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);
186 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
190 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
191 ICmpInst::Predicate Cond, Instruction &I);
192 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
194 Instruction *commonCastTransforms(CastInst &CI);
195 Instruction *commonIntCastTransforms(CastInst &CI);
196 Instruction *visitTrunc(CastInst &CI);
197 Instruction *visitZExt(CastInst &CI);
198 Instruction *visitSExt(CastInst &CI);
199 Instruction *visitFPTrunc(CastInst &CI);
200 Instruction *visitFPExt(CastInst &CI);
201 Instruction *visitFPToUI(CastInst &CI);
202 Instruction *visitFPToSI(CastInst &CI);
203 Instruction *visitUIToFP(CastInst &CI);
204 Instruction *visitSIToFP(CastInst &CI);
205 Instruction *visitPtrToInt(CastInst &CI);
206 Instruction *visitIntToPtr(CastInst &CI);
207 Instruction *visitBitCast(CastInst &CI);
208 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
210 Instruction *visitSelectInst(SelectInst &CI);
211 Instruction *visitCallInst(CallInst &CI);
212 Instruction *visitInvokeInst(InvokeInst &II);
213 Instruction *visitPHINode(PHINode &PN);
214 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
215 Instruction *visitAllocationInst(AllocationInst &AI);
216 Instruction *visitFreeInst(FreeInst &FI);
217 Instruction *visitLoadInst(LoadInst &LI);
218 Instruction *visitStoreInst(StoreInst &SI);
219 Instruction *visitBranchInst(BranchInst &BI);
220 Instruction *visitSwitchInst(SwitchInst &SI);
221 Instruction *visitInsertElementInst(InsertElementInst &IE);
222 Instruction *visitExtractElementInst(ExtractElementInst &EI);
223 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
225 // visitInstruction - Specify what to return for unhandled instructions...
226 Instruction *visitInstruction(Instruction &I) { return 0; }
229 Instruction *visitCallSite(CallSite CS);
230 bool transformConstExprCastCall(CallSite CS);
233 // InsertNewInstBefore - insert an instruction New before instruction Old
234 // in the program. Add the new instruction to the worklist.
236 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
237 assert(New && New->getParent() == 0 &&
238 "New instruction already inserted into a basic block!");
239 BasicBlock *BB = Old.getParent();
240 BB->getInstList().insert(&Old, New); // Insert inst
245 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
246 /// This also adds the cast to the worklist. Finally, this returns the
248 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
250 if (V->getType() == Ty) return V;
252 if (Constant *CV = dyn_cast<Constant>(V))
253 return ConstantExpr::getCast(opc, CV, Ty);
255 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
260 // ReplaceInstUsesWith - This method is to be used when an instruction is
261 // found to be dead, replacable with another preexisting expression. Here
262 // we add all uses of I to the worklist, replace all uses of I with the new
263 // value, then return I, so that the inst combiner will know that I was
266 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
267 AddUsersToWorkList(I); // Add all modified instrs to worklist
269 I.replaceAllUsesWith(V);
272 // If we are replacing the instruction with itself, this must be in a
273 // segment of unreachable code, so just clobber the instruction.
274 I.replaceAllUsesWith(UndefValue::get(I.getType()));
279 // UpdateValueUsesWith - This method is to be used when an value is
280 // found to be replacable with another preexisting expression or was
281 // updated. Here we add all uses of I to the worklist, replace all uses of
282 // I with the new value (unless the instruction was just updated), then
283 // return true, so that the inst combiner will know that I was modified.
285 bool UpdateValueUsesWith(Value *Old, Value *New) {
286 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
288 Old->replaceAllUsesWith(New);
289 if (Instruction *I = dyn_cast<Instruction>(Old))
291 if (Instruction *I = dyn_cast<Instruction>(New))
296 // EraseInstFromFunction - When dealing with an instruction that has side
297 // effects or produces a void value, we can't rely on DCE to delete the
298 // instruction. Instead, visit methods should return the value returned by
300 Instruction *EraseInstFromFunction(Instruction &I) {
301 assert(I.use_empty() && "Cannot erase instruction that is used!");
302 AddUsesToWorkList(I);
303 RemoveFromWorkList(&I);
305 return 0; // Don't do anything with FI
309 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
310 /// InsertBefore instruction. This is specialized a bit to avoid inserting
311 /// casts that are known to not do anything...
313 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
314 Value *V, const Type *DestTy,
315 Instruction *InsertBefore);
317 /// SimplifyCommutative - This performs a few simplifications for
318 /// commutative operators.
319 bool SimplifyCommutative(BinaryOperator &I);
321 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
322 /// most-complex to least-complex order.
323 bool SimplifyCompare(CmpInst &I);
325 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
326 /// on the demanded bits.
327 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
328 APInt& KnownZero, APInt& KnownOne,
331 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
332 uint64_t &UndefElts, unsigned Depth = 0);
334 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
335 // PHI node as operand #0, see if we can fold the instruction into the PHI
336 // (which is only possible if all operands to the PHI are constants).
337 Instruction *FoldOpIntoPhi(Instruction &I);
339 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
340 // operator and they all are only used by the PHI, PHI together their
341 // inputs, and do the operation once, to the result of the PHI.
342 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
343 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
346 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
347 ConstantInt *AndRHS, BinaryOperator &TheAnd);
349 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
350 bool isSub, Instruction &I);
351 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
352 bool isSigned, bool Inside, Instruction &IB);
353 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
354 Instruction *MatchBSwap(BinaryOperator &I);
356 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
359 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
362 // getComplexity: Assign a complexity or rank value to LLVM Values...
363 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
364 static unsigned getComplexity(Value *V) {
365 if (isa<Instruction>(V)) {
366 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
370 if (isa<Argument>(V)) return 3;
371 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
374 // isOnlyUse - Return true if this instruction will be deleted if we stop using
376 static bool isOnlyUse(Value *V) {
377 return V->hasOneUse() || isa<Constant>(V);
380 // getPromotedType - Return the specified type promoted as it would be to pass
381 // though a va_arg area...
382 static const Type *getPromotedType(const Type *Ty) {
383 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
384 if (ITy->getBitWidth() < 32)
385 return Type::Int32Ty;
386 } else if (Ty == Type::FloatTy)
387 return Type::DoubleTy;
391 /// getBitCastOperand - If the specified operand is a CastInst or a constant
392 /// expression bitcast, return the operand value, otherwise return null.
393 static Value *getBitCastOperand(Value *V) {
394 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
395 return I->getOperand(0);
396 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
397 if (CE->getOpcode() == Instruction::BitCast)
398 return CE->getOperand(0);
402 /// This function is a wrapper around CastInst::isEliminableCastPair. It
403 /// simply extracts arguments and returns what that function returns.
404 static Instruction::CastOps
405 isEliminableCastPair(
406 const CastInst *CI, ///< The first cast instruction
407 unsigned opcode, ///< The opcode of the second cast instruction
408 const Type *DstTy, ///< The target type for the second cast instruction
409 TargetData *TD ///< The target data for pointer size
412 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
413 const Type *MidTy = CI->getType(); // B from above
415 // Get the opcodes of the two Cast instructions
416 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
417 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
419 return Instruction::CastOps(
420 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
421 DstTy, TD->getIntPtrType()));
424 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
425 /// in any code being generated. It does not require codegen if V is simple
426 /// enough or if the cast can be folded into other casts.
427 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
428 const Type *Ty, TargetData *TD) {
429 if (V->getType() == Ty || isa<Constant>(V)) return false;
431 // If this is another cast that can be eliminated, it isn't codegen either.
432 if (const CastInst *CI = dyn_cast<CastInst>(V))
433 if (isEliminableCastPair(CI, opcode, Ty, TD))
438 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
439 /// InsertBefore instruction. This is specialized a bit to avoid inserting
440 /// casts that are known to not do anything...
442 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
443 Value *V, const Type *DestTy,
444 Instruction *InsertBefore) {
445 if (V->getType() == DestTy) return V;
446 if (Constant *C = dyn_cast<Constant>(V))
447 return ConstantExpr::getCast(opcode, C, DestTy);
449 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
452 // SimplifyCommutative - This performs a few simplifications for commutative
455 // 1. Order operands such that they are listed from right (least complex) to
456 // left (most complex). This puts constants before unary operators before
459 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
460 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
462 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
463 bool Changed = false;
464 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
465 Changed = !I.swapOperands();
467 if (!I.isAssociative()) return Changed;
468 Instruction::BinaryOps Opcode = I.getOpcode();
469 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
470 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
471 if (isa<Constant>(I.getOperand(1))) {
472 Constant *Folded = ConstantExpr::get(I.getOpcode(),
473 cast<Constant>(I.getOperand(1)),
474 cast<Constant>(Op->getOperand(1)));
475 I.setOperand(0, Op->getOperand(0));
476 I.setOperand(1, Folded);
478 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
479 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
480 isOnlyUse(Op) && isOnlyUse(Op1)) {
481 Constant *C1 = cast<Constant>(Op->getOperand(1));
482 Constant *C2 = cast<Constant>(Op1->getOperand(1));
484 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
485 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
486 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
490 I.setOperand(0, New);
491 I.setOperand(1, Folded);
498 /// SimplifyCompare - For a CmpInst this function just orders the operands
499 /// so that theyare listed from right (least complex) to left (most complex).
500 /// This puts constants before unary operators before binary operators.
501 bool InstCombiner::SimplifyCompare(CmpInst &I) {
502 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
505 // Compare instructions are not associative so there's nothing else we can do.
509 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
510 // if the LHS is a constant zero (which is the 'negate' form).
512 static inline Value *dyn_castNegVal(Value *V) {
513 if (BinaryOperator::isNeg(V))
514 return BinaryOperator::getNegArgument(V);
516 // Constants can be considered to be negated values if they can be folded.
517 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
518 return ConstantExpr::getNeg(C);
522 static inline Value *dyn_castNotVal(Value *V) {
523 if (BinaryOperator::isNot(V))
524 return BinaryOperator::getNotArgument(V);
526 // Constants can be considered to be not'ed values...
527 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
528 return ConstantInt::get(~C->getValue());
532 // dyn_castFoldableMul - If this value is a multiply that can be folded into
533 // other computations (because it has a constant operand), return the
534 // non-constant operand of the multiply, and set CST to point to the multiplier.
535 // Otherwise, return null.
537 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
538 if (V->hasOneUse() && V->getType()->isInteger())
539 if (Instruction *I = dyn_cast<Instruction>(V)) {
540 if (I->getOpcode() == Instruction::Mul)
541 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
542 return I->getOperand(0);
543 if (I->getOpcode() == Instruction::Shl)
544 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
545 // The multiplier is really 1 << CST.
546 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
547 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
548 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
549 return I->getOperand(0);
555 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
556 /// expression, return it.
557 static User *dyn_castGetElementPtr(Value *V) {
558 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
559 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
560 if (CE->getOpcode() == Instruction::GetElementPtr)
561 return cast<User>(V);
565 /// AddOne - Add one to a ConstantInt
566 static ConstantInt *AddOne(ConstantInt *C) {
567 APInt Val(C->getValue());
568 return ConstantInt::get(++Val);
570 /// SubOne - Subtract one from a ConstantInt
571 static ConstantInt *SubOne(ConstantInt *C) {
572 APInt Val(C->getValue());
573 return ConstantInt::get(--Val);
575 /// Add - Add two ConstantInts together
576 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
577 return ConstantInt::get(C1->getValue() + C2->getValue());
579 /// And - Bitwise AND two ConstantInts together
580 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
581 return ConstantInt::get(C1->getValue() & C2->getValue());
583 /// Subtract - Subtract one ConstantInt from another
584 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
585 return ConstantInt::get(C1->getValue() - C2->getValue());
587 /// Multiply - Multiply two ConstantInts together
588 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
589 return ConstantInt::get(C1->getValue() * C2->getValue());
592 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
593 /// known to be either zero or one and return them in the KnownZero/KnownOne
594 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
596 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
597 /// we cannot optimize based on the assumption that it is zero without changing
598 /// it to be an explicit zero. If we don't change it to zero, other code could
599 /// optimized based on the contradictory assumption that it is non-zero.
600 /// Because instcombine aggressively folds operations with undef args anyway,
601 /// this won't lose us code quality.
602 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
603 APInt& KnownOne, unsigned Depth = 0) {
604 assert(V && "No Value?");
605 assert(Depth <= 6 && "Limit Search Depth");
606 uint32_t BitWidth = Mask.getBitWidth();
607 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
608 KnownZero.getBitWidth() == BitWidth &&
609 KnownOne.getBitWidth() == BitWidth &&
610 "V, Mask, KnownOne and KnownZero should have same BitWidth");
611 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
612 // We know all of the bits for a constant!
613 KnownOne = CI->getValue() & Mask;
614 KnownZero = ~KnownOne & Mask;
618 if (Depth == 6 || Mask == 0)
619 return; // Limit search depth.
621 Instruction *I = dyn_cast<Instruction>(V);
624 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
625 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
627 switch (I->getOpcode()) {
628 case Instruction::And: {
629 // If either the LHS or the RHS are Zero, the result is zero.
630 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
631 APInt Mask2(Mask & ~KnownZero);
632 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
633 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
634 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
636 // Output known-1 bits are only known if set in both the LHS & RHS.
637 KnownOne &= KnownOne2;
638 // Output known-0 are known to be clear if zero in either the LHS | RHS.
639 KnownZero |= KnownZero2;
642 case Instruction::Or: {
643 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
644 APInt Mask2(Mask & ~KnownOne);
645 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
646 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
647 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
649 // Output known-0 bits are only known if clear in both the LHS & RHS.
650 KnownZero &= KnownZero2;
651 // Output known-1 are known to be set if set in either the LHS | RHS.
652 KnownOne |= KnownOne2;
655 case Instruction::Xor: {
656 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
657 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
658 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
659 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
661 // Output known-0 bits are known if clear or set in both the LHS & RHS.
662 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
663 // Output known-1 are known to be set if set in only one of the LHS, RHS.
664 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
665 KnownZero = KnownZeroOut;
668 case Instruction::Select:
669 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
670 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
671 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
672 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
674 // Only known if known in both the LHS and RHS.
675 KnownOne &= KnownOne2;
676 KnownZero &= KnownZero2;
678 case Instruction::FPTrunc:
679 case Instruction::FPExt:
680 case Instruction::FPToUI:
681 case Instruction::FPToSI:
682 case Instruction::SIToFP:
683 case Instruction::PtrToInt:
684 case Instruction::UIToFP:
685 case Instruction::IntToPtr:
686 return; // Can't work with floating point or pointers
687 case Instruction::Trunc: {
688 // All these have integer operands
689 uint32_t SrcBitWidth =
690 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
692 MaskIn.zext(SrcBitWidth);
693 KnownZero.zext(SrcBitWidth);
694 KnownOne.zext(SrcBitWidth);
695 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
696 KnownZero.trunc(BitWidth);
697 KnownOne.trunc(BitWidth);
700 case Instruction::BitCast: {
701 const Type *SrcTy = I->getOperand(0)->getType();
702 if (SrcTy->isInteger()) {
703 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
708 case Instruction::ZExt: {
709 // Compute the bits in the result that are not present in the input.
710 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
711 uint32_t SrcBitWidth = SrcTy->getBitWidth();
714 MaskIn.trunc(SrcBitWidth);
715 KnownZero.trunc(SrcBitWidth);
716 KnownOne.trunc(SrcBitWidth);
717 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
718 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
719 // The top bits are known to be zero.
720 KnownZero.zext(BitWidth);
721 KnownOne.zext(BitWidth);
722 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
725 case Instruction::SExt: {
726 // Compute the bits in the result that are not present in the input.
727 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
728 uint32_t SrcBitWidth = SrcTy->getBitWidth();
731 MaskIn.trunc(SrcBitWidth);
732 KnownZero.trunc(SrcBitWidth);
733 KnownOne.trunc(SrcBitWidth);
734 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
735 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
736 KnownZero.zext(BitWidth);
737 KnownOne.zext(BitWidth);
739 // If the sign bit of the input is known set or clear, then we know the
740 // top bits of the result.
741 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
742 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
743 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
744 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
747 case Instruction::Shl:
748 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
749 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
750 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
751 APInt Mask2(Mask.lshr(ShiftAmt));
752 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
753 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
754 KnownZero <<= ShiftAmt;
755 KnownOne <<= ShiftAmt;
756 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
760 case Instruction::LShr:
761 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
762 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
763 // Compute the new bits that are at the top now.
764 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
766 // Unsigned shift right.
767 APInt Mask2(Mask.shl(ShiftAmt));
768 ComputeMaskedBits(I->getOperand(0), Mask2, 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);
772 // high bits known zero.
773 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
777 case Instruction::AShr:
778 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
779 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
780 // Compute the new bits that are at the top now.
781 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
783 // Signed shift right.
784 APInt Mask2(Mask.shl(ShiftAmt));
785 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
786 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
787 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
788 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
790 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
791 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
792 KnownZero |= HighBits;
793 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
794 KnownOne |= HighBits;
801 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
802 /// this predicate to simplify operations downstream. Mask is known to be zero
803 /// for bits that V cannot have.
804 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
805 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
806 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
807 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
808 return (KnownZero & Mask) == Mask;
811 /// ShrinkDemandedConstant - Check to see if the specified operand of the
812 /// specified instruction is a constant integer. If so, check to see if there
813 /// are any bits set in the constant that are not demanded. If so, shrink the
814 /// constant and return true.
815 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
817 assert(I && "No instruction?");
818 assert(OpNo < I->getNumOperands() && "Operand index too large");
820 // If the operand is not a constant integer, nothing to do.
821 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
822 if (!OpC) return false;
824 // If there are no bits set that aren't demanded, nothing to do.
825 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
826 if ((~Demanded & OpC->getValue()) == 0)
829 // This instruction is producing bits that are not demanded. Shrink the RHS.
830 Demanded &= OpC->getValue();
831 I->setOperand(OpNo, ConstantInt::get(Demanded));
835 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
836 // set of known zero and one bits, compute the maximum and minimum values that
837 // could have the specified known zero and known one bits, returning them in
839 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
840 const APInt& KnownZero,
841 const APInt& KnownOne,
842 APInt& Min, APInt& Max) {
843 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
844 assert(KnownZero.getBitWidth() == BitWidth &&
845 KnownOne.getBitWidth() == BitWidth &&
846 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
847 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
848 APInt UnknownBits = ~(KnownZero|KnownOne);
850 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
851 // bit if it is unknown.
853 Max = KnownOne|UnknownBits;
855 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
857 Max.clear(BitWidth-1);
861 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
862 // a set of known zero and one bits, compute the maximum and minimum values that
863 // could have the specified known zero and known one bits, returning them in
865 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
866 const APInt& KnownZero,
867 const APInt& KnownOne,
870 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
871 assert(KnownZero.getBitWidth() == BitWidth &&
872 KnownOne.getBitWidth() == BitWidth &&
873 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
874 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
875 APInt UnknownBits = ~(KnownZero|KnownOne);
877 // The minimum value is when the unknown bits are all zeros.
879 // The maximum value is when the unknown bits are all ones.
880 Max = KnownOne|UnknownBits;
883 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
884 /// value based on the demanded bits. When this function is called, it is known
885 /// that only the bits set in DemandedMask of the result of V are ever used
886 /// downstream. Consequently, depending on the mask and V, it may be possible
887 /// to replace V with a constant or one of its operands. In such cases, this
888 /// function does the replacement and returns true. In all other cases, it
889 /// returns false after analyzing the expression and setting KnownOne and known
890 /// to be one in the expression. KnownZero contains all the bits that are known
891 /// to be zero in the expression. These are provided to potentially allow the
892 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
893 /// the expression. KnownOne and KnownZero always follow the invariant that
894 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
895 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
896 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
897 /// and KnownOne must all be the same.
898 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
899 APInt& KnownZero, APInt& KnownOne,
901 assert(V != 0 && "Null pointer of Value???");
902 assert(Depth <= 6 && "Limit Search Depth");
903 uint32_t BitWidth = DemandedMask.getBitWidth();
904 const IntegerType *VTy = cast<IntegerType>(V->getType());
905 assert(VTy->getBitWidth() == BitWidth &&
906 KnownZero.getBitWidth() == BitWidth &&
907 KnownOne.getBitWidth() == BitWidth &&
908 "Value *V, DemandedMask, KnownZero and KnownOne \
909 must have same BitWidth");
910 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
911 // We know all of the bits for a constant!
912 KnownOne = CI->getValue() & DemandedMask;
913 KnownZero = ~KnownOne & DemandedMask;
919 if (!V->hasOneUse()) { // Other users may use these bits.
920 if (Depth != 0) { // Not at the root.
921 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
922 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
925 // If this is the root being simplified, allow it to have multiple uses,
926 // just set the DemandedMask to all bits.
927 DemandedMask = APInt::getAllOnesValue(BitWidth);
928 } else if (DemandedMask == 0) { // Not demanding any bits from V.
929 if (V != UndefValue::get(VTy))
930 return UpdateValueUsesWith(V, UndefValue::get(VTy));
932 } else if (Depth == 6) { // Limit search depth.
936 Instruction *I = dyn_cast<Instruction>(V);
937 if (!I) return false; // Only analyze instructions.
939 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
940 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
941 switch (I->getOpcode()) {
943 case Instruction::And:
944 // If either the LHS or the RHS are Zero, the result is zero.
945 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
946 RHSKnownZero, RHSKnownOne, Depth+1))
948 assert((RHSKnownZero & RHSKnownOne) == 0 &&
949 "Bits known to be one AND zero?");
951 // If something is known zero on the RHS, the bits aren't demanded on the
953 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
954 LHSKnownZero, LHSKnownOne, Depth+1))
956 assert((LHSKnownZero & LHSKnownOne) == 0 &&
957 "Bits known to be one AND zero?");
959 // If all of the demanded bits are known 1 on one side, return the other.
960 // These bits cannot contribute to the result of the 'and'.
961 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
962 (DemandedMask & ~LHSKnownZero))
963 return UpdateValueUsesWith(I, I->getOperand(0));
964 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
965 (DemandedMask & ~RHSKnownZero))
966 return UpdateValueUsesWith(I, I->getOperand(1));
968 // If all of the demanded bits in the inputs are known zeros, return zero.
969 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
970 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
972 // If the RHS is a constant, see if we can simplify it.
973 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
974 return UpdateValueUsesWith(I, I);
976 // Output known-1 bits are only known if set in both the LHS & RHS.
977 RHSKnownOne &= LHSKnownOne;
978 // Output known-0 are known to be clear if zero in either the LHS | RHS.
979 RHSKnownZero |= LHSKnownZero;
981 case Instruction::Or:
982 // If either the LHS or the RHS are One, the result is One.
983 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
984 RHSKnownZero, RHSKnownOne, Depth+1))
986 assert((RHSKnownZero & RHSKnownOne) == 0 &&
987 "Bits known to be one AND zero?");
988 // If something is known one on the RHS, the bits aren't demanded on the
990 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
991 LHSKnownZero, LHSKnownOne, Depth+1))
993 assert((LHSKnownZero & LHSKnownOne) == 0 &&
994 "Bits known to be one AND zero?");
996 // If all of the demanded bits are known zero on one side, return the other.
997 // These bits cannot contribute to the result of the 'or'.
998 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
999 (DemandedMask & ~LHSKnownOne))
1000 return UpdateValueUsesWith(I, I->getOperand(0));
1001 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1002 (DemandedMask & ~RHSKnownOne))
1003 return UpdateValueUsesWith(I, I->getOperand(1));
1005 // If all of the potentially set bits on one side are known to be set on
1006 // the other side, just use the 'other' side.
1007 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1008 (DemandedMask & (~RHSKnownZero)))
1009 return UpdateValueUsesWith(I, I->getOperand(0));
1010 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1011 (DemandedMask & (~LHSKnownZero)))
1012 return UpdateValueUsesWith(I, I->getOperand(1));
1014 // If the RHS is a constant, see if we can simplify it.
1015 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1016 return UpdateValueUsesWith(I, I);
1018 // Output known-0 bits are only known if clear in both the LHS & RHS.
1019 RHSKnownZero &= LHSKnownZero;
1020 // Output known-1 are known to be set if set in either the LHS | RHS.
1021 RHSKnownOne |= LHSKnownOne;
1023 case Instruction::Xor: {
1024 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1025 RHSKnownZero, RHSKnownOne, Depth+1))
1027 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1028 "Bits known to be one AND zero?");
1029 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1030 LHSKnownZero, LHSKnownOne, Depth+1))
1032 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1033 "Bits known to be one AND zero?");
1035 // If all of the demanded bits are known zero on one side, return the other.
1036 // These bits cannot contribute to the result of the 'xor'.
1037 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1038 return UpdateValueUsesWith(I, I->getOperand(0));
1039 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1040 return UpdateValueUsesWith(I, I->getOperand(1));
1042 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1043 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1044 (RHSKnownOne & LHSKnownOne);
1045 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1046 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1047 (RHSKnownOne & LHSKnownZero);
1049 // If all of the demanded bits are known to be zero on one side or the
1050 // other, turn this into an *inclusive* or.
1051 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1052 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1054 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1056 InsertNewInstBefore(Or, *I);
1057 return UpdateValueUsesWith(I, Or);
1060 // If all of the demanded bits on one side are known, and all of the set
1061 // bits on that side are also known to be set on the other side, turn this
1062 // into an AND, as we know the bits will be cleared.
1063 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1064 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1066 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1067 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1069 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1070 InsertNewInstBefore(And, *I);
1071 return UpdateValueUsesWith(I, And);
1075 // If the RHS is a constant, see if we can simplify it.
1076 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1077 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1078 return UpdateValueUsesWith(I, I);
1080 RHSKnownZero = KnownZeroOut;
1081 RHSKnownOne = KnownOneOut;
1084 case Instruction::Select:
1085 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1086 RHSKnownZero, RHSKnownOne, Depth+1))
1088 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1089 LHSKnownZero, LHSKnownOne, Depth+1))
1091 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1092 "Bits known to be one AND zero?");
1093 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1094 "Bits known to be one AND zero?");
1096 // If the operands are constants, see if we can simplify them.
1097 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1098 return UpdateValueUsesWith(I, I);
1099 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1100 return UpdateValueUsesWith(I, I);
1102 // Only known if known in both the LHS and RHS.
1103 RHSKnownOne &= LHSKnownOne;
1104 RHSKnownZero &= LHSKnownZero;
1106 case Instruction::Trunc: {
1108 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1109 DemandedMask.zext(truncBf);
1110 RHSKnownZero.zext(truncBf);
1111 RHSKnownOne.zext(truncBf);
1112 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1113 RHSKnownZero, RHSKnownOne, Depth+1))
1115 DemandedMask.trunc(BitWidth);
1116 RHSKnownZero.trunc(BitWidth);
1117 RHSKnownOne.trunc(BitWidth);
1118 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1119 "Bits known to be one AND zero?");
1122 case Instruction::BitCast:
1123 if (!I->getOperand(0)->getType()->isInteger())
1126 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1127 RHSKnownZero, RHSKnownOne, Depth+1))
1129 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1130 "Bits known to be one AND zero?");
1132 case Instruction::ZExt: {
1133 // Compute the bits in the result that are not present in the input.
1134 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1135 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1137 DemandedMask.trunc(SrcBitWidth);
1138 RHSKnownZero.trunc(SrcBitWidth);
1139 RHSKnownOne.trunc(SrcBitWidth);
1140 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1141 RHSKnownZero, RHSKnownOne, Depth+1))
1143 DemandedMask.zext(BitWidth);
1144 RHSKnownZero.zext(BitWidth);
1145 RHSKnownOne.zext(BitWidth);
1146 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1147 "Bits known to be one AND zero?");
1148 // The top bits are known to be zero.
1149 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1152 case Instruction::SExt: {
1153 // Compute the bits in the result that are not present in the input.
1154 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1155 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1157 APInt InputDemandedBits = DemandedMask &
1158 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1160 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1161 // If any of the sign extended bits are demanded, we know that the sign
1163 if ((NewBits & DemandedMask) != 0)
1164 InputDemandedBits.set(SrcBitWidth-1);
1166 InputDemandedBits.trunc(SrcBitWidth);
1167 RHSKnownZero.trunc(SrcBitWidth);
1168 RHSKnownOne.trunc(SrcBitWidth);
1169 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1170 RHSKnownZero, RHSKnownOne, Depth+1))
1172 InputDemandedBits.zext(BitWidth);
1173 RHSKnownZero.zext(BitWidth);
1174 RHSKnownOne.zext(BitWidth);
1175 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1176 "Bits known to be one AND zero?");
1178 // If the sign bit of the input is known set or clear, then we know the
1179 // top bits of the result.
1181 // If the input sign bit is known zero, or if the NewBits are not demanded
1182 // convert this into a zero extension.
1183 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1185 // Convert to ZExt cast
1186 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1187 return UpdateValueUsesWith(I, NewCast);
1188 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1189 RHSKnownOne |= NewBits;
1193 case Instruction::Add: {
1194 // Figure out what the input bits are. If the top bits of the and result
1195 // are not demanded, then the add doesn't demand them from its input
1197 uint32_t NLZ = DemandedMask.countLeadingZeros();
1199 // If there is a constant on the RHS, there are a variety of xformations
1201 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1202 // If null, this should be simplified elsewhere. Some of the xforms here
1203 // won't work if the RHS is zero.
1207 // If the top bit of the output is demanded, demand everything from the
1208 // input. Otherwise, we demand all the input bits except NLZ top bits.
1209 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1211 // Find information about known zero/one bits in the input.
1212 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1213 LHSKnownZero, LHSKnownOne, Depth+1))
1216 // If the RHS of the add has bits set that can't affect the input, reduce
1218 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1219 return UpdateValueUsesWith(I, I);
1221 // Avoid excess work.
1222 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1225 // Turn it into OR if input bits are zero.
1226 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1228 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1230 InsertNewInstBefore(Or, *I);
1231 return UpdateValueUsesWith(I, Or);
1234 // We can say something about the output known-zero and known-one bits,
1235 // depending on potential carries from the input constant and the
1236 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1237 // bits set and the RHS constant is 0x01001, then we know we have a known
1238 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1240 // To compute this, we first compute the potential carry bits. These are
1241 // the bits which may be modified. I'm not aware of a better way to do
1243 const APInt& RHSVal = RHS->getValue();
1244 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1246 // Now that we know which bits have carries, compute the known-1/0 sets.
1248 // Bits are known one if they are known zero in one operand and one in the
1249 // other, and there is no input carry.
1250 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1251 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1253 // Bits are known zero if they are known zero in both operands and there
1254 // is no input carry.
1255 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1257 // If the high-bits of this ADD are not demanded, then it does not demand
1258 // the high bits of its LHS or RHS.
1259 if (DemandedMask[BitWidth-1] == 0) {
1260 // Right fill the mask of bits for this ADD to demand the most
1261 // significant bit and all those below it.
1262 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1263 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1264 LHSKnownZero, LHSKnownOne, Depth+1))
1266 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1267 LHSKnownZero, LHSKnownOne, Depth+1))
1273 case Instruction::Sub:
1274 // If the high-bits of this SUB are not demanded, then it does not demand
1275 // the high bits of its LHS or RHS.
1276 if (DemandedMask[BitWidth-1] == 0) {
1277 // Right fill the mask of bits for this SUB to demand the most
1278 // significant bit and all those below it.
1279 uint32_t NLZ = DemandedMask.countLeadingZeros();
1280 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1281 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1282 LHSKnownZero, LHSKnownOne, Depth+1))
1284 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1285 LHSKnownZero, LHSKnownOne, Depth+1))
1289 case Instruction::Shl:
1290 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1291 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1292 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1293 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1294 RHSKnownZero, RHSKnownOne, Depth+1))
1296 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1297 "Bits known to be one AND zero?");
1298 RHSKnownZero <<= ShiftAmt;
1299 RHSKnownOne <<= ShiftAmt;
1300 // low bits known zero.
1302 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1305 case Instruction::LShr:
1306 // For a logical shift right
1307 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1308 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1310 // Unsigned shift right.
1311 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1312 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1313 RHSKnownZero, RHSKnownOne, Depth+1))
1315 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1316 "Bits known to be one AND zero?");
1317 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1318 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1320 // Compute the new bits that are at the top now.
1321 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1322 RHSKnownZero |= HighBits; // high bits known zero.
1326 case Instruction::AShr:
1327 // If this is an arithmetic shift right and only the low-bit is set, we can
1328 // always convert this into a logical shr, even if the shift amount is
1329 // variable. The low bit of the shift cannot be an input sign bit unless
1330 // the shift amount is >= the size of the datatype, which is undefined.
1331 if (DemandedMask == 1) {
1332 // Perform the logical shift right.
1333 Value *NewVal = BinaryOperator::createLShr(
1334 I->getOperand(0), I->getOperand(1), I->getName());
1335 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1336 return UpdateValueUsesWith(I, NewVal);
1339 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1340 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1342 // Signed shift right.
1343 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1344 if (SimplifyDemandedBits(I->getOperand(0),
1346 RHSKnownZero, RHSKnownOne, Depth+1))
1348 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1349 "Bits known to be one AND zero?");
1350 // Compute the new bits that are at the top now.
1351 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1352 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1353 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1355 // Handle the sign bits.
1356 APInt SignBit(APInt::getSignBit(BitWidth));
1357 // Adjust to where it is now in the mask.
1358 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1360 // If the input sign bit is known to be zero, or if none of the top bits
1361 // are demanded, turn this into an unsigned shift right.
1362 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1363 (HighBits & ~DemandedMask) == HighBits) {
1364 // Perform the logical shift right.
1365 Value *NewVal = BinaryOperator::createLShr(
1366 I->getOperand(0), SA, I->getName());
1367 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1368 return UpdateValueUsesWith(I, NewVal);
1369 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1370 RHSKnownOne |= HighBits;
1376 // If the client is only demanding bits that we know, return the known
1378 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1379 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1384 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1385 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1386 /// actually used by the caller. This method analyzes which elements of the
1387 /// operand are undef and returns that information in UndefElts.
1389 /// If the information about demanded elements can be used to simplify the
1390 /// operation, the operation is simplified, then the resultant value is
1391 /// returned. This returns null if no change was made.
1392 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1393 uint64_t &UndefElts,
1395 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1396 assert(VWidth <= 64 && "Vector too wide to analyze!");
1397 uint64_t EltMask = ~0ULL >> (64-VWidth);
1398 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1399 "Invalid DemandedElts!");
1401 if (isa<UndefValue>(V)) {
1402 // If the entire vector is undefined, just return this info.
1403 UndefElts = EltMask;
1405 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1406 UndefElts = EltMask;
1407 return UndefValue::get(V->getType());
1411 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1412 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1413 Constant *Undef = UndefValue::get(EltTy);
1415 std::vector<Constant*> Elts;
1416 for (unsigned i = 0; i != VWidth; ++i)
1417 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1418 Elts.push_back(Undef);
1419 UndefElts |= (1ULL << i);
1420 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1421 Elts.push_back(Undef);
1422 UndefElts |= (1ULL << i);
1423 } else { // Otherwise, defined.
1424 Elts.push_back(CP->getOperand(i));
1427 // If we changed the constant, return it.
1428 Constant *NewCP = ConstantVector::get(Elts);
1429 return NewCP != CP ? NewCP : 0;
1430 } else if (isa<ConstantAggregateZero>(V)) {
1431 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1433 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1434 Constant *Zero = Constant::getNullValue(EltTy);
1435 Constant *Undef = UndefValue::get(EltTy);
1436 std::vector<Constant*> Elts;
1437 for (unsigned i = 0; i != VWidth; ++i)
1438 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1439 UndefElts = DemandedElts ^ EltMask;
1440 return ConstantVector::get(Elts);
1443 if (!V->hasOneUse()) { // Other users may use these bits.
1444 if (Depth != 0) { // Not at the root.
1445 // TODO: Just compute the UndefElts information recursively.
1449 } else if (Depth == 10) { // Limit search depth.
1453 Instruction *I = dyn_cast<Instruction>(V);
1454 if (!I) return false; // Only analyze instructions.
1456 bool MadeChange = false;
1457 uint64_t UndefElts2;
1459 switch (I->getOpcode()) {
1462 case Instruction::InsertElement: {
1463 // If this is a variable index, we don't know which element it overwrites.
1464 // demand exactly the same input as we produce.
1465 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1467 // Note that we can't propagate undef elt info, because we don't know
1468 // which elt is getting updated.
1469 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1470 UndefElts2, Depth+1);
1471 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1475 // If this is inserting an element that isn't demanded, remove this
1477 unsigned IdxNo = Idx->getZExtValue();
1478 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1479 return AddSoonDeadInstToWorklist(*I, 0);
1481 // Otherwise, the element inserted overwrites whatever was there, so the
1482 // input demanded set is simpler than the output set.
1483 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1484 DemandedElts & ~(1ULL << IdxNo),
1485 UndefElts, Depth+1);
1486 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1488 // The inserted element is defined.
1489 UndefElts |= 1ULL << IdxNo;
1493 case Instruction::And:
1494 case Instruction::Or:
1495 case Instruction::Xor:
1496 case Instruction::Add:
1497 case Instruction::Sub:
1498 case Instruction::Mul:
1499 // div/rem demand all inputs, because they don't want divide by zero.
1500 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1501 UndefElts, Depth+1);
1502 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1503 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1504 UndefElts2, Depth+1);
1505 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1507 // Output elements are undefined if both are undefined. Consider things
1508 // like undef&0. The result is known zero, not undef.
1509 UndefElts &= UndefElts2;
1512 case Instruction::Call: {
1513 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1515 switch (II->getIntrinsicID()) {
1518 // Binary vector operations that work column-wise. A dest element is a
1519 // function of the corresponding input elements from the two inputs.
1520 case Intrinsic::x86_sse_sub_ss:
1521 case Intrinsic::x86_sse_mul_ss:
1522 case Intrinsic::x86_sse_min_ss:
1523 case Intrinsic::x86_sse_max_ss:
1524 case Intrinsic::x86_sse2_sub_sd:
1525 case Intrinsic::x86_sse2_mul_sd:
1526 case Intrinsic::x86_sse2_min_sd:
1527 case Intrinsic::x86_sse2_max_sd:
1528 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1529 UndefElts, Depth+1);
1530 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1531 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1532 UndefElts2, Depth+1);
1533 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1535 // If only the low elt is demanded and this is a scalarizable intrinsic,
1536 // scalarize it now.
1537 if (DemandedElts == 1) {
1538 switch (II->getIntrinsicID()) {
1540 case Intrinsic::x86_sse_sub_ss:
1541 case Intrinsic::x86_sse_mul_ss:
1542 case Intrinsic::x86_sse2_sub_sd:
1543 case Intrinsic::x86_sse2_mul_sd:
1544 // TODO: Lower MIN/MAX/ABS/etc
1545 Value *LHS = II->getOperand(1);
1546 Value *RHS = II->getOperand(2);
1547 // Extract the element as scalars.
1548 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1549 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1551 switch (II->getIntrinsicID()) {
1552 default: assert(0 && "Case stmts out of sync!");
1553 case Intrinsic::x86_sse_sub_ss:
1554 case Intrinsic::x86_sse2_sub_sd:
1555 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1556 II->getName()), *II);
1558 case Intrinsic::x86_sse_mul_ss:
1559 case Intrinsic::x86_sse2_mul_sd:
1560 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1561 II->getName()), *II);
1566 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1568 InsertNewInstBefore(New, *II);
1569 AddSoonDeadInstToWorklist(*II, 0);
1574 // Output elements are undefined if both are undefined. Consider things
1575 // like undef&0. The result is known zero, not undef.
1576 UndefElts &= UndefElts2;
1582 return MadeChange ? I : 0;
1585 /// @returns true if the specified compare instruction is
1586 /// true when both operands are equal...
1587 /// @brief Determine if the ICmpInst returns true if both operands are equal
1588 static bool isTrueWhenEqual(ICmpInst &ICI) {
1589 ICmpInst::Predicate pred = ICI.getPredicate();
1590 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1591 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1592 pred == ICmpInst::ICMP_SLE;
1595 /// AssociativeOpt - Perform an optimization on an associative operator. This
1596 /// function is designed to check a chain of associative operators for a
1597 /// potential to apply a certain optimization. Since the optimization may be
1598 /// applicable if the expression was reassociated, this checks the chain, then
1599 /// reassociates the expression as necessary to expose the optimization
1600 /// opportunity. This makes use of a special Functor, which must define
1601 /// 'shouldApply' and 'apply' methods.
1603 template<typename Functor>
1604 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1605 unsigned Opcode = Root.getOpcode();
1606 Value *LHS = Root.getOperand(0);
1608 // Quick check, see if the immediate LHS matches...
1609 if (F.shouldApply(LHS))
1610 return F.apply(Root);
1612 // Otherwise, if the LHS is not of the same opcode as the root, return.
1613 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1614 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1615 // Should we apply this transform to the RHS?
1616 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1618 // If not to the RHS, check to see if we should apply to the LHS...
1619 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1620 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1624 // If the functor wants to apply the optimization to the RHS of LHSI,
1625 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1627 BasicBlock *BB = Root.getParent();
1629 // Now all of the instructions are in the current basic block, go ahead
1630 // and perform the reassociation.
1631 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1633 // First move the selected RHS to the LHS of the root...
1634 Root.setOperand(0, LHSI->getOperand(1));
1636 // Make what used to be the LHS of the root be the user of the root...
1637 Value *ExtraOperand = TmpLHSI->getOperand(1);
1638 if (&Root == TmpLHSI) {
1639 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1642 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1643 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1644 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1645 BasicBlock::iterator ARI = &Root; ++ARI;
1646 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1649 // Now propagate the ExtraOperand down the chain of instructions until we
1651 while (TmpLHSI != LHSI) {
1652 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1653 // Move the instruction to immediately before the chain we are
1654 // constructing to avoid breaking dominance properties.
1655 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1656 BB->getInstList().insert(ARI, NextLHSI);
1659 Value *NextOp = NextLHSI->getOperand(1);
1660 NextLHSI->setOperand(1, ExtraOperand);
1662 ExtraOperand = NextOp;
1665 // Now that the instructions are reassociated, have the functor perform
1666 // the transformation...
1667 return F.apply(Root);
1670 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1676 // AddRHS - Implements: X + X --> X << 1
1679 AddRHS(Value *rhs) : RHS(rhs) {}
1680 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1681 Instruction *apply(BinaryOperator &Add) const {
1682 return BinaryOperator::createShl(Add.getOperand(0),
1683 ConstantInt::get(Add.getType(), 1));
1687 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1689 struct AddMaskingAnd {
1691 AddMaskingAnd(Constant *c) : C2(c) {}
1692 bool shouldApply(Value *LHS) const {
1694 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1695 ConstantExpr::getAnd(C1, C2)->isNullValue();
1697 Instruction *apply(BinaryOperator &Add) const {
1698 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1702 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1704 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1705 if (Constant *SOC = dyn_cast<Constant>(SO))
1706 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1708 return IC->InsertNewInstBefore(CastInst::create(
1709 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1712 // Figure out if the constant is the left or the right argument.
1713 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1714 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1716 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1718 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1719 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1722 Value *Op0 = SO, *Op1 = ConstOperand;
1724 std::swap(Op0, Op1);
1726 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1727 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1728 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1729 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1730 SO->getName()+".cmp");
1732 assert(0 && "Unknown binary instruction type!");
1735 return IC->InsertNewInstBefore(New, I);
1738 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1739 // constant as the other operand, try to fold the binary operator into the
1740 // select arguments. This also works for Cast instructions, which obviously do
1741 // not have a second operand.
1742 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1744 // Don't modify shared select instructions
1745 if (!SI->hasOneUse()) return 0;
1746 Value *TV = SI->getOperand(1);
1747 Value *FV = SI->getOperand(2);
1749 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1750 // Bool selects with constant operands can be folded to logical ops.
1751 if (SI->getType() == Type::Int1Ty) return 0;
1753 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1754 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1756 return new SelectInst(SI->getCondition(), SelectTrueVal,
1763 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1764 /// node as operand #0, see if we can fold the instruction into the PHI (which
1765 /// is only possible if all operands to the PHI are constants).
1766 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1767 PHINode *PN = cast<PHINode>(I.getOperand(0));
1768 unsigned NumPHIValues = PN->getNumIncomingValues();
1769 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1771 // Check to see if all of the operands of the PHI are constants. If there is
1772 // one non-constant value, remember the BB it is. If there is more than one
1773 // or if *it* is a PHI, bail out.
1774 BasicBlock *NonConstBB = 0;
1775 for (unsigned i = 0; i != NumPHIValues; ++i)
1776 if (!isa<Constant>(PN->getIncomingValue(i))) {
1777 if (NonConstBB) return 0; // More than one non-const value.
1778 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1779 NonConstBB = PN->getIncomingBlock(i);
1781 // If the incoming non-constant value is in I's block, we have an infinite
1783 if (NonConstBB == I.getParent())
1787 // If there is exactly one non-constant value, we can insert a copy of the
1788 // operation in that block. However, if this is a critical edge, we would be
1789 // inserting the computation one some other paths (e.g. inside a loop). Only
1790 // do this if the pred block is unconditionally branching into the phi block.
1792 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1793 if (!BI || !BI->isUnconditional()) return 0;
1796 // Okay, we can do the transformation: create the new PHI node.
1797 PHINode *NewPN = new PHINode(I.getType(), "");
1798 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1799 InsertNewInstBefore(NewPN, *PN);
1800 NewPN->takeName(PN);
1802 // Next, add all of the operands to the PHI.
1803 if (I.getNumOperands() == 2) {
1804 Constant *C = cast<Constant>(I.getOperand(1));
1805 for (unsigned i = 0; i != NumPHIValues; ++i) {
1807 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1808 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1809 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1811 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1813 assert(PN->getIncomingBlock(i) == NonConstBB);
1814 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1815 InV = BinaryOperator::create(BO->getOpcode(),
1816 PN->getIncomingValue(i), C, "phitmp",
1817 NonConstBB->getTerminator());
1818 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1819 InV = CmpInst::create(CI->getOpcode(),
1821 PN->getIncomingValue(i), C, "phitmp",
1822 NonConstBB->getTerminator());
1824 assert(0 && "Unknown binop!");
1826 AddToWorkList(cast<Instruction>(InV));
1828 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1831 CastInst *CI = cast<CastInst>(&I);
1832 const Type *RetTy = CI->getType();
1833 for (unsigned i = 0; i != NumPHIValues; ++i) {
1835 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1836 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1838 assert(PN->getIncomingBlock(i) == NonConstBB);
1839 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1840 I.getType(), "phitmp",
1841 NonConstBB->getTerminator());
1842 AddToWorkList(cast<Instruction>(InV));
1844 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1847 return ReplaceInstUsesWith(I, NewPN);
1850 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1851 bool Changed = SimplifyCommutative(I);
1852 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1854 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1855 // X + undef -> undef
1856 if (isa<UndefValue>(RHS))
1857 return ReplaceInstUsesWith(I, RHS);
1860 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1861 if (RHSC->isNullValue())
1862 return ReplaceInstUsesWith(I, LHS);
1863 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1864 if (CFP->isExactlyValue(-0.0))
1865 return ReplaceInstUsesWith(I, LHS);
1868 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1869 // X + (signbit) --> X ^ signbit
1870 const APInt& Val = CI->getValue();
1871 uint32_t BitWidth = Val.getBitWidth();
1872 if (Val == APInt::getSignBit(BitWidth))
1873 return BinaryOperator::createXor(LHS, RHS);
1875 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1876 // (X & 254)+1 -> (X&254)|1
1877 if (!isa<VectorType>(I.getType())) {
1878 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1879 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1880 KnownZero, KnownOne))
1885 if (isa<PHINode>(LHS))
1886 if (Instruction *NV = FoldOpIntoPhi(I))
1889 ConstantInt *XorRHS = 0;
1891 if (isa<ConstantInt>(RHSC) &&
1892 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1893 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1894 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1896 uint32_t Size = TySizeBits / 2;
1897 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1898 APInt CFF80Val(-C0080Val);
1900 if (TySizeBits > Size) {
1901 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1902 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1903 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1904 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1905 // This is a sign extend if the top bits are known zero.
1906 if (!MaskedValueIsZero(XorLHS,
1907 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1908 Size = 0; // Not a sign ext, but can't be any others either.
1913 C0080Val = APIntOps::lshr(C0080Val, Size);
1914 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1915 } while (Size >= 1);
1917 // FIXME: This shouldn't be necessary. When the backends can handle types
1918 // with funny bit widths then this whole cascade of if statements should
1919 // be removed. It is just here to get the size of the "middle" type back
1920 // up to something that the back ends can handle.
1921 const Type *MiddleType = 0;
1924 case 32: MiddleType = Type::Int32Ty; break;
1925 case 16: MiddleType = Type::Int16Ty; break;
1926 case 8: MiddleType = Type::Int8Ty; break;
1929 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1930 InsertNewInstBefore(NewTrunc, I);
1931 return new SExtInst(NewTrunc, I.getType(), I.getName());
1937 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
1938 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1940 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1941 if (RHSI->getOpcode() == Instruction::Sub)
1942 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1943 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1945 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1946 if (LHSI->getOpcode() == Instruction::Sub)
1947 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1948 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1953 if (Value *V = dyn_castNegVal(LHS))
1954 return BinaryOperator::createSub(RHS, V);
1957 if (!isa<Constant>(RHS))
1958 if (Value *V = dyn_castNegVal(RHS))
1959 return BinaryOperator::createSub(LHS, V);
1963 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1964 if (X == RHS) // X*C + X --> X * (C+1)
1965 return BinaryOperator::createMul(RHS, AddOne(C2));
1967 // X*C1 + X*C2 --> X * (C1+C2)
1969 if (X == dyn_castFoldableMul(RHS, C1))
1970 return BinaryOperator::createMul(X, Add(C1, C2));
1973 // X + X*C --> X * (C+1)
1974 if (dyn_castFoldableMul(RHS, C2) == LHS)
1975 return BinaryOperator::createMul(LHS, AddOne(C2));
1977 // X + ~X --> -1 since ~X = -X-1
1978 if (dyn_castNotVal(LHS) == RHS ||
1979 dyn_castNotVal(RHS) == LHS)
1980 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
1983 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1984 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1985 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
1988 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1990 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
1991 return BinaryOperator::createSub(SubOne(CRHS), X);
1993 // (X & FF00) + xx00 -> (X+xx00) & FF00
1994 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1995 Constant *Anded = And(CRHS, C2);
1996 if (Anded == CRHS) {
1997 // See if all bits from the first bit set in the Add RHS up are included
1998 // in the mask. First, get the rightmost bit.
1999 const APInt& AddRHSV = CRHS->getValue();
2001 // Form a mask of all bits from the lowest bit added through the top.
2002 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2004 // See if the and mask includes all of these bits.
2005 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2007 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2008 // Okay, the xform is safe. Insert the new add pronto.
2009 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2010 LHS->getName()), I);
2011 return BinaryOperator::createAnd(NewAdd, C2);
2016 // Try to fold constant add into select arguments.
2017 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2018 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2022 // add (cast *A to intptrtype) B ->
2023 // cast (GEP (cast *A to sbyte*) B) ->
2026 CastInst *CI = dyn_cast<CastInst>(LHS);
2029 CI = dyn_cast<CastInst>(RHS);
2032 if (CI && CI->getType()->isSized() &&
2033 (CI->getType()->getPrimitiveSizeInBits() ==
2034 TD->getIntPtrType()->getPrimitiveSizeInBits())
2035 && isa<PointerType>(CI->getOperand(0)->getType())) {
2036 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2037 PointerType::get(Type::Int8Ty), I);
2038 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2039 return new PtrToIntInst(I2, CI->getType());
2043 return Changed ? &I : 0;
2046 // isSignBit - Return true if the value represented by the constant only has the
2047 // highest order bit set.
2048 static bool isSignBit(ConstantInt *CI) {
2049 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2050 return CI->getValue() == APInt::getSignBit(NumBits);
2053 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2054 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2056 if (Op0 == Op1) // sub X, X -> 0
2057 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2059 // If this is a 'B = x-(-A)', change to B = x+A...
2060 if (Value *V = dyn_castNegVal(Op1))
2061 return BinaryOperator::createAdd(Op0, V);
2063 if (isa<UndefValue>(Op0))
2064 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2065 if (isa<UndefValue>(Op1))
2066 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2068 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2069 // Replace (-1 - A) with (~A)...
2070 if (C->isAllOnesValue())
2071 return BinaryOperator::createNot(Op1);
2073 // C - ~X == X + (1+C)
2075 if (match(Op1, m_Not(m_Value(X))))
2076 return BinaryOperator::createAdd(X, AddOne(C));
2078 // -(X >>u 31) -> (X >>s 31)
2079 // -(X >>s 31) -> (X >>u 31)
2081 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2082 if (SI->getOpcode() == Instruction::LShr) {
2083 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2084 // Check to see if we are shifting out everything but the sign bit.
2085 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2086 SI->getType()->getPrimitiveSizeInBits()-1) {
2087 // Ok, the transformation is safe. Insert AShr.
2088 return BinaryOperator::create(Instruction::AShr,
2089 SI->getOperand(0), CU, SI->getName());
2093 else if (SI->getOpcode() == Instruction::AShr) {
2094 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2095 // Check to see if we are shifting out everything but the sign bit.
2096 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2097 SI->getType()->getPrimitiveSizeInBits()-1) {
2098 // Ok, the transformation is safe. Insert LShr.
2099 return BinaryOperator::createLShr(
2100 SI->getOperand(0), CU, SI->getName());
2106 // Try to fold constant sub into select arguments.
2107 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2108 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2111 if (isa<PHINode>(Op0))
2112 if (Instruction *NV = FoldOpIntoPhi(I))
2116 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2117 if (Op1I->getOpcode() == Instruction::Add &&
2118 !Op0->getType()->isFPOrFPVector()) {
2119 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2120 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2121 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2122 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2123 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2124 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2125 // C1-(X+C2) --> (C1-C2)-X
2126 return BinaryOperator::createSub(Subtract(CI1, CI2),
2127 Op1I->getOperand(0));
2131 if (Op1I->hasOneUse()) {
2132 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2133 // is not used by anyone else...
2135 if (Op1I->getOpcode() == Instruction::Sub &&
2136 !Op1I->getType()->isFPOrFPVector()) {
2137 // Swap the two operands of the subexpr...
2138 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2139 Op1I->setOperand(0, IIOp1);
2140 Op1I->setOperand(1, IIOp0);
2142 // Create the new top level add instruction...
2143 return BinaryOperator::createAdd(Op0, Op1);
2146 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2148 if (Op1I->getOpcode() == Instruction::And &&
2149 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2150 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2153 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2154 return BinaryOperator::createAnd(Op0, NewNot);
2157 // 0 - (X sdiv C) -> (X sdiv -C)
2158 if (Op1I->getOpcode() == Instruction::SDiv)
2159 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2160 if (CSI->isNullValue())
2161 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2162 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2163 ConstantExpr::getNeg(DivRHS));
2165 // X - X*C --> X * (1-C)
2166 ConstantInt *C2 = 0;
2167 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2168 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2169 return BinaryOperator::createMul(Op0, CP1);
2174 if (!Op0->getType()->isFPOrFPVector())
2175 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2176 if (Op0I->getOpcode() == Instruction::Add) {
2177 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2178 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2179 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2180 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2181 } else if (Op0I->getOpcode() == Instruction::Sub) {
2182 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2183 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2187 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2188 if (X == Op1) // X*C - X --> X * (C-1)
2189 return BinaryOperator::createMul(Op1, SubOne(C1));
2191 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2192 if (X == dyn_castFoldableMul(Op1, C2))
2193 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2198 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2199 /// really just returns true if the most significant (sign) bit is set.
2200 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2202 case ICmpInst::ICMP_SLT:
2203 // True if LHS s< RHS and RHS == 0
2204 return RHS->isNullValue();
2205 case ICmpInst::ICMP_SLE:
2206 // True if LHS s<= RHS and RHS == -1
2207 return RHS->isAllOnesValue();
2208 case ICmpInst::ICMP_UGE:
2209 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2210 return RHS->getValue() ==
2211 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2212 case ICmpInst::ICMP_UGT:
2213 // True if LHS u> RHS and RHS == high-bit-mask - 1
2214 return RHS->getValue() ==
2215 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2221 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2222 bool Changed = SimplifyCommutative(I);
2223 Value *Op0 = I.getOperand(0);
2225 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2226 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2228 // Simplify mul instructions with a constant RHS...
2229 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2230 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2232 // ((X << C1)*C2) == (X * (C2 << C1))
2233 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2234 if (SI->getOpcode() == Instruction::Shl)
2235 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2236 return BinaryOperator::createMul(SI->getOperand(0),
2237 ConstantExpr::getShl(CI, ShOp));
2239 if (CI->isNullValue())
2240 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2241 if (CI->equalsInt(1)) // X * 1 == X
2242 return ReplaceInstUsesWith(I, Op0);
2243 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2244 return BinaryOperator::createNeg(Op0, I.getName());
2246 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2247 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2248 return BinaryOperator::createShl(Op0,
2249 ConstantInt::get(Op0->getType(), Val.logBase2()));
2251 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2252 if (Op1F->isNullValue())
2253 return ReplaceInstUsesWith(I, Op1);
2255 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2256 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2257 if (Op1F->getValue() == 1.0)
2258 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2261 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2262 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2263 isa<ConstantInt>(Op0I->getOperand(1))) {
2264 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2265 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2267 InsertNewInstBefore(Add, I);
2268 Value *C1C2 = ConstantExpr::getMul(Op1,
2269 cast<Constant>(Op0I->getOperand(1)));
2270 return BinaryOperator::createAdd(Add, C1C2);
2274 // Try to fold constant mul into select arguments.
2275 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2276 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2279 if (isa<PHINode>(Op0))
2280 if (Instruction *NV = FoldOpIntoPhi(I))
2284 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2285 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2286 return BinaryOperator::createMul(Op0v, Op1v);
2288 // If one of the operands of the multiply is a cast from a boolean value, then
2289 // we know the bool is either zero or one, so this is a 'masking' multiply.
2290 // See if we can simplify things based on how the boolean was originally
2292 CastInst *BoolCast = 0;
2293 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2294 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2297 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2298 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2301 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2302 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2303 const Type *SCOpTy = SCIOp0->getType();
2305 // If the icmp is true iff the sign bit of X is set, then convert this
2306 // multiply into a shift/and combination.
2307 if (isa<ConstantInt>(SCIOp1) &&
2308 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2309 // Shift the X value right to turn it into "all signbits".
2310 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2311 SCOpTy->getPrimitiveSizeInBits()-1);
2313 InsertNewInstBefore(
2314 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2315 BoolCast->getOperand(0)->getName()+
2318 // If the multiply type is not the same as the source type, sign extend
2319 // or truncate to the multiply type.
2320 if (I.getType() != V->getType()) {
2321 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2322 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2323 Instruction::CastOps opcode =
2324 (SrcBits == DstBits ? Instruction::BitCast :
2325 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2326 V = InsertCastBefore(opcode, V, I.getType(), I);
2329 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2330 return BinaryOperator::createAnd(V, OtherOp);
2335 return Changed ? &I : 0;
2338 /// This function implements the transforms on div instructions that work
2339 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2340 /// used by the visitors to those instructions.
2341 /// @brief Transforms common to all three div instructions
2342 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2343 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2346 if (isa<UndefValue>(Op0))
2347 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2349 // X / undef -> undef
2350 if (isa<UndefValue>(Op1))
2351 return ReplaceInstUsesWith(I, Op1);
2353 // Handle cases involving: div X, (select Cond, Y, Z)
2354 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2355 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2356 // same basic block, then we replace the select with Y, and the condition
2357 // of the select with false (if the cond value is in the same BB). If the
2358 // select has uses other than the div, this allows them to be simplified
2359 // also. Note that div X, Y is just as good as div X, 0 (undef)
2360 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2361 if (ST->isNullValue()) {
2362 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2363 if (CondI && CondI->getParent() == I.getParent())
2364 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2365 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2366 I.setOperand(1, SI->getOperand(2));
2368 UpdateValueUsesWith(SI, SI->getOperand(2));
2372 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2373 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2374 if (ST->isNullValue()) {
2375 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2376 if (CondI && CondI->getParent() == I.getParent())
2377 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2378 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2379 I.setOperand(1, SI->getOperand(1));
2381 UpdateValueUsesWith(SI, SI->getOperand(1));
2389 /// This function implements the transforms common to both integer division
2390 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2391 /// division instructions.
2392 /// @brief Common integer divide transforms
2393 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2394 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2396 if (Instruction *Common = commonDivTransforms(I))
2399 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2401 if (RHS->equalsInt(1))
2402 return ReplaceInstUsesWith(I, Op0);
2404 // (X / C1) / C2 -> X / (C1*C2)
2405 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2406 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2407 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2408 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2409 Multiply(RHS, LHSRHS));
2412 if (!RHS->isZero()) { // avoid X udiv 0
2413 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2414 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2416 if (isa<PHINode>(Op0))
2417 if (Instruction *NV = FoldOpIntoPhi(I))
2422 // 0 / X == 0, we don't need to preserve faults!
2423 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2424 if (LHS->equalsInt(0))
2425 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2430 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2431 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2433 // Handle the integer div common cases
2434 if (Instruction *Common = commonIDivTransforms(I))
2437 // X udiv C^2 -> X >> C
2438 // Check to see if this is an unsigned division with an exact power of 2,
2439 // if so, convert to a right shift.
2440 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2441 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2442 return BinaryOperator::createLShr(Op0,
2443 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2446 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2447 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2448 if (RHSI->getOpcode() == Instruction::Shl &&
2449 isa<ConstantInt>(RHSI->getOperand(0))) {
2450 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2451 if (C1.isPowerOf2()) {
2452 Value *N = RHSI->getOperand(1);
2453 const Type *NTy = N->getType();
2454 if (uint32_t C2 = C1.logBase2()) {
2455 Constant *C2V = ConstantInt::get(NTy, C2);
2456 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2458 return BinaryOperator::createLShr(Op0, N);
2463 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2464 // where C1&C2 are powers of two.
2465 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2466 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2467 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2468 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2469 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2470 // Compute the shift amounts
2471 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2472 // Construct the "on true" case of the select
2473 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2474 Instruction *TSI = BinaryOperator::createLShr(
2475 Op0, TC, SI->getName()+".t");
2476 TSI = InsertNewInstBefore(TSI, I);
2478 // Construct the "on false" case of the select
2479 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2480 Instruction *FSI = BinaryOperator::createLShr(
2481 Op0, FC, SI->getName()+".f");
2482 FSI = InsertNewInstBefore(FSI, I);
2484 // construct the select instruction and return it.
2485 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2491 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2492 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2494 // Handle the integer div common cases
2495 if (Instruction *Common = commonIDivTransforms(I))
2498 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2500 if (RHS->isAllOnesValue())
2501 return BinaryOperator::createNeg(Op0);
2504 if (Value *LHSNeg = dyn_castNegVal(Op0))
2505 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2508 // If the sign bits of both operands are zero (i.e. we can prove they are
2509 // unsigned inputs), turn this into a udiv.
2510 if (I.getType()->isInteger()) {
2511 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2512 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2513 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2520 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2521 return commonDivTransforms(I);
2524 /// GetFactor - If we can prove that the specified value is at least a multiple
2525 /// of some factor, return that factor.
2526 static Constant *GetFactor(Value *V) {
2527 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2530 // Unless we can be tricky, we know this is a multiple of 1.
2531 Constant *Result = ConstantInt::get(V->getType(), 1);
2533 Instruction *I = dyn_cast<Instruction>(V);
2534 if (!I) return Result;
2536 if (I->getOpcode() == Instruction::Mul) {
2537 // Handle multiplies by a constant, etc.
2538 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2539 GetFactor(I->getOperand(1)));
2540 } else if (I->getOpcode() == Instruction::Shl) {
2541 // (X<<C) -> X * (1 << C)
2542 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2543 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2544 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2546 } else if (I->getOpcode() == Instruction::And) {
2547 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2548 // X & 0xFFF0 is known to be a multiple of 16.
2549 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2550 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2551 return ConstantExpr::getShl(Result,
2552 ConstantInt::get(Result->getType(), Zeros));
2554 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2555 // Only handle int->int casts.
2556 if (!CI->isIntegerCast())
2558 Value *Op = CI->getOperand(0);
2559 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2564 /// This function implements the transforms on rem instructions that work
2565 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2566 /// is used by the visitors to those instructions.
2567 /// @brief Transforms common to all three rem instructions
2568 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2569 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2571 // 0 % X == 0, we don't need to preserve faults!
2572 if (Constant *LHS = dyn_cast<Constant>(Op0))
2573 if (LHS->isNullValue())
2574 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2576 if (isa<UndefValue>(Op0)) // undef % X -> 0
2577 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2578 if (isa<UndefValue>(Op1))
2579 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2581 // Handle cases involving: rem X, (select Cond, Y, Z)
2582 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2583 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2584 // the same basic block, then we replace the select with Y, and the
2585 // condition of the select with false (if the cond value is in the same
2586 // BB). If the select has uses other than the div, this allows them to be
2588 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2589 if (ST->isNullValue()) {
2590 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2591 if (CondI && CondI->getParent() == I.getParent())
2592 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2593 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2594 I.setOperand(1, SI->getOperand(2));
2596 UpdateValueUsesWith(SI, SI->getOperand(2));
2599 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2600 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2601 if (ST->isNullValue()) {
2602 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2603 if (CondI && CondI->getParent() == I.getParent())
2604 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2605 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2606 I.setOperand(1, SI->getOperand(1));
2608 UpdateValueUsesWith(SI, SI->getOperand(1));
2616 /// This function implements the transforms common to both integer remainder
2617 /// instructions (urem and srem). It is called by the visitors to those integer
2618 /// remainder instructions.
2619 /// @brief Common integer remainder transforms
2620 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2621 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2623 if (Instruction *common = commonRemTransforms(I))
2626 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2627 // X % 0 == undef, we don't need to preserve faults!
2628 if (RHS->equalsInt(0))
2629 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2631 if (RHS->equalsInt(1)) // X % 1 == 0
2632 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2634 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2635 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2636 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2638 } else if (isa<PHINode>(Op0I)) {
2639 if (Instruction *NV = FoldOpIntoPhi(I))
2642 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2643 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2644 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2651 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2652 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2654 if (Instruction *common = commonIRemTransforms(I))
2657 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2658 // X urem C^2 -> X and C
2659 // Check to see if this is an unsigned remainder with an exact power of 2,
2660 // if so, convert to a bitwise and.
2661 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2662 if (C->getValue().isPowerOf2())
2663 return BinaryOperator::createAnd(Op0, SubOne(C));
2666 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2667 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2668 if (RHSI->getOpcode() == Instruction::Shl &&
2669 isa<ConstantInt>(RHSI->getOperand(0))) {
2670 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2671 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2672 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2674 return BinaryOperator::createAnd(Op0, Add);
2679 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2680 // where C1&C2 are powers of two.
2681 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2682 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2683 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2684 // STO == 0 and SFO == 0 handled above.
2685 if ((STO->getValue().isPowerOf2()) &&
2686 (SFO->getValue().isPowerOf2())) {
2687 Value *TrueAnd = InsertNewInstBefore(
2688 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2689 Value *FalseAnd = InsertNewInstBefore(
2690 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2691 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2699 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2700 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2702 if (Instruction *common = commonIRemTransforms(I))
2705 if (Value *RHSNeg = dyn_castNegVal(Op1))
2706 if (!isa<ConstantInt>(RHSNeg) ||
2707 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2709 AddUsesToWorkList(I);
2710 I.setOperand(1, RHSNeg);
2714 // If the top bits of both operands are zero (i.e. we can prove they are
2715 // unsigned inputs), turn this into a urem.
2716 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2717 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2718 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2719 return BinaryOperator::createURem(Op0, Op1, I.getName());
2725 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2726 return commonRemTransforms(I);
2729 // isMaxValueMinusOne - return true if this is Max-1
2730 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2731 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2733 // Calculate 0111111111..11111
2734 APInt Val(APInt::getSignedMaxValue(TypeBits));
2735 return C->getValue() == Val-1;
2737 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2740 // isMinValuePlusOne - return true if this is Min+1
2741 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2743 // Calculate 1111111111000000000000
2744 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2745 APInt Val(APInt::getSignedMinValue(TypeBits));
2746 return C->getValue() == Val+1;
2748 return C->getValue() == 1; // unsigned
2751 // isOneBitSet - Return true if there is exactly one bit set in the specified
2753 static bool isOneBitSet(const ConstantInt *CI) {
2754 return CI->getValue().isPowerOf2();
2757 // isHighOnes - Return true if the constant is of the form 1+0+.
2758 // This is the same as lowones(~X).
2759 static bool isHighOnes(const ConstantInt *CI) {
2760 return (~CI->getValue() + 1).isPowerOf2();
2763 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2764 /// are carefully arranged to allow folding of expressions such as:
2766 /// (A < B) | (A > B) --> (A != B)
2768 /// Note that this is only valid if the first and second predicates have the
2769 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2771 /// Three bits are used to represent the condition, as follows:
2776 /// <=> Value Definition
2777 /// 000 0 Always false
2784 /// 111 7 Always true
2786 static unsigned getICmpCode(const ICmpInst *ICI) {
2787 switch (ICI->getPredicate()) {
2789 case ICmpInst::ICMP_UGT: return 1; // 001
2790 case ICmpInst::ICMP_SGT: return 1; // 001
2791 case ICmpInst::ICMP_EQ: return 2; // 010
2792 case ICmpInst::ICMP_UGE: return 3; // 011
2793 case ICmpInst::ICMP_SGE: return 3; // 011
2794 case ICmpInst::ICMP_ULT: return 4; // 100
2795 case ICmpInst::ICMP_SLT: return 4; // 100
2796 case ICmpInst::ICMP_NE: return 5; // 101
2797 case ICmpInst::ICMP_ULE: return 6; // 110
2798 case ICmpInst::ICMP_SLE: return 6; // 110
2801 assert(0 && "Invalid ICmp predicate!");
2806 /// getICmpValue - This is the complement of getICmpCode, which turns an
2807 /// opcode and two operands into either a constant true or false, or a brand
2808 /// new /// ICmp instruction. The sign is passed in to determine which kind
2809 /// of predicate to use in new icmp instructions.
2810 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2812 default: assert(0 && "Illegal ICmp code!");
2813 case 0: return ConstantInt::getFalse();
2816 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2818 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2819 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2822 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2824 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2827 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2829 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2830 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2833 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2835 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2836 case 7: return ConstantInt::getTrue();
2840 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2841 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2842 (ICmpInst::isSignedPredicate(p1) &&
2843 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2844 (ICmpInst::isSignedPredicate(p2) &&
2845 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2849 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2850 struct FoldICmpLogical {
2853 ICmpInst::Predicate pred;
2854 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2855 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2856 pred(ICI->getPredicate()) {}
2857 bool shouldApply(Value *V) const {
2858 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2859 if (PredicatesFoldable(pred, ICI->getPredicate()))
2860 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2861 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2864 Instruction *apply(Instruction &Log) const {
2865 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2866 if (ICI->getOperand(0) != LHS) {
2867 assert(ICI->getOperand(1) == LHS);
2868 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2871 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2872 unsigned LHSCode = getICmpCode(ICI);
2873 unsigned RHSCode = getICmpCode(RHSICI);
2875 switch (Log.getOpcode()) {
2876 case Instruction::And: Code = LHSCode & RHSCode; break;
2877 case Instruction::Or: Code = LHSCode | RHSCode; break;
2878 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2879 default: assert(0 && "Illegal logical opcode!"); return 0;
2882 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2883 ICmpInst::isSignedPredicate(ICI->getPredicate());
2885 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2886 if (Instruction *I = dyn_cast<Instruction>(RV))
2888 // Otherwise, it's a constant boolean value...
2889 return IC.ReplaceInstUsesWith(Log, RV);
2892 } // end anonymous namespace
2894 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2895 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2896 // guaranteed to be a binary operator.
2897 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2899 ConstantInt *AndRHS,
2900 BinaryOperator &TheAnd) {
2901 Value *X = Op->getOperand(0);
2902 Constant *Together = 0;
2904 Together = And(AndRHS, OpRHS);
2906 switch (Op->getOpcode()) {
2907 case Instruction::Xor:
2908 if (Op->hasOneUse()) {
2909 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2910 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2911 InsertNewInstBefore(And, TheAnd);
2913 return BinaryOperator::createXor(And, Together);
2916 case Instruction::Or:
2917 if (Together == AndRHS) // (X | C) & C --> C
2918 return ReplaceInstUsesWith(TheAnd, AndRHS);
2920 if (Op->hasOneUse() && Together != OpRHS) {
2921 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2922 Instruction *Or = BinaryOperator::createOr(X, Together);
2923 InsertNewInstBefore(Or, TheAnd);
2925 return BinaryOperator::createAnd(Or, AndRHS);
2928 case Instruction::Add:
2929 if (Op->hasOneUse()) {
2930 // Adding a one to a single bit bit-field should be turned into an XOR
2931 // of the bit. First thing to check is to see if this AND is with a
2932 // single bit constant.
2933 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
2935 // If there is only one bit set...
2936 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2937 // Ok, at this point, we know that we are masking the result of the
2938 // ADD down to exactly one bit. If the constant we are adding has
2939 // no bits set below this bit, then we can eliminate the ADD.
2940 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
2942 // Check to see if any bits below the one bit set in AndRHSV are set.
2943 if ((AddRHS & (AndRHSV-1)) == 0) {
2944 // If not, the only thing that can effect the output of the AND is
2945 // the bit specified by AndRHSV. If that bit is set, the effect of
2946 // the XOR is to toggle the bit. If it is clear, then the ADD has
2948 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2949 TheAnd.setOperand(0, X);
2952 // Pull the XOR out of the AND.
2953 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
2954 InsertNewInstBefore(NewAnd, TheAnd);
2955 NewAnd->takeName(Op);
2956 return BinaryOperator::createXor(NewAnd, AndRHS);
2963 case Instruction::Shl: {
2964 // We know that the AND will not produce any of the bits shifted in, so if
2965 // the anded constant includes them, clear them now!
2967 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
2968 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
2969 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
2970 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
2972 if (CI->getValue() == ShlMask) {
2973 // Masking out bits that the shift already masks
2974 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2975 } else if (CI != AndRHS) { // Reducing bits set in and.
2976 TheAnd.setOperand(1, CI);
2981 case Instruction::LShr:
2983 // We know that the AND will not produce any of the bits shifted in, so if
2984 // the anded constant includes them, clear them now! This only applies to
2985 // unsigned shifts, because a signed shr may bring in set bits!
2987 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
2988 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
2989 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
2990 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
2992 if (CI->getValue() == ShrMask) {
2993 // Masking out bits that the shift already masks.
2994 return ReplaceInstUsesWith(TheAnd, Op);
2995 } else if (CI != AndRHS) {
2996 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3001 case Instruction::AShr:
3003 // See if this is shifting in some sign extension, then masking it out
3005 if (Op->hasOneUse()) {
3006 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3007 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3008 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3009 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3010 if (C == AndRHS) { // Masking out bits shifted in.
3011 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3012 // Make the argument unsigned.
3013 Value *ShVal = Op->getOperand(0);
3014 ShVal = InsertNewInstBefore(
3015 BinaryOperator::createLShr(ShVal, OpRHS,
3016 Op->getName()), TheAnd);
3017 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3026 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3027 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3028 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3029 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3030 /// insert new instructions.
3031 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3032 bool isSigned, bool Inside,
3034 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3035 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3036 "Lo is not <= Hi in range emission code!");
3039 if (Lo == Hi) // Trivially false.
3040 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3042 // V >= Min && V < Hi --> V < Hi
3043 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3044 ICmpInst::Predicate pred = (isSigned ?
3045 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3046 return new ICmpInst(pred, V, Hi);
3049 // Emit V-Lo <u Hi-Lo
3050 Constant *NegLo = ConstantExpr::getNeg(Lo);
3051 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3052 InsertNewInstBefore(Add, IB);
3053 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3054 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3057 if (Lo == Hi) // Trivially true.
3058 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3060 // V < Min || V >= Hi -> V > Hi-1
3061 Hi = SubOne(cast<ConstantInt>(Hi));
3062 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3063 ICmpInst::Predicate pred = (isSigned ?
3064 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3065 return new ICmpInst(pred, V, Hi);
3068 // Emit V-Lo >u Hi-1-Lo
3069 // Note that Hi has already had one subtracted from it, above.
3070 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3071 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3072 InsertNewInstBefore(Add, IB);
3073 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3074 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3077 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3078 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3079 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3080 // not, since all 1s are not contiguous.
3081 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3082 const APInt& V = Val->getValue();
3083 uint32_t BitWidth = Val->getType()->getBitWidth();
3084 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3086 // look for the first zero bit after the run of ones
3087 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3088 // look for the first non-zero bit
3089 ME = V.getActiveBits();
3093 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3094 /// where isSub determines whether the operator is a sub. If we can fold one of
3095 /// the following xforms:
3097 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3098 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3099 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3101 /// return (A +/- B).
3103 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3104 ConstantInt *Mask, bool isSub,
3106 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3107 if (!LHSI || LHSI->getNumOperands() != 2 ||
3108 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3110 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3112 switch (LHSI->getOpcode()) {
3114 case Instruction::And:
3115 if (And(N, Mask) == Mask) {
3116 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3117 if ((Mask->getValue().countLeadingZeros() +
3118 Mask->getValue().countPopulation()) ==
3119 Mask->getValue().getBitWidth())
3122 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3123 // part, we don't need any explicit masks to take them out of A. If that
3124 // is all N is, ignore it.
3125 uint32_t MB = 0, ME = 0;
3126 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3127 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3128 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3129 if (MaskedValueIsZero(RHS, Mask))
3134 case Instruction::Or:
3135 case Instruction::Xor:
3136 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3137 if ((Mask->getValue().countLeadingZeros() +
3138 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3139 && And(N, Mask)->isZero())
3146 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3148 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3149 return InsertNewInstBefore(New, I);
3152 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3153 bool Changed = SimplifyCommutative(I);
3154 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3156 if (isa<UndefValue>(Op1)) // X & undef -> 0
3157 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3161 return ReplaceInstUsesWith(I, Op1);
3163 // See if we can simplify any instructions used by the instruction whose sole
3164 // purpose is to compute bits we don't care about.
3165 if (!isa<VectorType>(I.getType())) {
3166 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3167 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3168 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3169 KnownZero, KnownOne))
3172 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3173 if (CP->isAllOnesValue())
3174 return ReplaceInstUsesWith(I, I.getOperand(0));
3178 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3179 const APInt& AndRHSMask = AndRHS->getValue();
3180 APInt NotAndRHS(~AndRHSMask);
3182 // Optimize a variety of ((val OP C1) & C2) combinations...
3183 if (isa<BinaryOperator>(Op0)) {
3184 Instruction *Op0I = cast<Instruction>(Op0);
3185 Value *Op0LHS = Op0I->getOperand(0);
3186 Value *Op0RHS = Op0I->getOperand(1);
3187 switch (Op0I->getOpcode()) {
3188 case Instruction::Xor:
3189 case Instruction::Or:
3190 // If the mask is only needed on one incoming arm, push it up.
3191 if (Op0I->hasOneUse()) {
3192 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3193 // Not masking anything out for the LHS, move to RHS.
3194 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3195 Op0RHS->getName()+".masked");
3196 InsertNewInstBefore(NewRHS, I);
3197 return BinaryOperator::create(
3198 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3200 if (!isa<Constant>(Op0RHS) &&
3201 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3202 // Not masking anything out for the RHS, move to LHS.
3203 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3204 Op0LHS->getName()+".masked");
3205 InsertNewInstBefore(NewLHS, I);
3206 return BinaryOperator::create(
3207 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3212 case Instruction::Add:
3213 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3214 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3215 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3216 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3217 return BinaryOperator::createAnd(V, AndRHS);
3218 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3219 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3222 case Instruction::Sub:
3223 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3224 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3225 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3226 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3227 return BinaryOperator::createAnd(V, AndRHS);
3231 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3232 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3234 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3235 // If this is an integer truncation or change from signed-to-unsigned, and
3236 // if the source is an and/or with immediate, transform it. This
3237 // frequently occurs for bitfield accesses.
3238 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3239 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3240 CastOp->getNumOperands() == 2)
3241 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3242 if (CastOp->getOpcode() == Instruction::And) {
3243 // Change: and (cast (and X, C1) to T), C2
3244 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3245 // This will fold the two constants together, which may allow
3246 // other simplifications.
3247 Instruction *NewCast = CastInst::createTruncOrBitCast(
3248 CastOp->getOperand(0), I.getType(),
3249 CastOp->getName()+".shrunk");
3250 NewCast = InsertNewInstBefore(NewCast, I);
3251 // trunc_or_bitcast(C1)&C2
3252 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3253 C3 = ConstantExpr::getAnd(C3, AndRHS);
3254 return BinaryOperator::createAnd(NewCast, C3);
3255 } else if (CastOp->getOpcode() == Instruction::Or) {
3256 // Change: and (cast (or X, C1) to T), C2
3257 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3258 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3259 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3260 return ReplaceInstUsesWith(I, AndRHS);
3265 // Try to fold constant and into select arguments.
3266 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3267 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3269 if (isa<PHINode>(Op0))
3270 if (Instruction *NV = FoldOpIntoPhi(I))
3274 Value *Op0NotVal = dyn_castNotVal(Op0);
3275 Value *Op1NotVal = dyn_castNotVal(Op1);
3277 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3278 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3280 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3281 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3282 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3283 I.getName()+".demorgan");
3284 InsertNewInstBefore(Or, I);
3285 return BinaryOperator::createNot(Or);
3289 Value *A = 0, *B = 0;
3290 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3291 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3292 return ReplaceInstUsesWith(I, Op1);
3293 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3294 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3295 return ReplaceInstUsesWith(I, Op0);
3297 if (Op0->hasOneUse() &&
3298 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3299 if (A == Op1) { // (A^B)&A -> A&(A^B)
3300 I.swapOperands(); // Simplify below
3301 std::swap(Op0, Op1);
3302 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3303 cast<BinaryOperator>(Op0)->swapOperands();
3304 I.swapOperands(); // Simplify below
3305 std::swap(Op0, Op1);
3308 if (Op1->hasOneUse() &&
3309 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3310 if (B == Op0) { // B&(A^B) -> B&(B^A)
3311 cast<BinaryOperator>(Op1)->swapOperands();
3314 if (A == Op0) { // A&(A^B) -> A & ~B
3315 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3316 InsertNewInstBefore(NotB, I);
3317 return BinaryOperator::createAnd(A, NotB);
3322 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3323 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3324 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3327 Value *LHSVal, *RHSVal;
3328 ConstantInt *LHSCst, *RHSCst;
3329 ICmpInst::Predicate LHSCC, RHSCC;
3330 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3331 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3332 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3333 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3334 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3335 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3336 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3337 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3338 // Ensure that the larger constant is on the RHS.
3339 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3340 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3341 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3342 ICmpInst *LHS = cast<ICmpInst>(Op0);
3343 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3344 std::swap(LHS, RHS);
3345 std::swap(LHSCst, RHSCst);
3346 std::swap(LHSCC, RHSCC);
3349 // At this point, we know we have have two icmp instructions
3350 // comparing a value against two constants and and'ing the result
3351 // together. Because of the above check, we know that we only have
3352 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3353 // (from the FoldICmpLogical check above), that the two constants
3354 // are not equal and that the larger constant is on the RHS
3355 assert(LHSCst != RHSCst && "Compares not folded above?");
3358 default: assert(0 && "Unknown integer condition code!");
3359 case ICmpInst::ICMP_EQ:
3361 default: assert(0 && "Unknown integer condition code!");
3362 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3363 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3364 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3365 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3366 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3367 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3368 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3369 return ReplaceInstUsesWith(I, LHS);
3371 case ICmpInst::ICMP_NE:
3373 default: assert(0 && "Unknown integer condition code!");
3374 case ICmpInst::ICMP_ULT:
3375 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3376 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3377 break; // (X != 13 & X u< 15) -> no change
3378 case ICmpInst::ICMP_SLT:
3379 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3380 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3381 break; // (X != 13 & X s< 15) -> no change
3382 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3383 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3384 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3385 return ReplaceInstUsesWith(I, RHS);
3386 case ICmpInst::ICMP_NE:
3387 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3388 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3389 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3390 LHSVal->getName()+".off");
3391 InsertNewInstBefore(Add, I);
3392 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3393 ConstantInt::get(Add->getType(), 1));
3395 break; // (X != 13 & X != 15) -> no change
3398 case ICmpInst::ICMP_ULT:
3400 default: assert(0 && "Unknown integer condition code!");
3401 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3402 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3403 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3404 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3406 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3407 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3408 return ReplaceInstUsesWith(I, LHS);
3409 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3413 case ICmpInst::ICMP_SLT:
3415 default: assert(0 && "Unknown integer condition code!");
3416 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3417 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3418 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3419 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3421 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3422 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3423 return ReplaceInstUsesWith(I, LHS);
3424 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3428 case ICmpInst::ICMP_UGT:
3430 default: assert(0 && "Unknown integer condition code!");
3431 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3432 return ReplaceInstUsesWith(I, LHS);
3433 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3434 return ReplaceInstUsesWith(I, RHS);
3435 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3437 case ICmpInst::ICMP_NE:
3438 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3439 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3440 break; // (X u> 13 & X != 15) -> no change
3441 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3442 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3444 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3448 case ICmpInst::ICMP_SGT:
3450 default: assert(0 && "Unknown integer condition code!");
3451 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3452 return ReplaceInstUsesWith(I, LHS);
3453 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3454 return ReplaceInstUsesWith(I, RHS);
3455 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3457 case ICmpInst::ICMP_NE:
3458 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3459 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3460 break; // (X s> 13 & X != 15) -> no change
3461 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3462 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3464 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3472 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3473 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3474 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3475 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3476 const Type *SrcTy = Op0C->getOperand(0)->getType();
3477 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3478 // Only do this if the casts both really cause code to be generated.
3479 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3481 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3483 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3484 Op1C->getOperand(0),
3486 InsertNewInstBefore(NewOp, I);
3487 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3491 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3492 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3493 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3494 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3495 SI0->getOperand(1) == SI1->getOperand(1) &&
3496 (SI0->hasOneUse() || SI1->hasOneUse())) {
3497 Instruction *NewOp =
3498 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3500 SI0->getName()), I);
3501 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3502 SI1->getOperand(1));
3506 return Changed ? &I : 0;
3509 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3510 /// in the result. If it does, and if the specified byte hasn't been filled in
3511 /// yet, fill it in and return false.
3512 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3513 Instruction *I = dyn_cast<Instruction>(V);
3514 if (I == 0) return true;
3516 // If this is an or instruction, it is an inner node of the bswap.
3517 if (I->getOpcode() == Instruction::Or)
3518 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3519 CollectBSwapParts(I->getOperand(1), ByteValues);
3521 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3522 // If this is a shift by a constant int, and it is "24", then its operand
3523 // defines a byte. We only handle unsigned types here.
3524 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3525 // Not shifting the entire input by N-1 bytes?
3526 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3527 8*(ByteValues.size()-1))
3531 if (I->getOpcode() == Instruction::Shl) {
3532 // X << 24 defines the top byte with the lowest of the input bytes.
3533 DestNo = ByteValues.size()-1;
3535 // X >>u 24 defines the low byte with the highest of the input bytes.
3539 // If the destination byte value is already defined, the values are or'd
3540 // together, which isn't a bswap (unless it's an or of the same bits).
3541 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3543 ByteValues[DestNo] = I->getOperand(0);
3547 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3549 Value *Shift = 0, *ShiftLHS = 0;
3550 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3551 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3552 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3554 Instruction *SI = cast<Instruction>(Shift);
3556 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3557 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3558 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3561 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3563 if (AndAmt->getValue().getActiveBits() > 64)
3565 uint64_t AndAmtVal = AndAmt->getZExtValue();
3566 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3567 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3569 // Unknown mask for bswap.
3570 if (DestByte == ByteValues.size()) return true;
3572 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3574 if (SI->getOpcode() == Instruction::Shl)
3575 SrcByte = DestByte - ShiftBytes;
3577 SrcByte = DestByte + ShiftBytes;
3579 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3580 if (SrcByte != ByteValues.size()-DestByte-1)
3583 // If the destination byte value is already defined, the values are or'd
3584 // together, which isn't a bswap (unless it's an or of the same bits).
3585 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3587 ByteValues[DestByte] = SI->getOperand(0);
3591 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3592 /// If so, insert the new bswap intrinsic and return it.
3593 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3594 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3595 if (!ITy || ITy->getBitWidth() % 16)
3596 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3598 /// ByteValues - For each byte of the result, we keep track of which value
3599 /// defines each byte.
3600 SmallVector<Value*, 8> ByteValues;
3601 ByteValues.resize(ITy->getBitWidth()/8);
3603 // Try to find all the pieces corresponding to the bswap.
3604 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3605 CollectBSwapParts(I.getOperand(1), ByteValues))
3608 // Check to see if all of the bytes come from the same value.
3609 Value *V = ByteValues[0];
3610 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3612 // Check to make sure that all of the bytes come from the same value.
3613 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3614 if (ByteValues[i] != V)
3616 const Type *Tys[] = { ITy, ITy };
3617 Module *M = I.getParent()->getParent()->getParent();
3618 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 2);
3619 return new CallInst(F, V);
3623 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3624 bool Changed = SimplifyCommutative(I);
3625 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3627 if (isa<UndefValue>(Op1)) // X | undef -> -1
3628 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
3632 return ReplaceInstUsesWith(I, Op0);
3634 // See if we can simplify any instructions used by the instruction whose sole
3635 // purpose is to compute bits we don't care about.
3636 if (!isa<VectorType>(I.getType())) {
3637 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3638 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3639 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3640 KnownZero, KnownOne))
3645 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3646 ConstantInt *C1 = 0; Value *X = 0;
3647 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3648 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3649 Instruction *Or = BinaryOperator::createOr(X, RHS);
3650 InsertNewInstBefore(Or, I);
3652 return BinaryOperator::createAnd(Or,
3653 ConstantInt::get(RHS->getValue() | C1->getValue()));
3656 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3657 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3658 Instruction *Or = BinaryOperator::createOr(X, RHS);
3659 InsertNewInstBefore(Or, I);
3661 return BinaryOperator::createXor(Or,
3662 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3665 // Try to fold constant and into select arguments.
3666 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3667 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3669 if (isa<PHINode>(Op0))
3670 if (Instruction *NV = FoldOpIntoPhi(I))
3674 Value *A = 0, *B = 0;
3675 ConstantInt *C1 = 0, *C2 = 0;
3677 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3678 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3679 return ReplaceInstUsesWith(I, Op1);
3680 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3681 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3682 return ReplaceInstUsesWith(I, Op0);
3684 // (A | B) | C and A | (B | C) -> bswap if possible.
3685 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3686 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3687 match(Op1, m_Or(m_Value(), m_Value())) ||
3688 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3689 match(Op1, m_Shift(m_Value(), m_Value())))) {
3690 if (Instruction *BSwap = MatchBSwap(I))
3694 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3695 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3696 MaskedValueIsZero(Op1, C1->getValue())) {
3697 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3698 InsertNewInstBefore(NOr, I);
3700 return BinaryOperator::createXor(NOr, C1);
3703 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3704 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3705 MaskedValueIsZero(Op0, C1->getValue())) {
3706 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3707 InsertNewInstBefore(NOr, I);
3709 return BinaryOperator::createXor(NOr, C1);
3712 // (A & C1)|(B & C2)
3713 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3714 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3716 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3717 return BinaryOperator::createAnd(A,
3718 ConstantInt::get(C1->getValue() | C2->getValue()));
3721 // If we have: ((V + N) & C1) | (V & C2)
3722 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3723 // replace with V+N.
3724 if (C1->getValue() == ~C2->getValue()) {
3725 Value *V1 = 0, *V2 = 0;
3726 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3727 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3728 // Add commutes, try both ways.
3729 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3730 return ReplaceInstUsesWith(I, A);
3731 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3732 return ReplaceInstUsesWith(I, A);
3734 // Or commutes, try both ways.
3735 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3736 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3737 // Add commutes, try both ways.
3738 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3739 return ReplaceInstUsesWith(I, B);
3740 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3741 return ReplaceInstUsesWith(I, B);
3746 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3747 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3748 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3749 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3750 SI0->getOperand(1) == SI1->getOperand(1) &&
3751 (SI0->hasOneUse() || SI1->hasOneUse())) {
3752 Instruction *NewOp =
3753 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3755 SI0->getName()), I);
3756 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3757 SI1->getOperand(1));
3761 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3762 if (A == Op1) // ~A | A == -1
3763 return ReplaceInstUsesWith(I,
3764 ConstantInt::getAllOnesValue(I.getType()));
3768 // Note, A is still live here!
3769 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3771 return ReplaceInstUsesWith(I,
3772 ConstantInt::getAllOnesValue(I.getType()));
3774 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3775 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3776 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3777 I.getName()+".demorgan"), I);
3778 return BinaryOperator::createNot(And);
3782 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3783 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3784 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3787 Value *LHSVal, *RHSVal;
3788 ConstantInt *LHSCst, *RHSCst;
3789 ICmpInst::Predicate LHSCC, RHSCC;
3790 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3791 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3792 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3793 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3794 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3795 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3796 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3797 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3798 // Ensure that the larger constant is on the RHS.
3799 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3800 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3801 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3802 ICmpInst *LHS = cast<ICmpInst>(Op0);
3803 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3804 std::swap(LHS, RHS);
3805 std::swap(LHSCst, RHSCst);
3806 std::swap(LHSCC, RHSCC);
3809 // At this point, we know we have have two icmp instructions
3810 // comparing a value against two constants and or'ing the result
3811 // together. Because of the above check, we know that we only have
3812 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3813 // FoldICmpLogical check above), that the two constants are not
3815 assert(LHSCst != RHSCst && "Compares not folded above?");
3818 default: assert(0 && "Unknown integer condition code!");
3819 case ICmpInst::ICMP_EQ:
3821 default: assert(0 && "Unknown integer condition code!");
3822 case ICmpInst::ICMP_EQ:
3823 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3824 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3825 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3826 LHSVal->getName()+".off");
3827 InsertNewInstBefore(Add, I);
3828 AddCST = Subtract(AddOne(RHSCst), LHSCst);
3829 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3831 break; // (X == 13 | X == 15) -> no change
3832 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3833 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3835 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3836 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3837 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3838 return ReplaceInstUsesWith(I, RHS);
3841 case ICmpInst::ICMP_NE:
3843 default: assert(0 && "Unknown integer condition code!");
3844 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3845 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3846 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3847 return ReplaceInstUsesWith(I, LHS);
3848 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3849 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3850 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3851 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3854 case ICmpInst::ICMP_ULT:
3856 default: assert(0 && "Unknown integer condition code!");
3857 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3859 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3860 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3862 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3864 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3865 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3866 return ReplaceInstUsesWith(I, RHS);
3867 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3871 case ICmpInst::ICMP_SLT:
3873 default: assert(0 && "Unknown integer condition code!");
3874 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3876 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3877 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3879 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3881 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3882 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3883 return ReplaceInstUsesWith(I, RHS);
3884 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3888 case ICmpInst::ICMP_UGT:
3890 default: assert(0 && "Unknown integer condition code!");
3891 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3892 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3893 return ReplaceInstUsesWith(I, LHS);
3894 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3896 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3897 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3898 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3899 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3903 case ICmpInst::ICMP_SGT:
3905 default: assert(0 && "Unknown integer condition code!");
3906 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3907 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3908 return ReplaceInstUsesWith(I, LHS);
3909 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3911 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3912 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3913 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3914 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3922 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3923 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3924 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3925 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3926 const Type *SrcTy = Op0C->getOperand(0)->getType();
3927 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3928 // Only do this if the casts both really cause code to be generated.
3929 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3931 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3933 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3934 Op1C->getOperand(0),
3936 InsertNewInstBefore(NewOp, I);
3937 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3942 return Changed ? &I : 0;
3945 // XorSelf - Implements: X ^ X --> 0
3948 XorSelf(Value *rhs) : RHS(rhs) {}
3949 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3950 Instruction *apply(BinaryOperator &Xor) const {
3956 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3957 bool Changed = SimplifyCommutative(I);
3958 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3960 if (isa<UndefValue>(Op1))
3961 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3963 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3964 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3965 assert(Result == &I && "AssociativeOpt didn't work?");
3966 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3969 // See if we can simplify any instructions used by the instruction whose sole
3970 // purpose is to compute bits we don't care about.
3971 if (!isa<VectorType>(I.getType())) {
3972 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3973 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3974 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3975 KnownZero, KnownOne))
3979 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3980 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
3981 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3982 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
3983 return new ICmpInst(ICI->getInversePredicate(),
3984 ICI->getOperand(0), ICI->getOperand(1));
3986 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3987 // ~(c-X) == X-c-1 == X+(-c-1)
3988 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3989 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3990 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3991 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3992 ConstantInt::get(I.getType(), 1));
3993 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3996 // ~(~X & Y) --> (X | ~Y)
3997 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3998 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3999 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4001 BinaryOperator::createNot(Op0I->getOperand(1),
4002 Op0I->getOperand(1)->getName()+".not");
4003 InsertNewInstBefore(NotY, I);
4004 return BinaryOperator::createOr(Op0NotVal, NotY);
4008 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4009 if (Op0I->getOpcode() == Instruction::Add) {
4010 // ~(X-c) --> (-c-1)-X
4011 if (RHS->isAllOnesValue()) {
4012 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4013 return BinaryOperator::createSub(
4014 ConstantExpr::getSub(NegOp0CI,
4015 ConstantInt::get(I.getType(), 1)),
4016 Op0I->getOperand(0));
4017 } else if (RHS->getValue().isSignBit()) {
4018 // (X + C) ^ signbit -> (X + C + signbit)
4019 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4020 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4023 } else if (Op0I->getOpcode() == Instruction::Or) {
4024 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4025 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4026 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4027 // Anything in both C1 and C2 is known to be zero, remove it from
4029 Constant *CommonBits = And(Op0CI, RHS);
4030 NewRHS = ConstantExpr::getAnd(NewRHS,
4031 ConstantExpr::getNot(CommonBits));
4032 AddToWorkList(Op0I);
4033 I.setOperand(0, Op0I->getOperand(0));
4034 I.setOperand(1, NewRHS);
4040 // Try to fold constant and into select arguments.
4041 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4042 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4044 if (isa<PHINode>(Op0))
4045 if (Instruction *NV = FoldOpIntoPhi(I))
4049 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4051 return ReplaceInstUsesWith(I,
4052 ConstantInt::getAllOnesValue(I.getType()));
4054 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4056 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
4059 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4062 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4063 if (A == Op0) { // B^(B|A) == (A|B)^B
4064 Op1I->swapOperands();
4066 std::swap(Op0, Op1);
4067 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4068 I.swapOperands(); // Simplified below.
4069 std::swap(Op0, Op1);
4071 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4072 if (Op0 == A) // A^(A^B) == B
4073 return ReplaceInstUsesWith(I, B);
4074 else if (Op0 == B) // A^(B^A) == B
4075 return ReplaceInstUsesWith(I, A);
4076 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4077 if (A == Op0) { // A^(A&B) -> A^(B&A)
4078 Op1I->swapOperands();
4081 if (B == Op0) { // A^(B&A) -> (B&A)^A
4082 I.swapOperands(); // Simplified below.
4083 std::swap(Op0, Op1);
4088 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4091 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4092 if (A == Op1) // (B|A)^B == (A|B)^B
4094 if (B == Op1) { // (A|B)^B == A & ~B
4096 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4097 return BinaryOperator::createAnd(A, NotB);
4099 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4100 if (Op1 == A) // (A^B)^A == B
4101 return ReplaceInstUsesWith(I, B);
4102 else if (Op1 == B) // (B^A)^A == B
4103 return ReplaceInstUsesWith(I, A);
4104 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4105 if (A == Op1) // (A&B)^A -> (B&A)^A
4107 if (B == Op1 && // (B&A)^A == ~B & A
4108 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4110 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4111 return BinaryOperator::createAnd(N, Op1);
4116 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4117 if (Op0I && Op1I && Op0I->isShift() &&
4118 Op0I->getOpcode() == Op1I->getOpcode() &&
4119 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4120 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4121 Instruction *NewOp =
4122 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4123 Op1I->getOperand(0),
4124 Op0I->getName()), I);
4125 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4126 Op1I->getOperand(1));
4130 Value *A, *B, *C, *D;
4131 // (A & B)^(A | B) -> A ^ B
4132 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4133 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4134 if ((A == C && B == D) || (A == D && B == C))
4135 return BinaryOperator::createXor(A, B);
4137 // (A | B)^(A & B) -> A ^ B
4138 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4139 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4140 if ((A == C && B == D) || (A == D && B == C))
4141 return BinaryOperator::createXor(A, B);
4145 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4146 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4147 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4148 // (X & Y)^(X & Y) -> (Y^Z) & X
4149 Value *X = 0, *Y = 0, *Z = 0;
4151 X = A, Y = B, Z = D;
4153 X = A, Y = B, Z = C;
4155 X = B, Y = A, Z = D;
4157 X = B, Y = A, Z = C;
4160 Instruction *NewOp =
4161 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4162 return BinaryOperator::createAnd(NewOp, X);
4167 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4168 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4169 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4172 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4173 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4174 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4175 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4176 const Type *SrcTy = Op0C->getOperand(0)->getType();
4177 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4178 // Only do this if the casts both really cause code to be generated.
4179 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4181 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4183 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4184 Op1C->getOperand(0),
4186 InsertNewInstBefore(NewOp, I);
4187 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4191 return Changed ? &I : 0;
4194 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4195 /// overflowed for this type.
4196 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4197 ConstantInt *In2, bool IsSigned = false) {
4198 Result = cast<ConstantInt>(Add(In1, In2));
4201 if (In2->getValue().isNegative())
4202 return Result->getValue().sgt(In1->getValue());
4204 return Result->getValue().slt(In1->getValue());
4206 return Result->getValue().ult(In1->getValue());
4209 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4210 /// code necessary to compute the offset from the base pointer (without adding
4211 /// in the base pointer). Return the result as a signed integer of intptr size.
4212 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4213 TargetData &TD = IC.getTargetData();
4214 gep_type_iterator GTI = gep_type_begin(GEP);
4215 const Type *IntPtrTy = TD.getIntPtrType();
4216 Value *Result = Constant::getNullValue(IntPtrTy);
4218 // Build a mask for high order bits.
4219 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4221 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4222 Value *Op = GEP->getOperand(i);
4223 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4224 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4225 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4226 if (!OpC->isNullValue()) {
4227 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4228 Scale = ConstantExpr::getMul(OpC, Scale);
4229 if (Constant *RC = dyn_cast<Constant>(Result))
4230 Result = ConstantExpr::getAdd(RC, Scale);
4232 // Emit an add instruction.
4233 Result = IC.InsertNewInstBefore(
4234 BinaryOperator::createAdd(Result, Scale,
4235 GEP->getName()+".offs"), I);
4239 // Convert to correct type.
4240 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4241 Op->getName()+".c"), I);
4243 // We'll let instcombine(mul) convert this to a shl if possible.
4244 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4245 GEP->getName()+".idx"), I);
4247 // Emit an add instruction.
4248 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4249 GEP->getName()+".offs"), I);
4255 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4256 /// else. At this point we know that the GEP is on the LHS of the comparison.
4257 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4258 ICmpInst::Predicate Cond,
4260 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4262 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4263 if (isa<PointerType>(CI->getOperand(0)->getType()))
4264 RHS = CI->getOperand(0);
4266 Value *PtrBase = GEPLHS->getOperand(0);
4267 if (PtrBase == RHS) {
4268 // As an optimization, we don't actually have to compute the actual value of
4269 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4270 // each index is zero or not.
4271 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4272 Instruction *InVal = 0;
4273 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4274 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4276 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4277 if (isa<UndefValue>(C)) // undef index -> undef.
4278 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4279 if (C->isNullValue())
4281 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4282 EmitIt = false; // This is indexing into a zero sized array?
4283 } else if (isa<ConstantInt>(C))
4284 return ReplaceInstUsesWith(I, // No comparison is needed here.
4285 ConstantInt::get(Type::Int1Ty,
4286 Cond == ICmpInst::ICMP_NE));
4291 new ICmpInst(Cond, GEPLHS->getOperand(i),
4292 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4296 InVal = InsertNewInstBefore(InVal, I);
4297 InsertNewInstBefore(Comp, I);
4298 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4299 InVal = BinaryOperator::createOr(InVal, Comp);
4300 else // True if all are equal
4301 InVal = BinaryOperator::createAnd(InVal, Comp);
4309 // No comparison is needed here, all indexes = 0
4310 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4311 Cond == ICmpInst::ICMP_EQ));
4314 // Only lower this if the icmp is the only user of the GEP or if we expect
4315 // the result to fold to a constant!
4316 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4317 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4318 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4319 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4320 Constant::getNullValue(Offset->getType()));
4322 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4323 // If the base pointers are different, but the indices are the same, just
4324 // compare the base pointer.
4325 if (PtrBase != GEPRHS->getOperand(0)) {
4326 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4327 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4328 GEPRHS->getOperand(0)->getType();
4330 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4331 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4332 IndicesTheSame = false;
4336 // If all indices are the same, just compare the base pointers.
4338 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4339 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4341 // Otherwise, the base pointers are different and the indices are
4342 // different, bail out.
4346 // If one of the GEPs has all zero indices, recurse.
4347 bool AllZeros = true;
4348 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4349 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4350 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4355 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4356 ICmpInst::getSwappedPredicate(Cond), I);
4358 // If the other GEP has all zero indices, recurse.
4360 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4361 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4362 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4367 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4369 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4370 // If the GEPs only differ by one index, compare it.
4371 unsigned NumDifferences = 0; // Keep track of # differences.
4372 unsigned DiffOperand = 0; // The operand that differs.
4373 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4374 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4375 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4376 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4377 // Irreconcilable differences.
4381 if (NumDifferences++) break;
4386 if (NumDifferences == 0) // SAME GEP?
4387 return ReplaceInstUsesWith(I, // No comparison is needed here.
4388 ConstantInt::get(Type::Int1Ty,
4389 Cond == ICmpInst::ICMP_EQ));
4390 else if (NumDifferences == 1) {
4391 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4392 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4393 // Make sure we do a signed comparison here.
4394 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4398 // Only lower this if the icmp is the only user of the GEP or if we expect
4399 // the result to fold to a constant!
4400 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4401 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4402 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4403 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4404 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4405 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4411 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4412 bool Changed = SimplifyCompare(I);
4413 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4415 // Fold trivial predicates.
4416 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4417 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4418 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4419 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4421 // Simplify 'fcmp pred X, X'
4423 switch (I.getPredicate()) {
4424 default: assert(0 && "Unknown predicate!");
4425 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4426 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4427 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4428 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4429 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4430 case FCmpInst::FCMP_OLT: // True if ordered and less than
4431 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4432 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4434 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4435 case FCmpInst::FCMP_ULT: // True if unordered or less than
4436 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4437 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4438 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4439 I.setPredicate(FCmpInst::FCMP_UNO);
4440 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4443 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4444 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4445 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4446 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4447 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4448 I.setPredicate(FCmpInst::FCMP_ORD);
4449 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4454 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4455 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4457 // Handle fcmp with constant RHS
4458 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4459 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4460 switch (LHSI->getOpcode()) {
4461 case Instruction::PHI:
4462 if (Instruction *NV = FoldOpIntoPhi(I))
4465 case Instruction::Select:
4466 // If either operand of the select is a constant, we can fold the
4467 // comparison into the select arms, which will cause one to be
4468 // constant folded and the select turned into a bitwise or.
4469 Value *Op1 = 0, *Op2 = 0;
4470 if (LHSI->hasOneUse()) {
4471 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4472 // Fold the known value into the constant operand.
4473 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4474 // Insert a new FCmp of the other select operand.
4475 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4476 LHSI->getOperand(2), RHSC,
4478 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4479 // Fold the known value into the constant operand.
4480 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4481 // Insert a new FCmp of the other select operand.
4482 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4483 LHSI->getOperand(1), RHSC,
4489 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4494 return Changed ? &I : 0;
4497 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4498 bool Changed = SimplifyCompare(I);
4499 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4500 const Type *Ty = Op0->getType();
4504 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4505 isTrueWhenEqual(I)));
4507 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4508 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4510 // icmp of GlobalValues can never equal each other as long as they aren't
4511 // external weak linkage type.
4512 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4513 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4514 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4515 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4516 !isTrueWhenEqual(I)));
4518 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4519 // addresses never equal each other! We already know that Op0 != Op1.
4520 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4521 isa<ConstantPointerNull>(Op0)) &&
4522 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4523 isa<ConstantPointerNull>(Op1)))
4524 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4525 !isTrueWhenEqual(I)));
4527 // icmp's with boolean values can always be turned into bitwise operations
4528 if (Ty == Type::Int1Ty) {
4529 switch (I.getPredicate()) {
4530 default: assert(0 && "Invalid icmp instruction!");
4531 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4532 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4533 InsertNewInstBefore(Xor, I);
4534 return BinaryOperator::createNot(Xor);
4536 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4537 return BinaryOperator::createXor(Op0, Op1);
4539 case ICmpInst::ICMP_UGT:
4540 case ICmpInst::ICMP_SGT:
4541 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4543 case ICmpInst::ICMP_ULT:
4544 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4545 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4546 InsertNewInstBefore(Not, I);
4547 return BinaryOperator::createAnd(Not, Op1);
4549 case ICmpInst::ICMP_UGE:
4550 case ICmpInst::ICMP_SGE:
4551 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4553 case ICmpInst::ICMP_ULE:
4554 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4555 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4556 InsertNewInstBefore(Not, I);
4557 return BinaryOperator::createOr(Not, Op1);
4562 // See if we are doing a comparison between a constant and an instruction that
4563 // can be folded into the comparison.
4564 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4565 switch (I.getPredicate()) {
4567 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4568 if (CI->isMinValue(false))
4569 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4570 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4571 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4572 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4573 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4576 case ICmpInst::ICMP_SLT:
4577 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4578 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4579 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4580 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4581 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4582 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4585 case ICmpInst::ICMP_UGT:
4586 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4587 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4588 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4589 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4590 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4591 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4594 case ICmpInst::ICMP_SGT:
4595 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4596 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4597 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4598 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4599 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4600 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4603 case ICmpInst::ICMP_ULE:
4604 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4605 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4606 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4607 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4608 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4609 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4612 case ICmpInst::ICMP_SLE:
4613 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4614 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4615 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4616 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4617 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4618 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4621 case ICmpInst::ICMP_UGE:
4622 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4623 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4624 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4625 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4626 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4627 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4630 case ICmpInst::ICMP_SGE:
4631 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4632 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4633 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4634 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4635 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4636 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4640 // If we still have a icmp le or icmp ge instruction, turn it into the
4641 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4642 // already been handled above, this requires little checking.
4644 switch (I.getPredicate()) {
4646 case ICmpInst::ICMP_ULE:
4647 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4648 case ICmpInst::ICMP_SLE:
4649 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4650 case ICmpInst::ICMP_UGE:
4651 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4652 case ICmpInst::ICMP_SGE:
4653 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4656 // See if we can fold the comparison based on bits known to be zero or one
4658 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4659 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4660 if (SimplifyDemandedBits(Op0, APInt::getAllOnesValue(BitWidth),
4661 KnownZero, KnownOne, 0))
4664 // Given the known and unknown bits, compute a range that the LHS could be
4666 if ((KnownOne | KnownZero) != 0) {
4667 // Compute the Min, Max and RHS values based on the known bits. For the
4668 // EQ and NE we use unsigned values.
4669 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4670 const APInt& RHSVal = CI->getValue();
4671 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4672 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4675 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4678 switch (I.getPredicate()) { // LE/GE have been folded already.
4679 default: assert(0 && "Unknown icmp opcode!");
4680 case ICmpInst::ICMP_EQ:
4681 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4682 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4684 case ICmpInst::ICMP_NE:
4685 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4686 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4688 case ICmpInst::ICMP_ULT:
4689 if (Max.ult(RHSVal))
4690 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4691 if (Min.ugt(RHSVal))
4692 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4694 case ICmpInst::ICMP_UGT:
4695 if (Min.ugt(RHSVal))
4696 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4697 if (Max.ult(RHSVal))
4698 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4700 case ICmpInst::ICMP_SLT:
4701 if (Max.slt(RHSVal))
4702 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4703 if (Min.sgt(RHSVal))
4704 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4706 case ICmpInst::ICMP_SGT:
4707 if (Min.sgt(RHSVal))
4708 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4709 if (Max.slt(RHSVal))
4710 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4715 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4716 // instruction, see if that instruction also has constants so that the
4717 // instruction can be folded into the icmp
4718 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4719 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
4723 // Handle icmp with constant (but not simple integer constant) RHS
4724 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4725 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4726 switch (LHSI->getOpcode()) {
4727 case Instruction::GetElementPtr:
4728 if (RHSC->isNullValue()) {
4729 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
4730 bool isAllZeros = true;
4731 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4732 if (!isa<Constant>(LHSI->getOperand(i)) ||
4733 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4738 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
4739 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4743 case Instruction::PHI:
4744 if (Instruction *NV = FoldOpIntoPhi(I))
4747 case Instruction::Select: {
4748 // If either operand of the select is a constant, we can fold the
4749 // comparison into the select arms, which will cause one to be
4750 // constant folded and the select turned into a bitwise or.
4751 Value *Op1 = 0, *Op2 = 0;
4752 if (LHSI->hasOneUse()) {
4753 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4754 // Fold the known value into the constant operand.
4755 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4756 // Insert a new ICmp of the other select operand.
4757 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4758 LHSI->getOperand(2), RHSC,
4760 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4761 // Fold the known value into the constant operand.
4762 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4763 // Insert a new ICmp of the other select operand.
4764 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4765 LHSI->getOperand(1), RHSC,
4771 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4774 case Instruction::Malloc:
4775 // If we have (malloc != null), and if the malloc has a single use, we
4776 // can assume it is successful and remove the malloc.
4777 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
4778 AddToWorkList(LHSI);
4779 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4780 !isTrueWhenEqual(I)));
4786 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4787 if (User *GEP = dyn_castGetElementPtr(Op0))
4788 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
4790 if (User *GEP = dyn_castGetElementPtr(Op1))
4791 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
4792 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
4795 // Test to see if the operands of the icmp are casted versions of other
4796 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
4798 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
4799 if (isa<PointerType>(Op0->getType()) &&
4800 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
4801 // We keep moving the cast from the left operand over to the right
4802 // operand, where it can often be eliminated completely.
4803 Op0 = CI->getOperand(0);
4805 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
4806 // so eliminate it as well.
4807 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
4808 Op1 = CI2->getOperand(0);
4810 // If Op1 is a constant, we can fold the cast into the constant.
4811 if (Op0->getType() != Op1->getType())
4812 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4813 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
4815 // Otherwise, cast the RHS right before the icmp
4816 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
4818 return new ICmpInst(I.getPredicate(), Op0, Op1);
4822 if (isa<CastInst>(Op0)) {
4823 // Handle the special case of: icmp (cast bool to X), <cst>
4824 // This comes up when you have code like
4827 // For generality, we handle any zero-extension of any operand comparison
4828 // with a constant or another cast from the same type.
4829 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
4830 if (Instruction *R = visitICmpInstWithCastAndCast(I))
4834 if (I.isEquality()) {
4835 Value *A, *B, *C, *D;
4836 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4837 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4838 Value *OtherVal = A == Op1 ? B : A;
4839 return new ICmpInst(I.getPredicate(), OtherVal,
4840 Constant::getNullValue(A->getType()));
4843 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4844 // A^c1 == C^c2 --> A == C^(c1^c2)
4845 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
4846 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
4847 if (Op1->hasOneUse()) {
4848 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
4849 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
4850 return new ICmpInst(I.getPredicate(), A,
4851 InsertNewInstBefore(Xor, I));
4854 // A^B == A^D -> B == D
4855 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
4856 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
4857 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
4858 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
4862 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
4863 (A == Op0 || B == Op0)) {
4864 // A == (A^B) -> B == 0
4865 Value *OtherVal = A == Op0 ? B : A;
4866 return new ICmpInst(I.getPredicate(), OtherVal,
4867 Constant::getNullValue(A->getType()));
4869 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
4870 // (A-B) == A -> B == 0
4871 return new ICmpInst(I.getPredicate(), B,
4872 Constant::getNullValue(B->getType()));
4874 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
4875 // A == (A-B) -> B == 0
4876 return new ICmpInst(I.getPredicate(), B,
4877 Constant::getNullValue(B->getType()));
4880 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4881 if (Op0->hasOneUse() && Op1->hasOneUse() &&
4882 match(Op0, m_And(m_Value(A), m_Value(B))) &&
4883 match(Op1, m_And(m_Value(C), m_Value(D)))) {
4884 Value *X = 0, *Y = 0, *Z = 0;
4887 X = B; Y = D; Z = A;
4888 } else if (A == D) {
4889 X = B; Y = C; Z = A;
4890 } else if (B == C) {
4891 X = A; Y = D; Z = B;
4892 } else if (B == D) {
4893 X = A; Y = C; Z = B;
4896 if (X) { // Build (X^Y) & Z
4897 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
4898 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
4899 I.setOperand(0, Op1);
4900 I.setOperand(1, Constant::getNullValue(Op1->getType()));
4905 return Changed ? &I : 0;
4908 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
4910 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
4913 const APInt &RHSV = RHS->getValue();
4915 switch (LHSI->getOpcode()) {
4916 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
4917 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4918 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
4920 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
4921 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
4922 Value *CompareVal = LHSI->getOperand(0);
4924 // If the sign bit of the XorCST is not set, there is no change to
4925 // the operation, just stop using the Xor.
4926 if (!XorCST->getValue().isNegative()) {
4927 ICI.setOperand(0, CompareVal);
4928 AddToWorkList(LHSI);
4932 // Was the old condition true if the operand is positive?
4933 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
4935 // If so, the new one isn't.
4936 isTrueIfPositive ^= true;
4938 if (isTrueIfPositive)
4939 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
4941 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
4945 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
4946 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4947 LHSI->getOperand(0)->hasOneUse()) {
4948 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4950 // If the LHS is an AND of a truncating cast, we can widen the
4951 // and/compare to be the input width without changing the value
4952 // produced, eliminating a cast.
4953 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
4954 // We can do this transformation if either the AND constant does not
4955 // have its sign bit set or if it is an equality comparison.
4956 // Extending a relational comparison when we're checking the sign
4957 // bit would not work.
4958 if (Cast->hasOneUse() &&
4959 (ICI.isEquality() || AndCST->getValue().isPositive() &&
4960 RHSV.isPositive())) {
4962 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
4963 APInt NewCST = AndCST->getValue();
4964 NewCST.zext(BitWidth);
4966 NewCI.zext(BitWidth);
4967 Instruction *NewAnd =
4968 BinaryOperator::createAnd(Cast->getOperand(0),
4969 ConstantInt::get(NewCST),LHSI->getName());
4970 InsertNewInstBefore(NewAnd, ICI);
4971 return new ICmpInst(ICI.getPredicate(), NewAnd,
4972 ConstantInt::get(NewCI));
4976 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4977 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4978 // happens a LOT in code produced by the C front-end, for bitfield
4980 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
4981 if (Shift && !Shift->isShift())
4985 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4986 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4987 const Type *AndTy = AndCST->getType(); // Type of the and.
4989 // We can fold this as long as we can't shift unknown bits
4990 // into the mask. This can only happen with signed shift
4991 // rights, as they sign-extend.
4993 bool CanFold = Shift->isLogicalShift();
4995 // To test for the bad case of the signed shr, see if any
4996 // of the bits shifted in could be tested after the mask.
4997 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
4998 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5000 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5001 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5002 AndCST->getValue()) == 0)
5008 if (Shift->getOpcode() == Instruction::Shl)
5009 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5011 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5013 // Check to see if we are shifting out any of the bits being
5015 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5016 // If we shifted bits out, the fold is not going to work out.
5017 // As a special case, check to see if this means that the
5018 // result is always true or false now.
5019 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5020 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5021 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5022 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5024 ICI.setOperand(1, NewCst);
5025 Constant *NewAndCST;
5026 if (Shift->getOpcode() == Instruction::Shl)
5027 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5029 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5030 LHSI->setOperand(1, NewAndCST);
5031 LHSI->setOperand(0, Shift->getOperand(0));
5032 AddToWorkList(Shift); // Shift is dead.
5033 AddUsesToWorkList(ICI);
5039 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5040 // preferable because it allows the C<<Y expression to be hoisted out
5041 // of a loop if Y is invariant and X is not.
5042 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5043 ICI.isEquality() && !Shift->isArithmeticShift() &&
5044 isa<Instruction>(Shift->getOperand(0))) {
5047 if (Shift->getOpcode() == Instruction::LShr) {
5048 NS = BinaryOperator::createShl(AndCST,
5049 Shift->getOperand(1), "tmp");
5051 // Insert a logical shift.
5052 NS = BinaryOperator::createLShr(AndCST,
5053 Shift->getOperand(1), "tmp");
5055 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5057 // Compute X & (C << Y).
5058 Instruction *NewAnd =
5059 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5060 InsertNewInstBefore(NewAnd, ICI);
5062 ICI.setOperand(0, NewAnd);
5068 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
5069 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5070 if (ICI.isEquality()) {
5071 uint32_t TypeBits = RHSV.getBitWidth();
5073 // Check that the shift amount is in range. If not, don't perform
5074 // undefined shifts. When the shift is visited it will be
5076 if (ShAmt->uge(TypeBits))
5079 // If we are comparing against bits always shifted out, the
5080 // comparison cannot succeed.
5082 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5083 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5084 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5085 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5086 return ReplaceInstUsesWith(ICI, Cst);
5089 if (LHSI->hasOneUse()) {
5090 // Otherwise strength reduce the shift into an and.
5091 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5093 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5096 BinaryOperator::createAnd(LHSI->getOperand(0),
5097 Mask, LHSI->getName()+".mask");
5098 Value *And = InsertNewInstBefore(AndI, ICI);
5099 return new ICmpInst(ICI.getPredicate(), And,
5100 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5106 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5107 case Instruction::AShr:
5108 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5109 if (ICI.isEquality()) {
5110 // Check that the shift amount is in range. If not, don't perform
5111 // undefined shifts. When the shift is visited it will be
5113 uint32_t TypeBits = RHSV.getBitWidth();
5114 if (ShAmt->uge(TypeBits))
5116 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5118 // If we are comparing against bits always shifted out, the
5119 // comparison cannot succeed.
5120 APInt Comp = RHSV << ShAmtVal;
5121 if (LHSI->getOpcode() == Instruction::LShr)
5122 Comp = Comp.lshr(ShAmtVal);
5124 Comp = Comp.ashr(ShAmtVal);
5126 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5127 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5128 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5129 return ReplaceInstUsesWith(ICI, Cst);
5132 if (LHSI->hasOneUse() || RHSV == 0) {
5133 // Otherwise strength reduce the shift into an and.
5134 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5135 Constant *Mask = ConstantInt::get(Val);
5138 BinaryOperator::createAnd(LHSI->getOperand(0),
5139 Mask, LHSI->getName()+".mask");
5140 Value *And = InsertNewInstBefore(AndI, ICI);
5141 return new ICmpInst(ICI.getPredicate(), And,
5142 ConstantExpr::getShl(RHS, ShAmt));
5148 case Instruction::SDiv:
5149 case Instruction::UDiv:
5150 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5151 // Fold this div into the comparison, producing a range check.
5152 // Determine, based on the divide type, what the range is being
5153 // checked. If there is an overflow on the low or high side, remember
5154 // it, otherwise compute the range [low, hi) bounding the new value.
5155 // See: InsertRangeTest above for the kinds of replacements possible.
5156 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5157 // FIXME: If the operand types don't match the type of the divide
5158 // then don't attempt this transform. The code below doesn't have the
5159 // logic to deal with a signed divide and an unsigned compare (and
5160 // vice versa). This is because (x /s C1) <s C2 produces different
5161 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5162 // (x /u C1) <u C2. Simply casting the operands and result won't
5163 // work. :( The if statement below tests that condition and bails
5165 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
5166 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5168 if (DivRHS->isZero())
5169 break; // Don't hack on div by zero
5171 // Initialize the variables that will indicate the nature of the
5173 bool LoOverflow = false, HiOverflow = false;
5174 ConstantInt *LoBound = 0, *HiBound = 0;
5176 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5177 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5178 // C2 (CI). By solving for X we can turn this into a range check
5179 // instead of computing a divide.
5180 ConstantInt *Prod = Multiply(RHS, DivRHS);
5182 // Determine if the product overflows by seeing if the product is
5183 // not equal to the divide. Make sure we do the same kind of divide
5184 // as in the LHS instruction that we're folding.
5185 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5186 ConstantExpr::getUDiv(Prod, DivRHS)) != RHS;
5188 // Get the ICmp opcode
5189 ICmpInst::Predicate predicate = ICI.getPredicate();
5191 if (!DivIsSigned) { // udiv
5193 LoOverflow = ProdOV;
5194 HiOverflow = ProdOV ||
5195 AddWithOverflow(HiBound, LoBound, DivRHS, false);
5196 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5197 if (RHSV == 0) { // (X / pos) op 0
5199 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5201 } else if (RHSV.isPositive()) { // (X / pos) op pos
5203 LoOverflow = ProdOV;
5204 HiOverflow = ProdOV ||
5205 AddWithOverflow(HiBound, Prod, DivRHS, true);
5206 } else { // (X / pos) op neg
5207 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5208 LoOverflow = AddWithOverflow(LoBound, Prod,
5209 cast<ConstantInt>(DivRHSH), true);
5210 HiBound = AddOne(Prod);
5211 HiOverflow = ProdOV;
5213 } else { // Divisor is < 0.
5214 if (RHSV == 0) { // (X / neg) op 0
5215 LoBound = AddOne(DivRHS);
5216 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5217 if (HiBound == DivRHS)
5218 LoBound = 0; // - INTMIN = INTMIN
5219 } else if (RHSV.isPositive()) { // (X / neg) op pos
5220 HiOverflow = LoOverflow = ProdOV;
5222 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS),
5224 HiBound = AddOne(Prod);
5225 } else { // (X / neg) op neg
5227 LoOverflow = HiOverflow = ProdOV;
5228 HiBound = Subtract(Prod, DivRHS);
5231 // Dividing by a negate swaps the condition.
5232 predicate = ICmpInst::getSwappedPredicate(predicate);
5236 Value *X = LHSI->getOperand(0);
5237 switch (predicate) {
5238 default: assert(0 && "Unhandled icmp opcode!");
5239 case ICmpInst::ICMP_EQ:
5240 if (LoOverflow && HiOverflow)
5241 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5242 else if (HiOverflow)
5243 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5244 ICmpInst::ICMP_UGE, X, LoBound);
5245 else if (LoOverflow)
5246 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5247 ICmpInst::ICMP_ULT, X, HiBound);
5249 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5251 case ICmpInst::ICMP_NE:
5252 if (LoOverflow && HiOverflow)
5253 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5254 else if (HiOverflow)
5255 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5256 ICmpInst::ICMP_ULT, X, LoBound);
5257 else if (LoOverflow)
5258 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5259 ICmpInst::ICMP_UGE, X, HiBound);
5261 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5263 case ICmpInst::ICMP_ULT:
5264 case ICmpInst::ICMP_SLT:
5266 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5267 return new ICmpInst(predicate, X, LoBound);
5268 case ICmpInst::ICMP_UGT:
5269 case ICmpInst::ICMP_SGT:
5271 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5272 if (predicate == ICmpInst::ICMP_UGT)
5273 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5275 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5282 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5283 if (ICI.isEquality()) {
5284 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5286 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5287 // the second operand is a constant, simplify a bit.
5288 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5289 switch (BO->getOpcode()) {
5290 case Instruction::SRem:
5291 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5292 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5293 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5294 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5295 Instruction *NewRem =
5296 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5298 InsertNewInstBefore(NewRem, ICI);
5299 return new ICmpInst(ICI.getPredicate(), NewRem,
5300 Constant::getNullValue(BO->getType()));
5304 case Instruction::Add:
5305 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5306 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5307 if (BO->hasOneUse())
5308 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5309 Subtract(RHS, BOp1C));
5310 } else if (RHSV == 0) {
5311 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5312 // efficiently invertible, or if the add has just this one use.
5313 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5315 if (Value *NegVal = dyn_castNegVal(BOp1))
5316 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5317 else if (Value *NegVal = dyn_castNegVal(BOp0))
5318 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5319 else if (BO->hasOneUse()) {
5320 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5321 InsertNewInstBefore(Neg, ICI);
5323 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5327 case Instruction::Xor:
5328 // For the xor case, we can xor two constants together, eliminating
5329 // the explicit xor.
5330 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5331 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5332 ConstantExpr::getXor(RHS, BOC));
5335 case Instruction::Sub:
5336 // Replace (([sub|xor] A, B) != 0) with (A != B)
5338 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5342 case Instruction::Or:
5343 // If bits are being or'd in that are not present in the constant we
5344 // are comparing against, then the comparison could never succeed!
5345 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5346 Constant *NotCI = ConstantExpr::getNot(RHS);
5347 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5348 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5353 case Instruction::And:
5354 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5355 // If bits are being compared against that are and'd out, then the
5356 // comparison can never succeed!
5357 if ((RHSV & ~BOC->getValue()) != 0)
5358 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5361 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5362 if (RHS == BOC && RHSV.isPowerOf2())
5363 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5364 ICmpInst::ICMP_NE, LHSI,
5365 Constant::getNullValue(RHS->getType()));
5367 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5368 if (isSignBit(BOC)) {
5369 Value *X = BO->getOperand(0);
5370 Constant *Zero = Constant::getNullValue(X->getType());
5371 ICmpInst::Predicate pred = isICMP_NE ?
5372 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5373 return new ICmpInst(pred, X, Zero);
5376 // ((X & ~7) == 0) --> X < 8
5377 if (RHSV == 0 && isHighOnes(BOC)) {
5378 Value *X = BO->getOperand(0);
5379 Constant *NegX = ConstantExpr::getNeg(BOC);
5380 ICmpInst::Predicate pred = isICMP_NE ?
5381 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5382 return new ICmpInst(pred, X, NegX);
5387 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5388 // Handle icmp {eq|ne} <intrinsic>, intcst.
5389 if (II->getIntrinsicID() == Intrinsic::bswap) {
5391 ICI.setOperand(0, II->getOperand(1));
5392 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5396 } else { // Not a ICMP_EQ/ICMP_NE
5397 // If the LHS is a cast from an integral value of the same size, then
5398 // since we know the RHS is a constant, try to simlify.
5399 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5400 Value *CastOp = Cast->getOperand(0);
5401 const Type *SrcTy = CastOp->getType();
5402 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5403 if (SrcTy->isInteger() &&
5404 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5405 // If this is an unsigned comparison, try to make the comparison use
5406 // smaller constant values.
5407 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5408 // X u< 128 => X s> -1
5409 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5410 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5411 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5412 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5413 // X u> 127 => X s< 0
5414 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5415 Constant::getNullValue(SrcTy));
5423 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5424 /// We only handle extending casts so far.
5426 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5427 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5428 Value *LHSCIOp = LHSCI->getOperand(0);
5429 const Type *SrcTy = LHSCIOp->getType();
5430 const Type *DestTy = LHSCI->getType();
5433 // We only handle extension cast instructions, so far. Enforce this.
5434 if (LHSCI->getOpcode() != Instruction::ZExt &&
5435 LHSCI->getOpcode() != Instruction::SExt)
5438 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5439 bool isSignedCmp = ICI.isSignedPredicate();
5441 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5442 // Not an extension from the same type?
5443 RHSCIOp = CI->getOperand(0);
5444 if (RHSCIOp->getType() != LHSCIOp->getType())
5447 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5448 // and the other is a zext), then we can't handle this.
5449 if (CI->getOpcode() != LHSCI->getOpcode())
5452 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5453 // then we can't handle this.
5454 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5457 // Okay, just insert a compare of the reduced operands now!
5458 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5461 // If we aren't dealing with a constant on the RHS, exit early
5462 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5466 // Compute the constant that would happen if we truncated to SrcTy then
5467 // reextended to DestTy.
5468 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5469 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5471 // If the re-extended constant didn't change...
5473 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5474 // For example, we might have:
5475 // %A = sext short %X to uint
5476 // %B = icmp ugt uint %A, 1330
5477 // It is incorrect to transform this into
5478 // %B = icmp ugt short %X, 1330
5479 // because %A may have negative value.
5481 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5482 // OR operation is EQ/NE.
5483 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5484 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5489 // The re-extended constant changed so the constant cannot be represented
5490 // in the shorter type. Consequently, we cannot emit a simple comparison.
5492 // First, handle some easy cases. We know the result cannot be equal at this
5493 // point so handle the ICI.isEquality() cases
5494 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5495 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5496 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5497 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5499 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5500 // should have been folded away previously and not enter in here.
5503 // We're performing a signed comparison.
5504 if (cast<ConstantInt>(CI)->getValue().isNegative())
5505 Result = ConstantInt::getFalse(); // X < (small) --> false
5507 Result = ConstantInt::getTrue(); // X < (large) --> true
5509 // We're performing an unsigned comparison.
5511 // We're performing an unsigned comp with a sign extended value.
5512 // This is true if the input is >= 0. [aka >s -1]
5513 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5514 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5515 NegOne, ICI.getName()), ICI);
5517 // Unsigned extend & unsigned compare -> always true.
5518 Result = ConstantInt::getTrue();
5522 // Finally, return the value computed.
5523 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5524 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5525 return ReplaceInstUsesWith(ICI, Result);
5527 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5528 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5529 "ICmp should be folded!");
5530 if (Constant *CI = dyn_cast<Constant>(Result))
5531 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5533 return BinaryOperator::createNot(Result);
5537 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5538 return commonShiftTransforms(I);
5541 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5542 return commonShiftTransforms(I);
5545 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5546 return commonShiftTransforms(I);
5549 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5550 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5551 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5553 // shl X, 0 == X and shr X, 0 == X
5554 // shl 0, X == 0 and shr 0, X == 0
5555 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5556 Op0 == Constant::getNullValue(Op0->getType()))
5557 return ReplaceInstUsesWith(I, Op0);
5559 if (isa<UndefValue>(Op0)) {
5560 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5561 return ReplaceInstUsesWith(I, Op0);
5562 else // undef << X -> 0, undef >>u X -> 0
5563 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5565 if (isa<UndefValue>(Op1)) {
5566 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5567 return ReplaceInstUsesWith(I, Op0);
5568 else // X << undef, X >>u undef -> 0
5569 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5572 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5573 if (I.getOpcode() == Instruction::AShr)
5574 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5575 if (CSI->isAllOnesValue())
5576 return ReplaceInstUsesWith(I, CSI);
5578 // Try to fold constant and into select arguments.
5579 if (isa<Constant>(Op0))
5580 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5581 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5584 // See if we can turn a signed shr into an unsigned shr.
5585 if (I.isArithmeticShift()) {
5586 if (MaskedValueIsZero(Op0,
5587 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5588 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5592 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5593 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5598 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5599 BinaryOperator &I) {
5600 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5602 // See if we can simplify any instructions used by the instruction whose sole
5603 // purpose is to compute bits we don't care about.
5604 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5605 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5606 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5607 KnownZero, KnownOne))
5610 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5611 // of a signed value.
5613 if (Op1->uge(TypeBits)) {
5614 if (I.getOpcode() != Instruction::AShr)
5615 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5617 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5622 // ((X*C1) << C2) == (X * (C1 << C2))
5623 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5624 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5625 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5626 return BinaryOperator::createMul(BO->getOperand(0),
5627 ConstantExpr::getShl(BOOp, Op1));
5629 // Try to fold constant and into select arguments.
5630 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5631 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5633 if (isa<PHINode>(Op0))
5634 if (Instruction *NV = FoldOpIntoPhi(I))
5637 if (Op0->hasOneUse()) {
5638 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5639 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5642 switch (Op0BO->getOpcode()) {
5644 case Instruction::Add:
5645 case Instruction::And:
5646 case Instruction::Or:
5647 case Instruction::Xor: {
5648 // These operators commute.
5649 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5650 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5651 match(Op0BO->getOperand(1),
5652 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5653 Instruction *YS = BinaryOperator::createShl(
5654 Op0BO->getOperand(0), Op1,
5656 InsertNewInstBefore(YS, I); // (Y << C)
5658 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5659 Op0BO->getOperand(1)->getName());
5660 InsertNewInstBefore(X, I); // (X + (Y << C))
5661 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5662 return BinaryOperator::createAnd(X, ConstantInt::get(
5663 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5666 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5667 Value *Op0BOOp1 = Op0BO->getOperand(1);
5668 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5670 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5671 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5673 Instruction *YS = BinaryOperator::createShl(
5674 Op0BO->getOperand(0), Op1,
5676 InsertNewInstBefore(YS, I); // (Y << C)
5678 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5679 V1->getName()+".mask");
5680 InsertNewInstBefore(XM, I); // X & (CC << C)
5682 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5687 case Instruction::Sub: {
5688 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5689 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5690 match(Op0BO->getOperand(0),
5691 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5692 Instruction *YS = BinaryOperator::createShl(
5693 Op0BO->getOperand(1), Op1,
5695 InsertNewInstBefore(YS, I); // (Y << C)
5697 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5698 Op0BO->getOperand(0)->getName());
5699 InsertNewInstBefore(X, I); // (X + (Y << C))
5700 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5701 return BinaryOperator::createAnd(X, ConstantInt::get(
5702 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5705 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5706 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5707 match(Op0BO->getOperand(0),
5708 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5709 m_ConstantInt(CC))) && V2 == Op1 &&
5710 cast<BinaryOperator>(Op0BO->getOperand(0))
5711 ->getOperand(0)->hasOneUse()) {
5712 Instruction *YS = BinaryOperator::createShl(
5713 Op0BO->getOperand(1), Op1,
5715 InsertNewInstBefore(YS, I); // (Y << C)
5717 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5718 V1->getName()+".mask");
5719 InsertNewInstBefore(XM, I); // X & (CC << C)
5721 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5729 // If the operand is an bitwise operator with a constant RHS, and the
5730 // shift is the only use, we can pull it out of the shift.
5731 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5732 bool isValid = true; // Valid only for And, Or, Xor
5733 bool highBitSet = false; // Transform if high bit of constant set?
5735 switch (Op0BO->getOpcode()) {
5736 default: isValid = false; break; // Do not perform transform!
5737 case Instruction::Add:
5738 isValid = isLeftShift;
5740 case Instruction::Or:
5741 case Instruction::Xor:
5744 case Instruction::And:
5749 // If this is a signed shift right, and the high bit is modified
5750 // by the logical operation, do not perform the transformation.
5751 // The highBitSet boolean indicates the value of the high bit of
5752 // the constant which would cause it to be modified for this
5755 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
5756 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
5760 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5762 Instruction *NewShift =
5763 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
5764 InsertNewInstBefore(NewShift, I);
5765 NewShift->takeName(Op0BO);
5767 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5774 // Find out if this is a shift of a shift by a constant.
5775 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
5776 if (ShiftOp && !ShiftOp->isShift())
5779 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5780 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5781 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
5782 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
5783 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
5784 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
5785 Value *X = ShiftOp->getOperand(0);
5787 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5788 if (AmtSum > TypeBits)
5791 const IntegerType *Ty = cast<IntegerType>(I.getType());
5793 // Check for (X << c1) << c2 and (X >> c1) >> c2
5794 if (I.getOpcode() == ShiftOp->getOpcode()) {
5795 return BinaryOperator::create(I.getOpcode(), X,
5796 ConstantInt::get(Ty, AmtSum));
5797 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
5798 I.getOpcode() == Instruction::AShr) {
5799 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
5800 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
5801 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
5802 I.getOpcode() == Instruction::LShr) {
5803 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
5804 Instruction *Shift =
5805 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
5806 InsertNewInstBefore(Shift, I);
5808 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
5809 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5812 // Okay, if we get here, one shift must be left, and the other shift must be
5813 // right. See if the amounts are equal.
5814 if (ShiftAmt1 == ShiftAmt2) {
5815 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
5816 if (I.getOpcode() == Instruction::Shl) {
5817 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
5818 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
5820 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
5821 if (I.getOpcode() == Instruction::LShr) {
5822 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
5823 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
5825 // We can simplify ((X << C) >>s C) into a trunc + sext.
5826 // NOTE: we could do this for any C, but that would make 'unusual' integer
5827 // types. For now, just stick to ones well-supported by the code
5829 const Type *SExtType = 0;
5830 switch (Ty->getBitWidth() - ShiftAmt1) {
5837 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
5842 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
5843 InsertNewInstBefore(NewTrunc, I);
5844 return new SExtInst(NewTrunc, Ty);
5846 // Otherwise, we can't handle it yet.
5847 } else if (ShiftAmt1 < ShiftAmt2) {
5848 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
5850 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
5851 if (I.getOpcode() == Instruction::Shl) {
5852 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5853 ShiftOp->getOpcode() == Instruction::AShr);
5854 Instruction *Shift =
5855 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5856 InsertNewInstBefore(Shift, I);
5858 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
5859 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5862 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
5863 if (I.getOpcode() == Instruction::LShr) {
5864 assert(ShiftOp->getOpcode() == Instruction::Shl);
5865 Instruction *Shift =
5866 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
5867 InsertNewInstBefore(Shift, I);
5869 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
5870 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5873 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
5875 assert(ShiftAmt2 < ShiftAmt1);
5876 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
5878 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
5879 if (I.getOpcode() == Instruction::Shl) {
5880 assert(ShiftOp->getOpcode() == Instruction::LShr ||
5881 ShiftOp->getOpcode() == Instruction::AShr);
5882 Instruction *Shift =
5883 BinaryOperator::create(ShiftOp->getOpcode(), X,
5884 ConstantInt::get(Ty, ShiftDiff));
5885 InsertNewInstBefore(Shift, I);
5887 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
5888 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5891 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
5892 if (I.getOpcode() == Instruction::LShr) {
5893 assert(ShiftOp->getOpcode() == Instruction::Shl);
5894 Instruction *Shift =
5895 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
5896 InsertNewInstBefore(Shift, I);
5898 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
5899 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
5902 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
5909 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5910 /// expression. If so, decompose it, returning some value X, such that Val is
5913 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5915 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5916 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5917 Offset = CI->getZExtValue();
5919 return ConstantInt::get(Type::Int32Ty, 0);
5920 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5921 if (I->getNumOperands() == 2) {
5922 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5923 if (I->getOpcode() == Instruction::Shl) {
5924 // This is a value scaled by '1 << the shift amt'.
5925 Scale = 1U << CUI->getZExtValue();
5927 return I->getOperand(0);
5928 } else if (I->getOpcode() == Instruction::Mul) {
5929 // This value is scaled by 'CUI'.
5930 Scale = CUI->getZExtValue();
5932 return I->getOperand(0);
5933 } else if (I->getOpcode() == Instruction::Add) {
5934 // We have X+C. Check to see if we really have (X*C2)+C1,
5935 // where C1 is divisible by C2.
5938 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5939 Offset += CUI->getZExtValue();
5940 if (SubScale > 1 && (Offset % SubScale == 0)) {
5949 // Otherwise, we can't look past this.
5956 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5957 /// try to eliminate the cast by moving the type information into the alloc.
5958 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5959 AllocationInst &AI) {
5960 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5961 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5963 // Remove any uses of AI that are dead.
5964 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5966 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5967 Instruction *User = cast<Instruction>(*UI++);
5968 if (isInstructionTriviallyDead(User)) {
5969 while (UI != E && *UI == User)
5970 ++UI; // If this instruction uses AI more than once, don't break UI.
5973 DOUT << "IC: DCE: " << *User;
5974 EraseInstFromFunction(*User);
5978 // Get the type really allocated and the type casted to.
5979 const Type *AllocElTy = AI.getAllocatedType();
5980 const Type *CastElTy = PTy->getElementType();
5981 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5983 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
5984 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
5985 if (CastElTyAlign < AllocElTyAlign) return 0;
5987 // If the allocation has multiple uses, only promote it if we are strictly
5988 // increasing the alignment of the resultant allocation. If we keep it the
5989 // same, we open the door to infinite loops of various kinds.
5990 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5992 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5993 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5994 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5996 // See if we can satisfy the modulus by pulling a scale out of the array
5998 unsigned ArraySizeScale;
6000 Value *NumElements = // See if the array size is a decomposable linear expr.
6001 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6003 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6005 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6006 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6008 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6013 // If the allocation size is constant, form a constant mul expression
6014 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6015 if (isa<ConstantInt>(NumElements))
6016 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6017 // otherwise multiply the amount and the number of elements
6018 else if (Scale != 1) {
6019 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6020 Amt = InsertNewInstBefore(Tmp, AI);
6024 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6025 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6026 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6027 Amt = InsertNewInstBefore(Tmp, AI);
6030 AllocationInst *New;
6031 if (isa<MallocInst>(AI))
6032 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6034 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6035 InsertNewInstBefore(New, AI);
6038 // If the allocation has multiple uses, insert a cast and change all things
6039 // that used it to use the new cast. This will also hack on CI, but it will
6041 if (!AI.hasOneUse()) {
6042 AddUsesToWorkList(AI);
6043 // New is the allocation instruction, pointer typed. AI is the original
6044 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6045 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6046 InsertNewInstBefore(NewCast, AI);
6047 AI.replaceAllUsesWith(NewCast);
6049 return ReplaceInstUsesWith(CI, New);
6052 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6053 /// and return it as type Ty without inserting any new casts and without
6054 /// changing the computed value. This is used by code that tries to decide
6055 /// whether promoting or shrinking integer operations to wider or smaller types
6056 /// will allow us to eliminate a truncate or extend.
6058 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6059 /// extension operation if Ty is larger.
6060 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6061 int &NumCastsRemoved) {
6062 // We can always evaluate constants in another type.
6063 if (isa<ConstantInt>(V))
6066 Instruction *I = dyn_cast<Instruction>(V);
6067 if (!I) return false;
6069 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6071 switch (I->getOpcode()) {
6072 case Instruction::Add:
6073 case Instruction::Sub:
6074 case Instruction::And:
6075 case Instruction::Or:
6076 case Instruction::Xor:
6077 if (!I->hasOneUse()) return false;
6078 // These operators can all arbitrarily be extended or truncated.
6079 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
6080 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
6082 case Instruction::Shl:
6083 if (!I->hasOneUse()) return false;
6084 // If we are truncating the result of this SHL, and if it's a shift of a
6085 // constant amount, we can always perform a SHL in a smaller type.
6086 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6087 uint32_t BitWidth = Ty->getBitWidth();
6088 if (BitWidth < OrigTy->getBitWidth() &&
6089 CI->getLimitedValue(BitWidth) < BitWidth)
6090 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6093 case Instruction::LShr:
6094 if (!I->hasOneUse()) return false;
6095 // If this is a truncate of a logical shr, we can truncate it to a smaller
6096 // lshr iff we know that the bits we would otherwise be shifting in are
6098 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6099 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6100 uint32_t BitWidth = Ty->getBitWidth();
6101 if (BitWidth < OrigBitWidth &&
6102 MaskedValueIsZero(I->getOperand(0),
6103 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6104 CI->getLimitedValue(BitWidth) < BitWidth) {
6105 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
6109 case Instruction::Trunc:
6110 case Instruction::ZExt:
6111 case Instruction::SExt:
6112 // If this is a cast from the destination type, we can trivially eliminate
6113 // it, and this will remove a cast overall.
6114 if (I->getOperand(0)->getType() == Ty) {
6115 // If the first operand is itself a cast, and is eliminable, do not count
6116 // this as an eliminable cast. We would prefer to eliminate those two
6118 if (isa<CastInst>(I->getOperand(0)))
6126 // TODO: Can handle more cases here.
6133 /// EvaluateInDifferentType - Given an expression that
6134 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6135 /// evaluate the expression.
6136 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6138 if (Constant *C = dyn_cast<Constant>(V))
6139 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6141 // Otherwise, it must be an instruction.
6142 Instruction *I = cast<Instruction>(V);
6143 Instruction *Res = 0;
6144 switch (I->getOpcode()) {
6145 case Instruction::Add:
6146 case Instruction::Sub:
6147 case Instruction::And:
6148 case Instruction::Or:
6149 case Instruction::Xor:
6150 case Instruction::AShr:
6151 case Instruction::LShr:
6152 case Instruction::Shl: {
6153 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6154 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6155 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6156 LHS, RHS, I->getName());
6159 case Instruction::Trunc:
6160 case Instruction::ZExt:
6161 case Instruction::SExt:
6162 case Instruction::BitCast:
6163 // If the source type of the cast is the type we're trying for then we can
6164 // just return the source. There's no need to insert it because its not new.
6165 if (I->getOperand(0)->getType() == Ty)
6166 return I->getOperand(0);
6168 // Some other kind of cast, which shouldn't happen, so just ..
6171 // TODO: Can handle more cases here.
6172 assert(0 && "Unreachable!");
6176 return InsertNewInstBefore(Res, *I);
6179 /// @brief Implement the transforms common to all CastInst visitors.
6180 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6181 Value *Src = CI.getOperand(0);
6183 // Casting undef to anything results in undef so might as just replace it and
6184 // get rid of the cast.
6185 if (isa<UndefValue>(Src)) // cast undef -> undef
6186 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
6188 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
6189 // eliminate it now.
6190 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6191 if (Instruction::CastOps opc =
6192 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6193 // The first cast (CSrc) is eliminable so we need to fix up or replace
6194 // the second cast (CI). CSrc will then have a good chance of being dead.
6195 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6199 // If casting the result of a getelementptr instruction with no offset, turn
6200 // this into a cast of the original pointer!
6202 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6203 bool AllZeroOperands = true;
6204 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
6205 if (!isa<Constant>(GEP->getOperand(i)) ||
6206 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
6207 AllZeroOperands = false;
6210 if (AllZeroOperands) {
6211 // Changing the cast operand is usually not a good idea but it is safe
6212 // here because the pointer operand is being replaced with another
6213 // pointer operand so the opcode doesn't need to change.
6214 CI.setOperand(0, GEP->getOperand(0));
6219 // If we are casting a malloc or alloca to a pointer to a type of the same
6220 // size, rewrite the allocation instruction to allocate the "right" type.
6221 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6222 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6225 // If we are casting a select then fold the cast into the select
6226 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6227 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6230 // If we are casting a PHI then fold the cast into the PHI
6231 if (isa<PHINode>(Src))
6232 if (Instruction *NV = FoldOpIntoPhi(CI))
6238 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6239 /// integer types. This function implements the common transforms for all those
6241 /// @brief Implement the transforms common to CastInst with integer operands
6242 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6243 if (Instruction *Result = commonCastTransforms(CI))
6246 Value *Src = CI.getOperand(0);
6247 const Type *SrcTy = Src->getType();
6248 const Type *DestTy = CI.getType();
6249 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6250 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6252 // See if we can simplify any instructions used by the LHS whose sole
6253 // purpose is to compute bits we don't care about.
6254 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6255 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6256 KnownZero, KnownOne))
6259 // If the source isn't an instruction or has more than one use then we
6260 // can't do anything more.
6261 Instruction *SrcI = dyn_cast<Instruction>(Src);
6262 if (!SrcI || !Src->hasOneUse())
6265 // Attempt to propagate the cast into the instruction for int->int casts.
6266 int NumCastsRemoved = 0;
6267 if (!isa<BitCastInst>(CI) &&
6268 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6270 // If this cast is a truncate, evaluting in a different type always
6271 // eliminates the cast, so it is always a win. If this is a noop-cast
6272 // this just removes a noop cast which isn't pointful, but simplifies
6273 // the code. If this is a zero-extension, we need to do an AND to
6274 // maintain the clear top-part of the computation, so we require that
6275 // the input have eliminated at least one cast. If this is a sign
6276 // extension, we insert two new casts (to do the extension) so we
6277 // require that two casts have been eliminated.
6279 switch (CI.getOpcode()) {
6281 // All the others use floating point so we shouldn't actually
6282 // get here because of the check above.
6283 assert(0 && "Unknown cast type");
6284 case Instruction::Trunc:
6287 case Instruction::ZExt:
6288 DoXForm = NumCastsRemoved >= 1;
6290 case Instruction::SExt:
6291 DoXForm = NumCastsRemoved >= 2;
6293 case Instruction::BitCast:
6299 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6300 CI.getOpcode() == Instruction::SExt);
6301 assert(Res->getType() == DestTy);
6302 switch (CI.getOpcode()) {
6303 default: assert(0 && "Unknown cast type!");
6304 case Instruction::Trunc:
6305 case Instruction::BitCast:
6306 // Just replace this cast with the result.
6307 return ReplaceInstUsesWith(CI, Res);
6308 case Instruction::ZExt: {
6309 // We need to emit an AND to clear the high bits.
6310 assert(SrcBitSize < DestBitSize && "Not a zext?");
6311 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6313 return BinaryOperator::createAnd(Res, C);
6315 case Instruction::SExt:
6316 // We need to emit a cast to truncate, then a cast to sext.
6317 return CastInst::create(Instruction::SExt,
6318 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6324 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6325 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6327 switch (SrcI->getOpcode()) {
6328 case Instruction::Add:
6329 case Instruction::Mul:
6330 case Instruction::And:
6331 case Instruction::Or:
6332 case Instruction::Xor:
6333 // If we are discarding information, rewrite.
6334 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6335 // Don't insert two casts if they cannot be eliminated. We allow
6336 // two casts to be inserted if the sizes are the same. This could
6337 // only be converting signedness, which is a noop.
6338 if (DestBitSize == SrcBitSize ||
6339 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6340 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6341 Instruction::CastOps opcode = CI.getOpcode();
6342 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6343 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6344 return BinaryOperator::create(
6345 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6349 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6350 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6351 SrcI->getOpcode() == Instruction::Xor &&
6352 Op1 == ConstantInt::getTrue() &&
6353 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6354 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6355 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6358 case Instruction::SDiv:
6359 case Instruction::UDiv:
6360 case Instruction::SRem:
6361 case Instruction::URem:
6362 // If we are just changing the sign, rewrite.
6363 if (DestBitSize == SrcBitSize) {
6364 // Don't insert two casts if they cannot be eliminated. We allow
6365 // two casts to be inserted if the sizes are the same. This could
6366 // only be converting signedness, which is a noop.
6367 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6368 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6369 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6371 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6373 return BinaryOperator::create(
6374 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6379 case Instruction::Shl:
6380 // Allow changing the sign of the source operand. Do not allow
6381 // changing the size of the shift, UNLESS the shift amount is a
6382 // constant. We must not change variable sized shifts to a smaller
6383 // size, because it is undefined to shift more bits out than exist
6385 if (DestBitSize == SrcBitSize ||
6386 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6387 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6388 Instruction::BitCast : Instruction::Trunc);
6389 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6390 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6391 return BinaryOperator::createShl(Op0c, Op1c);
6394 case Instruction::AShr:
6395 // If this is a signed shr, and if all bits shifted in are about to be
6396 // truncated off, turn it into an unsigned shr to allow greater
6398 if (DestBitSize < SrcBitSize &&
6399 isa<ConstantInt>(Op1)) {
6400 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6401 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6402 // Insert the new logical shift right.
6403 return BinaryOperator::createLShr(Op0, Op1);
6408 case Instruction::ICmp:
6409 // If we are just checking for a icmp eq of a single bit and casting it
6410 // to an integer, then shift the bit to the appropriate place and then
6411 // cast to integer to avoid the comparison.
6412 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6413 const APInt& Op1CV = Op1C->getValue();
6414 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6415 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6416 // cast (X == 1) to int --> X iff X has only the low bit set.
6417 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6418 // cast (X != 0) to int --> X iff X has only the low bit set.
6419 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6420 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6421 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6422 if (Op1CV == 0 || Op1CV.isPowerOf2()) {
6423 // If Op1C some other power of two, convert:
6424 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6425 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6426 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6427 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6429 // This only works for EQ and NE
6430 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6431 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6434 APInt KnownZeroMask(KnownZero ^ TypeMask);
6435 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6436 bool isNE = pred == ICmpInst::ICMP_NE;
6437 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6438 // (X&4) == 2 --> false
6439 // (X&4) != 2 --> true
6440 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6441 Res = ConstantExpr::getZExt(Res, CI.getType());
6442 return ReplaceInstUsesWith(CI, Res);
6445 uint32_t ShiftAmt = KnownZeroMask.logBase2();
6448 // Perform a logical shr by shiftamt.
6449 // Insert the shift to put the result in the low bit.
6450 In = InsertNewInstBefore(
6451 BinaryOperator::createLShr(In,
6452 ConstantInt::get(In->getType(), ShiftAmt),
6453 In->getName()+".lobit"), CI);
6456 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6457 Constant *One = ConstantInt::get(In->getType(), 1);
6458 In = BinaryOperator::createXor(In, One, "tmp");
6459 InsertNewInstBefore(cast<Instruction>(In), CI);
6462 if (CI.getType() == In->getType())
6463 return ReplaceInstUsesWith(CI, In);
6465 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6474 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6475 if (Instruction *Result = commonIntCastTransforms(CI))
6478 Value *Src = CI.getOperand(0);
6479 const Type *Ty = CI.getType();
6480 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6481 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6483 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6484 switch (SrcI->getOpcode()) {
6486 case Instruction::LShr:
6487 // We can shrink lshr to something smaller if we know the bits shifted in
6488 // are already zeros.
6489 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6490 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6492 // Get a mask for the bits shifting in.
6493 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6494 Value* SrcIOp0 = SrcI->getOperand(0);
6495 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6496 if (ShAmt >= DestBitWidth) // All zeros.
6497 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6499 // Okay, we can shrink this. Truncate the input, then return a new
6501 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6502 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6504 return BinaryOperator::createLShr(V1, V2);
6506 } else { // This is a variable shr.
6508 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6509 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6510 // loop-invariant and CSE'd.
6511 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6512 Value *One = ConstantInt::get(SrcI->getType(), 1);
6514 Value *V = InsertNewInstBefore(
6515 BinaryOperator::createShl(One, SrcI->getOperand(1),
6517 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6518 SrcI->getOperand(0),
6520 Value *Zero = Constant::getNullValue(V->getType());
6521 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6531 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6532 // If one of the common conversion will work ..
6533 if (Instruction *Result = commonIntCastTransforms(CI))
6536 Value *Src = CI.getOperand(0);
6538 // If this is a cast of a cast
6539 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6540 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6541 // types and if the sizes are just right we can convert this into a logical
6542 // 'and' which will be much cheaper than the pair of casts.
6543 if (isa<TruncInst>(CSrc)) {
6544 // Get the sizes of the types involved
6545 Value *A = CSrc->getOperand(0);
6546 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6547 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6548 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6549 // If we're actually extending zero bits and the trunc is a no-op
6550 if (MidSize < DstSize && SrcSize == DstSize) {
6551 // Replace both of the casts with an And of the type mask.
6552 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6553 Constant *AndConst = ConstantInt::get(AndValue);
6555 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6556 // Unfortunately, if the type changed, we need to cast it back.
6557 if (And->getType() != CI.getType()) {
6558 And->setName(CSrc->getName()+".mask");
6559 InsertNewInstBefore(And, CI);
6560 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6570 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6571 return commonIntCastTransforms(CI);
6574 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6575 return commonCastTransforms(CI);
6578 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6579 return commonCastTransforms(CI);
6582 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6583 return commonCastTransforms(CI);
6586 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6587 return commonCastTransforms(CI);
6590 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6591 return commonCastTransforms(CI);
6594 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6595 return commonCastTransforms(CI);
6598 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6599 return commonCastTransforms(CI);
6602 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6603 return commonCastTransforms(CI);
6606 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6608 // If the operands are integer typed then apply the integer transforms,
6609 // otherwise just apply the common ones.
6610 Value *Src = CI.getOperand(0);
6611 const Type *SrcTy = Src->getType();
6612 const Type *DestTy = CI.getType();
6614 if (SrcTy->isInteger() && DestTy->isInteger()) {
6615 if (Instruction *Result = commonIntCastTransforms(CI))
6618 if (Instruction *Result = commonCastTransforms(CI))
6623 // Get rid of casts from one type to the same type. These are useless and can
6624 // be replaced by the operand.
6625 if (DestTy == Src->getType())
6626 return ReplaceInstUsesWith(CI, Src);
6628 // If the source and destination are pointers, and this cast is equivalent to
6629 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6630 // This can enhance SROA and other transforms that want type-safe pointers.
6631 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6632 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6633 const Type *DstElTy = DstPTy->getElementType();
6634 const Type *SrcElTy = SrcPTy->getElementType();
6636 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6637 unsigned NumZeros = 0;
6638 while (SrcElTy != DstElTy &&
6639 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6640 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6641 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6645 // If we found a path from the src to dest, create the getelementptr now.
6646 if (SrcElTy == DstElTy) {
6647 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
6648 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
6653 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6654 if (SVI->hasOneUse()) {
6655 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6656 // a bitconvert to a vector with the same # elts.
6657 if (isa<VectorType>(DestTy) &&
6658 cast<VectorType>(DestTy)->getNumElements() ==
6659 SVI->getType()->getNumElements()) {
6661 // If either of the operands is a cast from CI.getType(), then
6662 // evaluating the shuffle in the casted destination's type will allow
6663 // us to eliminate at least one cast.
6664 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6665 Tmp->getOperand(0)->getType() == DestTy) ||
6666 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6667 Tmp->getOperand(0)->getType() == DestTy)) {
6668 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6669 SVI->getOperand(0), DestTy, &CI);
6670 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6671 SVI->getOperand(1), DestTy, &CI);
6672 // Return a new shuffle vector. Use the same element ID's, as we
6673 // know the vector types match #elts.
6674 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6682 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6684 /// %D = select %cond, %C, %A
6686 /// %C = select %cond, %B, 0
6689 /// Assuming that the specified instruction is an operand to the select, return
6690 /// a bitmask indicating which operands of this instruction are foldable if they
6691 /// equal the other incoming value of the select.
6693 static unsigned GetSelectFoldableOperands(Instruction *I) {
6694 switch (I->getOpcode()) {
6695 case Instruction::Add:
6696 case Instruction::Mul:
6697 case Instruction::And:
6698 case Instruction::Or:
6699 case Instruction::Xor:
6700 return 3; // Can fold through either operand.
6701 case Instruction::Sub: // Can only fold on the amount subtracted.
6702 case Instruction::Shl: // Can only fold on the shift amount.
6703 case Instruction::LShr:
6704 case Instruction::AShr:
6707 return 0; // Cannot fold
6711 /// GetSelectFoldableConstant - For the same transformation as the previous
6712 /// function, return the identity constant that goes into the select.
6713 static Constant *GetSelectFoldableConstant(Instruction *I) {
6714 switch (I->getOpcode()) {
6715 default: assert(0 && "This cannot happen!"); abort();
6716 case Instruction::Add:
6717 case Instruction::Sub:
6718 case Instruction::Or:
6719 case Instruction::Xor:
6720 case Instruction::Shl:
6721 case Instruction::LShr:
6722 case Instruction::AShr:
6723 return Constant::getNullValue(I->getType());
6724 case Instruction::And:
6725 return ConstantInt::getAllOnesValue(I->getType());
6726 case Instruction::Mul:
6727 return ConstantInt::get(I->getType(), 1);
6731 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6732 /// have the same opcode and only one use each. Try to simplify this.
6733 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6735 if (TI->getNumOperands() == 1) {
6736 // If this is a non-volatile load or a cast from the same type,
6739 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6742 return 0; // unknown unary op.
6745 // Fold this by inserting a select from the input values.
6746 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6747 FI->getOperand(0), SI.getName()+".v");
6748 InsertNewInstBefore(NewSI, SI);
6749 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6753 // Only handle binary operators here.
6754 if (!isa<BinaryOperator>(TI))
6757 // Figure out if the operations have any operands in common.
6758 Value *MatchOp, *OtherOpT, *OtherOpF;
6760 if (TI->getOperand(0) == FI->getOperand(0)) {
6761 MatchOp = TI->getOperand(0);
6762 OtherOpT = TI->getOperand(1);
6763 OtherOpF = FI->getOperand(1);
6764 MatchIsOpZero = true;
6765 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6766 MatchOp = TI->getOperand(1);
6767 OtherOpT = TI->getOperand(0);
6768 OtherOpF = FI->getOperand(0);
6769 MatchIsOpZero = false;
6770 } else if (!TI->isCommutative()) {
6772 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6773 MatchOp = TI->getOperand(0);
6774 OtherOpT = TI->getOperand(1);
6775 OtherOpF = FI->getOperand(0);
6776 MatchIsOpZero = true;
6777 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6778 MatchOp = TI->getOperand(1);
6779 OtherOpT = TI->getOperand(0);
6780 OtherOpF = FI->getOperand(1);
6781 MatchIsOpZero = true;
6786 // If we reach here, they do have operations in common.
6787 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6788 OtherOpF, SI.getName()+".v");
6789 InsertNewInstBefore(NewSI, SI);
6791 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6793 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6795 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6797 assert(0 && "Shouldn't get here");
6801 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6802 Value *CondVal = SI.getCondition();
6803 Value *TrueVal = SI.getTrueValue();
6804 Value *FalseVal = SI.getFalseValue();
6806 // select true, X, Y -> X
6807 // select false, X, Y -> Y
6808 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6809 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
6811 // select C, X, X -> X
6812 if (TrueVal == FalseVal)
6813 return ReplaceInstUsesWith(SI, TrueVal);
6815 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6816 return ReplaceInstUsesWith(SI, FalseVal);
6817 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6818 return ReplaceInstUsesWith(SI, TrueVal);
6819 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6820 if (isa<Constant>(TrueVal))
6821 return ReplaceInstUsesWith(SI, TrueVal);
6823 return ReplaceInstUsesWith(SI, FalseVal);
6826 if (SI.getType() == Type::Int1Ty) {
6827 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
6828 if (C->getZExtValue()) {
6829 // Change: A = select B, true, C --> A = or B, C
6830 return BinaryOperator::createOr(CondVal, FalseVal);
6832 // Change: A = select B, false, C --> A = and !B, C
6834 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6835 "not."+CondVal->getName()), SI);
6836 return BinaryOperator::createAnd(NotCond, FalseVal);
6838 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
6839 if (C->getZExtValue() == false) {
6840 // Change: A = select B, C, false --> A = and B, C
6841 return BinaryOperator::createAnd(CondVal, TrueVal);
6843 // Change: A = select B, C, true --> A = or !B, C
6845 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6846 "not."+CondVal->getName()), SI);
6847 return BinaryOperator::createOr(NotCond, TrueVal);
6852 // Selecting between two integer constants?
6853 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6854 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6855 // select C, 1, 0 -> cast C to int
6856 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
6857 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6858 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
6859 // select C, 0, 1 -> cast !C to int
6861 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6862 "not."+CondVal->getName()), SI);
6863 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6866 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6868 // (x <s 0) ? -1 : 0 -> ashr x, 31
6869 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6870 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
6871 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6872 bool CanXForm = false;
6873 if (IC->isSignedPredicate())
6874 CanXForm = CmpCst->isZero() &&
6875 IC->getPredicate() == ICmpInst::ICMP_SLT;
6877 uint32_t Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6878 CanXForm = CmpCst->getValue() == APInt::getSignedMaxValue(Bits) &&
6879 IC->getPredicate() == ICmpInst::ICMP_UGT;
6883 // The comparison constant and the result are not neccessarily the
6884 // same width. Make an all-ones value by inserting a AShr.
6885 Value *X = IC->getOperand(0);
6886 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
6887 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
6888 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
6890 InsertNewInstBefore(SRA, SI);
6892 // Finally, convert to the type of the select RHS. We figure out
6893 // if this requires a SExt, Trunc or BitCast based on the sizes.
6894 Instruction::CastOps opc = Instruction::BitCast;
6895 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
6896 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
6897 if (SRASize < SISize)
6898 opc = Instruction::SExt;
6899 else if (SRASize > SISize)
6900 opc = Instruction::Trunc;
6901 return CastInst::create(opc, SRA, SI.getType());
6906 // If one of the constants is zero (we know they can't both be) and we
6907 // have a fcmp instruction with zero, and we have an 'and' with the
6908 // non-constant value, eliminate this whole mess. This corresponds to
6909 // cases like this: ((X & 27) ? 27 : 0)
6910 if (TrueValC->isZero() || FalseValC->isZero())
6911 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6912 cast<Constant>(IC->getOperand(1))->isNullValue())
6913 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6914 if (ICA->getOpcode() == Instruction::And &&
6915 isa<ConstantInt>(ICA->getOperand(1)) &&
6916 (ICA->getOperand(1) == TrueValC ||
6917 ICA->getOperand(1) == FalseValC) &&
6918 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6919 // Okay, now we know that everything is set up, we just don't
6920 // know whether we have a icmp_ne or icmp_eq and whether the
6921 // true or false val is the zero.
6922 bool ShouldNotVal = !TrueValC->isZero();
6923 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6926 V = InsertNewInstBefore(BinaryOperator::create(
6927 Instruction::Xor, V, ICA->getOperand(1)), SI);
6928 return ReplaceInstUsesWith(SI, V);
6933 // See if we are selecting two values based on a comparison of the two values.
6934 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6935 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6936 // Transform (X == Y) ? X : Y -> Y
6937 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6938 return ReplaceInstUsesWith(SI, FalseVal);
6939 // Transform (X != Y) ? X : Y -> X
6940 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6941 return ReplaceInstUsesWith(SI, TrueVal);
6942 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6944 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6945 // Transform (X == Y) ? Y : X -> X
6946 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6947 return ReplaceInstUsesWith(SI, FalseVal);
6948 // Transform (X != Y) ? Y : X -> Y
6949 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6950 return ReplaceInstUsesWith(SI, TrueVal);
6951 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6955 // See if we are selecting two values based on a comparison of the two values.
6956 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6957 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6958 // Transform (X == Y) ? X : Y -> Y
6959 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6960 return ReplaceInstUsesWith(SI, FalseVal);
6961 // Transform (X != Y) ? X : Y -> X
6962 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6963 return ReplaceInstUsesWith(SI, TrueVal);
6964 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6966 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6967 // Transform (X == Y) ? Y : X -> X
6968 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6969 return ReplaceInstUsesWith(SI, FalseVal);
6970 // Transform (X != Y) ? Y : X -> Y
6971 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6972 return ReplaceInstUsesWith(SI, TrueVal);
6973 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6977 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6978 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6979 if (TI->hasOneUse() && FI->hasOneUse()) {
6980 Instruction *AddOp = 0, *SubOp = 0;
6982 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6983 if (TI->getOpcode() == FI->getOpcode())
6984 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6987 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6988 // even legal for FP.
6989 if (TI->getOpcode() == Instruction::Sub &&
6990 FI->getOpcode() == Instruction::Add) {
6991 AddOp = FI; SubOp = TI;
6992 } else if (FI->getOpcode() == Instruction::Sub &&
6993 TI->getOpcode() == Instruction::Add) {
6994 AddOp = TI; SubOp = FI;
6998 Value *OtherAddOp = 0;
6999 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7000 OtherAddOp = AddOp->getOperand(1);
7001 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7002 OtherAddOp = AddOp->getOperand(0);
7006 // So at this point we know we have (Y -> OtherAddOp):
7007 // select C, (add X, Y), (sub X, Z)
7008 Value *NegVal; // Compute -Z
7009 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7010 NegVal = ConstantExpr::getNeg(C);
7012 NegVal = InsertNewInstBefore(
7013 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7016 Value *NewTrueOp = OtherAddOp;
7017 Value *NewFalseOp = NegVal;
7019 std::swap(NewTrueOp, NewFalseOp);
7020 Instruction *NewSel =
7021 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7023 NewSel = InsertNewInstBefore(NewSel, SI);
7024 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7029 // See if we can fold the select into one of our operands.
7030 if (SI.getType()->isInteger()) {
7031 // See the comment above GetSelectFoldableOperands for a description of the
7032 // transformation we are doing here.
7033 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7034 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7035 !isa<Constant>(FalseVal))
7036 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7037 unsigned OpToFold = 0;
7038 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7040 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7045 Constant *C = GetSelectFoldableConstant(TVI);
7046 Instruction *NewSel =
7047 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7048 InsertNewInstBefore(NewSel, SI);
7049 NewSel->takeName(TVI);
7050 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7051 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7053 assert(0 && "Unknown instruction!!");
7058 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7059 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7060 !isa<Constant>(TrueVal))
7061 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7062 unsigned OpToFold = 0;
7063 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7065 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7070 Constant *C = GetSelectFoldableConstant(FVI);
7071 Instruction *NewSel =
7072 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7073 InsertNewInstBefore(NewSel, SI);
7074 NewSel->takeName(FVI);
7075 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7076 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7078 assert(0 && "Unknown instruction!!");
7083 if (BinaryOperator::isNot(CondVal)) {
7084 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7085 SI.setOperand(1, FalseVal);
7086 SI.setOperand(2, TrueVal);
7093 /// GetKnownAlignment - If the specified pointer has an alignment that we can
7094 /// determine, return it, otherwise return 0.
7095 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
7096 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7097 unsigned Align = GV->getAlignment();
7098 if (Align == 0 && TD)
7099 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7101 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7102 unsigned Align = AI->getAlignment();
7103 if (Align == 0 && TD) {
7104 if (isa<AllocaInst>(AI))
7105 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7106 else if (isa<MallocInst>(AI)) {
7107 // Malloc returns maximally aligned memory.
7108 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7111 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7114 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7118 } else if (isa<BitCastInst>(V) ||
7119 (isa<ConstantExpr>(V) &&
7120 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7121 User *CI = cast<User>(V);
7122 if (isa<PointerType>(CI->getOperand(0)->getType()))
7123 return GetKnownAlignment(CI->getOperand(0), TD);
7125 } else if (isa<GetElementPtrInst>(V) ||
7126 (isa<ConstantExpr>(V) &&
7127 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
7128 User *GEPI = cast<User>(V);
7129 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
7130 if (BaseAlignment == 0) return 0;
7132 // If all indexes are zero, it is just the alignment of the base pointer.
7133 bool AllZeroOperands = true;
7134 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7135 if (!isa<Constant>(GEPI->getOperand(i)) ||
7136 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7137 AllZeroOperands = false;
7140 if (AllZeroOperands)
7141 return BaseAlignment;
7143 // Otherwise, if the base alignment is >= the alignment we expect for the
7144 // base pointer type, then we know that the resultant pointer is aligned at
7145 // least as much as its type requires.
7148 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7149 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7150 if (TD->getABITypeAlignment(PtrTy->getElementType())
7152 const Type *GEPTy = GEPI->getType();
7153 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7154 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
7162 /// visitCallInst - CallInst simplification. This mostly only handles folding
7163 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7164 /// the heavy lifting.
7166 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7167 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7168 if (!II) return visitCallSite(&CI);
7170 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7172 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7173 bool Changed = false;
7175 // memmove/cpy/set of zero bytes is a noop.
7176 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7177 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7179 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7180 if (CI->getZExtValue() == 1) {
7181 // Replace the instruction with just byte operations. We would
7182 // transform other cases to loads/stores, but we don't know if
7183 // alignment is sufficient.
7187 // If we have a memmove and the source operation is a constant global,
7188 // then the source and dest pointers can't alias, so we can change this
7189 // into a call to memcpy.
7190 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7191 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7192 if (GVSrc->isConstant()) {
7193 Module *M = CI.getParent()->getParent()->getParent();
7195 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7197 Name = "llvm.memcpy.i32";
7199 Name = "llvm.memcpy.i64";
7200 Constant *MemCpy = M->getOrInsertFunction(Name,
7201 CI.getCalledFunction()->getFunctionType());
7202 CI.setOperand(0, MemCpy);
7207 // If we can determine a pointer alignment that is bigger than currently
7208 // set, update the alignment.
7209 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7210 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7211 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7212 unsigned Align = std::min(Alignment1, Alignment2);
7213 if (MI->getAlignment()->getZExtValue() < Align) {
7214 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7217 } else if (isa<MemSetInst>(MI)) {
7218 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7219 if (MI->getAlignment()->getZExtValue() < Alignment) {
7220 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7225 if (Changed) return II;
7227 switch (II->getIntrinsicID()) {
7229 case Intrinsic::ppc_altivec_lvx:
7230 case Intrinsic::ppc_altivec_lvxl:
7231 case Intrinsic::x86_sse_loadu_ps:
7232 case Intrinsic::x86_sse2_loadu_pd:
7233 case Intrinsic::x86_sse2_loadu_dq:
7234 // Turn PPC lvx -> load if the pointer is known aligned.
7235 // Turn X86 loadups -> load if the pointer is known aligned.
7236 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7237 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7238 PointerType::get(II->getType()), CI);
7239 return new LoadInst(Ptr);
7242 case Intrinsic::ppc_altivec_stvx:
7243 case Intrinsic::ppc_altivec_stvxl:
7244 // Turn stvx -> store if the pointer is known aligned.
7245 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7246 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7247 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7249 return new StoreInst(II->getOperand(1), Ptr);
7252 case Intrinsic::x86_sse_storeu_ps:
7253 case Intrinsic::x86_sse2_storeu_pd:
7254 case Intrinsic::x86_sse2_storeu_dq:
7255 case Intrinsic::x86_sse2_storel_dq:
7256 // Turn X86 storeu -> store if the pointer is known aligned.
7257 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7258 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7259 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7261 return new StoreInst(II->getOperand(2), Ptr);
7265 case Intrinsic::x86_sse_cvttss2si: {
7266 // These intrinsics only demands the 0th element of its input vector. If
7267 // we can simplify the input based on that, do so now.
7269 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7271 II->setOperand(1, V);
7277 case Intrinsic::ppc_altivec_vperm:
7278 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7279 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7280 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7282 // Check that all of the elements are integer constants or undefs.
7283 bool AllEltsOk = true;
7284 for (unsigned i = 0; i != 16; ++i) {
7285 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7286 !isa<UndefValue>(Mask->getOperand(i))) {
7293 // Cast the input vectors to byte vectors.
7294 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7295 II->getOperand(1), Mask->getType(), CI);
7296 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7297 II->getOperand(2), Mask->getType(), CI);
7298 Value *Result = UndefValue::get(Op0->getType());
7300 // Only extract each element once.
7301 Value *ExtractedElts[32];
7302 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7304 for (unsigned i = 0; i != 16; ++i) {
7305 if (isa<UndefValue>(Mask->getOperand(i)))
7307 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7308 Idx &= 31; // Match the hardware behavior.
7310 if (ExtractedElts[Idx] == 0) {
7312 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7313 InsertNewInstBefore(Elt, CI);
7314 ExtractedElts[Idx] = Elt;
7317 // Insert this value into the result vector.
7318 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7319 InsertNewInstBefore(cast<Instruction>(Result), CI);
7321 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7326 case Intrinsic::stackrestore: {
7327 // If the save is right next to the restore, remove the restore. This can
7328 // happen when variable allocas are DCE'd.
7329 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7330 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7331 BasicBlock::iterator BI = SS;
7333 return EraseInstFromFunction(CI);
7337 // If the stack restore is in a return/unwind block and if there are no
7338 // allocas or calls between the restore and the return, nuke the restore.
7339 TerminatorInst *TI = II->getParent()->getTerminator();
7340 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7341 BasicBlock::iterator BI = II;
7342 bool CannotRemove = false;
7343 for (++BI; &*BI != TI; ++BI) {
7344 if (isa<AllocaInst>(BI) ||
7345 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7346 CannotRemove = true;
7351 return EraseInstFromFunction(CI);
7358 return visitCallSite(II);
7361 // InvokeInst simplification
7363 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7364 return visitCallSite(&II);
7367 // visitCallSite - Improvements for call and invoke instructions.
7369 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7370 bool Changed = false;
7372 // If the callee is a constexpr cast of a function, attempt to move the cast
7373 // to the arguments of the call/invoke.
7374 if (transformConstExprCastCall(CS)) return 0;
7376 Value *Callee = CS.getCalledValue();
7378 if (Function *CalleeF = dyn_cast<Function>(Callee))
7379 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7380 Instruction *OldCall = CS.getInstruction();
7381 // If the call and callee calling conventions don't match, this call must
7382 // be unreachable, as the call is undefined.
7383 new StoreInst(ConstantInt::getTrue(),
7384 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7385 if (!OldCall->use_empty())
7386 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7387 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7388 return EraseInstFromFunction(*OldCall);
7392 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7393 // This instruction is not reachable, just remove it. We insert a store to
7394 // undef so that we know that this code is not reachable, despite the fact
7395 // that we can't modify the CFG here.
7396 new StoreInst(ConstantInt::getTrue(),
7397 UndefValue::get(PointerType::get(Type::Int1Ty)),
7398 CS.getInstruction());
7400 if (!CS.getInstruction()->use_empty())
7401 CS.getInstruction()->
7402 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7404 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7405 // Don't break the CFG, insert a dummy cond branch.
7406 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7407 ConstantInt::getTrue(), II);
7409 return EraseInstFromFunction(*CS.getInstruction());
7412 const PointerType *PTy = cast<PointerType>(Callee->getType());
7413 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7414 if (FTy->isVarArg()) {
7415 // See if we can optimize any arguments passed through the varargs area of
7417 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7418 E = CS.arg_end(); I != E; ++I)
7419 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7420 // If this cast does not effect the value passed through the varargs
7421 // area, we can eliminate the use of the cast.
7422 Value *Op = CI->getOperand(0);
7423 if (CI->isLosslessCast()) {
7430 return Changed ? CS.getInstruction() : 0;
7433 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7434 // attempt to move the cast to the arguments of the call/invoke.
7436 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7437 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7438 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7439 if (CE->getOpcode() != Instruction::BitCast ||
7440 !isa<Function>(CE->getOperand(0)))
7442 Function *Callee = cast<Function>(CE->getOperand(0));
7443 Instruction *Caller = CS.getInstruction();
7445 // Okay, this is a cast from a function to a different type. Unless doing so
7446 // would cause a type conversion of one of our arguments, change this call to
7447 // be a direct call with arguments casted to the appropriate types.
7449 const FunctionType *FT = Callee->getFunctionType();
7450 const Type *OldRetTy = Caller->getType();
7452 // Check to see if we are changing the return type...
7453 if (OldRetTy != FT->getReturnType()) {
7454 if (Callee->isDeclaration() && !Caller->use_empty() &&
7455 // Conversion is ok if changing from pointer to int of same size.
7456 !(isa<PointerType>(FT->getReturnType()) &&
7457 TD->getIntPtrType() == OldRetTy))
7458 return false; // Cannot transform this return value.
7460 // If the callsite is an invoke instruction, and the return value is used by
7461 // a PHI node in a successor, we cannot change the return type of the call
7462 // because there is no place to put the cast instruction (without breaking
7463 // the critical edge). Bail out in this case.
7464 if (!Caller->use_empty())
7465 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7466 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7468 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7469 if (PN->getParent() == II->getNormalDest() ||
7470 PN->getParent() == II->getUnwindDest())
7474 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7475 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7477 CallSite::arg_iterator AI = CS.arg_begin();
7478 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7479 const Type *ParamTy = FT->getParamType(i);
7480 const Type *ActTy = (*AI)->getType();
7481 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7482 //Some conversions are safe even if we do not have a body.
7483 //Either we can cast directly, or we can upconvert the argument
7484 bool isConvertible = ActTy == ParamTy ||
7485 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7486 (ParamTy->isInteger() && ActTy->isInteger() &&
7487 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7488 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7489 && c->getValue().isStrictlyPositive());
7490 if (Callee->isDeclaration() && !isConvertible) return false;
7492 // Most other conversions can be done if we have a body, even if these
7493 // lose information, e.g. int->short.
7494 // Some conversions cannot be done at all, e.g. float to pointer.
7495 // Logic here parallels CastInst::getCastOpcode (the design there
7496 // requires legality checks like this be done before calling it).
7497 if (ParamTy->isInteger()) {
7498 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7499 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7502 if (!ActTy->isInteger() && !ActTy->isFloatingPoint() &&
7503 !isa<PointerType>(ActTy))
7505 } else if (ParamTy->isFloatingPoint()) {
7506 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7507 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7510 if (!ActTy->isInteger() && !ActTy->isFloatingPoint())
7512 } else if (const VectorType *VParamTy = dyn_cast<VectorType>(ParamTy)) {
7513 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7514 if (VActTy->getBitWidth() != VParamTy->getBitWidth())
7517 if (VParamTy->getBitWidth() != ActTy->getPrimitiveSizeInBits())
7519 } else if (isa<PointerType>(ParamTy)) {
7520 if (!ActTy->isInteger() && !isa<PointerType>(ActTy))
7527 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7528 Callee->isDeclaration())
7529 return false; // Do not delete arguments unless we have a function body...
7531 // Okay, we decided that this is a safe thing to do: go ahead and start
7532 // inserting cast instructions as necessary...
7533 std::vector<Value*> Args;
7534 Args.reserve(NumActualArgs);
7536 AI = CS.arg_begin();
7537 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7538 const Type *ParamTy = FT->getParamType(i);
7539 if ((*AI)->getType() == ParamTy) {
7540 Args.push_back(*AI);
7542 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7543 false, ParamTy, false);
7544 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7545 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7549 // If the function takes more arguments than the call was taking, add them
7551 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7552 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7554 // If we are removing arguments to the function, emit an obnoxious warning...
7555 if (FT->getNumParams() < NumActualArgs)
7556 if (!FT->isVarArg()) {
7557 cerr << "WARNING: While resolving call to function '"
7558 << Callee->getName() << "' arguments were dropped!\n";
7560 // Add all of the arguments in their promoted form to the arg list...
7561 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7562 const Type *PTy = getPromotedType((*AI)->getType());
7563 if (PTy != (*AI)->getType()) {
7564 // Must promote to pass through va_arg area!
7565 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7567 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7568 InsertNewInstBefore(Cast, *Caller);
7569 Args.push_back(Cast);
7571 Args.push_back(*AI);
7576 if (FT->getReturnType() == Type::VoidTy)
7577 Caller->setName(""); // Void type should not have a name.
7580 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7581 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7582 &Args[0], Args.size(), Caller->getName(), Caller);
7583 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7585 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
7586 if (cast<CallInst>(Caller)->isTailCall())
7587 cast<CallInst>(NC)->setTailCall();
7588 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7591 // Insert a cast of the return type as necessary.
7593 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7594 if (NV->getType() != Type::VoidTy) {
7595 const Type *CallerTy = Caller->getType();
7596 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7598 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7600 // If this is an invoke instruction, we should insert it after the first
7601 // non-phi, instruction in the normal successor block.
7602 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7603 BasicBlock::iterator I = II->getNormalDest()->begin();
7604 while (isa<PHINode>(I)) ++I;
7605 InsertNewInstBefore(NC, *I);
7607 // Otherwise, it's a call, just insert cast right after the call instr
7608 InsertNewInstBefore(NC, *Caller);
7610 AddUsersToWorkList(*Caller);
7612 NV = UndefValue::get(Caller->getType());
7616 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7617 Caller->replaceAllUsesWith(NV);
7618 Caller->eraseFromParent();
7619 RemoveFromWorkList(Caller);
7623 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7624 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7625 /// and a single binop.
7626 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7627 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7628 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
7629 isa<CmpInst>(FirstInst));
7630 unsigned Opc = FirstInst->getOpcode();
7631 Value *LHSVal = FirstInst->getOperand(0);
7632 Value *RHSVal = FirstInst->getOperand(1);
7634 const Type *LHSType = LHSVal->getType();
7635 const Type *RHSType = RHSVal->getType();
7637 // Scan to see if all operands are the same opcode, all have one use, and all
7638 // kill their operands (i.e. the operands have one use).
7639 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7640 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7641 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7642 // Verify type of the LHS matches so we don't fold cmp's of different
7643 // types or GEP's with different index types.
7644 I->getOperand(0)->getType() != LHSType ||
7645 I->getOperand(1)->getType() != RHSType)
7648 // If they are CmpInst instructions, check their predicates
7649 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7650 if (cast<CmpInst>(I)->getPredicate() !=
7651 cast<CmpInst>(FirstInst)->getPredicate())
7654 // Keep track of which operand needs a phi node.
7655 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7656 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7659 // Otherwise, this is safe to transform, determine if it is profitable.
7661 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7662 // Indexes are often folded into load/store instructions, so we don't want to
7663 // hide them behind a phi.
7664 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7667 Value *InLHS = FirstInst->getOperand(0);
7668 Value *InRHS = FirstInst->getOperand(1);
7669 PHINode *NewLHS = 0, *NewRHS = 0;
7671 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7672 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7673 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7674 InsertNewInstBefore(NewLHS, PN);
7679 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7680 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7681 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7682 InsertNewInstBefore(NewRHS, PN);
7686 // Add all operands to the new PHIs.
7687 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7689 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7690 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7693 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7694 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7698 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7699 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7700 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7701 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7704 assert(isa<GetElementPtrInst>(FirstInst));
7705 return new GetElementPtrInst(LHSVal, RHSVal);
7709 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7710 /// of the block that defines it. This means that it must be obvious the value
7711 /// of the load is not changed from the point of the load to the end of the
7714 /// Finally, it is safe, but not profitable, to sink a load targetting a
7715 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
7717 static bool isSafeToSinkLoad(LoadInst *L) {
7718 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7720 for (++BBI; BBI != E; ++BBI)
7721 if (BBI->mayWriteToMemory())
7724 // Check for non-address taken alloca. If not address-taken already, it isn't
7725 // profitable to do this xform.
7726 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
7727 bool isAddressTaken = false;
7728 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
7730 if (isa<LoadInst>(UI)) continue;
7731 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
7732 // If storing TO the alloca, then the address isn't taken.
7733 if (SI->getOperand(1) == AI) continue;
7735 isAddressTaken = true;
7739 if (!isAddressTaken)
7747 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7748 // operator and they all are only used by the PHI, PHI together their
7749 // inputs, and do the operation once, to the result of the PHI.
7750 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7751 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7753 // Scan the instruction, looking for input operations that can be folded away.
7754 // If all input operands to the phi are the same instruction (e.g. a cast from
7755 // the same type or "+42") we can pull the operation through the PHI, reducing
7756 // code size and simplifying code.
7757 Constant *ConstantOp = 0;
7758 const Type *CastSrcTy = 0;
7759 bool isVolatile = false;
7760 if (isa<CastInst>(FirstInst)) {
7761 CastSrcTy = FirstInst->getOperand(0)->getType();
7762 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
7763 // Can fold binop, compare or shift here if the RHS is a constant,
7764 // otherwise call FoldPHIArgBinOpIntoPHI.
7765 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7766 if (ConstantOp == 0)
7767 return FoldPHIArgBinOpIntoPHI(PN);
7768 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7769 isVolatile = LI->isVolatile();
7770 // We can't sink the load if the loaded value could be modified between the
7771 // load and the PHI.
7772 if (LI->getParent() != PN.getIncomingBlock(0) ||
7773 !isSafeToSinkLoad(LI))
7775 } else if (isa<GetElementPtrInst>(FirstInst)) {
7776 if (FirstInst->getNumOperands() == 2)
7777 return FoldPHIArgBinOpIntoPHI(PN);
7778 // Can't handle general GEPs yet.
7781 return 0; // Cannot fold this operation.
7784 // Check to see if all arguments are the same operation.
7785 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7786 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7787 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7788 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7791 if (I->getOperand(0)->getType() != CastSrcTy)
7792 return 0; // Cast operation must match.
7793 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7794 // We can't sink the load if the loaded value could be modified between
7795 // the load and the PHI.
7796 if (LI->isVolatile() != isVolatile ||
7797 LI->getParent() != PN.getIncomingBlock(i) ||
7798 !isSafeToSinkLoad(LI))
7800 } else if (I->getOperand(1) != ConstantOp) {
7805 // Okay, they are all the same operation. Create a new PHI node of the
7806 // correct type, and PHI together all of the LHS's of the instructions.
7807 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7808 PN.getName()+".in");
7809 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7811 Value *InVal = FirstInst->getOperand(0);
7812 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7814 // Add all operands to the new PHI.
7815 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7816 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7817 if (NewInVal != InVal)
7819 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7824 // The new PHI unions all of the same values together. This is really
7825 // common, so we handle it intelligently here for compile-time speed.
7829 InsertNewInstBefore(NewPN, PN);
7833 // Insert and return the new operation.
7834 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7835 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7836 else if (isa<LoadInst>(FirstInst))
7837 return new LoadInst(PhiVal, "", isVolatile);
7838 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7839 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7840 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7841 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7842 PhiVal, ConstantOp);
7844 assert(0 && "Unknown operation");
7848 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7850 static bool DeadPHICycle(PHINode *PN,
7851 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
7852 if (PN->use_empty()) return true;
7853 if (!PN->hasOneUse()) return false;
7855 // Remember this node, and if we find the cycle, return.
7856 if (!PotentiallyDeadPHIs.insert(PN))
7859 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7860 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7865 // PHINode simplification
7867 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7868 // If LCSSA is around, don't mess with Phi nodes
7869 if (MustPreserveLCSSA) return 0;
7871 if (Value *V = PN.hasConstantValue())
7872 return ReplaceInstUsesWith(PN, V);
7874 // If all PHI operands are the same operation, pull them through the PHI,
7875 // reducing code size.
7876 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7877 PN.getIncomingValue(0)->hasOneUse())
7878 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7881 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7882 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7883 // PHI)... break the cycle.
7884 if (PN.hasOneUse()) {
7885 Instruction *PHIUser = cast<Instruction>(PN.use_back());
7886 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
7887 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
7888 PotentiallyDeadPHIs.insert(&PN);
7889 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7890 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7893 // If this phi has a single use, and if that use just computes a value for
7894 // the next iteration of a loop, delete the phi. This occurs with unused
7895 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
7896 // common case here is good because the only other things that catch this
7897 // are induction variable analysis (sometimes) and ADCE, which is only run
7899 if (PHIUser->hasOneUse() &&
7900 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
7901 PHIUser->use_back() == &PN) {
7902 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7909 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7910 Instruction *InsertPoint,
7912 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7913 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7914 // We must cast correctly to the pointer type. Ensure that we
7915 // sign extend the integer value if it is smaller as this is
7916 // used for address computation.
7917 Instruction::CastOps opcode =
7918 (VTySize < PtrSize ? Instruction::SExt :
7919 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7920 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7924 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7925 Value *PtrOp = GEP.getOperand(0);
7926 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7927 // If so, eliminate the noop.
7928 if (GEP.getNumOperands() == 1)
7929 return ReplaceInstUsesWith(GEP, PtrOp);
7931 if (isa<UndefValue>(GEP.getOperand(0)))
7932 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7934 bool HasZeroPointerIndex = false;
7935 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7936 HasZeroPointerIndex = C->isNullValue();
7938 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7939 return ReplaceInstUsesWith(GEP, PtrOp);
7941 // Keep track of whether all indices are zero constants integers.
7942 bool AllZeroIndices = true;
7944 // Eliminate unneeded casts for indices.
7945 bool MadeChange = false;
7947 gep_type_iterator GTI = gep_type_begin(GEP);
7948 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
7949 // Track whether this GEP has all zero indices, if so, it doesn't move the
7950 // input pointer, it just changes its type.
7951 if (AllZeroIndices) {
7952 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(i)))
7953 AllZeroIndices = CI->isNullValue();
7955 AllZeroIndices = false;
7957 if (isa<SequentialType>(*GTI)) {
7958 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7959 if (CI->getOpcode() == Instruction::ZExt ||
7960 CI->getOpcode() == Instruction::SExt) {
7961 const Type *SrcTy = CI->getOperand(0)->getType();
7962 // We can eliminate a cast from i32 to i64 iff the target
7963 // is a 32-bit pointer target.
7964 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7966 GEP.setOperand(i, CI->getOperand(0));
7970 // If we are using a wider index than needed for this platform, shrink it
7971 // to what we need. If the incoming value needs a cast instruction,
7972 // insert it. This explicit cast can make subsequent optimizations more
7974 Value *Op = GEP.getOperand(i);
7975 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
7976 if (Constant *C = dyn_cast<Constant>(Op)) {
7977 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7980 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7982 GEP.setOperand(i, Op);
7987 if (MadeChange) return &GEP;
7989 // If this GEP instruction doesn't move the pointer, and if the input operand
7990 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
7991 // real input to the dest type.
7992 if (AllZeroIndices && isa<BitCastInst>(GEP.getOperand(0)))
7993 return new BitCastInst(cast<BitCastInst>(GEP.getOperand(0))->getOperand(0),
7996 // Combine Indices - If the source pointer to this getelementptr instruction
7997 // is a getelementptr instruction, combine the indices of the two
7998 // getelementptr instructions into a single instruction.
8000 SmallVector<Value*, 8> SrcGEPOperands;
8001 if (User *Src = dyn_castGetElementPtr(PtrOp))
8002 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8004 if (!SrcGEPOperands.empty()) {
8005 // Note that if our source is a gep chain itself that we wait for that
8006 // chain to be resolved before we perform this transformation. This
8007 // avoids us creating a TON of code in some cases.
8009 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8010 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8011 return 0; // Wait until our source is folded to completion.
8013 SmallVector<Value*, 8> Indices;
8015 // Find out whether the last index in the source GEP is a sequential idx.
8016 bool EndsWithSequential = false;
8017 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8018 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8019 EndsWithSequential = !isa<StructType>(*I);
8021 // Can we combine the two pointer arithmetics offsets?
8022 if (EndsWithSequential) {
8023 // Replace: gep (gep %P, long B), long A, ...
8024 // With: T = long A+B; gep %P, T, ...
8026 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8027 if (SO1 == Constant::getNullValue(SO1->getType())) {
8029 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8032 // If they aren't the same type, convert both to an integer of the
8033 // target's pointer size.
8034 if (SO1->getType() != GO1->getType()) {
8035 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8036 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8037 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8038 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8040 unsigned PS = TD->getPointerSize();
8041 if (TD->getTypeSize(SO1->getType()) == PS) {
8042 // Convert GO1 to SO1's type.
8043 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8045 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8046 // Convert SO1 to GO1's type.
8047 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8049 const Type *PT = TD->getIntPtrType();
8050 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8051 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8055 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8056 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8058 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8059 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8063 // Recycle the GEP we already have if possible.
8064 if (SrcGEPOperands.size() == 2) {
8065 GEP.setOperand(0, SrcGEPOperands[0]);
8066 GEP.setOperand(1, Sum);
8069 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8070 SrcGEPOperands.end()-1);
8071 Indices.push_back(Sum);
8072 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8074 } else if (isa<Constant>(*GEP.idx_begin()) &&
8075 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8076 SrcGEPOperands.size() != 1) {
8077 // Otherwise we can do the fold if the first index of the GEP is a zero
8078 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8079 SrcGEPOperands.end());
8080 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8083 if (!Indices.empty())
8084 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
8085 Indices.size(), GEP.getName());
8087 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8088 // GEP of global variable. If all of the indices for this GEP are
8089 // constants, we can promote this to a constexpr instead of an instruction.
8091 // Scan for nonconstants...
8092 SmallVector<Constant*, 8> Indices;
8093 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8094 for (; I != E && isa<Constant>(*I); ++I)
8095 Indices.push_back(cast<Constant>(*I));
8097 if (I == E) { // If they are all constants...
8098 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8099 &Indices[0],Indices.size());
8101 // Replace all uses of the GEP with the new constexpr...
8102 return ReplaceInstUsesWith(GEP, CE);
8104 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8105 if (!isa<PointerType>(X->getType())) {
8106 // Not interesting. Source pointer must be a cast from pointer.
8107 } else if (HasZeroPointerIndex) {
8108 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8109 // into : GEP [10 x ubyte]* X, long 0, ...
8111 // This occurs when the program declares an array extern like "int X[];"
8113 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8114 const PointerType *XTy = cast<PointerType>(X->getType());
8115 if (const ArrayType *XATy =
8116 dyn_cast<ArrayType>(XTy->getElementType()))
8117 if (const ArrayType *CATy =
8118 dyn_cast<ArrayType>(CPTy->getElementType()))
8119 if (CATy->getElementType() == XATy->getElementType()) {
8120 // At this point, we know that the cast source type is a pointer
8121 // to an array of the same type as the destination pointer
8122 // array. Because the array type is never stepped over (there
8123 // is a leading zero) we can fold the cast into this GEP.
8124 GEP.setOperand(0, X);
8127 } else if (GEP.getNumOperands() == 2) {
8128 // Transform things like:
8129 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8130 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8131 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8132 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8133 if (isa<ArrayType>(SrcElTy) &&
8134 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8135 TD->getTypeSize(ResElTy)) {
8136 Value *V = InsertNewInstBefore(
8137 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8138 GEP.getOperand(1), GEP.getName()), GEP);
8139 // V and GEP are both pointer types --> BitCast
8140 return new BitCastInst(V, GEP.getType());
8143 // Transform things like:
8144 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8145 // (where tmp = 8*tmp2) into:
8146 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8148 if (isa<ArrayType>(SrcElTy) &&
8149 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8150 uint64_t ArrayEltSize =
8151 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8153 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8154 // allow either a mul, shift, or constant here.
8156 ConstantInt *Scale = 0;
8157 if (ArrayEltSize == 1) {
8158 NewIdx = GEP.getOperand(1);
8159 Scale = ConstantInt::get(NewIdx->getType(), 1);
8160 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8161 NewIdx = ConstantInt::get(CI->getType(), 1);
8163 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8164 if (Inst->getOpcode() == Instruction::Shl &&
8165 isa<ConstantInt>(Inst->getOperand(1))) {
8166 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
8167 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
8168 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
8169 NewIdx = Inst->getOperand(0);
8170 } else if (Inst->getOpcode() == Instruction::Mul &&
8171 isa<ConstantInt>(Inst->getOperand(1))) {
8172 Scale = cast<ConstantInt>(Inst->getOperand(1));
8173 NewIdx = Inst->getOperand(0);
8177 // If the index will be to exactly the right offset with the scale taken
8178 // out, perform the transformation.
8179 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8180 if (isa<ConstantInt>(Scale))
8181 Scale = ConstantInt::get(Scale->getType(),
8182 Scale->getZExtValue() / ArrayEltSize);
8183 if (Scale->getZExtValue() != 1) {
8184 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8186 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8187 NewIdx = InsertNewInstBefore(Sc, GEP);
8190 // Insert the new GEP instruction.
8191 Instruction *NewGEP =
8192 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8193 NewIdx, GEP.getName());
8194 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8195 // The NewGEP must be pointer typed, so must the old one -> BitCast
8196 return new BitCastInst(NewGEP, GEP.getType());
8205 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8206 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8207 if (AI.isArrayAllocation()) // Check C != 1
8208 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8210 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8211 AllocationInst *New = 0;
8213 // Create and insert the replacement instruction...
8214 if (isa<MallocInst>(AI))
8215 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8217 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8218 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8221 InsertNewInstBefore(New, AI);
8223 // Scan to the end of the allocation instructions, to skip over a block of
8224 // allocas if possible...
8226 BasicBlock::iterator It = New;
8227 while (isa<AllocationInst>(*It)) ++It;
8229 // Now that I is pointing to the first non-allocation-inst in the block,
8230 // insert our getelementptr instruction...
8232 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8233 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8234 New->getName()+".sub", It);
8236 // Now make everything use the getelementptr instead of the original
8238 return ReplaceInstUsesWith(AI, V);
8239 } else if (isa<UndefValue>(AI.getArraySize())) {
8240 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8243 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8244 // Note that we only do this for alloca's, because malloc should allocate and
8245 // return a unique pointer, even for a zero byte allocation.
8246 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8247 TD->getTypeSize(AI.getAllocatedType()) == 0)
8248 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8253 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8254 Value *Op = FI.getOperand(0);
8256 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8257 if (CastInst *CI = dyn_cast<CastInst>(Op))
8258 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8259 FI.setOperand(0, CI->getOperand(0));
8263 // free undef -> unreachable.
8264 if (isa<UndefValue>(Op)) {
8265 // Insert a new store to null because we cannot modify the CFG here.
8266 new StoreInst(ConstantInt::getTrue(),
8267 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8268 return EraseInstFromFunction(FI);
8271 // If we have 'free null' delete the instruction. This can happen in stl code
8272 // when lots of inlining happens.
8273 if (isa<ConstantPointerNull>(Op))
8274 return EraseInstFromFunction(FI);
8280 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8281 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8282 User *CI = cast<User>(LI.getOperand(0));
8283 Value *CastOp = CI->getOperand(0);
8285 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8286 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8287 const Type *SrcPTy = SrcTy->getElementType();
8289 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8290 isa<VectorType>(DestPTy)) {
8291 // If the source is an array, the code below will not succeed. Check to
8292 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8294 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8295 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8296 if (ASrcTy->getNumElements() != 0) {
8298 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8299 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8300 SrcTy = cast<PointerType>(CastOp->getType());
8301 SrcPTy = SrcTy->getElementType();
8304 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8305 isa<VectorType>(SrcPTy)) &&
8306 // Do not allow turning this into a load of an integer, which is then
8307 // casted to a pointer, this pessimizes pointer analysis a lot.
8308 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8309 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8310 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8312 // Okay, we are casting from one integer or pointer type to another of
8313 // the same size. Instead of casting the pointer before the load, cast
8314 // the result of the loaded value.
8315 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8317 LI.isVolatile()),LI);
8318 // Now cast the result of the load.
8319 return new BitCastInst(NewLoad, LI.getType());
8326 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8327 /// from this value cannot trap. If it is not obviously safe to load from the
8328 /// specified pointer, we do a quick local scan of the basic block containing
8329 /// ScanFrom, to determine if the address is already accessed.
8330 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8331 // If it is an alloca or global variable, it is always safe to load from.
8332 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8334 // Otherwise, be a little bit agressive by scanning the local block where we
8335 // want to check to see if the pointer is already being loaded or stored
8336 // from/to. If so, the previous load or store would have already trapped,
8337 // so there is no harm doing an extra load (also, CSE will later eliminate
8338 // the load entirely).
8339 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8344 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8345 if (LI->getOperand(0) == V) return true;
8346 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8347 if (SI->getOperand(1) == V) return true;
8353 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8354 Value *Op = LI.getOperand(0);
8356 // load (cast X) --> cast (load X) iff safe
8357 if (isa<CastInst>(Op))
8358 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8361 // None of the following transforms are legal for volatile loads.
8362 if (LI.isVolatile()) return 0;
8364 if (&LI.getParent()->front() != &LI) {
8365 BasicBlock::iterator BBI = &LI; --BBI;
8366 // If the instruction immediately before this is a store to the same
8367 // address, do a simple form of store->load forwarding.
8368 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8369 if (SI->getOperand(1) == LI.getOperand(0))
8370 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8371 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8372 if (LIB->getOperand(0) == LI.getOperand(0))
8373 return ReplaceInstUsesWith(LI, LIB);
8376 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8377 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8378 isa<UndefValue>(GEPI->getOperand(0))) {
8379 // Insert a new store to null instruction before the load to indicate
8380 // that this code is not reachable. We do this instead of inserting
8381 // an unreachable instruction directly because we cannot modify the
8383 new StoreInst(UndefValue::get(LI.getType()),
8384 Constant::getNullValue(Op->getType()), &LI);
8385 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8388 if (Constant *C = dyn_cast<Constant>(Op)) {
8389 // load null/undef -> undef
8390 if ((C->isNullValue() || isa<UndefValue>(C))) {
8391 // Insert a new store to null instruction before the load to indicate that
8392 // this code is not reachable. We do this instead of inserting an
8393 // unreachable instruction directly because we cannot modify the CFG.
8394 new StoreInst(UndefValue::get(LI.getType()),
8395 Constant::getNullValue(Op->getType()), &LI);
8396 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8399 // Instcombine load (constant global) into the value loaded.
8400 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8401 if (GV->isConstant() && !GV->isDeclaration())
8402 return ReplaceInstUsesWith(LI, GV->getInitializer());
8404 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8405 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8406 if (CE->getOpcode() == Instruction::GetElementPtr) {
8407 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8408 if (GV->isConstant() && !GV->isDeclaration())
8410 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8411 return ReplaceInstUsesWith(LI, V);
8412 if (CE->getOperand(0)->isNullValue()) {
8413 // Insert a new store to null instruction before the load to indicate
8414 // that this code is not reachable. We do this instead of inserting
8415 // an unreachable instruction directly because we cannot modify the
8417 new StoreInst(UndefValue::get(LI.getType()),
8418 Constant::getNullValue(Op->getType()), &LI);
8419 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8422 } else if (CE->isCast()) {
8423 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8428 if (Op->hasOneUse()) {
8429 // Change select and PHI nodes to select values instead of addresses: this
8430 // helps alias analysis out a lot, allows many others simplifications, and
8431 // exposes redundancy in the code.
8433 // Note that we cannot do the transformation unless we know that the
8434 // introduced loads cannot trap! Something like this is valid as long as
8435 // the condition is always false: load (select bool %C, int* null, int* %G),
8436 // but it would not be valid if we transformed it to load from null
8439 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8440 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8441 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8442 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8443 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8444 SI->getOperand(1)->getName()+".val"), LI);
8445 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8446 SI->getOperand(2)->getName()+".val"), LI);
8447 return new SelectInst(SI->getCondition(), V1, V2);
8450 // load (select (cond, null, P)) -> load P
8451 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8452 if (C->isNullValue()) {
8453 LI.setOperand(0, SI->getOperand(2));
8457 // load (select (cond, P, null)) -> load P
8458 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8459 if (C->isNullValue()) {
8460 LI.setOperand(0, SI->getOperand(1));
8468 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8470 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8471 User *CI = cast<User>(SI.getOperand(1));
8472 Value *CastOp = CI->getOperand(0);
8474 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8475 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8476 const Type *SrcPTy = SrcTy->getElementType();
8478 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8479 // If the source is an array, the code below will not succeed. Check to
8480 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8482 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8483 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8484 if (ASrcTy->getNumElements() != 0) {
8486 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8487 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8488 SrcTy = cast<PointerType>(CastOp->getType());
8489 SrcPTy = SrcTy->getElementType();
8492 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8493 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8494 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8496 // Okay, we are casting from one integer or pointer type to another of
8497 // the same size. Instead of casting the pointer before
8498 // the store, cast the value to be stored.
8500 Value *SIOp0 = SI.getOperand(0);
8501 Instruction::CastOps opcode = Instruction::BitCast;
8502 const Type* CastSrcTy = SIOp0->getType();
8503 const Type* CastDstTy = SrcPTy;
8504 if (isa<PointerType>(CastDstTy)) {
8505 if (CastSrcTy->isInteger())
8506 opcode = Instruction::IntToPtr;
8507 } else if (isa<IntegerType>(CastDstTy)) {
8508 if (isa<PointerType>(SIOp0->getType()))
8509 opcode = Instruction::PtrToInt;
8511 if (Constant *C = dyn_cast<Constant>(SIOp0))
8512 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8514 NewCast = IC.InsertNewInstBefore(
8515 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8517 return new StoreInst(NewCast, CastOp);
8524 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8525 Value *Val = SI.getOperand(0);
8526 Value *Ptr = SI.getOperand(1);
8528 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8529 EraseInstFromFunction(SI);
8534 // If the RHS is an alloca with a single use, zapify the store, making the
8536 if (Ptr->hasOneUse()) {
8537 if (isa<AllocaInst>(Ptr)) {
8538 EraseInstFromFunction(SI);
8543 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8544 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8545 GEP->getOperand(0)->hasOneUse()) {
8546 EraseInstFromFunction(SI);
8552 // Do really simple DSE, to catch cases where there are several consequtive
8553 // stores to the same location, separated by a few arithmetic operations. This
8554 // situation often occurs with bitfield accesses.
8555 BasicBlock::iterator BBI = &SI;
8556 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8560 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8561 // Prev store isn't volatile, and stores to the same location?
8562 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8565 EraseInstFromFunction(*PrevSI);
8571 // If this is a load, we have to stop. However, if the loaded value is from
8572 // the pointer we're loading and is producing the pointer we're storing,
8573 // then *this* store is dead (X = load P; store X -> P).
8574 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8575 if (LI == Val && LI->getOperand(0) == Ptr) {
8576 EraseInstFromFunction(SI);
8580 // Otherwise, this is a load from some other location. Stores before it
8585 // Don't skip over loads or things that can modify memory.
8586 if (BBI->mayWriteToMemory())
8591 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8593 // store X, null -> turns into 'unreachable' in SimplifyCFG
8594 if (isa<ConstantPointerNull>(Ptr)) {
8595 if (!isa<UndefValue>(Val)) {
8596 SI.setOperand(0, UndefValue::get(Val->getType()));
8597 if (Instruction *U = dyn_cast<Instruction>(Val))
8598 AddToWorkList(U); // Dropped a use.
8601 return 0; // Do not modify these!
8604 // store undef, Ptr -> noop
8605 if (isa<UndefValue>(Val)) {
8606 EraseInstFromFunction(SI);
8611 // If the pointer destination is a cast, see if we can fold the cast into the
8613 if (isa<CastInst>(Ptr))
8614 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8616 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8618 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8622 // If this store is the last instruction in the basic block, and if the block
8623 // ends with an unconditional branch, try to move it to the successor block.
8625 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8626 if (BI->isUnconditional()) {
8627 // Check to see if the successor block has exactly two incoming edges. If
8628 // so, see if the other predecessor contains a store to the same location.
8629 // if so, insert a PHI node (if needed) and move the stores down.
8630 BasicBlock *Dest = BI->getSuccessor(0);
8632 pred_iterator PI = pred_begin(Dest);
8633 BasicBlock *Other = 0;
8634 if (*PI != BI->getParent())
8637 if (PI != pred_end(Dest)) {
8638 if (*PI != BI->getParent())
8643 if (++PI != pred_end(Dest))
8646 if (Other) { // If only one other pred...
8647 BBI = Other->getTerminator();
8648 // Make sure this other block ends in an unconditional branch and that
8649 // there is an instruction before the branch.
8650 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8651 BBI != Other->begin()) {
8653 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8655 // If this instruction is a store to the same location.
8656 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8657 // Okay, we know we can perform this transformation. Insert a PHI
8658 // node now if we need it.
8659 Value *MergedVal = OtherStore->getOperand(0);
8660 if (MergedVal != SI.getOperand(0)) {
8661 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8662 PN->reserveOperandSpace(2);
8663 PN->addIncoming(SI.getOperand(0), SI.getParent());
8664 PN->addIncoming(OtherStore->getOperand(0), Other);
8665 MergedVal = InsertNewInstBefore(PN, Dest->front());
8668 // Advance to a place where it is safe to insert the new store and
8670 BBI = Dest->begin();
8671 while (isa<PHINode>(BBI)) ++BBI;
8672 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8673 OtherStore->isVolatile()), *BBI);
8675 // Nuke the old stores.
8676 EraseInstFromFunction(SI);
8677 EraseInstFromFunction(*OtherStore);
8689 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8690 // Change br (not X), label True, label False to: br X, label False, True
8692 BasicBlock *TrueDest;
8693 BasicBlock *FalseDest;
8694 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8695 !isa<Constant>(X)) {
8696 // Swap Destinations and condition...
8698 BI.setSuccessor(0, FalseDest);
8699 BI.setSuccessor(1, TrueDest);
8703 // Cannonicalize fcmp_one -> fcmp_oeq
8704 FCmpInst::Predicate FPred; Value *Y;
8705 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8706 TrueDest, FalseDest)))
8707 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8708 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8709 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8710 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8711 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
8712 NewSCC->takeName(I);
8713 // Swap Destinations and condition...
8714 BI.setCondition(NewSCC);
8715 BI.setSuccessor(0, FalseDest);
8716 BI.setSuccessor(1, TrueDest);
8717 RemoveFromWorkList(I);
8718 I->eraseFromParent();
8719 AddToWorkList(NewSCC);
8723 // Cannonicalize icmp_ne -> icmp_eq
8724 ICmpInst::Predicate IPred;
8725 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8726 TrueDest, FalseDest)))
8727 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8728 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8729 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8730 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8731 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8732 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
8733 NewSCC->takeName(I);
8734 // Swap Destinations and condition...
8735 BI.setCondition(NewSCC);
8736 BI.setSuccessor(0, FalseDest);
8737 BI.setSuccessor(1, TrueDest);
8738 RemoveFromWorkList(I);
8739 I->eraseFromParent();;
8740 AddToWorkList(NewSCC);
8747 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8748 Value *Cond = SI.getCondition();
8749 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8750 if (I->getOpcode() == Instruction::Add)
8751 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8752 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8753 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8754 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8756 SI.setOperand(0, I->getOperand(0));
8764 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8765 /// is to leave as a vector operation.
8766 static bool CheapToScalarize(Value *V, bool isConstant) {
8767 if (isa<ConstantAggregateZero>(V))
8769 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
8770 if (isConstant) return true;
8771 // If all elts are the same, we can extract.
8772 Constant *Op0 = C->getOperand(0);
8773 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8774 if (C->getOperand(i) != Op0)
8778 Instruction *I = dyn_cast<Instruction>(V);
8779 if (!I) return false;
8781 // Insert element gets simplified to the inserted element or is deleted if
8782 // this is constant idx extract element and its a constant idx insertelt.
8783 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8784 isa<ConstantInt>(I->getOperand(2)))
8786 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8788 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8789 if (BO->hasOneUse() &&
8790 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8791 CheapToScalarize(BO->getOperand(1), isConstant)))
8793 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8794 if (CI->hasOneUse() &&
8795 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8796 CheapToScalarize(CI->getOperand(1), isConstant)))
8802 /// Read and decode a shufflevector mask.
8804 /// It turns undef elements into values that are larger than the number of
8805 /// elements in the input.
8806 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8807 unsigned NElts = SVI->getType()->getNumElements();
8808 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8809 return std::vector<unsigned>(NElts, 0);
8810 if (isa<UndefValue>(SVI->getOperand(2)))
8811 return std::vector<unsigned>(NElts, 2*NElts);
8813 std::vector<unsigned> Result;
8814 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
8815 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8816 if (isa<UndefValue>(CP->getOperand(i)))
8817 Result.push_back(NElts*2); // undef -> 8
8819 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8823 /// FindScalarElement - Given a vector and an element number, see if the scalar
8824 /// value is already around as a register, for example if it were inserted then
8825 /// extracted from the vector.
8826 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8827 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
8828 const VectorType *PTy = cast<VectorType>(V->getType());
8829 unsigned Width = PTy->getNumElements();
8830 if (EltNo >= Width) // Out of range access.
8831 return UndefValue::get(PTy->getElementType());
8833 if (isa<UndefValue>(V))
8834 return UndefValue::get(PTy->getElementType());
8835 else if (isa<ConstantAggregateZero>(V))
8836 return Constant::getNullValue(PTy->getElementType());
8837 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
8838 return CP->getOperand(EltNo);
8839 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8840 // If this is an insert to a variable element, we don't know what it is.
8841 if (!isa<ConstantInt>(III->getOperand(2)))
8843 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8845 // If this is an insert to the element we are looking for, return the
8848 return III->getOperand(1);
8850 // Otherwise, the insertelement doesn't modify the value, recurse on its
8852 return FindScalarElement(III->getOperand(0), EltNo);
8853 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8854 unsigned InEl = getShuffleMask(SVI)[EltNo];
8856 return FindScalarElement(SVI->getOperand(0), InEl);
8857 else if (InEl < Width*2)
8858 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8860 return UndefValue::get(PTy->getElementType());
8863 // Otherwise, we don't know.
8867 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8869 // If packed val is undef, replace extract with scalar undef.
8870 if (isa<UndefValue>(EI.getOperand(0)))
8871 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8873 // If packed val is constant 0, replace extract with scalar 0.
8874 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8875 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8877 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
8878 // If packed val is constant with uniform operands, replace EI
8879 // with that operand
8880 Constant *op0 = C->getOperand(0);
8881 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8882 if (C->getOperand(i) != op0) {
8887 return ReplaceInstUsesWith(EI, op0);
8890 // If extracting a specified index from the vector, see if we can recursively
8891 // find a previously computed scalar that was inserted into the vector.
8892 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8893 // This instruction only demands the single element from the input vector.
8894 // If the input vector has a single use, simplify it based on this use
8896 uint64_t IndexVal = IdxC->getZExtValue();
8897 if (EI.getOperand(0)->hasOneUse()) {
8899 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8902 EI.setOperand(0, V);
8907 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8908 return ReplaceInstUsesWith(EI, Elt);
8911 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8912 if (I->hasOneUse()) {
8913 // Push extractelement into predecessor operation if legal and
8914 // profitable to do so
8915 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8916 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8917 if (CheapToScalarize(BO, isConstantElt)) {
8918 ExtractElementInst *newEI0 =
8919 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8920 EI.getName()+".lhs");
8921 ExtractElementInst *newEI1 =
8922 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8923 EI.getName()+".rhs");
8924 InsertNewInstBefore(newEI0, EI);
8925 InsertNewInstBefore(newEI1, EI);
8926 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8928 } else if (isa<LoadInst>(I)) {
8929 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8930 PointerType::get(EI.getType()), EI);
8931 GetElementPtrInst *GEP =
8932 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8933 InsertNewInstBefore(GEP, EI);
8934 return new LoadInst(GEP);
8937 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8938 // Extracting the inserted element?
8939 if (IE->getOperand(2) == EI.getOperand(1))
8940 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8941 // If the inserted and extracted elements are constants, they must not
8942 // be the same value, extract from the pre-inserted value instead.
8943 if (isa<Constant>(IE->getOperand(2)) &&
8944 isa<Constant>(EI.getOperand(1))) {
8945 AddUsesToWorkList(EI);
8946 EI.setOperand(0, IE->getOperand(0));
8949 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8950 // If this is extracting an element from a shufflevector, figure out where
8951 // it came from and extract from the appropriate input element instead.
8952 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8953 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8955 if (SrcIdx < SVI->getType()->getNumElements())
8956 Src = SVI->getOperand(0);
8957 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8958 SrcIdx -= SVI->getType()->getNumElements();
8959 Src = SVI->getOperand(1);
8961 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8963 return new ExtractElementInst(Src, SrcIdx);
8970 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8971 /// elements from either LHS or RHS, return the shuffle mask and true.
8972 /// Otherwise, return false.
8973 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8974 std::vector<Constant*> &Mask) {
8975 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8976 "Invalid CollectSingleShuffleElements");
8977 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
8979 if (isa<UndefValue>(V)) {
8980 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8982 } else if (V == LHS) {
8983 for (unsigned i = 0; i != NumElts; ++i)
8984 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8986 } else if (V == RHS) {
8987 for (unsigned i = 0; i != NumElts; ++i)
8988 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8990 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8991 // If this is an insert of an extract from some other vector, include it.
8992 Value *VecOp = IEI->getOperand(0);
8993 Value *ScalarOp = IEI->getOperand(1);
8994 Value *IdxOp = IEI->getOperand(2);
8996 if (!isa<ConstantInt>(IdxOp))
8998 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9000 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9001 // Okay, we can handle this if the vector we are insertinting into is
9003 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9004 // If so, update the mask to reflect the inserted undef.
9005 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9008 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9009 if (isa<ConstantInt>(EI->getOperand(1)) &&
9010 EI->getOperand(0)->getType() == V->getType()) {
9011 unsigned ExtractedIdx =
9012 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9014 // This must be extracting from either LHS or RHS.
9015 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9016 // Okay, we can handle this if the vector we are insertinting into is
9018 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9019 // If so, update the mask to reflect the inserted value.
9020 if (EI->getOperand(0) == LHS) {
9021 Mask[InsertedIdx & (NumElts-1)] =
9022 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9024 assert(EI->getOperand(0) == RHS);
9025 Mask[InsertedIdx & (NumElts-1)] =
9026 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9035 // TODO: Handle shufflevector here!
9040 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9041 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9042 /// that computes V and the LHS value of the shuffle.
9043 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9045 assert(isa<VectorType>(V->getType()) &&
9046 (RHS == 0 || V->getType() == RHS->getType()) &&
9047 "Invalid shuffle!");
9048 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9050 if (isa<UndefValue>(V)) {
9051 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9053 } else if (isa<ConstantAggregateZero>(V)) {
9054 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9056 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9057 // If this is an insert of an extract from some other vector, include it.
9058 Value *VecOp = IEI->getOperand(0);
9059 Value *ScalarOp = IEI->getOperand(1);
9060 Value *IdxOp = IEI->getOperand(2);
9062 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9063 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9064 EI->getOperand(0)->getType() == V->getType()) {
9065 unsigned ExtractedIdx =
9066 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9067 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9069 // Either the extracted from or inserted into vector must be RHSVec,
9070 // otherwise we'd end up with a shuffle of three inputs.
9071 if (EI->getOperand(0) == RHS || RHS == 0) {
9072 RHS = EI->getOperand(0);
9073 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9074 Mask[InsertedIdx & (NumElts-1)] =
9075 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9080 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9081 // Everything but the extracted element is replaced with the RHS.
9082 for (unsigned i = 0; i != NumElts; ++i) {
9083 if (i != InsertedIdx)
9084 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9089 // If this insertelement is a chain that comes from exactly these two
9090 // vectors, return the vector and the effective shuffle.
9091 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9092 return EI->getOperand(0);
9097 // TODO: Handle shufflevector here!
9099 // Otherwise, can't do anything fancy. Return an identity vector.
9100 for (unsigned i = 0; i != NumElts; ++i)
9101 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9105 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9106 Value *VecOp = IE.getOperand(0);
9107 Value *ScalarOp = IE.getOperand(1);
9108 Value *IdxOp = IE.getOperand(2);
9110 // If the inserted element was extracted from some other vector, and if the
9111 // indexes are constant, try to turn this into a shufflevector operation.
9112 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9113 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9114 EI->getOperand(0)->getType() == IE.getType()) {
9115 unsigned NumVectorElts = IE.getType()->getNumElements();
9116 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9117 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9119 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9120 return ReplaceInstUsesWith(IE, VecOp);
9122 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9123 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9125 // If we are extracting a value from a vector, then inserting it right
9126 // back into the same place, just use the input vector.
9127 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9128 return ReplaceInstUsesWith(IE, VecOp);
9130 // We could theoretically do this for ANY input. However, doing so could
9131 // turn chains of insertelement instructions into a chain of shufflevector
9132 // instructions, and right now we do not merge shufflevectors. As such,
9133 // only do this in a situation where it is clear that there is benefit.
9134 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9135 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9136 // the values of VecOp, except then one read from EIOp0.
9137 // Build a new shuffle mask.
9138 std::vector<Constant*> Mask;
9139 if (isa<UndefValue>(VecOp))
9140 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9142 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9143 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9146 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9147 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9148 ConstantVector::get(Mask));
9151 // If this insertelement isn't used by some other insertelement, turn it
9152 // (and any insertelements it points to), into one big shuffle.
9153 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9154 std::vector<Constant*> Mask;
9156 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9157 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9158 // We now have a shuffle of LHS, RHS, Mask.
9159 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9168 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9169 Value *LHS = SVI.getOperand(0);
9170 Value *RHS = SVI.getOperand(1);
9171 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9173 bool MadeChange = false;
9175 // Undefined shuffle mask -> undefined value.
9176 if (isa<UndefValue>(SVI.getOperand(2)))
9177 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9179 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9180 // the undef, change them to undefs.
9181 if (isa<UndefValue>(SVI.getOperand(1))) {
9182 // Scan to see if there are any references to the RHS. If so, replace them
9183 // with undef element refs and set MadeChange to true.
9184 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9185 if (Mask[i] >= e && Mask[i] != 2*e) {
9192 // Remap any references to RHS to use LHS.
9193 std::vector<Constant*> Elts;
9194 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9196 Elts.push_back(UndefValue::get(Type::Int32Ty));
9198 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9200 SVI.setOperand(2, ConstantVector::get(Elts));
9204 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9205 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9206 if (LHS == RHS || isa<UndefValue>(LHS)) {
9207 if (isa<UndefValue>(LHS) && LHS == RHS) {
9208 // shuffle(undef,undef,mask) -> undef.
9209 return ReplaceInstUsesWith(SVI, LHS);
9212 // Remap any references to RHS to use LHS.
9213 std::vector<Constant*> Elts;
9214 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9216 Elts.push_back(UndefValue::get(Type::Int32Ty));
9218 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9219 (Mask[i] < e && isa<UndefValue>(LHS)))
9220 Mask[i] = 2*e; // Turn into undef.
9222 Mask[i] &= (e-1); // Force to LHS.
9223 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9226 SVI.setOperand(0, SVI.getOperand(1));
9227 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9228 SVI.setOperand(2, ConstantVector::get(Elts));
9229 LHS = SVI.getOperand(0);
9230 RHS = SVI.getOperand(1);
9234 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9235 bool isLHSID = true, isRHSID = true;
9237 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9238 if (Mask[i] >= e*2) continue; // Ignore undef values.
9239 // Is this an identity shuffle of the LHS value?
9240 isLHSID &= (Mask[i] == i);
9242 // Is this an identity shuffle of the RHS value?
9243 isRHSID &= (Mask[i]-e == i);
9246 // Eliminate identity shuffles.
9247 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9248 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9250 // If the LHS is a shufflevector itself, see if we can combine it with this
9251 // one without producing an unusual shuffle. Here we are really conservative:
9252 // we are absolutely afraid of producing a shuffle mask not in the input
9253 // program, because the code gen may not be smart enough to turn a merged
9254 // shuffle into two specific shuffles: it may produce worse code. As such,
9255 // we only merge two shuffles if the result is one of the two input shuffle
9256 // masks. In this case, merging the shuffles just removes one instruction,
9257 // which we know is safe. This is good for things like turning:
9258 // (splat(splat)) -> splat.
9259 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9260 if (isa<UndefValue>(RHS)) {
9261 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9263 std::vector<unsigned> NewMask;
9264 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9266 NewMask.push_back(2*e);
9268 NewMask.push_back(LHSMask[Mask[i]]);
9270 // If the result mask is equal to the src shuffle or this shuffle mask, do
9272 if (NewMask == LHSMask || NewMask == Mask) {
9273 std::vector<Constant*> Elts;
9274 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9275 if (NewMask[i] >= e*2) {
9276 Elts.push_back(UndefValue::get(Type::Int32Ty));
9278 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9281 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9282 LHSSVI->getOperand(1),
9283 ConstantVector::get(Elts));
9288 return MadeChange ? &SVI : 0;
9294 /// TryToSinkInstruction - Try to move the specified instruction from its
9295 /// current block into the beginning of DestBlock, which can only happen if it's
9296 /// safe to move the instruction past all of the instructions between it and the
9297 /// end of its block.
9298 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9299 assert(I->hasOneUse() && "Invariants didn't hold!");
9301 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9302 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9304 // Do not sink alloca instructions out of the entry block.
9305 if (isa<AllocaInst>(I) && I->getParent() ==
9306 &DestBlock->getParent()->getEntryBlock())
9309 // We can only sink load instructions if there is nothing between the load and
9310 // the end of block that could change the value.
9311 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9312 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9314 if (Scan->mayWriteToMemory())
9318 BasicBlock::iterator InsertPos = DestBlock->begin();
9319 while (isa<PHINode>(InsertPos)) ++InsertPos;
9321 I->moveBefore(InsertPos);
9327 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9328 /// all reachable code to the worklist.
9330 /// This has a couple of tricks to make the code faster and more powerful. In
9331 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9332 /// them to the worklist (this significantly speeds up instcombine on code where
9333 /// many instructions are dead or constant). Additionally, if we find a branch
9334 /// whose condition is a known constant, we only visit the reachable successors.
9336 static void AddReachableCodeToWorklist(BasicBlock *BB,
9337 SmallPtrSet<BasicBlock*, 64> &Visited,
9339 const TargetData *TD) {
9340 std::vector<BasicBlock*> Worklist;
9341 Worklist.push_back(BB);
9343 while (!Worklist.empty()) {
9344 BB = Worklist.back();
9345 Worklist.pop_back();
9347 // We have now visited this block! If we've already been here, ignore it.
9348 if (!Visited.insert(BB)) continue;
9350 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9351 Instruction *Inst = BBI++;
9353 // DCE instruction if trivially dead.
9354 if (isInstructionTriviallyDead(Inst)) {
9356 DOUT << "IC: DCE: " << *Inst;
9357 Inst->eraseFromParent();
9361 // ConstantProp instruction if trivially constant.
9362 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9363 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9364 Inst->replaceAllUsesWith(C);
9366 Inst->eraseFromParent();
9370 IC.AddToWorkList(Inst);
9373 // Recursively visit successors. If this is a branch or switch on a
9374 // constant, only visit the reachable successor.
9375 TerminatorInst *TI = BB->getTerminator();
9376 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9377 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9378 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9379 Worklist.push_back(BI->getSuccessor(!CondVal));
9382 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9383 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9384 // See if this is an explicit destination.
9385 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9386 if (SI->getCaseValue(i) == Cond) {
9387 Worklist.push_back(SI->getSuccessor(i));
9391 // Otherwise it is the default destination.
9392 Worklist.push_back(SI->getSuccessor(0));
9397 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9398 Worklist.push_back(TI->getSuccessor(i));
9402 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
9403 bool Changed = false;
9404 TD = &getAnalysis<TargetData>();
9406 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
9407 << F.getNameStr() << "\n");
9410 // Do a depth-first traversal of the function, populate the worklist with
9411 // the reachable instructions. Ignore blocks that are not reachable. Keep
9412 // track of which blocks we visit.
9413 SmallPtrSet<BasicBlock*, 64> Visited;
9414 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
9416 // Do a quick scan over the function. If we find any blocks that are
9417 // unreachable, remove any instructions inside of them. This prevents
9418 // the instcombine code from having to deal with some bad special cases.
9419 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9420 if (!Visited.count(BB)) {
9421 Instruction *Term = BB->getTerminator();
9422 while (Term != BB->begin()) { // Remove instrs bottom-up
9423 BasicBlock::iterator I = Term; --I;
9425 DOUT << "IC: DCE: " << *I;
9428 if (!I->use_empty())
9429 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9430 I->eraseFromParent();
9435 while (!Worklist.empty()) {
9436 Instruction *I = RemoveOneFromWorkList();
9437 if (I == 0) continue; // skip null values.
9439 // Check to see if we can DCE the instruction.
9440 if (isInstructionTriviallyDead(I)) {
9441 // Add operands to the worklist.
9442 if (I->getNumOperands() < 4)
9443 AddUsesToWorkList(*I);
9446 DOUT << "IC: DCE: " << *I;
9448 I->eraseFromParent();
9449 RemoveFromWorkList(I);
9453 // Instruction isn't dead, see if we can constant propagate it.
9454 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9455 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9457 // Add operands to the worklist.
9458 AddUsesToWorkList(*I);
9459 ReplaceInstUsesWith(*I, C);
9462 I->eraseFromParent();
9463 RemoveFromWorkList(I);
9467 // See if we can trivially sink this instruction to a successor basic block.
9468 if (I->hasOneUse()) {
9469 BasicBlock *BB = I->getParent();
9470 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9471 if (UserParent != BB) {
9472 bool UserIsSuccessor = false;
9473 // See if the user is one of our successors.
9474 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9475 if (*SI == UserParent) {
9476 UserIsSuccessor = true;
9480 // If the user is one of our immediate successors, and if that successor
9481 // only has us as a predecessors (we'd have to split the critical edge
9482 // otherwise), we can keep going.
9483 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9484 next(pred_begin(UserParent)) == pred_end(UserParent))
9485 // Okay, the CFG is simple enough, try to sink this instruction.
9486 Changed |= TryToSinkInstruction(I, UserParent);
9490 // Now that we have an instruction, try combining it to simplify it...
9494 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
9495 if (Instruction *Result = visit(*I)) {
9497 // Should we replace the old instruction with a new one?
9499 DOUT << "IC: Old = " << *I
9500 << " New = " << *Result;
9502 // Everything uses the new instruction now.
9503 I->replaceAllUsesWith(Result);
9505 // Push the new instruction and any users onto the worklist.
9506 AddToWorkList(Result);
9507 AddUsersToWorkList(*Result);
9509 // Move the name to the new instruction first.
9510 Result->takeName(I);
9512 // Insert the new instruction into the basic block...
9513 BasicBlock *InstParent = I->getParent();
9514 BasicBlock::iterator InsertPos = I;
9516 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9517 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9520 InstParent->getInstList().insert(InsertPos, Result);
9522 // Make sure that we reprocess all operands now that we reduced their
9524 AddUsesToWorkList(*I);
9526 // Instructions can end up on the worklist more than once. Make sure
9527 // we do not process an instruction that has been deleted.
9528 RemoveFromWorkList(I);
9530 // Erase the old instruction.
9531 InstParent->getInstList().erase(I);
9534 DOUT << "IC: Mod = " << OrigI
9538 // If the instruction was modified, it's possible that it is now dead.
9539 // if so, remove it.
9540 if (isInstructionTriviallyDead(I)) {
9541 // Make sure we process all operands now that we are reducing their
9543 AddUsesToWorkList(*I);
9545 // Instructions may end up in the worklist more than once. Erase all
9546 // occurrences of this instruction.
9547 RemoveFromWorkList(I);
9548 I->eraseFromParent();
9551 AddUsersToWorkList(*I);
9558 assert(WorklistMap.empty() && "Worklist empty, but map not?");
9563 bool InstCombiner::runOnFunction(Function &F) {
9564 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
9566 bool EverMadeChange = false;
9568 // Iterate while there is work to do.
9569 unsigned Iteration = 0;
9570 while (DoOneIteration(F, Iteration++))
9571 EverMadeChange = true;
9572 return EverMadeChange;
9575 FunctionPass *llvm::createInstructionCombiningPass() {
9576 return new InstCombiner();