBefore(X, I); // (X + (Y << C))
6368 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
6369 C2 = ConstantExpr::getShl(C2, Op1);
6370 return BinaryOperator::createAnd(X, C2);
6373 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6374 Value *Op0BOOp1 = Op0BO->getOperand(1);
6375 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6377 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6378 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6380 Instruction *YS = BinaryOperator::createShl(
6381 Op0BO->getOperand(0), Op1,
6383 InsertNewInstBefore(YS, I); // (Y << C)
6385 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6386 V1->getName()+".mask");
6387 InsertNewInstBefore(XM, I); // X & (CC << C)
6389 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6394 case Instruction::Sub: {
6395 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6396 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6397 match(Op0BO->getOperand(0),
6398 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6399 Instruction *YS = BinaryOperator::createShl(
6400 Op0BO->getOperand(1), Op1,
6402 InsertNewInstBefore(YS, I); // (Y << C)
6404 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6405 Op0BO->getOperand(0)->getName());
6406 InsertNewInstBefore(X, I); // (X + (Y << C))
6407 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
6408 C2 = ConstantExpr::getShl(C2, Op1);
6409 return BinaryOperator::createAnd(X, C2);
6412 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6413 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6414 match(Op0BO->getOperand(0),
6415 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6416 m_ConstantInt(CC))) && V2 == Op1 &&
6417 cast<BinaryOperator>(Op0BO->getOperand(0))
6418 ->getOperand(0)->hasOneUse()) {
6419 Instruction *YS = BinaryOperator::createShl(
6420 Op0BO->getOperand(1), Op1,
6422 InsertNewInstBefore(YS, I); // (Y << C)
6424 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6425 V1->getName()+".mask");
6426 InsertNewInstBefore(XM, I); // X & (CC << C)
6428 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6436 // If the operand is an bitwise operator with a constant RHS, and the
6437 // shift is the only use, we can pull it out of the shift.
6438 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6439 bool isValid = true; // Valid only for And, Or, Xor
6440 bool highBitSet = false; // Transform if high bit of constant set?
6442 switch (Op0BO->getOpcode()) {
6443 default: isValid = false; break; // Do not perform transform!
6444 case Instruction::Add:
6445 isValid = isLeftShift;
6447 case Instruction::Or:
6448 case Instruction::Xor:
6451 case Instruction::And:
6456 // If this is a signed shift right, and the high bit is modified
6457 // by the logical operation, do not perform the transformation.
6458 // The highBitSet boolean indicates the value of the high bit of
6459 // the constant which would cause it to be modified for this
6462 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
6463 uint64_t Val = Op0C->getZExtValue();
6464 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
6468 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6470 Instruction *NewShift =
6471 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6472 InsertNewInstBefore(NewShift, I);
6473 NewShift->takeName(Op0BO);
6475 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6482 // Find out if this is a shift of a shift by a constant.
6483 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6484 if (ShiftOp && !ShiftOp->isShift())
6487 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6488 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6489 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
6490 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
6491 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6492 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6493 Value *X = ShiftOp->getOperand(0);
6495 unsigned AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6496 if (AmtSum > I.getType()->getPrimitiveSizeInBits())
6497 AmtSum = I.getType()->getPrimitiveSizeInBits();
6499 const IntegerType *Ty = cast<IntegerType>(I.getType());
6501 // Check for (X << c1) << c2 and (X >> c1) >> c2
6502 if (I.getOpcode() == ShiftOp->getOpcode()) {
6503 return BinaryOperator::create(I.getOpcode(), X,
6504 ConstantInt::get(Ty, AmtSum));
6505 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6506 I.getOpcode() == Instruction::AShr) {
6507 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6508 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6509 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6510 I.getOpcode() == Instruction::LShr) {
6511 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6512 Instruction *Shift =
6513 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6514 InsertNewInstBefore(Shift, I);
6516 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
6517 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6520 // Okay, if we get here, one shift must be left, and the other shift must be
6521 // right. See if the amounts are equal.
6522 if (ShiftAmt1 == ShiftAmt2) {
6523 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6524 if (I.getOpcode() == Instruction::Shl) {
6525 uint64_t Mask = Ty->getBitMask() << ShiftAmt1;
6526 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
6528 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6529 if (I.getOpcode() == Instruction::LShr) {
6530 uint64_t Mask = Ty->getBitMask() >> ShiftAmt1;
6531 return BinaryOperator::createAnd(X, ConstantInt::get(Ty, Mask));
6533 // We can simplify ((X << C) >>s C) into a trunc + sext.
6534 // NOTE: we could do this for any C, but that would make 'unusual' integer
6535 // types. For now, just stick to ones well-supported by the code
6537 const Type *SExtType = 0;
6538 switch (Ty->getBitWidth() - ShiftAmt1) {
6539 case 8 : SExtType = Type::Int8Ty; break;
6540 case 16: SExtType = Type::Int16Ty; break;
6541 case 32: SExtType = Type::Int32Ty; break;
6545 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6546 InsertNewInstBefore(NewTrunc, I);
6547 return new SExtInst(NewTrunc, Ty);
6549 // Otherwise, we can't handle it yet.
6550 } else if (ShiftAmt1 < ShiftAmt2) {
6551 unsigned ShiftDiff = ShiftAmt2-ShiftAmt1;
6553 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6554 if (I.getOpcode() == Instruction::Shl) {
6555 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6556 ShiftOp->getOpcode() == Instruction::AShr);
6557 Instruction *Shift =
6558 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6559 InsertNewInstBefore(Shift, I);
6561 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
6562 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6565 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6566 if (I.getOpcode() == Instruction::LShr) {
6567 assert(ShiftOp->getOpcode() == Instruction::Shl);
6568 Instruction *Shift =
6569 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6570 InsertNewInstBefore(Shift, I);
6572 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
6573 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6576 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6578 assert(ShiftAmt2 < ShiftAmt1);
6579 unsigned ShiftDiff = ShiftAmt1-ShiftAmt2;
6581 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6582 if (I.getOpcode() == Instruction::Shl) {
6583 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6584 ShiftOp->getOpcode() == Instruction::AShr);
6585 Instruction *Shift =
6586 BinaryOperator::create(ShiftOp->getOpcode(), X,
6587 ConstantInt::get(Ty, ShiftDiff));
6588 InsertNewInstBefore(Shift, I);
6590 uint64_t Mask = Ty->getBitMask() << ShiftAmt2;
6591 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6594 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6595 if (I.getOpcode() == Instruction::LShr) {
6596 assert(ShiftOp->getOpcode() == Instruction::Shl);
6597 Instruction *Shift =
6598 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6599 InsertNewInstBefore(Shift, I);
6601 uint64_t Mask = Ty->getBitMask() >> ShiftAmt2;
6602 return BinaryOperator::createAnd(Shift, ConstantInt::get(Ty, Mask));
6605 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6612 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6613 /// expression. If so, decompose it, returning some value X, such that Val is
6616 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6618 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6619 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6620 Offset = CI->getZExtValue();
6622 return ConstantInt::get(Type::Int32Ty, 0);
6623 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
6624 if (I->getNumOperands() == 2) {
6625 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6626 if (I->getOpcode() == Instruction::Shl) {
6627 // This is a value scaled by '1 << the shift amt'.
6628 Scale = 1U << CUI->getZExtValue();
6630 return I->getOperand(0);
6631 } else if (I->getOpcode() == Instruction::Mul) {
6632 // This value is scaled by 'CUI'.
6633 Scale = CUI->getZExtValue();
6635 return I->getOperand(0);
6636 } else if (I->getOpcode() == Instruction::Add) {
6637 // We have X+C. Check to see if we really have (X*C2)+C1,
6638 // where C1 is divisible by C2.
6641 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6642 Offset += CUI->getZExtValue();
6643 if (SubScale > 1 && (Offset % SubScale == 0)) {
6652 // Otherwise, we can't look past this.
6659 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6660 /// try to eliminate the cast by moving the type information into the alloc.
6661 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
6662 AllocationInst &AI) {
6663 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
6664 if (!PTy) return 0; // Not casting the allocation to a pointer type.
6666 // Remove any uses of AI that are dead.
6667 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6669 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6670 Instruction *User = cast<Instruction>(*UI++);
6671 if (isInstructionTriviallyDead(User)) {
6672 while (UI != E && *UI == User)
6673 ++UI; // If this instruction uses AI more than once, don't break UI.
6676 DOUT << "IC: DCE: " << *User;
6677 EraseInstFromFunction(*User);
6681 // Get the type really allocated and the type casted to.
6682 const Type *AllocElTy = AI.getAllocatedType();
6683 const Type *CastElTy = PTy->getElementType();
6684 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6686 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6687 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6688 if (CastElTyAlign < AllocElTyAlign) return 0;
6690 // If the allocation has multiple uses, only promote it if we are strictly
6691 // increasing the alignment of the resultant allocation. If we keep it the
6692 // same, we open the door to infinite loops of various kinds.
6693 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6695 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
6696 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
6697 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6699 // See if we can satisfy the modulus by pulling a scale out of the array
6701 unsigned ArraySizeScale, ArrayOffset;
6702 Value *NumElements = // See if the array size is a decomposable linear expr.
6703 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6705 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6707 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6708 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6710 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6715 // If the allocation size is constant, form a constant mul expression
6716 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6717 if (isa<ConstantInt>(NumElements))
6718 Amt = ConstantExpr::getMul(
6719 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6720 // otherwise multiply the amount and the number of elements
6721 else if (Scale != 1) {
6722 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6723 Amt = InsertNewInstBefore(Tmp, AI);
6727 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6728 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
6729 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6730 Amt = InsertNewInstBefore(Tmp, AI);
6733 AllocationInst *New;
6734 if (isa<MallocInst>(AI))
6735 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6737 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6738 InsertNewInstBefore(New, AI);
6741 // If the allocation has multiple uses, insert a cast and change all things
6742 // that used it to use the new cast. This will also hack on CI, but it will
6744 if (!AI.hasOneUse()) {
6745 AddUsesToWorkList(AI);
6746 // New is the allocation instruction, pointer typed. AI is the original
6747 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6748 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6749 InsertNewInstBefore(NewCast, AI);
6750 AI.replaceAllUsesWith(NewCast);
6752 return ReplaceInstUsesWith(CI, New);
6755 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6756 /// and return it as type Ty without inserting any new casts and without
6757 /// changing the computed value. This is used by code that tries to decide
6758 /// whether promoting or shrinking integer operations to wider or smaller types
6759 /// will allow us to eliminate a truncate or extend.
6761 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6762 /// extension operation if Ty is larger.
6763 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6764 int &NumCastsRemoved) {
6765 // We can always evaluate constants in another type.
6766 if (isa<ConstantInt>(V))
6769 Instruction *I = dyn_cast<Instruction>(V);
6770 if (!I) return false;
6772 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6774 switch (I->getOpcode()) {
6775 case Instruction::Add:
6776 case Instruction::Sub:
6777 case Instruction::And:
6778 case Instruction::Or:
6779 case Instruction::Xor:
6780 if (!I->hasOneUse()) return false;
6781 // These operators can all arbitrarily be extended or truncated.
6782 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
6783 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
6785 case Instruction::Shl:
6786 if (!I->hasOneUse()) return false;
6787 // If we are truncating the result of this SHL, and if it's a shift of a
6788 // constant amount, we can always perform a SHL in a smaller type.
6789 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6790 if (Ty->getBitWidth() < OrigTy->getBitWidth() &&
6791 CI->getZExtValue() < Ty->getBitWidth())
6792 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6795 case Instruction::LShr:
6796 if (!I->hasOneUse()) return false;
6797 // If this is a truncate of a logical shr, we can truncate it to a smaller
6798 // lshr iff we know that the bits we would otherwise be shifting in are
6800 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6801 if (Ty->getBitWidth() < OrigTy->getBitWidth() &&
6802 MaskedValueIsZero(I->getOperand(0),
6803 OrigTy->getBitMask() & ~Ty->getBitMask()) &&
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 uint64_t KnownZero = 0, KnownOne = 0;
6955 if (SimplifyDemandedBits(&CI, cast<IntegerType>(DestTy)->getBitMask(),
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?");
7012 ConstantInt::get(Type::Int64Ty, (1ULL << SrcBitSize)-1);
7013 if (DestBitSize < 64)
7014 C = ConstantExpr::getTrunc(C, DestTy);
7015 return BinaryOperator::createAnd(Res, C);
7017 case Instruction::SExt:
7018 // We need to emit a cast to truncate, then a cast to sext.
7019 return CastInst::create(Instruction::SExt,
7020 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7026 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7027 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7029 switch (SrcI->getOpcode()) {
7030 case Instruction::Add:
7031 case Instruction::Mul:
7032 case Instruction::And:
7033 case Instruction::Or:
7034 case Instruction::Xor:
7035 // If we are discarding information, or just changing the sign,
7037 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7038 // Don't insert two casts if they cannot be eliminated. We allow
7039 // two casts to be inserted if the sizes are the same. This could
7040 // only be converting signedness, which is a noop.
7041 if (DestBitSize == SrcBitSize ||
7042 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7043 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7044 Instruction::CastOps opcode = CI.getOpcode();
7045 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7046 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7047 return BinaryOperator::create(
7048 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7052 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7053 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7054 SrcI->getOpcode() == Instruction::Xor &&
7055 Op1 == ConstantInt::getTrue() &&
7056 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7057 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7058 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
7061 case Instruction::SDiv:
7062 case Instruction::UDiv:
7063 case Instruction::SRem:
7064 case Instruction::URem:
7065 // If we are just changing the sign, rewrite.
7066 if (DestBitSize == SrcBitSize) {
7067 // Don't insert two casts if they cannot be eliminated. We allow
7068 // two casts to be inserted if the sizes are the same. This could
7069 // only be converting signedness, which is a noop.
7070 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7071 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7072 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7074 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7076 return BinaryOperator::create(
7077 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7082 case Instruction::Shl:
7083 // Allow changing the sign of the source operand. Do not allow
7084 // changing the size of the shift, UNLESS the shift amount is a
7085 // constant. We must not change variable sized shifts to a smaller
7086 // size, because it is undefined to shift more bits out than exist
7088 if (DestBitSize == SrcBitSize ||
7089 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7090 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7091 Instruction::BitCast : Instruction::Trunc);
7092 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7093 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7094 return BinaryOperator::createShl(Op0c, Op1c);
7097 case Instruction::AShr:
7098 // If this is a signed shr, and if all bits shifted in are about to be
7099 // truncated off, turn it into an unsigned shr to allow greater
7101 if (DestBitSize < SrcBitSize &&
7102 isa<ConstantInt>(Op1)) {
7103 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
7104 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7105 // Insert the new logical shift right.
7106 return BinaryOperator::createLShr(Op0, Op1);
7111 case Instruction::ICmp:
7112 // If we are just checking for a icmp eq of a single bit and casting it
7113 // to an integer, then shift the bit to the appropriate place and then
7114 // cast to integer to avoid the comparison.
7115 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
7116 uint64_t Op1CV = Op1C->getZExtValue();
7117 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
7118 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
7119 // cast (X == 1) to int --> X iff X has only the low bit set.
7120 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
7121 // cast (X != 0) to int --> X iff X has only the low bit set.
7122 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
7123 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
7124 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
7125 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
7126 // If Op1C some other power of two, convert:
7127 uint64_t KnownZero, KnownOne;
7128 uint64_t TypeMask = Op1C->getType()->getBitMask();
7129 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
7131 // This only works for EQ and NE
7132 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
7133 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
7136 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
7137 bool isNE = pred == ICmpInst::ICMP_NE;
7138 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
7139 // (X&4) == 2 --> false
7140 // (X&4) != 2 --> true
7141 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7142 Res = ConstantExpr::getZExt(Res, CI.getType());
7143 return ReplaceInstUsesWith(CI, Res);
7146 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
7149 // Perform a logical shr by shiftamt.
7150 // Insert the shift to put the result in the low bit.
7151 In = InsertNewInstBefore(
7152 BinaryOperator::createLShr(In,
7153 ConstantInt::get(In->getType(), ShiftAmt),
7154 In->getName()+".lobit"), CI);
7157 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7158 Constant *One = ConstantInt::get(In->getType(), 1);
7159 In = BinaryOperator::createXor(In, One, "tmp");
7160 InsertNewInstBefore(cast<Instruction>(In), CI);
7163 if (CI.getType() == In->getType())
7164 return ReplaceInstUsesWith(CI, In);
7166 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7175 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
7176 if (Instruction *Result = commonIntCastTransforms(CI))
7179 Value *Src = CI.getOperand(0);
7180 const Type *Ty = CI.getType();
7181 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
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 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
7194 Value* SrcIOp0 = SrcI->getOperand(0);
7195 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7196 if (ShAmt >= DestBitWidth) // All zeros.
7197 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7199 // Okay, we can shrink this. Truncate the input, then return a new
7201 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7202 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7204 return BinaryOperator::createLShr(V1, V2);
7206 } else { // This is a variable shr.
7208 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7209 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7210 // loop-invariant and CSE'd.
7211 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7212 Value *One = ConstantInt::get(SrcI->getType(), 1);
7214 Value *V = InsertNewInstBefore(
7215 BinaryOperator::createShl(One, SrcI->getOperand(1),
7217 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7218 SrcI->getOperand(0),
7220 Value *Zero = Constant::getNullValue(V->getType());
7221 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7231 Instruction *InstCombiner::visitZExt(CastInst &CI) {
7232 // If one of the common conversion will work ..
7233 if (Instruction *Result = commonIntCastTransforms(CI))
7236 Value *Src = CI.getOperand(0);
7238 // If this is a cast of a cast
7239 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7240 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7241 // types and if the sizes are just right we can convert this into a logical
7242 // 'and' which will be much cheaper than the pair of casts.
7243 if (isa<TruncInst>(CSrc)) {
7244 // Get the sizes of the types involved
7245 Value *A = CSrc->getOperand(0);
7246 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
7247 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7248 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7249 // If we're actually extending zero bits and the trunc is a no-op
7250 if (MidSize < DstSize && SrcSize == DstSize) {
7251 // Replace both of the casts with an And of the type mask.
7252 uint64_t AndValue = cast<IntegerType>(CSrc->getType())->getBitMask();
7253 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
7255 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7256 // Unfortunately, if the type changed, we need to cast it back.
7257 if (And->getType() != CI.getType()) {
7258 And->setName(CSrc->getName()+".mask");
7259 InsertNewInstBefore(And, CI);
7260 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7270 Instruction *InstCombiner::visitSExt(CastInst &CI) {
7271 return commonIntCastTransforms(CI);
7274 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
7275 return commonCastTransforms(CI);
7278 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7279 return commonCastTransforms(CI);
7282 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7283 return commonCastTransforms(CI);
7286 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7287 return commonCastTransforms(CI);
7290 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7291 return commonCastTransforms(CI);
7294 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7295 return commonCastTransforms(CI);
7298 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7299 return commonCastTransforms(CI);
7302 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
7303 return commonCastTransforms(CI);
7306 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
7308 // If the operands are integer typed then apply the integer transforms,
7309 // otherwise just apply the common ones.
7310 Value *Src = CI.getOperand(0);
7311 const Type *SrcTy = Src->getType();
7312 const Type *DestTy = CI.getType();
7314 if (SrcTy->isInteger() && DestTy->isInteger()) {
7315 if (Instruction *Result = commonIntCastTransforms(CI))
7318 if (Instruction *Result = commonCastTransforms(CI))
7323 // Get rid of casts from one type to the same type. These are useless and can
7324 // be replaced by the operand.
7325 if (DestTy == Src->getType())
7326 return ReplaceInstUsesWith(CI, Src);
7328 // If the source and destination are pointers, and this cast is equivalent to
7329 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
7330 // This can enhance SROA and other transforms that want type-safe pointers.
7331 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7332 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
7333 const Type *DstElTy = DstPTy->getElementType();
7334 const Type *SrcElTy = SrcPTy->getElementType();
7336 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7337 unsigned NumZeros = 0;
7338 while (SrcElTy != DstElTy &&
7339 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7340 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7341 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7345 // If we found a path from the src to dest, create the getelementptr now.
7346 if (SrcElTy == DstElTy) {
7347 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7348 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
7353 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7354 if (SVI->hasOneUse()) {
7355 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7356 // a bitconvert to a vector with the same # elts.
7357 if (isa<VectorType>(DestTy) &&
7358 cast<VectorType>(DestTy)->getNumElements() ==
7359 SVI->getType()->getNumElements()) {
7361 // If either of the operands is a cast from CI.getType(), then
7362 // evaluating the shuffle in the casted destination's type will allow
7363 // us to eliminate at least one cast.
7364 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7365 Tmp->getOperand(0)->getType() == DestTy) ||
7366 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7367 Tmp->getOperand(0)->getType() == DestTy)) {
7368 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7369 SVI->getOperand(0), DestTy, &CI);
7370 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7371 SVI->getOperand(1), DestTy, &CI);
7372 // Return a new shuffle vector. Use the same element ID's, as we
7373 // know the vector types match #elts.
7374 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7382 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7384 /// %D = select %cond, %C, %A
7386 /// %C = select %cond, %B, 0
7389 /// Assuming that the specified instruction is an operand to the select, return
7390 /// a bitmask indicating which operands of this instruction are foldable if they
7391 /// equal the other incoming value of the select.
7393 static unsigned GetSelectFoldableOperands(Instruction *I) {
7394 switch (I->getOpcode()) {
7395 case Instruction::Add:
7396 case Instruction::Mul:
7397 case Instruction::And:
7398 case Instruction::Or:
7399 case Instruction::Xor:
7400 return 3; // Can fold through either operand.
7401 case Instruction::Sub: // Can only fold on the amount subtracted.
7402 case Instruction::Shl: // Can only fold on the shift amount.
7403 case Instruction::LShr:
7404 case Instruction::AShr:
7407 return 0; // Cannot fold
7411 /// GetSelectFoldableConstant - For the same transformation as the previous
7412 /// function, return the identity constant that goes into the select.
7413 static Constant *GetSelectFoldableConstant(Instruction *I) {
7414 switch (I->getOpcode()) {
7415 default: assert(0 && "This cannot happen!"); abort();
7416 case Instruction::Add:
7417 case Instruction::Sub:
7418 case Instruction::Or:
7419 case Instruction::Xor:
7420 case Instruction::Shl:
7421 case Instruction::LShr:
7422 case Instruction::AShr:
7423 return Constant::getNullValue(I->getType());
7424 case Instruction::And:
7425 return ConstantInt::getAllOnesValue(I->getType());
7426 case Instruction::Mul:
7427 return ConstantInt::get(I->getType(), 1);
7431 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7432 /// have the same opcode and only one use each. Try to simplify this.
7433 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7435 if (TI->getNumOperands() == 1) {
7436 // If this is a non-volatile load or a cast from the same type,
7439 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7442 return 0; // unknown unary op.
7445 // Fold this by inserting a select from the input values.
7446 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7447 FI->getOperand(0), SI.getName()+".v");
7448 InsertNewInstBefore(NewSI, SI);
7449 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7453 // Only handle binary operators here.
7454 if (!isa<BinaryOperator>(TI))
7457 // Figure out if the operations have any operands in common.
7458 Value *MatchOp, *OtherOpT, *OtherOpF;
7460 if (TI->getOperand(0) == FI->getOperand(0)) {
7461 MatchOp = TI->getOperand(0);
7462 OtherOpT = TI->getOperand(1);
7463 OtherOpF = FI->getOperand(1);
7464 MatchIsOpZero = true;
7465 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7466 MatchOp = TI->getOperand(1);
7467 OtherOpT = TI->getOperand(0);
7468 OtherOpF = FI->getOperand(0);
7469 MatchIsOpZero = false;
7470 } else if (!TI->isCommutative()) {
7472 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7473 MatchOp = TI->getOperand(0);
7474 OtherOpT = TI->getOperand(1);
7475 OtherOpF = FI->getOperand(0);
7476 MatchIsOpZero = true;
7477 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7478 MatchOp = TI->getOperand(1);
7479 OtherOpT = TI->getOperand(0);
7480 OtherOpF = FI->getOperand(1);
7481 MatchIsOpZero = true;
7486 // If we reach here, they do have operations in common.
7487 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7488 OtherOpF, SI.getName()+".v");
7489 InsertNewInstBefore(NewSI, SI);
7491 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7493 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7495 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7497 assert(0 && "Shouldn't get here");
7501 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7502 Value *CondVal = SI.getCondition();
7503 Value *TrueVal = SI.getTrueValue();
7504 Value *FalseVal = SI.getFalseValue();
7506 // select true, X, Y -> X
7507 // select false, X, Y -> Y
7508 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7509 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7511 // select C, X, X -> X
7512 if (TrueVal == FalseVal)
7513 return ReplaceInstUsesWith(SI, TrueVal);
7515 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7516 return ReplaceInstUsesWith(SI, FalseVal);
7517 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7518 return ReplaceInstUsesWith(SI, TrueVal);
7519 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7520 if (isa<Constant>(TrueVal))
7521 return ReplaceInstUsesWith(SI, TrueVal);
7523 return ReplaceInstUsesWith(SI, FalseVal);
7526 if (SI.getType() == Type::Int1Ty) {
7527 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7528 if (C->getZExtValue()) {
7529 // Change: A = select B, true, C --> A = or B, C
7530 return BinaryOperator::createOr(CondVal, FalseVal);
7532 // Change: A = select B, false, C --> A = and !B, C
7534 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7535 "not."+CondVal->getName()), SI);
7536 return BinaryOperator::createAnd(NotCond, FalseVal);
7538 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7539 if (C->getZExtValue() == false) {
7540 // Change: A = select B, C, false --> A = and B, C
7541 return BinaryOperator::createAnd(CondVal, TrueVal);
7543 // Change: A = select B, C, true --> A = or !B, C
7545 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7546 "not."+CondVal->getName()), SI);
7547 return BinaryOperator::createOr(NotCond, TrueVal);
7552 // Selecting between two integer constants?
7553 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7554 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7555 // select C, 1, 0 -> cast C to int
7556 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
7557 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7558 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
7559 // select C, 0, 1 -> cast !C to int
7561 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7562 "not."+CondVal->getName()), SI);
7563 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7566 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7568 // (x <s 0) ? -1 : 0 -> ashr x, 31
7569 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
7570 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
7571 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7572 bool CanXForm = false;
7573 if (IC->isSignedPredicate())
7574 CanXForm = CmpCst->isNullValue() &&
7575 IC->getPredicate() == ICmpInst::ICMP_SLT;
7577 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
7578 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
7579 IC->getPredicate() == ICmpInst::ICMP_UGT;
7583 // The comparison constant and the result are not neccessarily the
7584 // same width. Make an all-ones value by inserting a AShr.
7585 Value *X = IC->getOperand(0);
7586 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
7587 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7588 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7590 InsertNewInstBefore(SRA, SI);
7592 // Finally, convert to the type of the select RHS. We figure out
7593 // if this requires a SExt, Trunc or BitCast based on the sizes.
7594 Instruction::CastOps opc = Instruction::BitCast;
7595 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
7596 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
7597 if (SRASize < SISize)
7598 opc = Instruction::SExt;
7599 else if (SRASize > SISize)
7600 opc = Instruction::Trunc;
7601 return CastInst::create(opc, SRA, SI.getType());
7606 // If one of the constants is zero (we know they can't both be) and we
7607 // have a fcmp instruction with zero, and we have an 'and' with the
7608 // non-constant value, eliminate this whole mess. This corresponds to
7609 // cases like this: ((X & 27) ? 27 : 0)
7610 if (TrueValC->isNullValue() || FalseValC->isNullValue())
7611 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7612 cast<Constant>(IC->getOperand(1))->isNullValue())
7613 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7614 if (ICA->getOpcode() == Instruction::And &&
7615 isa<ConstantInt>(ICA->getOperand(1)) &&
7616 (ICA->getOperand(1) == TrueValC ||
7617 ICA->getOperand(1) == FalseValC) &&
7618 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7619 // Okay, now we know that everything is set up, we just don't
7620 // know whether we have a icmp_ne or icmp_eq and whether the
7621 // true or false val is the zero.
7622 bool ShouldNotVal = !TrueValC->isNullValue();
7623 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7626 V = InsertNewInstBefore(BinaryOperator::create(
7627 Instruction::Xor, V, ICA->getOperand(1)), SI);
7628 return ReplaceInstUsesWith(SI, V);
7633 // See if we are selecting two values based on a comparison of the two values.
7634 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7635 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7636 // Transform (X == Y) ? X : Y -> Y
7637 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7638 return ReplaceInstUsesWith(SI, FalseVal);
7639 // Transform (X != Y) ? X : Y -> X
7640 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7641 return ReplaceInstUsesWith(SI, TrueVal);
7642 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7644 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7645 // Transform (X == Y) ? Y : X -> X
7646 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7647 return ReplaceInstUsesWith(SI, FalseVal);
7648 // Transform (X != Y) ? Y : X -> Y
7649 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7650 return ReplaceInstUsesWith(SI, TrueVal);
7651 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7655 // See if we are selecting two values based on a comparison of the two values.
7656 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7657 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7658 // Transform (X == Y) ? X : Y -> Y
7659 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7660 return ReplaceInstUsesWith(SI, FalseVal);
7661 // Transform (X != Y) ? X : Y -> X
7662 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7663 return ReplaceInstUsesWith(SI, TrueVal);
7664 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7666 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7667 // Transform (X == Y) ? Y : X -> X
7668 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7669 return ReplaceInstUsesWith(SI, FalseVal);
7670 // Transform (X != Y) ? Y : X -> Y
7671 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7672 return ReplaceInstUsesWith(SI, TrueVal);
7673 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7677 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7678 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7679 if (TI->hasOneUse() && FI->hasOneUse()) {
7680 Instruction *AddOp = 0, *SubOp = 0;
7682 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7683 if (TI->getOpcode() == FI->getOpcode())
7684 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7687 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7688 // even legal for FP.
7689 if (TI->getOpcode() == Instruction::Sub &&
7690 FI->getOpcode() == Instruction::Add) {
7691 AddOp = FI; SubOp = TI;
7692 } else if (FI->getOpcode() == Instruction::Sub &&
7693 TI->getOpcode() == Instruction::Add) {
7694 AddOp = TI; SubOp = FI;
7698 Value *OtherAddOp = 0;
7699 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7700 OtherAddOp = AddOp->getOperand(1);
7701 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7702 OtherAddOp = AddOp->getOperand(0);
7706 // So at this point we know we have (Y -> OtherAddOp):
7707 // select C, (add X, Y), (sub X, Z)
7708 Value *NegVal; // Compute -Z
7709 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7710 NegVal = ConstantExpr::getNeg(C);
7712 NegVal = InsertNewInstBefore(
7713 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7716 Value *NewTrueOp = OtherAddOp;
7717 Value *NewFalseOp = NegVal;
7719 std::swap(NewTrueOp, NewFalseOp);
7720 Instruction *NewSel =
7721 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7723 NewSel = InsertNewInstBefore(NewSel, SI);
7724 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7729 // See if we can fold the select into one of our operands.
7730 if (SI.getType()->isInteger()) {
7731 // See the comment above GetSelectFoldableOperands for a description of the
7732 // transformation we are doing here.
7733 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7734 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7735 !isa<Constant>(FalseVal))
7736 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7737 unsigned OpToFold = 0;
7738 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7740 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7745 Constant *C = GetSelectFoldableConstant(TVI);
7746 Instruction *NewSel =
7747 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7748 InsertNewInstBefore(NewSel, SI);
7749 NewSel->takeName(TVI);
7750 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7751 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7753 assert(0 && "Unknown instruction!!");
7758 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7759 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7760 !isa<Constant>(TrueVal))
7761 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7762 unsigned OpToFold = 0;
7763 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7765 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7770 Constant *C = GetSelectFoldableConstant(FVI);
7771 Instruction *NewSel =
7772 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7773 InsertNewInstBefore(NewSel, SI);
7774 NewSel->takeName(FVI);
7775 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7776 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7778 assert(0 && "Unknown instruction!!");
7783 if (BinaryOperator::isNot(CondVal)) {
7784 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7785 SI.setOperand(1, FalseVal);
7786 SI.setOperand(2, TrueVal);
7793 /// GetKnownAlignment - If the specified pointer has an alignment that we can
7794 /// determine, return it, otherwise return 0.
7795 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
7796 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7797 unsigned Align = GV->getAlignment();
7798 if (Align == 0 && TD)
7799 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7801 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7802 unsigned Align = AI->getAlignment();
7803 if (Align == 0 && TD) {
7804 if (isa<AllocaInst>(AI))
7805 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7806 else if (isa<MallocInst>(AI)) {
7807 // Malloc returns maximally aligned memory.
7808 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7811 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7814 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7818 } else if (isa<BitCastInst>(V) ||
7819 (isa<ConstantExpr>(V) &&
7820 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7821 User *CI = cast<User>(V);
7822 if (isa<PointerType>(CI->getOperand(0)->getType()))
7823 return GetKnownAlignment(CI->getOperand(0), TD);
7825 } else if (isa<GetElementPtrInst>(V) ||
7826 (isa<ConstantExpr>(V) &&
7827 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
7828 User *GEPI = cast<User>(V);
7829 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
7830 if (BaseAlignment == 0) return 0;
7832 // If all indexes are zero, it is just the alignment of the base pointer.
7833 bool AllZeroOperands = true;
7834 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7835 if (!isa<Constant>(GEPI->getOperand(i)) ||
7836 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7837 AllZeroOperands = false;
7840 if (AllZeroOperands)
7841 return BaseAlignment;
7843 // Otherwise, if the base alignment is >= the alignment we expect for the
7844 // base pointer type, then we know that the resultant pointer is aligned at
7845 // least as much as its type requires.
7848 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7849 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7850 if (TD->getABITypeAlignment(PtrTy->getElementType())
7852 const Type *GEPTy = GEPI->getType();
7853 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7854 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
7862 /// visitCallInst - CallInst simplification. This mostly only handles folding
7863 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7864 /// the heavy lifting.
7866 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7867 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7868 if (!II) return visitCallSite(&CI);
7870 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7872 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7873 bool Changed = false;
7875 // memmove/cpy/set of zero bytes is a noop.
7876 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7877 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7879 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7880 if (CI->getZExtValue() == 1) {
7881 // Replace the instruction with just byte operations. We would
7882 // transform other cases to loads/stores, but we don't know if
7883 // alignment is sufficient.
7887 // If we have a memmove and the source operation is a constant global,
7888 // then the source and dest pointers can't alias, so we can change this
7889 // into a call to memcpy.
7890 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7891 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7892 if (GVSrc->isConstant()) {
7893 Module *M = CI.getParent()->getParent()->getParent();
7895 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7897 Name = "llvm.memcpy.i32";
7899 Name = "llvm.memcpy.i64";
7900 Constant *MemCpy = M->getOrInsertFunction(Name,
7901 CI.getCalledFunction()->getFunctionType());
7902 CI.setOperand(0, MemCpy);
7907 // If we can determine a pointer alignment that is bigger than currently
7908 // set, update the alignment.
7909 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7910 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7911 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7912 unsigned Align = std::min(Alignment1, Alignment2);
7913 if (MI->getAlignment()->getZExtValue() < Align) {
7914 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7917 } else if (isa<MemSetInst>(MI)) {
7918 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7919 if (MI->getAlignment()->getZExtValue() < Alignment) {
7920 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7925 if (Changed) return II;
7927 switch (II->getIntrinsicID()) {
7929 case Intrinsic::ppc_altivec_lvx:
7930 case Intrinsic::ppc_altivec_lvxl:
7931 case Intrinsic::x86_sse_loadu_ps:
7932 case Intrinsic::x86_sse2_loadu_pd:
7933 case Intrinsic::x86_sse2_loadu_dq:
7934 // Turn PPC lvx -> load if the pointer is known aligned.
7935 // Turn X86 loadups -> load if the pointer is known aligned.
7936 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7937 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7938 PointerType::get(II->getType()), CI);
7939 return new LoadInst(Ptr);
7942 case Intrinsic::ppc_altivec_stvx:
7943 case Intrinsic::ppc_altivec_stvxl:
7944 // Turn stvx -> store if the pointer is known aligned.
7945 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7946 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7947 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7949 return new StoreInst(II->getOperand(1), Ptr);
7952 case Intrinsic::x86_sse_storeu_ps:
7953 case Intrinsic::x86_sse2_storeu_pd:
7954 case Intrinsic::x86_sse2_storeu_dq:
7955 case Intrinsic::x86_sse2_storel_dq:
7956 // Turn X86 storeu -> store if the pointer is known aligned.
7957 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7958 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7959 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7961 return new StoreInst(II->getOperand(2), Ptr);
7965 case Intrinsic::x86_sse_cvttss2si: {
7966 // These intrinsics only demands the 0th element of its input vector. If
7967 // we can simplify the input based on that, do so now.
7969 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7971 II->setOperand(1, V);
7977 case Intrinsic::ppc_altivec_vperm:
7978 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7979 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7980 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7982 // Check that all of the elements are integer constants or undefs.
7983 bool AllEltsOk = true;
7984 for (unsigned i = 0; i != 16; ++i) {
7985 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7986 !isa<UndefValue>(Mask->getOperand(i))) {
7993 // Cast the input vectors to byte vectors.
7994 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7995 II->getOperand(1), Mask->getType(), CI);
7996 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7997 II->getOperand(2), Mask->getType(), CI);
7998 Value *Result = UndefValue::get(Op0->getType());
8000 // Only extract each element once.
8001 Value *ExtractedElts[32];
8002 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8004 for (unsigned i = 0; i != 16; ++i) {
8005 if (isa<UndefValue>(Mask->getOperand(i)))
8007 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8008 Idx &= 31; // Match the hardware behavior.
8010 if (ExtractedElts[Idx] == 0) {
8012 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8013 InsertNewInstBefore(Elt, CI);
8014 ExtractedElts[Idx] = Elt;
8017 // Insert this value into the result vector.
8018 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
8019 InsertNewInstBefore(cast<Instruction>(Result), CI);
8021 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8026 case Intrinsic::stackrestore: {
8027 // If the save is right next to the restore, remove the restore. This can
8028 // happen when variable allocas are DCE'd.
8029 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8030 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8031 BasicBlock::iterator BI = SS;
8033 return EraseInstFromFunction(CI);
8037 // If the stack restore is in a return/unwind block and if there are no
8038 // allocas or calls between the restore and the return, nuke the restore.
8039 TerminatorInst *TI = II->getParent()->getTerminator();
8040 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
8041 BasicBlock::iterator BI = II;
8042 bool CannotRemove = false;
8043 for (++BI; &*BI != TI; ++BI) {
8044 if (isa<AllocaInst>(BI) ||
8045 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
8046 CannotRemove = true;
8051 return EraseInstFromFunction(CI);
8058 return visitCallSite(II);
8061 // InvokeInst simplification
8063 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8064 return visitCallSite(&II);
8067 // visitCallSite - Improvements for call and invoke instructions.
8069 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8070 bool Changed = false;
8072 // If the callee is a constexpr cast of a function, attempt to move the cast
8073 // to the arguments of the call/invoke.
8074 if (transformConstExprCastCall(CS)) return 0;
8076 Value *Callee = CS.getCalledValue();
8078 if (Function *CalleeF = dyn_cast<Function>(Callee))
8079 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8080 Instruction *OldCall = CS.getInstruction();
8081 // If the call and callee calling conventions don't match, this call must
8082 // be unreachable, as the call is undefined.
8083 new StoreInst(ConstantInt::getTrue(),
8084 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
8085 if (!OldCall->use_empty())
8086 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8087 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8088 return EraseInstFromFunction(*OldCall);
8092 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8093 // This instruction is not reachable, just remove it. We insert a store to
8094 // undef so that we know that this code is not reachable, despite the fact
8095 // that we can't modify the CFG here.
8096 new StoreInst(ConstantInt::getTrue(),
8097 UndefValue::get(PointerType::get(Type::Int1Ty)),
8098 CS.getInstruction());
8100 if (!CS.getInstruction()->use_empty())
8101 CS.getInstruction()->
8102 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8104 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8105 // Don't break the CFG, insert a dummy cond branch.
8106 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8107 ConstantInt::getTrue(), II);
8109 return EraseInstFromFunction(*CS.getInstruction());
8112 const PointerType *PTy = cast<PointerType>(Callee->getType());
8113 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8114 if (FTy->isVarArg()) {
8115 // See if we can optimize any arguments passed through the varargs area of
8117 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8118 E = CS.arg_end(); I != E; ++I)
8119 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8120 // If this cast does not effect the value passed through the varargs
8121 // area, we can eliminate the use of the cast.
8122 Value *Op = CI->getOperand(0);
8123 if (CI->isLosslessCast()) {
8130 return Changed ? CS.getInstruction() : 0;
8133 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8134 // attempt to move the cast to the arguments of the call/invoke.
8136 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8137 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8138 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8139 if (CE->getOpcode() != Instruction::BitCast ||
8140 !isa<Function>(CE->getOperand(0)))
8142 Function *Callee = cast<Function>(CE->getOperand(0));
8143 Instruction *Caller = CS.getInstruction();
8145 // Okay, this is a cast from a function to a different type. Unless doing so
8146 // would cause a type conversion of one of our arguments, change this call to
8147 // be a direct call with arguments casted to the appropriate types.
8149 const FunctionType *FT = Callee->getFunctionType();
8150 const Type *OldRetTy = Caller->getType();
8152 // Check to see if we are changing the return type...
8153 if (OldRetTy != FT->getReturnType()) {
8154 if (Callee->isDeclaration() && !Caller->use_empty() &&
8155 // Conversion is ok if changing from pointer to int of same size.
8156 !(isa<PointerType>(FT->getReturnType()) &&
8157 TD->getIntPtrType() == OldRetTy))
8158 return false; // Cannot transform this return value.
8160 // If the callsite is an invoke instruction, and the return value is used by
8161 // a PHI node in a successor, we cannot change the return type of the call
8162 // because there is no place to put the cast instruction (without breaking
8163 // the critical edge). Bail out in this case.
8164 if (!Caller->use_empty())
8165 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8166 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8168 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8169 if (PN->getParent() == II->getNormalDest() ||
8170 PN->getParent() == II->getUnwindDest())
8174 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8175 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8177 CallSite::arg_iterator AI = CS.arg_begin();
8178 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8179 const Type *ParamTy = FT->getParamType(i);
8180 const Type *ActTy = (*AI)->getType();
8181 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8182 //Either we can cast directly, or we can upconvert the argument
8183 bool isConvertible = ActTy == ParamTy ||
8184 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8185 (ParamTy->isInteger() && ActTy->isInteger() &&
8186 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8187 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8188 && c->getSExtValue() > 0);
8189 if (Callee->isDeclaration() && !isConvertible) return false;
8192 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8193 Callee->isDeclaration())
8194 return false; // Do not delete arguments unless we have a function body...
8196 // Okay, we decided that this is a safe thing to do: go ahead and start
8197 // inserting cast instructions as necessary...
8198 std::vector<Value*> Args;
8199 Args.reserve(NumActualArgs);
8201 AI = CS.arg_begin();
8202 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8203 const Type *ParamTy = FT->getParamType(i);
8204 if ((*AI)->getType() == ParamTy) {
8205 Args.push_back(*AI);
8207 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8208 false, ParamTy, false);
8209 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8210 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8214 // If the function takes more arguments than the call was taking, add them
8216 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8217 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8219 // If we are removing arguments to the function, emit an obnoxious warning...
8220 if (FT->getNumParams() < NumActualArgs)
8221 if (!FT->isVarArg()) {
8222 cerr << "WARNING: While resolving call to function '"
8223 << Callee->getName() << "' arguments were dropped!\n";
8225 // Add all of the arguments in their promoted form to the arg list...
8226 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8227 const Type *PTy = getPromotedType((*AI)->getType());
8228 if (PTy != (*AI)->getType()) {
8229 // Must promote to pass through va_arg area!
8230 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8232 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8233 InsertNewInstBefore(Cast, *Caller);
8234 Args.push_back(Cast);
8236 Args.push_back(*AI);
8241 if (FT->getReturnType() == Type::VoidTy)
8242 Caller->setName(""); // Void type should not have a name.
8245 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8246 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8247 &Args[0], Args.size(), Caller->getName(), Caller);
8248 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
8250 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
8251 if (cast<CallInst>(Caller)->isTailCall())
8252 cast<CallInst>(NC)->setTailCall();
8253 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8256 // Insert a cast of the return type as necessary.
8258 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
8259 if (NV->getType() != Type::VoidTy) {
8260 const Type *CallerTy = Caller->getType();
8261 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8263 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
8265 // If this is an invoke instruction, we should insert it after the first
8266 // non-phi, instruction in the normal successor block.
8267 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8268 BasicBlock::iterator I = II->getNormalDest()->begin();
8269 while (isa<PHINode>(I)) ++I;
8270 InsertNewInstBefore(NC, *I);
8272 // Otherwise, it's a call, just insert cast right after the call instr
8273 InsertNewInstBefore(NC, *Caller);
8275 AddUsersToWorkList(*Caller);
8277 NV = UndefValue::get(Caller->getType());
8281 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8282 Caller->replaceAllUsesWith(NV);
8283 Caller->eraseFromParent();
8284 RemoveFromWorkList(Caller);
8288 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8289 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8290 /// and a single binop.
8291 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8292 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8293 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8294 isa<CmpInst>(FirstInst));
8295 unsigned Opc = FirstInst->getOpcode();
8296 Value *LHSVal = FirstInst->getOperand(0);
8297 Value *RHSVal = FirstInst->getOperand(1);
8299 const Type *LHSType = LHSVal->getType();
8300 const Type *RHSType = RHSVal->getType();
8302 // Scan to see if all operands are the same opcode, all have one use, and all
8303 // kill their operands (i.e. the operands have one use).
8304 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8305 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8306 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8307 // Verify type of the LHS matches so we don't fold cmp's of different
8308 // types or GEP's with different index types.
8309 I->getOperand(0)->getType() != LHSType ||
8310 I->getOperand(1)->getType() != RHSType)
8313 // If they are CmpInst instructions, check their predicates
8314 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8315 if (cast<CmpInst>(I)->getPredicate() !=
8316 cast<CmpInst>(FirstInst)->getPredicate())
8319 // Keep track of which operand needs a phi node.
8320 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8321 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8324 // Otherwise, this is safe to transform, determine if it is profitable.
8326 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8327 // Indexes are often folded into load/store instructions, so we don't want to
8328 // hide them behind a phi.
8329 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8332 Value *InLHS = FirstInst->getOperand(0);
8333 Value *InRHS = FirstInst->getOperand(1);
8334 PHINode *NewLHS = 0, *NewRHS = 0;
8336 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8337 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8338 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8339 InsertNewInstBefore(NewLHS, PN);
8344 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8345 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8346 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8347 InsertNewInstBefore(NewRHS, PN);
8351 // Add all operands to the new PHIs.
8352 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8354 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8355 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8358 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8359 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8363 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8364 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8365 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8366 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8369 assert(isa<GetElementPtrInst>(FirstInst));
8370 return new GetElementPtrInst(LHSVal, RHSVal);
8374 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8375 /// of the block that defines it. This means that it must be obvious the value
8376 /// of the load is not changed from the point of the load to the end of the
8379 /// Finally, it is safe, but not profitable, to sink a load targetting a
8380 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8382 static bool isSafeToSinkLoad(LoadInst *L) {
8383 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8385 for (++BBI; BBI != E; ++BBI)
8386 if (BBI->mayWriteToMemory())
8389 // Check for non-address taken alloca. If not address-taken already, it isn't
8390 // profitable to do this xform.
8391 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8392 bool isAddressTaken = false;
8393 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8395 if (isa<LoadInst>(UI)) continue;
8396 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8397 // If storing TO the alloca, then the address isn't taken.
8398 if (SI->getOperand(1) == AI) continue;
8400 isAddressTaken = true;
8404 if (!isAddressTaken)
8412 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8413 // operator and they all are only used by the PHI, PHI together their
8414 // inputs, and do the operation once, to the result of the PHI.
8415 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8416 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8418 // Scan the instruction, looking for input operations that can be folded away.
8419 // If all input operands to the phi are the same instruction (e.g. a cast from
8420 // the same type or "+42") we can pull the operation through the PHI, reducing
8421 // code size and simplifying code.
8422 Constant *ConstantOp = 0;
8423 const Type *CastSrcTy = 0;
8424 bool isVolatile = false;
8425 if (isa<CastInst>(FirstInst)) {
8426 CastSrcTy = FirstInst->getOperand(0)->getType();
8427 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8428 // Can fold binop, compare or shift here if the RHS is a constant,
8429 // otherwise call FoldPHIArgBinOpIntoPHI.
8430 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8431 if (ConstantOp == 0)
8432 return FoldPHIArgBinOpIntoPHI(PN);
8433 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8434 isVolatile = LI->isVolatile();
8435 // We can't sink the load if the loaded value could be modified between the
8436 // load and the PHI.
8437 if (LI->getParent() != PN.getIncomingBlock(0) ||
8438 !isSafeToSinkLoad(LI))
8440 } else if (isa<GetElementPtrInst>(FirstInst)) {
8441 if (FirstInst->getNumOperands() == 2)
8442 return FoldPHIArgBinOpIntoPHI(PN);
8443 // Can't handle general GEPs yet.
8446 return 0; // Cannot fold this operation.
8449 // Check to see if all arguments are the same operation.
8450 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8451 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8452 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8453 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8456 if (I->getOperand(0)->getType() != CastSrcTy)
8457 return 0; // Cast operation must match.
8458 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8459 // We can't sink the load if the loaded value could be modified between
8460 // the load and the PHI.
8461 if (LI->isVolatile() != isVolatile ||
8462 LI->getParent() != PN.getIncomingBlock(i) ||
8463 !isSafeToSinkLoad(LI))
8465 } else if (I->getOperand(1) != ConstantOp) {
8470 // Okay, they are all the same operation. Create a new PHI node of the
8471 // correct type, and PHI together all of the LHS's of the instructions.
8472 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8473 PN.getName()+".in");
8474 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8476 Value *InVal = FirstInst->getOperand(0);
8477 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8479 // Add all operands to the new PHI.
8480 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8481 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8482 if (NewInVal != InVal)
8484 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8489 // The new PHI unions all of the same values together. This is really
8490 // common, so we handle it intelligently here for compile-time speed.
8494 InsertNewInstBefore(NewPN, PN);
8498 // Insert and return the new operation.
8499 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8500 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8501 else if (isa<LoadInst>(FirstInst))
8502 return new LoadInst(PhiVal, "", isVolatile);
8503 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8504 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8505 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8506 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8507 PhiVal, ConstantOp);
8509 assert(0 && "Unknown operation");
8513 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8515 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
8516 if (PN->use_empty()) return true;
8517 if (!PN->hasOneUse()) return false;
8519 // Remember this node, and if we find the cycle, return.
8520 if (!PotentiallyDeadPHIs.insert(PN).second)
8523 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8524 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8529 // PHINode simplification
8531 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8532 // If LCSSA is around, don't mess with Phi nodes
8533 if (MustPreserveLCSSA) return 0;
8535 if (Value *V = PN.hasConstantValue())
8536 return ReplaceInstUsesWith(PN, V);
8538 // If all PHI operands are the same operation, pull them through the PHI,
8539 // reducing code size.
8540 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8541 PN.getIncomingValue(0)->hasOneUse())
8542 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8545 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8546 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8547 // PHI)... break the cycle.
8548 if (PN.hasOneUse()) {
8549 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8550 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8551 std::set<PHINode*> PotentiallyDeadPHIs;
8552 PotentiallyDeadPHIs.insert(&PN);
8553 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8554 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8557 // If this phi has a single use, and if that use just computes a value for
8558 // the next iteration of a loop, delete the phi. This occurs with unused
8559 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8560 // common case here is good because the only other things that catch this
8561 // are induction variable analysis (sometimes) and ADCE, which is only run
8563 if (PHIUser->hasOneUse() &&
8564 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8565 PHIUser->use_back() == &PN) {
8566 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8573 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8574 Instruction *InsertPoint,
8576 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8577 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8578 // We must cast correctly to the pointer type. Ensure that we
8579 // sign extend the integer value if it is smaller as this is
8580 // used for address computation.
8581 Instruction::CastOps opcode =
8582 (VTySize < PtrSize ? Instruction::SExt :
8583 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8584 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8588 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8589 Value *PtrOp = GEP.getOperand(0);
8590 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
8591 // If so, eliminate the noop.
8592 if (GEP.getNumOperands() == 1)
8593 return ReplaceInstUsesWith(GEP, PtrOp);
8595 if (isa<UndefValue>(GEP.getOperand(0)))
8596 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8598 bool HasZeroPointerIndex = false;
8599 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8600 HasZeroPointerIndex = C->isNullValue();
8602 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8603 return ReplaceInstUsesWith(GEP, PtrOp);
8605 // Eliminate unneeded casts for indices.
8606 bool MadeChange = false;
8607 gep_type_iterator GTI = gep_type_begin(GEP);
8608 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
8609 if (isa<SequentialType>(*GTI)) {
8610 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8611 if (CI->getOpcode() == Instruction::ZExt ||
8612 CI->getOpcode() == Instruction::SExt) {
8613 const Type *SrcTy = CI->getOperand(0)->getType();
8614 // We can eliminate a cast from i32 to i64 iff the target
8615 // is a 32-bit pointer target.
8616 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8618 GEP.setOperand(i, CI->getOperand(0));
8622 // If we are using a wider index than needed for this platform, shrink it
8623 // to what we need. If the incoming value needs a cast instruction,
8624 // insert it. This explicit cast can make subsequent optimizations more
8626 Value *Op = GEP.getOperand(i);
8627 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
8628 if (Constant *C = dyn_cast<Constant>(Op)) {
8629 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8632 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8634 GEP.setOperand(i, Op);
8638 if (MadeChange) return &GEP;
8640 // Combine Indices - If the source pointer to this getelementptr instruction
8641 // is a getelementptr instruction, combine the indices of the two
8642 // getelementptr instructions into a single instruction.
8644 SmallVector<Value*, 8> SrcGEPOperands;
8645 if (User *Src = dyn_castGetElementPtr(PtrOp))
8646 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8648 if (!SrcGEPOperands.empty()) {
8649 // Note that if our source is a gep chain itself that we wait for that
8650 // chain to be resolved before we perform this transformation. This
8651 // avoids us creating a TON of code in some cases.
8653 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8654 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8655 return 0; // Wait until our source is folded to completion.
8657 SmallVector<Value*, 8> Indices;
8659 // Find out whether the last index in the source GEP is a sequential idx.
8660 bool EndsWithSequential = false;
8661 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8662 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8663 EndsWithSequential = !isa<StructType>(*I);
8665 // Can we combine the two pointer arithmetics offsets?
8666 if (EndsWithSequential) {
8667 // Replace: gep (gep %P, long B), long A, ...
8668 // With: T = long A+B; gep %P, T, ...
8670 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8671 if (SO1 == Constant::getNullValue(SO1->getType())) {
8673 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8676 // If they aren't the same type, convert both to an integer of the
8677 // target's pointer size.
8678 if (SO1->getType() != GO1->getType()) {
8679 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8680 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8681 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8682 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8684 unsigned PS = TD->getPointerSize();
8685 if (TD->getTypeSize(SO1->getType()) == PS) {
8686 // Convert GO1 to SO1's type.
8687 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8689 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8690 // Convert SO1 to GO1's type.
8691 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8693 const Type *PT = TD->getIntPtrType();
8694 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8695 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8699 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8700 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8702 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8703 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8707 // Recycle the GEP we already have if possible.
8708 if (SrcGEPOperands.size() == 2) {
8709 GEP.setOperand(0, SrcGEPOperands[0]);
8710 GEP.setOperand(1, Sum);
8713 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8714 SrcGEPOperands.end()-1);
8715 Indices.push_back(Sum);
8716 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8718 } else if (isa<Constant>(*GEP.idx_begin()) &&
8719 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8720 SrcGEPOperands.size() != 1) {
8721 // Otherwise we can do the fold if the first index of the GEP is a zero
8722 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8723 SrcGEPOperands.end());
8724 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8727 if (!Indices.empty())
8728 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
8729 Indices.size(), GEP.getName());
8731 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8732 // GEP of global variable. If all of the indices for this GEP are
8733 // constants, we can promote this to a constexpr instead of an instruction.
8735 // Scan for nonconstants...
8736 SmallVector<Constant*, 8> Indices;
8737 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8738 for (; I != E && isa<Constant>(*I); ++I)
8739 Indices.push_back(cast<Constant>(*I));
8741 if (I == E) { // If they are all constants...
8742 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8743 &Indices[0],Indices.size());
8745 // Replace all uses of the GEP with the new constexpr...
8746 return ReplaceInstUsesWith(GEP, CE);
8748 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8749 if (!isa<PointerType>(X->getType())) {
8750 // Not interesting. Source pointer must be a cast from pointer.
8751 } else if (HasZeroPointerIndex) {
8752 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8753 // into : GEP [10 x ubyte]* X, long 0, ...
8755 // This occurs when the program declares an array extern like "int X[];"
8757 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8758 const PointerType *XTy = cast<PointerType>(X->getType());
8759 if (const ArrayType *XATy =
8760 dyn_cast<ArrayType>(XTy->getElementType()))
8761 if (const ArrayType *CATy =
8762 dyn_cast<ArrayType>(CPTy->getElementType()))
8763 if (CATy->getElementType() == XATy->getElementType()) {
8764 // At this point, we know that the cast source type is a pointer
8765 // to an array of the same type as the destination pointer
8766 // array. Because the array type is never stepped over (there
8767 // is a leading zero) we can fold the cast into this GEP.
8768 GEP.setOperand(0, X);
8771 } else if (GEP.getNumOperands() == 2) {
8772 // Transform things like:
8773 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8774 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8775 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8776 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8777 if (isa<ArrayType>(SrcElTy) &&
8778 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8779 TD->getTypeSize(ResElTy)) {
8780 Value *V = InsertNewInstBefore(
8781 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8782 GEP.getOperand(1), GEP.getName()), GEP);
8783 // V and GEP are both pointer types --> BitCast
8784 return new BitCastInst(V, GEP.getType());
8787 // Transform things like:
8788 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8789 // (where tmp = 8*tmp2) into:
8790 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8792 if (isa<ArrayType>(SrcElTy) &&
8793 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8794 uint64_t ArrayEltSize =
8795 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8797 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8798 // allow either a mul, shift, or constant here.
8800 ConstantInt *Scale = 0;
8801 if (ArrayEltSize == 1) {
8802 NewIdx = GEP.getOperand(1);
8803 Scale = ConstantInt::get(NewIdx->getType(), 1);
8804 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8805 NewIdx = ConstantInt::get(CI->getType(), 1);
8807 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8808 if (Inst->getOpcode() == Instruction::Shl &&
8809 isa<ConstantInt>(Inst->getOperand(1))) {
8811 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
8812 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
8813 NewIdx = Inst->getOperand(0);
8814 } else if (Inst->getOpcode() == Instruction::Mul &&
8815 isa<ConstantInt>(Inst->getOperand(1))) {
8816 Scale = cast<ConstantInt>(Inst->getOperand(1));
8817 NewIdx = Inst->getOperand(0);
8821 // If the index will be to exactly the right offset with the scale taken
8822 // out, perform the transformation.
8823 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8824 if (isa<ConstantInt>(Scale))
8825 Scale = ConstantInt::get(Scale->getType(),
8826 Scale->getZExtValue() / ArrayEltSize);
8827 if (Scale->getZExtValue() != 1) {
8828 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8830 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8831 NewIdx = InsertNewInstBefore(Sc, GEP);
8834 // Insert the new GEP instruction.
8835 Instruction *NewGEP =
8836 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8837 NewIdx, GEP.getName());
8838 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8839 // The NewGEP must be pointer typed, so must the old one -> BitCast
8840 return new BitCastInst(NewGEP, GEP.getType());
8849 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8850 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8851 if (AI.isArrayAllocation()) // Check C != 1
8852 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8854 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8855 AllocationInst *New = 0;
8857 // Create and insert the replacement instruction...
8858 if (isa<MallocInst>(AI))
8859 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8861 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8862 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8865 InsertNewInstBefore(New, AI);
8867 // Scan to the end of the allocation instructions, to skip over a block of
8868 // allocas if possible...
8870 BasicBlock::iterator It = New;
8871 while (isa<AllocationInst>(*It)) ++It;
8873 // Now that I is pointing to the first non-allocation-inst in the block,
8874 // insert our getelementptr instruction...
8876 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8877 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8878 New->getName()+".sub", It);
8880 // Now make everything use the getelementptr instead of the original
8882 return ReplaceInstUsesWith(AI, V);
8883 } else if (isa<UndefValue>(AI.getArraySize())) {
8884 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8887 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8888 // Note that we only do this for alloca's, because malloc should allocate and
8889 // return a unique pointer, even for a zero byte allocation.
8890 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8891 TD->getTypeSize(AI.getAllocatedType()) == 0)
8892 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8897 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8898 Value *Op = FI.getOperand(0);
8900 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8901 if (CastInst *CI = dyn_cast<CastInst>(Op))
8902 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8903 FI.setOperand(0, CI->getOperand(0));
8907 // free undef -> unreachable.
8908 if (isa<UndefValue>(Op)) {
8909 // Insert a new store to null because we cannot modify the CFG here.
8910 new StoreInst(ConstantInt::getTrue(),
8911 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8912 return EraseInstFromFunction(FI);
8915 // If we have 'free null' delete the instruction. This can happen in stl code
8916 // when lots of inlining happens.
8917 if (isa<ConstantPointerNull>(Op))
8918 return EraseInstFromFunction(FI);
8924 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8925 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8926 User *CI = cast<User>(LI.getOperand(0));
8927 Value *CastOp = CI->getOperand(0);
8929 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8930 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8931 const Type *SrcPTy = SrcTy->getElementType();
8933 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8934 isa<VectorType>(DestPTy)) {
8935 // If the source is an array, the code below will not succeed. Check to
8936 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8938 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8939 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8940 if (ASrcTy->getNumElements() != 0) {
8942 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8943 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8944 SrcTy = cast<PointerType>(CastOp->getType());
8945 SrcPTy = SrcTy->getElementType();
8948 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8949 isa<VectorType>(SrcPTy)) &&
8950 // Do not allow turning this into a load of an integer, which is then
8951 // casted to a pointer, this pessimizes pointer analysis a lot.
8952 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8953 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8954 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8956 // Okay, we are casting from one integer or pointer type to another of
8957 // the same size. Instead of casting the pointer before the load, cast
8958 // the result of the loaded value.
8959 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8961 LI.isVolatile()),LI);
8962 // Now cast the result of the load.
8963 return new BitCastInst(NewLoad, LI.getType());
8970 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8971 /// from this value cannot trap. If it is not obviously safe to load from the
8972 /// specified pointer, we do a quick local scan of the basic block containing
8973 /// ScanFrom, to determine if the address is already accessed.
8974 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8975 // If it is an alloca or global variable, it is always safe to load from.
8976 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8978 // Otherwise, be a little bit agressive by scanning the local block where we
8979 // want to check to see if the pointer is already being loaded or stored
8980 // from/to. If so, the previous load or store would have already trapped,
8981 // so there is no harm doing an extra load (also, CSE will later eliminate
8982 // the load entirely).
8983 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8988 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8989 if (LI->getOperand(0) == V) return true;
8990 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8991 if (SI->getOperand(1) == V) return true;
8997 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8998 Value *Op = LI.getOperand(0);
9000 // load (cast X) --> cast (load X) iff safe
9001 if (isa<CastInst>(Op))
9002 if (Instruction *Res = InstCombineLoadCast(*this, LI))
9005 // None of the following transforms are legal for volatile loads.
9006 if (LI.isVolatile()) return 0;
9008 if (&LI.getParent()->front() != &LI) {
9009 BasicBlock::iterator BBI = &LI; --BBI;
9010 // If the instruction immediately before this is a store to the same
9011 // address, do a simple form of store->load forwarding.
9012 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9013 if (SI->getOperand(1) == LI.getOperand(0))
9014 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9015 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9016 if (LIB->getOperand(0) == LI.getOperand(0))
9017 return ReplaceInstUsesWith(LI, LIB);
9020 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
9021 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
9022 isa<UndefValue>(GEPI->getOperand(0))) {
9023 // Insert a new store to null instruction before the load to indicate
9024 // that this code is not reachable. We do this instead of inserting
9025 // an unreachable instruction directly because we cannot modify the
9027 new StoreInst(UndefValue::get(LI.getType()),
9028 Constant::getNullValue(Op->getType()), &LI);
9029 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9032 if (Constant *C = dyn_cast<Constant>(Op)) {
9033 // load null/undef -> undef
9034 if ((C->isNullValue() || isa<UndefValue>(C))) {
9035 // Insert a new store to null instruction before the load to indicate that
9036 // this code is not reachable. We do this instead of inserting an
9037 // unreachable instruction directly because we cannot modify the CFG.
9038 new StoreInst(UndefValue::get(LI.getType()),
9039 Constant::getNullValue(Op->getType()), &LI);
9040 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9043 // Instcombine load (constant global) into the value loaded.
9044 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9045 if (GV->isConstant() && !GV->isDeclaration())
9046 return ReplaceInstUsesWith(LI, GV->getInitializer());
9048 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9049 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9050 if (CE->getOpcode() == Instruction::GetElementPtr) {
9051 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9052 if (GV->isConstant() && !GV->isDeclaration())
9054 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9055 return ReplaceInstUsesWith(LI, V);
9056 if (CE->getOperand(0)->isNullValue()) {
9057 // Insert a new store to null instruction before the load to indicate
9058 // that this code is not reachable. We do this instead of inserting
9059 // an unreachable instruction directly because we cannot modify the
9061 new StoreInst(UndefValue::get(LI.getType()),
9062 Constant::getNullValue(Op->getType()), &LI);
9063 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9066 } else if (CE->isCast()) {
9067 if (Instruction *Res = InstCombineLoadCast(*this, LI))
9072 if (Op->hasOneUse()) {
9073 // Change select and PHI nodes to select values instead of addresses: this
9074 // helps alias analysis out a lot, allows many others simplifications, and
9075 // exposes redundancy in the code.
9077 // Note that we cannot do the transformation unless we know that the
9078 // introduced loads cannot trap! Something like this is valid as long as
9079 // the condition is always false: load (select bool %C, int* null, int* %G),
9080 // but it would not be valid if we transformed it to load from null
9083 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9084 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9085 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9086 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9087 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9088 SI->getOperand(1)->getName()+".val"), LI);
9089 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9090 SI->getOperand(2)->getName()+".val"), LI);
9091 return new SelectInst(SI->getCondition(), V1, V2);
9094 // load (select (cond, null, P)) -> load P
9095 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9096 if (C->isNullValue()) {
9097 LI.setOperand(0, SI->getOperand(2));
9101 // load (select (cond, P, null)) -> load P
9102 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9103 if (C->isNullValue()) {
9104 LI.setOperand(0, SI->getOperand(1));
9112 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9114 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9115 User *CI = cast<User>(SI.getOperand(1));
9116 Value *CastOp = CI->getOperand(0);
9118 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9119 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9120 const Type *SrcPTy = SrcTy->getElementType();
9122 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9123 // If the source is an array, the code below will not succeed. Check to
9124 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9126 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9127 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9128 if (ASrcTy->getNumElements() != 0) {
9130 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9131 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9132 SrcTy = cast<PointerType>(CastOp->getType());
9133 SrcPTy = SrcTy->getElementType();
9136 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9137 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9138 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9140 // Okay, we are casting from one integer or pointer type to another of
9141 // the same size. Instead of casting the pointer before
9142 // the store, cast the value to be stored.
9144 Value *SIOp0 = SI.getOperand(0);
9145 Instruction::CastOps opcode = Instruction::BitCast;
9146 const Type* CastSrcTy = SIOp0->getType();
9147 const Type* CastDstTy = SrcPTy;
9148 if (isa<PointerType>(CastDstTy)) {
9149 if (CastSrcTy->isInteger())
9150 opcode = Instruction::IntToPtr;
9151 } else if (isa<IntegerType>(CastDstTy)) {
9152 if (isa<PointerType>(SIOp0->getType()))
9153 opcode = Instruction::PtrToInt;
9155 if (Constant *C = dyn_cast<Constant>(SIOp0))
9156 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9158 NewCast = IC.InsertNewInstBefore(
9159 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9161 return new StoreInst(NewCast, CastOp);
9168 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9169 Value *Val = SI.getOperand(0);
9170 Value *Ptr = SI.getOperand(1);
9172 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9173 EraseInstFromFunction(SI);
9178 // If the RHS is an alloca with a single use, zapify the store, making the
9180 if (Ptr->hasOneUse()) {
9181 if (isa<AllocaInst>(Ptr)) {
9182 EraseInstFromFunction(SI);
9187 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9188 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9189 GEP->getOperand(0)->hasOneUse()) {
9190 EraseInstFromFunction(SI);
9196 // Do really simple DSE, to catch cases where there are several consequtive
9197 // stores to the same location, separated by a few arithmetic operations. This
9198 // situation often occurs with bitfield accesses.
9199 BasicBlock::iterator BBI = &SI;
9200 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9204 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9205 // Prev store isn't volatile, and stores to the same location?
9206 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9209 EraseInstFromFunction(*PrevSI);
9215 // If this is a load, we have to stop. However, if the loaded value is from
9216 // the pointer we're loading and is producing the pointer we're storing,
9217 // then *this* store is dead (X = load P; store X -> P).
9218 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9219 if (LI == Val && LI->getOperand(0) == Ptr) {
9220 EraseInstFromFunction(SI);
9224 // Otherwise, this is a load from some other location. Stores before it
9229 // Don't skip over loads or things that can modify memory.
9230 if (BBI->mayWriteToMemory())
9235 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9237 // store X, null -> turns into 'unreachable' in SimplifyCFG
9238 if (isa<ConstantPointerNull>(Ptr)) {
9239 if (!isa<UndefValue>(Val)) {
9240 SI.setOperand(0, UndefValue::get(Val->getType()));
9241 if (Instruction *U = dyn_cast<Instruction>(Val))
9242 AddToWorkList(U); // Dropped a use.
9245 return 0; // Do not modify these!
9248 // store undef, Ptr -> noop
9249 if (isa<UndefValue>(Val)) {
9250 EraseInstFromFunction(SI);
9255 // If the pointer destination is a cast, see if we can fold the cast into the
9257 if (isa<CastInst>(Ptr))
9258 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9260 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9262 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9266 // If this store is the last instruction in the basic block, and if the block
9267 // ends with an unconditional branch, try to move it to the successor block.
9269 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9270 if (BI->isUnconditional()) {
9271 // Check to see if the successor block has exactly two incoming edges. If
9272 // so, see if the other predecessor contains a store to the same location.
9273 // if so, insert a PHI node (if needed) and move the stores down.
9274 BasicBlock *Dest = BI->getSuccessor(0);
9276 pred_iterator PI = pred_begin(Dest);
9277 BasicBlock *Other = 0;
9278 if (*PI != BI->getParent())
9281 if (PI != pred_end(Dest)) {
9282 if (*PI != BI->getParent())
9287 if (++PI != pred_end(Dest))
9290 if (Other) { // If only one other pred...
9291 BBI = Other->getTerminator();
9292 // Make sure this other block ends in an unconditional branch and that
9293 // there is an instruction before the branch.
9294 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
9295 BBI != Other->begin()) {
9297 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
9299 // If this instruction is a store to the same location.
9300 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
9301 // Okay, we know we can perform this transformation. Insert a PHI
9302 // node now if we need it.
9303 Value *MergedVal = OtherStore->getOperand(0);
9304 if (MergedVal != SI.getOperand(0)) {
9305 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9306 PN->reserveOperandSpace(2);
9307 PN->addIncoming(SI.getOperand(0), SI.getParent());
9308 PN->addIncoming(OtherStore->getOperand(0), Other);
9309 MergedVal = InsertNewInstBefore(PN, Dest->front());
9312 // Advance to a place where it is safe to insert the new store and
9314 BBI = Dest->begin();
9315 while (isa<PHINode>(BBI)) ++BBI;
9316 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9317 OtherStore->isVolatile()), *BBI);
9319 // Nuke the old stores.
9320 EraseInstFromFunction(SI);
9321 EraseInstFromFunction(*OtherStore);
9333 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9334 // Change br (not X), label True, label False to: br X, label False, True
9336 BasicBlock *TrueDest;
9337 BasicBlock *FalseDest;
9338 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9339 !isa<Constant>(X)) {
9340 // Swap Destinations and condition...
9342 BI.setSuccessor(0, FalseDest);
9343 BI.setSuccessor(1, TrueDest);
9347 // Cannonicalize fcmp_one -> fcmp_oeq
9348 FCmpInst::Predicate FPred; Value *Y;
9349 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9350 TrueDest, FalseDest)))
9351 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9352 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9353 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9354 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9355 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9356 NewSCC->takeName(I);
9357 // Swap Destinations and condition...
9358 BI.setCondition(NewSCC);
9359 BI.setSuccessor(0, FalseDest);
9360 BI.setSuccessor(1, TrueDest);
9361 RemoveFromWorkList(I);
9362 I->eraseFromParent();
9363 AddToWorkList(NewSCC);
9367 // Cannonicalize icmp_ne -> icmp_eq
9368 ICmpInst::Predicate IPred;
9369 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9370 TrueDest, FalseDest)))
9371 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9372 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9373 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9374 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9375 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9376 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9377 NewSCC->takeName(I);
9378 // Swap Destinations and condition...
9379 BI.setCondition(NewSCC);
9380 BI.setSuccessor(0, FalseDest);
9381 BI.setSuccessor(1, TrueDest);
9382 RemoveFromWorkList(I);
9383 I->eraseFromParent();;
9384 AddToWorkList(NewSCC);
9391 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9392 Value *Cond = SI.getCondition();
9393 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9394 if (I->getOpcode() == Instruction::Add)
9395 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9396 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9397 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9398 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9400 SI.setOperand(0, I->getOperand(0));
9408 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9409 /// is to leave as a vector operation.
9410 static bool CheapToScalarize(Value *V, bool isConstant) {
9411 if (isa<ConstantAggregateZero>(V))
9413 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9414 if (isConstant) return true;
9415 // If all elts are the same, we can extract.
9416 Constant *Op0 = C->getOperand(0);
9417 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9418 if (C->getOperand(i) != Op0)
9422 Instruction *I = dyn_cast<Instruction>(V);
9423 if (!I) return false;
9425 // Insert element gets simplified to the inserted element or is deleted if
9426 // this is constant idx extract element and its a constant idx insertelt.
9427 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9428 isa<ConstantInt>(I->getOperand(2)))
9430 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9432 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9433 if (BO->hasOneUse() &&
9434 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9435 CheapToScalarize(BO->getOperand(1), isConstant)))
9437 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9438 if (CI->hasOneUse() &&
9439 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9440 CheapToScalarize(CI->getOperand(1), isConstant)))
9446 /// Read and decode a shufflevector mask.
9448 /// It turns undef elements into values that are larger than the number of
9449 /// elements in the input.
9450 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9451 unsigned NElts = SVI->getType()->getNumElements();
9452 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9453 return std::vector<unsigned>(NElts, 0);
9454 if (isa<UndefValue>(SVI->getOperand(2)))
9455 return std::vector<unsigned>(NElts, 2*NElts);
9457 std::vector<unsigned> Result;
9458 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9459 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9460 if (isa<UndefValue>(CP->getOperand(i)))
9461 Result.push_back(NElts*2); // undef -> 8
9463 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9467 /// FindScalarElement - Given a vector and an element number, see if the scalar
9468 /// value is already around as a register, for example if it were inserted then
9469 /// extracted from the vector.
9470 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9471 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9472 const VectorType *PTy = cast<VectorType>(V->getType());
9473 unsigned Width = PTy->getNumElements();
9474 if (EltNo >= Width) // Out of range access.
9475 return UndefValue::get(PTy->getElementType());
9477 if (isa<UndefValue>(V))
9478 return UndefValue::get(PTy->getElementType());
9479 else if (isa<ConstantAggregateZero>(V))
9480 return Constant::getNullValue(PTy->getElementType());
9481 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9482 return CP->getOperand(EltNo);
9483 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9484 // If this is an insert to a variable element, we don't know what it is.
9485 if (!isa<ConstantInt>(III->getOperand(2)))
9487 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9489 // If this is an insert to the element we are looking for, return the
9492 return III->getOperand(1);
9494 // Otherwise, the insertelement doesn't modify the value, recurse on its
9496 return FindScalarElement(III->getOperand(0), EltNo);
9497 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9498 unsigned InEl = getShuffleMask(SVI)[EltNo];
9500 return FindScalarElement(SVI->getOperand(0), InEl);
9501 else if (InEl < Width*2)
9502 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9504 return UndefValue::get(PTy->getElementType());
9507 // Otherwise, we don't know.
9511 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9513 // If packed val is undef, replace extract with scalar undef.
9514 if (isa<UndefValue>(EI.getOperand(0)))
9515 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9517 // If packed val is constant 0, replace extract with scalar 0.
9518 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9519 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9521 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9522 // If packed val is constant with uniform operands, replace EI
9523 // with that operand
9524 Constant *op0 = C->getOperand(0);
9525 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9526 if (C->getOperand(i) != op0) {
9531 return ReplaceInstUsesWith(EI, op0);
9534 // If extracting a specified index from the vector, see if we can recursively
9535 // find a previously computed scalar that was inserted into the vector.
9536 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9537 // This instruction only demands the single element from the input vector.
9538 // If the input vector has a single use, simplify it based on this use
9540 uint64_t IndexVal = IdxC->getZExtValue();
9541 if (EI.getOperand(0)->hasOneUse()) {
9543 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9546 EI.setOperand(0, V);
9551 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9552 return ReplaceInstUsesWith(EI, Elt);
9555 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9556 if (I->hasOneUse()) {
9557 // Push extractelement into predecessor operation if legal and
9558 // profitable to do so
9559 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9560 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9561 if (CheapToScalarize(BO, isConstantElt)) {
9562 ExtractElementInst *newEI0 =
9563 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9564 EI.getName()+".lhs");
9565 ExtractElementInst *newEI1 =
9566 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9567 EI.getName()+".rhs");
9568 InsertNewInstBefore(newEI0, EI);
9569 InsertNewInstBefore(newEI1, EI);
9570 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9572 } else if (isa<LoadInst>(I)) {
9573 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9574 PointerType::get(EI.getType()), EI);
9575 GetElementPtrInst *GEP =
9576 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9577 InsertNewInstBefore(GEP, EI);
9578 return new LoadInst(GEP);
9581 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9582 // Extracting the inserted element?
9583 if (IE->getOperand(2) == EI.getOperand(1))
9584 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9585 // If the inserted and extracted elements are constants, they must not
9586 // be the same value, extract from the pre-inserted value instead.
9587 if (isa<Constant>(IE->getOperand(2)) &&
9588 isa<Constant>(EI.getOperand(1))) {
9589 AddUsesToWorkList(EI);
9590 EI.setOperand(0, IE->getOperand(0));
9593 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9594 // If this is extracting an element from a shufflevector, figure out where
9595 // it came from and extract from the appropriate input element instead.
9596 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9597 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9599 if (SrcIdx < SVI->getType()->getNumElements())
9600 Src = SVI->getOperand(0);
9601 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9602 SrcIdx -= SVI->getType()->getNumElements();
9603 Src = SVI->getOperand(1);
9605 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9607 return new ExtractElementInst(Src, SrcIdx);
9614 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9615 /// elements from either LHS or RHS, return the shuffle mask and true.
9616 /// Otherwise, return false.
9617 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9618 std::vector<Constant*> &Mask) {
9619 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9620 "Invalid CollectSingleShuffleElements");
9621 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9623 if (isa<UndefValue>(V)) {
9624 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9626 } else if (V == LHS) {
9627 for (unsigned i = 0; i != NumElts; ++i)
9628 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9630 } else if (V == RHS) {
9631 for (unsigned i = 0; i != NumElts; ++i)
9632 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9634 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9635 // If this is an insert of an extract from some other vector, include it.
9636 Value *VecOp = IEI->getOperand(0);
9637 Value *ScalarOp = IEI->getOperand(1);
9638 Value *IdxOp = IEI->getOperand(2);
9640 if (!isa<ConstantInt>(IdxOp))
9642 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9644 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9645 // Okay, we can handle this if the vector we are insertinting into is
9647 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9648 // If so, update the mask to reflect the inserted undef.
9649 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9652 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9653 if (isa<ConstantInt>(EI->getOperand(1)) &&
9654 EI->getOperand(0)->getType() == V->getType()) {
9655 unsigned ExtractedIdx =
9656 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9658 // This must be extracting from either LHS or RHS.
9659 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9660 // Okay, we can handle this if the vector we are insertinting into is
9662 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9663 // If so, update the mask to reflect the inserted value.
9664 if (EI->getOperand(0) == LHS) {
9665 Mask[InsertedIdx & (NumElts-1)] =
9666 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9668 assert(EI->getOperand(0) == RHS);
9669 Mask[InsertedIdx & (NumElts-1)] =
9670 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9679 // TODO: Handle shufflevector here!
9684 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9685 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9686 /// that computes V and the LHS value of the shuffle.
9687 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9689 assert(isa<VectorType>(V->getType()) &&
9690 (RHS == 0 || V->getType() == RHS->getType()) &&
9691 "Invalid shuffle!");
9692 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9694 if (isa<UndefValue>(V)) {
9695 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9697 } else if (isa<ConstantAggregateZero>(V)) {
9698 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9700 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9701 // If this is an insert of an extract from some other vector, include it.
9702 Value *VecOp = IEI->getOperand(0);
9703 Value *ScalarOp = IEI->getOperand(1);
9704 Value *IdxOp = IEI->getOperand(2);
9706 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9707 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9708 EI->getOperand(0)->getType() == V->getType()) {
9709 unsigned ExtractedIdx =
9710 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9711 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9713 // Either the extracted from or inserted into vector must be RHSVec,
9714 // otherwise we'd end up with a shuffle of three inputs.
9715 if (EI->getOperand(0) == RHS || RHS == 0) {
9716 RHS = EI->getOperand(0);
9717 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9718 Mask[InsertedIdx & (NumElts-1)] =
9719 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9724 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9725 // Everything but the extracted element is replaced with the RHS.
9726 for (unsigned i = 0; i != NumElts; ++i) {
9727 if (i != InsertedIdx)
9728 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9733 // If this insertelement is a chain that comes from exactly these two
9734 // vectors, return the vector and the effective shuffle.
9735 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9736 return EI->getOperand(0);
9741 // TODO: Handle shufflevector here!
9743 // Otherwise, can't do anything fancy. Return an identity vector.
9744 for (unsigned i = 0; i != NumElts; ++i)
9745 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9749 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9750 Value *VecOp = IE.getOperand(0);
9751 Value *ScalarOp = IE.getOperand(1);
9752 Value *IdxOp = IE.getOperand(2);
9754 // If the inserted element was extracted from some other vector, and if the
9755 // indexes are constant, try to turn this into a shufflevector operation.
9756 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9757 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9758 EI->getOperand(0)->getType() == IE.getType()) {
9759 unsigned NumVectorElts = IE.getType()->getNumElements();
9760 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9761 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9763 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9764 return ReplaceInstUsesWith(IE, VecOp);
9766 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9767 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9769 // If we are extracting a value from a vector, then inserting it right
9770 // back into the same place, just use the input vector.
9771 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9772 return ReplaceInstUsesWith(IE, VecOp);
9774 // We could theoretically do this for ANY input. However, doing so could
9775 // turn chains of insertelement instructions into a chain of shufflevector
9776 // instructions, and right now we do not merge shufflevectors. As such,
9777 // only do this in a situation where it is clear that there is benefit.
9778 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9779 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9780 // the values of VecOp, except then one read from EIOp0.
9781 // Build a new shuffle mask.
9782 std::vector<Constant*> Mask;
9783 if (isa<UndefValue>(VecOp))
9784 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9786 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9787 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9790 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9791 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9792 ConstantVector::get(Mask));
9795 // If this insertelement isn't used by some other insertelement, turn it
9796 // (and any insertelements it points to), into one big shuffle.
9797 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9798 std::vector<Constant*> Mask;
9800 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9801 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9802 // We now have a shuffle of LHS, RHS, Mask.
9803 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9812 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9813 Value *LHS = SVI.getOperand(0);
9814 Value *RHS = SVI.getOperand(1);
9815 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9817 bool MadeChange = false;
9819 // Undefined shuffle mask -> undefined value.
9820 if (isa<UndefValue>(SVI.getOperand(2)))
9821 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9823 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9824 // the undef, change them to undefs.
9825 if (isa<UndefValue>(SVI.getOperand(1))) {
9826 // Scan to see if there are any references to the RHS. If so, replace them
9827 // with undef element refs and set MadeChange to true.
9828 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9829 if (Mask[i] >= e && Mask[i] != 2*e) {
9836 // Remap any references to RHS to use LHS.
9837 std::vector<Constant*> Elts;
9838 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9840 Elts.push_back(UndefValue::get(Type::Int32Ty));
9842 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9844 SVI.setOperand(2, ConstantVector::get(Elts));
9848 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9849 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9850 if (LHS == RHS || isa<UndefValue>(LHS)) {
9851 if (isa<UndefValue>(LHS) && LHS == RHS) {
9852 // shuffle(undef,undef,mask) -> undef.
9853 return ReplaceInstUsesWith(SVI, LHS);
9856 // Remap any references to RHS to use LHS.
9857 std::vector<Constant*> Elts;
9858 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9860 Elts.push_back(UndefValue::get(Type::Int32Ty));
9862 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9863 (Mask[i] < e && isa<UndefValue>(LHS)))
9864 Mask[i] = 2*e; // Turn into undef.
9866 Mask[i] &= (e-1); // Force to LHS.
9867 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9870 SVI.setOperand(0, SVI.getOperand(1));
9871 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9872 SVI.setOperand(2, ConstantVector::get(Elts));
9873 LHS = SVI.getOperand(0);
9874 RHS = SVI.getOperand(1);
9878 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9879 bool isLHSID = true, isRHSID = true;
9881 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9882 if (Mask[i] >= e*2) continue; // Ignore undef values.
9883 // Is this an identity shuffle of the LHS value?
9884 isLHSID &= (Mask[i] == i);
9886 // Is this an identity shuffle of the RHS value?
9887 isRHSID &= (Mask[i]-e == i);
9890 // Eliminate identity shuffles.
9891 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9892 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9894 // If the LHS is a shufflevector itself, see if we can combine it with this
9895 // one without producing an unusual shuffle. Here we are really conservative:
9896 // we are absolutely afraid of producing a shuffle mask not in the input
9897 // program, because the code gen may not be smart enough to turn a merged
9898 // shuffle into two specific shuffles: it may produce worse code. As such,
9899 // we only merge two shuffles if the result is one of the two input shuffle
9900 // masks. In this case, merging the shuffles just removes one instruction,
9901 // which we know is safe. This is good for things like turning:
9902 // (splat(splat)) -> splat.
9903 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9904 if (isa<UndefValue>(RHS)) {
9905 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9907 std::vector<unsigned> NewMask;
9908 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9910 NewMask.push_back(2*e);
9912 NewMask.push_back(LHSMask[Mask[i]]);
9914 // If the result mask is equal to the src shuffle or this shuffle mask, do
9916 if (NewMask == LHSMask || NewMask == Mask) {
9917 std::vector<Constant*> Elts;
9918 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9919 if (NewMask[i] >= e*2) {
9920 Elts.push_back(UndefValue::get(Type::Int32Ty));
9922 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9925 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9926 LHSSVI->getOperand(1),
9927 ConstantVector::get(Elts));
9932 return MadeChange ? &SVI : 0;
9938 /// TryToSinkInstruction - Try to move the specified instruction from its
9939 /// current block into the beginning of DestBlock, which can only happen if it's
9940 /// safe to move the instruction past all of the instructions between it and the
9941 /// end of its block.
9942 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9943 assert(I->hasOneUse() && "Invariants didn't hold!");
9945 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9946 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9948 // Do not sink alloca instructions out of the entry block.
9949 if (isa<AllocaInst>(I) && I->getParent() ==
9950 &DestBlock->getParent()->getEntryBlock())
9953 // We can only sink load instructions if there is nothing between the load and
9954 // the end of block that could change the value.
9955 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9956 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9958 if (Scan->mayWriteToMemory())
9962 BasicBlock::iterator InsertPos = DestBlock->begin();
9963 while (isa<PHINode>(InsertPos)) ++InsertPos;
9965 I->moveBefore(InsertPos);
9971 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9972 /// all reachable code to the worklist.
9974 /// This has a couple of tricks to make the code faster and more powerful. In
9975 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9976 /// them to the worklist (this significantly speeds up instcombine on code where
9977 /// many instructions are dead or constant). Additionally, if we find a branch
9978 /// whose condition is a known constant, we only visit the reachable successors.
9980 static void AddReachableCodeToWorklist(BasicBlock *BB,
9981 SmallPtrSet<BasicBlock*, 64> &Visited,
9983 const TargetData *TD) {
9984 // We have now visited this block! If we've already been here, bail out.
9985 if (!Visited.insert(BB)) return;
9987 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9988 Instruction *Inst = BBI++;
9990 // DCE instruction if trivially dead.
9991 if (isInstructionTriviallyDead(Inst)) {
9993 DOUT << "IC: DCE: " << *Inst;
9994 Inst->eraseFromParent();
9998 // ConstantProp instruction if trivially constant.
9999 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10000 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10001 Inst->replaceAllUsesWith(C);
10003 Inst->eraseFromParent();
10007 IC.AddToWorkList(Inst);
10010 // Recursively visit successors. If this is a branch or switch on a constant,
10011 // only visit the reachable successor.
10012 TerminatorInst *TI = BB->getTerminator();
10013 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10014 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10015 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10016 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, IC, TD);
10019 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10020 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10021 // See if this is an explicit destination.
10022 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10023 if (SI->getCaseValue(i) == Cond) {
10024 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, IC, TD);
10028 // Otherwise it is the default destination.
10029 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, IC, TD);
10034 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10035 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, IC, TD);
10038 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10039 bool Changed = false;
10040 TD = &getAnalysis<TargetData>();
10042 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10043 << F.getNameStr() << "\n");
10046 // Do a depth-first traversal of the function, populate the worklist with
10047 // the reachable instructions. Ignore blocks that are not reachable. Keep
10048 // track of which blocks we visit.
10049 SmallPtrSet<BasicBlock*, 64> Visited;
10050 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10052 // Do a quick scan over the function. If we find any blocks that are
10053 // unreachable, remove any instructions inside of them. This prevents
10054 // the instcombine code from having to deal with some bad special cases.
10055 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10056 if (!Visited.count(BB)) {
10057 Instruction *Term = BB->getTerminator();
10058 while (Term != BB->begin()) { // Remove instrs bottom-up
10059 BasicBlock::iterator I = Term; --I;
10061 DOUT << "IC: DCE: " << *I;
10064 if (!I->use_empty())
10065 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10066 I->eraseFromParent();
10071 while (!Worklist.empty()) {
10072 Instruction *I = RemoveOneFromWorkList();
10073 if (I == 0) continue; // skip null values.
10075 // Check to see if we can DCE the instruction.
10076 if (isInstructionTriviallyDead(I)) {
10077 // Add operands to the worklist.
10078 if (I->getNumOperands() < 4)
10079 AddUsesToWorkList(*I);
10082 DOUT << "IC: DCE: " << *I;
10084 I->eraseFromParent();
10085 RemoveFromWorkList(I);
10089 // Instruction isn't dead, see if we can constant propagate it.
10090 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10091 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10093 // Add operands to the worklist.
10094 AddUsesToWorkList(*I);
10095 ReplaceInstUsesWith(*I, C);
10098 I->eraseFromParent();
10099 RemoveFromWorkList(I);
10103 // See if we can trivially sink this instruction to a successor basic block.
10104 if (I->hasOneUse()) {
10105 BasicBlock *BB = I->getParent();
10106 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10107 if (UserParent != BB) {
10108 bool UserIsSuccessor = false;
10109 // See if the user is one of our successors.
10110 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10111 if (*SI == UserParent) {
10112 UserIsSuccessor = true;
10116 // If the user is one of our immediate successors, and if that successor
10117 // only has us as a predecessors (we'd have to split the critical edge
10118 // otherwise), we can keep going.
10119 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10120 next(pred_begin(UserParent)) == pred_end(UserParent))
10121 // Okay, the CFG is simple enough, try to sink this instruction.
10122 Changed |= TryToSinkInstruction(I, UserParent);
10126 // Now that we have an instruction, try combining it to simplify it...
10127 if (Instruction *Result = visit(*I)) {
10129 // Should we replace the old instruction with a new one?
10131 DOUT << "IC: Old = " << *I
10132 << " New = " << *Result;
10134 // Everything uses the new instruction now.
10135 I->replaceAllUsesWith(Result);
10137 // Push the new instruction and any users onto the worklist.
10138 AddToWorkList(Result);
10139 AddUsersToWorkList(*Result);
10141 // Move the name to the new instruction first.
10142 Result->takeName(I);
10144 // Insert the new instruction into the basic block...
10145 BasicBlock *InstParent = I->getParent();
10146 BasicBlock::iterator InsertPos = I;
10148 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10149 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10152 InstParent->getInstList().insert(InsertPos, Result);
10154 // Make sure that we reprocess all operands now that we reduced their
10156 AddUsesToWorkList(*I);
10158 // Instructions can end up on the worklist more than once. Make sure
10159 // we do not process an instruction that has been deleted.
10160 RemoveFromWorkList(I);
10162 // Erase the old instruction.
10163 InstParent->getInstList().erase(I);
10165 DOUT << "IC: MOD = " << *I;
10167 // If the instruction was modified, it's possible that it is now dead.
10168 // if so, remove it.
10169 if (isInstructionTriviallyDead(I)) {
10170 // Make sure we process all operands now that we are reducing their
10172 AddUsesToWorkList(*I);
10174 // Instructions may end up in the worklist more than once. Erase all
10175 // occurrences of this instruction.
10176 RemoveFromWorkList(I);
10177 I->eraseFromParent();
10180 AddUsersToWorkList(*I);
10187 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10192 bool InstCombiner::runOnFunction(Function &F) {
10193 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10195 bool EverMadeChange = false;
10197 // Iterate while there is work to do.
10198 unsigned Iteration = 0;
10199 while (DoOneIteration(F, Iteration++))
10200 EverMadeChange = true;
10201 return EverMadeChange;
10204 FunctionPass *llvm::createInstructionCombiningPass() {
10205 return new InstCombiner();