1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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/ParameterAttributes.h"
43 #include "llvm/Analysis/ConstantFolding.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
63 using namespace llvm::PatternMatch;
65 STATISTIC(NumCombined , "Number of insts combined");
66 STATISTIC(NumConstProp, "Number of constant folds");
67 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
68 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
69 STATISTIC(NumSunkInst , "Number of instructions sunk");
72 class VISIBILITY_HIDDEN InstCombiner
73 : public FunctionPass,
74 public InstVisitor<InstCombiner, Instruction*> {
75 // Worklist of all of the instructions that need to be simplified.
76 std::vector<Instruction*> Worklist;
77 DenseMap<Instruction*, unsigned> WorklistMap;
79 bool MustPreserveLCSSA;
81 static char ID; // Pass identification, replacement for typeid
82 InstCombiner() : FunctionPass((intptr_t)&ID) {}
84 /// AddToWorkList - Add the specified instruction to the worklist if it
85 /// isn't already in it.
86 void AddToWorkList(Instruction *I) {
87 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
88 Worklist.push_back(I);
91 // RemoveFromWorkList - remove I from the worklist if it exists.
92 void RemoveFromWorkList(Instruction *I) {
93 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
94 if (It == WorklistMap.end()) return; // Not in worklist.
96 // Don't bother moving everything down, just null out the slot.
97 Worklist[It->second] = 0;
99 WorklistMap.erase(It);
102 Instruction *RemoveOneFromWorkList() {
103 Instruction *I = Worklist.back();
105 WorklistMap.erase(I);
110 /// AddUsersToWorkList - When an instruction is simplified, add all users of
111 /// the instruction to the work lists because they might get more simplified
114 void AddUsersToWorkList(Value &I) {
115 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
117 AddToWorkList(cast<Instruction>(*UI));
120 /// AddUsesToWorkList - When an instruction is simplified, add operands to
121 /// the work lists because they might get more simplified now.
123 void AddUsesToWorkList(Instruction &I) {
124 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
125 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
129 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
130 /// dead. Add all of its operands to the worklist, turning them into
131 /// undef's to reduce the number of uses of those instructions.
133 /// Return the specified operand before it is turned into an undef.
135 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
136 Value *R = I.getOperand(op);
138 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
139 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
141 // Set the operand to undef to drop the use.
142 I.setOperand(i, UndefValue::get(Op->getType()));
149 virtual bool runOnFunction(Function &F);
151 bool DoOneIteration(Function &F, unsigned ItNum);
153 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
154 AU.addRequired<TargetData>();
155 AU.addPreservedID(LCSSAID);
156 AU.setPreservesCFG();
159 TargetData &getTargetData() const { return *TD; }
161 // Visitation implementation - Implement instruction combining for different
162 // instruction types. The semantics are as follows:
164 // null - No change was made
165 // I - Change was made, I is still valid, I may be dead though
166 // otherwise - Change was made, replace I with returned instruction
168 Instruction *visitAdd(BinaryOperator &I);
169 Instruction *visitSub(BinaryOperator &I);
170 Instruction *visitMul(BinaryOperator &I);
171 Instruction *visitURem(BinaryOperator &I);
172 Instruction *visitSRem(BinaryOperator &I);
173 Instruction *visitFRem(BinaryOperator &I);
174 Instruction *commonRemTransforms(BinaryOperator &I);
175 Instruction *commonIRemTransforms(BinaryOperator &I);
176 Instruction *commonDivTransforms(BinaryOperator &I);
177 Instruction *commonIDivTransforms(BinaryOperator &I);
178 Instruction *visitUDiv(BinaryOperator &I);
179 Instruction *visitSDiv(BinaryOperator &I);
180 Instruction *visitFDiv(BinaryOperator &I);
181 Instruction *visitAnd(BinaryOperator &I);
182 Instruction *visitOr (BinaryOperator &I);
183 Instruction *visitXor(BinaryOperator &I);
184 Instruction *visitShl(BinaryOperator &I);
185 Instruction *visitAShr(BinaryOperator &I);
186 Instruction *visitLShr(BinaryOperator &I);
187 Instruction *commonShiftTransforms(BinaryOperator &I);
188 Instruction *visitFCmpInst(FCmpInst &I);
189 Instruction *visitICmpInst(ICmpInst &I);
190 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
191 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
194 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
195 ConstantInt *DivRHS);
197 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
198 ICmpInst::Predicate Cond, Instruction &I);
199 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
201 Instruction *commonCastTransforms(CastInst &CI);
202 Instruction *commonIntCastTransforms(CastInst &CI);
203 Instruction *commonPointerCastTransforms(CastInst &CI);
204 Instruction *visitTrunc(TruncInst &CI);
205 Instruction *visitZExt(ZExtInst &CI);
206 Instruction *visitSExt(SExtInst &CI);
207 Instruction *visitFPTrunc(FPTruncInst &CI);
208 Instruction *visitFPExt(CastInst &CI);
209 Instruction *visitFPToUI(CastInst &CI);
210 Instruction *visitFPToSI(CastInst &CI);
211 Instruction *visitUIToFP(CastInst &CI);
212 Instruction *visitSIToFP(CastInst &CI);
213 Instruction *visitPtrToInt(CastInst &CI);
214 Instruction *visitIntToPtr(IntToPtrInst &CI);
215 Instruction *visitBitCast(BitCastInst &CI);
216 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
218 Instruction *visitSelectInst(SelectInst &CI);
219 Instruction *visitCallInst(CallInst &CI);
220 Instruction *visitInvokeInst(InvokeInst &II);
221 Instruction *visitPHINode(PHINode &PN);
222 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
223 Instruction *visitAllocationInst(AllocationInst &AI);
224 Instruction *visitFreeInst(FreeInst &FI);
225 Instruction *visitLoadInst(LoadInst &LI);
226 Instruction *visitStoreInst(StoreInst &SI);
227 Instruction *visitBranchInst(BranchInst &BI);
228 Instruction *visitSwitchInst(SwitchInst &SI);
229 Instruction *visitInsertElementInst(InsertElementInst &IE);
230 Instruction *visitExtractElementInst(ExtractElementInst &EI);
231 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
233 // visitInstruction - Specify what to return for unhandled instructions...
234 Instruction *visitInstruction(Instruction &I) { return 0; }
237 Instruction *visitCallSite(CallSite CS);
238 bool transformConstExprCastCall(CallSite CS);
239 Instruction *transformCallThroughTrampoline(CallSite CS);
242 // InsertNewInstBefore - insert an instruction New before instruction Old
243 // in the program. Add the new instruction to the worklist.
245 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
246 assert(New && New->getParent() == 0 &&
247 "New instruction already inserted into a basic block!");
248 BasicBlock *BB = Old.getParent();
249 BB->getInstList().insert(&Old, New); // Insert inst
254 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
255 /// This also adds the cast to the worklist. Finally, this returns the
257 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
259 if (V->getType() == Ty) return V;
261 if (Constant *CV = dyn_cast<Constant>(V))
262 return ConstantExpr::getCast(opc, CV, Ty);
264 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
269 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
270 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
274 // ReplaceInstUsesWith - This method is to be used when an instruction is
275 // found to be dead, replacable with another preexisting expression. Here
276 // we add all uses of I to the worklist, replace all uses of I with the new
277 // value, then return I, so that the inst combiner will know that I was
280 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
281 AddUsersToWorkList(I); // Add all modified instrs to worklist
283 I.replaceAllUsesWith(V);
286 // If we are replacing the instruction with itself, this must be in a
287 // segment of unreachable code, so just clobber the instruction.
288 I.replaceAllUsesWith(UndefValue::get(I.getType()));
293 // UpdateValueUsesWith - This method is to be used when an value is
294 // found to be replacable with another preexisting expression or was
295 // updated. Here we add all uses of I to the worklist, replace all uses of
296 // I with the new value (unless the instruction was just updated), then
297 // return true, so that the inst combiner will know that I was modified.
299 bool UpdateValueUsesWith(Value *Old, Value *New) {
300 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
302 Old->replaceAllUsesWith(New);
303 if (Instruction *I = dyn_cast<Instruction>(Old))
305 if (Instruction *I = dyn_cast<Instruction>(New))
310 // EraseInstFromFunction - When dealing with an instruction that has side
311 // effects or produces a void value, we can't rely on DCE to delete the
312 // instruction. Instead, visit methods should return the value returned by
314 Instruction *EraseInstFromFunction(Instruction &I) {
315 assert(I.use_empty() && "Cannot erase instruction that is used!");
316 AddUsesToWorkList(I);
317 RemoveFromWorkList(&I);
319 return 0; // Don't do anything with FI
323 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
324 /// InsertBefore instruction. This is specialized a bit to avoid inserting
325 /// casts that are known to not do anything...
327 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
328 Value *V, const Type *DestTy,
329 Instruction *InsertBefore);
331 /// SimplifyCommutative - This performs a few simplifications for
332 /// commutative operators.
333 bool SimplifyCommutative(BinaryOperator &I);
335 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
336 /// most-complex to least-complex order.
337 bool SimplifyCompare(CmpInst &I);
339 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
340 /// on the demanded bits.
341 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
342 APInt& KnownZero, APInt& KnownOne,
345 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
346 uint64_t &UndefElts, unsigned Depth = 0);
348 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
349 // PHI node as operand #0, see if we can fold the instruction into the PHI
350 // (which is only possible if all operands to the PHI are constants).
351 Instruction *FoldOpIntoPhi(Instruction &I);
353 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
354 // operator and they all are only used by the PHI, PHI together their
355 // inputs, and do the operation once, to the result of the PHI.
356 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
357 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
360 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
361 ConstantInt *AndRHS, BinaryOperator &TheAnd);
363 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
364 bool isSub, Instruction &I);
365 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
366 bool isSigned, bool Inside, Instruction &IB);
367 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
368 Instruction *MatchBSwap(BinaryOperator &I);
369 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
370 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
373 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
376 char InstCombiner::ID = 0;
377 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
380 // getComplexity: Assign a complexity or rank value to LLVM Values...
381 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
382 static unsigned getComplexity(Value *V) {
383 if (isa<Instruction>(V)) {
384 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
388 if (isa<Argument>(V)) return 3;
389 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
392 // isOnlyUse - Return true if this instruction will be deleted if we stop using
394 static bool isOnlyUse(Value *V) {
395 return V->hasOneUse() || isa<Constant>(V);
398 // getPromotedType - Return the specified type promoted as it would be to pass
399 // though a va_arg area...
400 static const Type *getPromotedType(const Type *Ty) {
401 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
402 if (ITy->getBitWidth() < 32)
403 return Type::Int32Ty;
408 /// getBitCastOperand - If the specified operand is a CastInst or a constant
409 /// expression bitcast, return the operand value, otherwise return null.
410 static Value *getBitCastOperand(Value *V) {
411 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
412 return I->getOperand(0);
413 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
414 if (CE->getOpcode() == Instruction::BitCast)
415 return CE->getOperand(0);
419 /// This function is a wrapper around CastInst::isEliminableCastPair. It
420 /// simply extracts arguments and returns what that function returns.
421 static Instruction::CastOps
422 isEliminableCastPair(
423 const CastInst *CI, ///< The first cast instruction
424 unsigned opcode, ///< The opcode of the second cast instruction
425 const Type *DstTy, ///< The target type for the second cast instruction
426 TargetData *TD ///< The target data for pointer size
429 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
430 const Type *MidTy = CI->getType(); // B from above
432 // Get the opcodes of the two Cast instructions
433 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
434 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
436 return Instruction::CastOps(
437 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
438 DstTy, TD->getIntPtrType()));
441 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
442 /// in any code being generated. It does not require codegen if V is simple
443 /// enough or if the cast can be folded into other casts.
444 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
445 const Type *Ty, TargetData *TD) {
446 if (V->getType() == Ty || isa<Constant>(V)) return false;
448 // If this is another cast that can be eliminated, it isn't codegen either.
449 if (const CastInst *CI = dyn_cast<CastInst>(V))
450 if (isEliminableCastPair(CI, opcode, Ty, TD))
455 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
456 /// InsertBefore instruction. This is specialized a bit to avoid inserting
457 /// casts that are known to not do anything...
459 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
460 Value *V, const Type *DestTy,
461 Instruction *InsertBefore) {
462 if (V->getType() == DestTy) return V;
463 if (Constant *C = dyn_cast<Constant>(V))
464 return ConstantExpr::getCast(opcode, C, DestTy);
466 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
469 // SimplifyCommutative - This performs a few simplifications for commutative
472 // 1. Order operands such that they are listed from right (least complex) to
473 // left (most complex). This puts constants before unary operators before
476 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
477 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
479 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
480 bool Changed = false;
481 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
482 Changed = !I.swapOperands();
484 if (!I.isAssociative()) return Changed;
485 Instruction::BinaryOps Opcode = I.getOpcode();
486 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
487 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
488 if (isa<Constant>(I.getOperand(1))) {
489 Constant *Folded = ConstantExpr::get(I.getOpcode(),
490 cast<Constant>(I.getOperand(1)),
491 cast<Constant>(Op->getOperand(1)));
492 I.setOperand(0, Op->getOperand(0));
493 I.setOperand(1, Folded);
495 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
496 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
497 isOnlyUse(Op) && isOnlyUse(Op1)) {
498 Constant *C1 = cast<Constant>(Op->getOperand(1));
499 Constant *C2 = cast<Constant>(Op1->getOperand(1));
501 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
502 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
503 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
507 I.setOperand(0, New);
508 I.setOperand(1, Folded);
515 /// SimplifyCompare - For a CmpInst this function just orders the operands
516 /// so that theyare listed from right (least complex) to left (most complex).
517 /// This puts constants before unary operators before binary operators.
518 bool InstCombiner::SimplifyCompare(CmpInst &I) {
519 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
522 // Compare instructions are not associative so there's nothing else we can do.
526 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
527 // if the LHS is a constant zero (which is the 'negate' form).
529 static inline Value *dyn_castNegVal(Value *V) {
530 if (BinaryOperator::isNeg(V))
531 return BinaryOperator::getNegArgument(V);
533 // Constants can be considered to be negated values if they can be folded.
534 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
535 return ConstantExpr::getNeg(C);
539 static inline Value *dyn_castNotVal(Value *V) {
540 if (BinaryOperator::isNot(V))
541 return BinaryOperator::getNotArgument(V);
543 // Constants can be considered to be not'ed values...
544 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
545 return ConstantInt::get(~C->getValue());
549 // dyn_castFoldableMul - If this value is a multiply that can be folded into
550 // other computations (because it has a constant operand), return the
551 // non-constant operand of the multiply, and set CST to point to the multiplier.
552 // Otherwise, return null.
554 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
555 if (V->hasOneUse() && V->getType()->isInteger())
556 if (Instruction *I = dyn_cast<Instruction>(V)) {
557 if (I->getOpcode() == Instruction::Mul)
558 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
559 return I->getOperand(0);
560 if (I->getOpcode() == Instruction::Shl)
561 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
562 // The multiplier is really 1 << CST.
563 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
564 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
565 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
566 return I->getOperand(0);
572 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
573 /// expression, return it.
574 static User *dyn_castGetElementPtr(Value *V) {
575 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
576 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
577 if (CE->getOpcode() == Instruction::GetElementPtr)
578 return cast<User>(V);
582 /// AddOne - Add one to a ConstantInt
583 static ConstantInt *AddOne(ConstantInt *C) {
584 APInt Val(C->getValue());
585 return ConstantInt::get(++Val);
587 /// SubOne - Subtract one from a ConstantInt
588 static ConstantInt *SubOne(ConstantInt *C) {
589 APInt Val(C->getValue());
590 return ConstantInt::get(--Val);
592 /// Add - Add two ConstantInts together
593 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
594 return ConstantInt::get(C1->getValue() + C2->getValue());
596 /// And - Bitwise AND two ConstantInts together
597 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
598 return ConstantInt::get(C1->getValue() & C2->getValue());
600 /// Subtract - Subtract one ConstantInt from another
601 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
602 return ConstantInt::get(C1->getValue() - C2->getValue());
604 /// Multiply - Multiply two ConstantInts together
605 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
606 return ConstantInt::get(C1->getValue() * C2->getValue());
608 /// MultiplyOverflows - True if the multiply can not be expressed in an int
610 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
611 uint32_t W = C1->getBitWidth();
612 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
621 APInt MulExt = LHSExt * RHSExt;
624 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
625 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
626 return MulExt.slt(Min) || MulExt.sgt(Max);
628 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
631 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
632 /// known to be either zero or one and return them in the KnownZero/KnownOne
633 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
635 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
636 /// we cannot optimize based on the assumption that it is zero without changing
637 /// it to be an explicit zero. If we don't change it to zero, other code could
638 /// optimized based on the contradictory assumption that it is non-zero.
639 /// Because instcombine aggressively folds operations with undef args anyway,
640 /// this won't lose us code quality.
641 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
642 APInt& KnownOne, unsigned Depth = 0) {
643 assert(V && "No Value?");
644 assert(Depth <= 6 && "Limit Search Depth");
645 uint32_t BitWidth = Mask.getBitWidth();
646 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
647 KnownZero.getBitWidth() == BitWidth &&
648 KnownOne.getBitWidth() == BitWidth &&
649 "V, Mask, KnownOne and KnownZero should have same BitWidth");
650 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
651 // We know all of the bits for a constant!
652 KnownOne = CI->getValue() & Mask;
653 KnownZero = ~KnownOne & Mask;
657 if (Depth == 6 || Mask == 0)
658 return; // Limit search depth.
660 Instruction *I = dyn_cast<Instruction>(V);
663 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
664 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
666 switch (I->getOpcode()) {
667 case Instruction::And: {
668 // If either the LHS or the RHS are Zero, the result is zero.
669 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
670 APInt Mask2(Mask & ~KnownZero);
671 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
672 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
673 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
675 // Output known-1 bits are only known if set in both the LHS & RHS.
676 KnownOne &= KnownOne2;
677 // Output known-0 are known to be clear if zero in either the LHS | RHS.
678 KnownZero |= KnownZero2;
681 case Instruction::Or: {
682 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
683 APInt Mask2(Mask & ~KnownOne);
684 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
685 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
686 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
688 // Output known-0 bits are only known if clear in both the LHS & RHS.
689 KnownZero &= KnownZero2;
690 // Output known-1 are known to be set if set in either the LHS | RHS.
691 KnownOne |= KnownOne2;
694 case Instruction::Xor: {
695 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
696 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
697 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
698 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
700 // Output known-0 bits are known if clear or set in both the LHS & RHS.
701 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
702 // Output known-1 are known to be set if set in only one of the LHS, RHS.
703 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
704 KnownZero = KnownZeroOut;
707 case Instruction::Select:
708 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
709 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
710 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
711 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
713 // Only known if known in both the LHS and RHS.
714 KnownOne &= KnownOne2;
715 KnownZero &= KnownZero2;
717 case Instruction::FPTrunc:
718 case Instruction::FPExt:
719 case Instruction::FPToUI:
720 case Instruction::FPToSI:
721 case Instruction::SIToFP:
722 case Instruction::PtrToInt:
723 case Instruction::UIToFP:
724 case Instruction::IntToPtr:
725 return; // Can't work with floating point or pointers
726 case Instruction::Trunc: {
727 // All these have integer operands
728 uint32_t SrcBitWidth =
729 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
731 MaskIn.zext(SrcBitWidth);
732 KnownZero.zext(SrcBitWidth);
733 KnownOne.zext(SrcBitWidth);
734 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
735 KnownZero.trunc(BitWidth);
736 KnownOne.trunc(BitWidth);
739 case Instruction::BitCast: {
740 const Type *SrcTy = I->getOperand(0)->getType();
741 if (SrcTy->isInteger()) {
742 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
747 case Instruction::ZExt: {
748 // Compute the bits in the result that are not present in the input.
749 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
750 uint32_t SrcBitWidth = SrcTy->getBitWidth();
753 MaskIn.trunc(SrcBitWidth);
754 KnownZero.trunc(SrcBitWidth);
755 KnownOne.trunc(SrcBitWidth);
756 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
757 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
758 // The top bits are known to be zero.
759 KnownZero.zext(BitWidth);
760 KnownOne.zext(BitWidth);
761 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
764 case Instruction::SExt: {
765 // Compute the bits in the result that are not present in the input.
766 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
767 uint32_t SrcBitWidth = SrcTy->getBitWidth();
770 MaskIn.trunc(SrcBitWidth);
771 KnownZero.trunc(SrcBitWidth);
772 KnownOne.trunc(SrcBitWidth);
773 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
774 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
775 KnownZero.zext(BitWidth);
776 KnownOne.zext(BitWidth);
778 // If the sign bit of the input is known set or clear, then we know the
779 // top bits of the result.
780 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
781 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
782 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
783 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
786 case Instruction::Shl:
787 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
788 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
789 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
790 APInt Mask2(Mask.lshr(ShiftAmt));
791 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
792 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
793 KnownZero <<= ShiftAmt;
794 KnownOne <<= ShiftAmt;
795 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
799 case Instruction::LShr:
800 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
801 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
802 // Compute the new bits that are at the top now.
803 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
805 // Unsigned shift right.
806 APInt Mask2(Mask.shl(ShiftAmt));
807 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
808 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
809 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
810 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
811 // high bits known zero.
812 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
816 case Instruction::AShr:
817 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
818 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
819 // Compute the new bits that are at the top now.
820 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
822 // Signed shift right.
823 APInt Mask2(Mask.shl(ShiftAmt));
824 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
825 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
826 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
827 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
829 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
830 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
831 KnownZero |= HighBits;
832 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
833 KnownOne |= HighBits;
840 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
841 /// this predicate to simplify operations downstream. Mask is known to be zero
842 /// for bits that V cannot have.
843 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
844 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
845 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
846 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
847 return (KnownZero & Mask) == Mask;
850 /// ShrinkDemandedConstant - Check to see if the specified operand of the
851 /// specified instruction is a constant integer. If so, check to see if there
852 /// are any bits set in the constant that are not demanded. If so, shrink the
853 /// constant and return true.
854 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
856 assert(I && "No instruction?");
857 assert(OpNo < I->getNumOperands() && "Operand index too large");
859 // If the operand is not a constant integer, nothing to do.
860 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
861 if (!OpC) return false;
863 // If there are no bits set that aren't demanded, nothing to do.
864 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
865 if ((~Demanded & OpC->getValue()) == 0)
868 // This instruction is producing bits that are not demanded. Shrink the RHS.
869 Demanded &= OpC->getValue();
870 I->setOperand(OpNo, ConstantInt::get(Demanded));
874 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
875 // set of known zero and one bits, compute the maximum and minimum values that
876 // could have the specified known zero and known one bits, returning them in
878 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
879 const APInt& KnownZero,
880 const APInt& KnownOne,
881 APInt& Min, APInt& Max) {
882 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
883 assert(KnownZero.getBitWidth() == BitWidth &&
884 KnownOne.getBitWidth() == BitWidth &&
885 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
886 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
887 APInt UnknownBits = ~(KnownZero|KnownOne);
889 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
890 // bit if it is unknown.
892 Max = KnownOne|UnknownBits;
894 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
896 Max.clear(BitWidth-1);
900 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
901 // a set of known zero and one bits, compute the maximum and minimum values that
902 // could have the specified known zero and known one bits, returning them in
904 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
905 const APInt &KnownZero,
906 const APInt &KnownOne,
907 APInt &Min, APInt &Max) {
908 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
909 assert(KnownZero.getBitWidth() == BitWidth &&
910 KnownOne.getBitWidth() == BitWidth &&
911 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
912 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
913 APInt UnknownBits = ~(KnownZero|KnownOne);
915 // The minimum value is when the unknown bits are all zeros.
917 // The maximum value is when the unknown bits are all ones.
918 Max = KnownOne|UnknownBits;
921 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
922 /// value based on the demanded bits. When this function is called, it is known
923 /// that only the bits set in DemandedMask of the result of V are ever used
924 /// downstream. Consequently, depending on the mask and V, it may be possible
925 /// to replace V with a constant or one of its operands. In such cases, this
926 /// function does the replacement and returns true. In all other cases, it
927 /// returns false after analyzing the expression and setting KnownOne and known
928 /// to be one in the expression. KnownZero contains all the bits that are known
929 /// to be zero in the expression. These are provided to potentially allow the
930 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
931 /// the expression. KnownOne and KnownZero always follow the invariant that
932 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
933 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
934 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
935 /// and KnownOne must all be the same.
936 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
937 APInt& KnownZero, APInt& KnownOne,
939 assert(V != 0 && "Null pointer of Value???");
940 assert(Depth <= 6 && "Limit Search Depth");
941 uint32_t BitWidth = DemandedMask.getBitWidth();
942 const IntegerType *VTy = cast<IntegerType>(V->getType());
943 assert(VTy->getBitWidth() == BitWidth &&
944 KnownZero.getBitWidth() == BitWidth &&
945 KnownOne.getBitWidth() == BitWidth &&
946 "Value *V, DemandedMask, KnownZero and KnownOne \
947 must have same BitWidth");
948 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
949 // We know all of the bits for a constant!
950 KnownOne = CI->getValue() & DemandedMask;
951 KnownZero = ~KnownOne & DemandedMask;
957 if (!V->hasOneUse()) { // Other users may use these bits.
958 if (Depth != 0) { // Not at the root.
959 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
960 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
963 // If this is the root being simplified, allow it to have multiple uses,
964 // just set the DemandedMask to all bits.
965 DemandedMask = APInt::getAllOnesValue(BitWidth);
966 } else if (DemandedMask == 0) { // Not demanding any bits from V.
967 if (V != UndefValue::get(VTy))
968 return UpdateValueUsesWith(V, UndefValue::get(VTy));
970 } else if (Depth == 6) { // Limit search depth.
974 Instruction *I = dyn_cast<Instruction>(V);
975 if (!I) return false; // Only analyze instructions.
977 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
978 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
979 switch (I->getOpcode()) {
981 case Instruction::And:
982 // If either the LHS or the RHS are Zero, the result is zero.
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?");
989 // If something is known zero on the RHS, the bits aren't demanded on the
991 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
992 LHSKnownZero, LHSKnownOne, Depth+1))
994 assert((LHSKnownZero & LHSKnownOne) == 0 &&
995 "Bits known to be one AND zero?");
997 // If all of the demanded bits are known 1 on one side, return the other.
998 // These bits cannot contribute to the result of the 'and'.
999 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1000 (DemandedMask & ~LHSKnownZero))
1001 return UpdateValueUsesWith(I, I->getOperand(0));
1002 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1003 (DemandedMask & ~RHSKnownZero))
1004 return UpdateValueUsesWith(I, I->getOperand(1));
1006 // If all of the demanded bits in the inputs are known zeros, return zero.
1007 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1008 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1010 // If the RHS is a constant, see if we can simplify it.
1011 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1012 return UpdateValueUsesWith(I, I);
1014 // Output known-1 bits are only known if set in both the LHS & RHS.
1015 RHSKnownOne &= LHSKnownOne;
1016 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1017 RHSKnownZero |= LHSKnownZero;
1019 case Instruction::Or:
1020 // If either the LHS or the RHS are One, the result is One.
1021 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1022 RHSKnownZero, RHSKnownOne, Depth+1))
1024 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1025 "Bits known to be one AND zero?");
1026 // If something is known one on the RHS, the bits aren't demanded on the
1028 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1029 LHSKnownZero, LHSKnownOne, Depth+1))
1031 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1032 "Bits known to be one AND zero?");
1034 // If all of the demanded bits are known zero on one side, return the other.
1035 // These bits cannot contribute to the result of the 'or'.
1036 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1037 (DemandedMask & ~LHSKnownOne))
1038 return UpdateValueUsesWith(I, I->getOperand(0));
1039 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1040 (DemandedMask & ~RHSKnownOne))
1041 return UpdateValueUsesWith(I, I->getOperand(1));
1043 // If all of the potentially set bits on one side are known to be set on
1044 // the other side, just use the 'other' side.
1045 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1046 (DemandedMask & (~RHSKnownZero)))
1047 return UpdateValueUsesWith(I, I->getOperand(0));
1048 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1049 (DemandedMask & (~LHSKnownZero)))
1050 return UpdateValueUsesWith(I, I->getOperand(1));
1052 // If the RHS is a constant, see if we can simplify it.
1053 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1054 return UpdateValueUsesWith(I, I);
1056 // Output known-0 bits are only known if clear in both the LHS & RHS.
1057 RHSKnownZero &= LHSKnownZero;
1058 // Output known-1 are known to be set if set in either the LHS | RHS.
1059 RHSKnownOne |= LHSKnownOne;
1061 case Instruction::Xor: {
1062 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1063 RHSKnownZero, RHSKnownOne, Depth+1))
1065 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1066 "Bits known to be one AND zero?");
1067 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1068 LHSKnownZero, LHSKnownOne, Depth+1))
1070 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1071 "Bits known to be one AND zero?");
1073 // If all of the demanded bits are known zero on one side, return the other.
1074 // These bits cannot contribute to the result of the 'xor'.
1075 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1076 return UpdateValueUsesWith(I, I->getOperand(0));
1077 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1078 return UpdateValueUsesWith(I, I->getOperand(1));
1080 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1081 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1082 (RHSKnownOne & LHSKnownOne);
1083 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1084 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1085 (RHSKnownOne & LHSKnownZero);
1087 // If all of the demanded bits are known to be zero on one side or the
1088 // other, turn this into an *inclusive* or.
1089 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1090 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1092 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1094 InsertNewInstBefore(Or, *I);
1095 return UpdateValueUsesWith(I, Or);
1098 // If all of the demanded bits on one side are known, and all of the set
1099 // bits on that side are also known to be set on the other side, turn this
1100 // into an AND, as we know the bits will be cleared.
1101 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1102 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1104 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1105 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1107 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1108 InsertNewInstBefore(And, *I);
1109 return UpdateValueUsesWith(I, And);
1113 // If the RHS is a constant, see if we can simplify it.
1114 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1115 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1116 return UpdateValueUsesWith(I, I);
1118 RHSKnownZero = KnownZeroOut;
1119 RHSKnownOne = KnownOneOut;
1122 case Instruction::Select:
1123 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1124 RHSKnownZero, RHSKnownOne, Depth+1))
1126 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1127 LHSKnownZero, LHSKnownOne, Depth+1))
1129 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1130 "Bits known to be one AND zero?");
1131 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1132 "Bits known to be one AND zero?");
1134 // If the operands are constants, see if we can simplify them.
1135 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1136 return UpdateValueUsesWith(I, I);
1137 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1138 return UpdateValueUsesWith(I, I);
1140 // Only known if known in both the LHS and RHS.
1141 RHSKnownOne &= LHSKnownOne;
1142 RHSKnownZero &= LHSKnownZero;
1144 case Instruction::Trunc: {
1146 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1147 DemandedMask.zext(truncBf);
1148 RHSKnownZero.zext(truncBf);
1149 RHSKnownOne.zext(truncBf);
1150 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1151 RHSKnownZero, RHSKnownOne, Depth+1))
1153 DemandedMask.trunc(BitWidth);
1154 RHSKnownZero.trunc(BitWidth);
1155 RHSKnownOne.trunc(BitWidth);
1156 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1157 "Bits known to be one AND zero?");
1160 case Instruction::BitCast:
1161 if (!I->getOperand(0)->getType()->isInteger())
1164 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1165 RHSKnownZero, RHSKnownOne, Depth+1))
1167 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1168 "Bits known to be one AND zero?");
1170 case Instruction::ZExt: {
1171 // Compute the bits in the result that are not present in the input.
1172 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1173 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1175 DemandedMask.trunc(SrcBitWidth);
1176 RHSKnownZero.trunc(SrcBitWidth);
1177 RHSKnownOne.trunc(SrcBitWidth);
1178 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1179 RHSKnownZero, RHSKnownOne, Depth+1))
1181 DemandedMask.zext(BitWidth);
1182 RHSKnownZero.zext(BitWidth);
1183 RHSKnownOne.zext(BitWidth);
1184 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1185 "Bits known to be one AND zero?");
1186 // The top bits are known to be zero.
1187 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1190 case Instruction::SExt: {
1191 // Compute the bits in the result that are not present in the input.
1192 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1193 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1195 APInt InputDemandedBits = DemandedMask &
1196 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1198 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1199 // If any of the sign extended bits are demanded, we know that the sign
1201 if ((NewBits & DemandedMask) != 0)
1202 InputDemandedBits.set(SrcBitWidth-1);
1204 InputDemandedBits.trunc(SrcBitWidth);
1205 RHSKnownZero.trunc(SrcBitWidth);
1206 RHSKnownOne.trunc(SrcBitWidth);
1207 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1208 RHSKnownZero, RHSKnownOne, Depth+1))
1210 InputDemandedBits.zext(BitWidth);
1211 RHSKnownZero.zext(BitWidth);
1212 RHSKnownOne.zext(BitWidth);
1213 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1214 "Bits known to be one AND zero?");
1216 // If the sign bit of the input is known set or clear, then we know the
1217 // top bits of the result.
1219 // If the input sign bit is known zero, or if the NewBits are not demanded
1220 // convert this into a zero extension.
1221 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1223 // Convert to ZExt cast
1224 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1225 return UpdateValueUsesWith(I, NewCast);
1226 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1227 RHSKnownOne |= NewBits;
1231 case Instruction::Add: {
1232 // Figure out what the input bits are. If the top bits of the and result
1233 // are not demanded, then the add doesn't demand them from its input
1235 uint32_t NLZ = DemandedMask.countLeadingZeros();
1237 // If there is a constant on the RHS, there are a variety of xformations
1239 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1240 // If null, this should be simplified elsewhere. Some of the xforms here
1241 // won't work if the RHS is zero.
1245 // If the top bit of the output is demanded, demand everything from the
1246 // input. Otherwise, we demand all the input bits except NLZ top bits.
1247 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1249 // Find information about known zero/one bits in the input.
1250 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1251 LHSKnownZero, LHSKnownOne, Depth+1))
1254 // If the RHS of the add has bits set that can't affect the input, reduce
1256 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1257 return UpdateValueUsesWith(I, I);
1259 // Avoid excess work.
1260 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1263 // Turn it into OR if input bits are zero.
1264 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1266 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1268 InsertNewInstBefore(Or, *I);
1269 return UpdateValueUsesWith(I, Or);
1272 // We can say something about the output known-zero and known-one bits,
1273 // depending on potential carries from the input constant and the
1274 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1275 // bits set and the RHS constant is 0x01001, then we know we have a known
1276 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1278 // To compute this, we first compute the potential carry bits. These are
1279 // the bits which may be modified. I'm not aware of a better way to do
1281 const APInt& RHSVal = RHS->getValue();
1282 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1284 // Now that we know which bits have carries, compute the known-1/0 sets.
1286 // Bits are known one if they are known zero in one operand and one in the
1287 // other, and there is no input carry.
1288 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1289 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1291 // Bits are known zero if they are known zero in both operands and there
1292 // is no input carry.
1293 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1295 // If the high-bits of this ADD are not demanded, then it does not demand
1296 // the high bits of its LHS or RHS.
1297 if (DemandedMask[BitWidth-1] == 0) {
1298 // Right fill the mask of bits for this ADD to demand the most
1299 // significant bit and all those below it.
1300 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1301 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1302 LHSKnownZero, LHSKnownOne, Depth+1))
1304 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1305 LHSKnownZero, LHSKnownOne, Depth+1))
1311 case Instruction::Sub:
1312 // If the high-bits of this SUB are not demanded, then it does not demand
1313 // the high bits of its LHS or RHS.
1314 if (DemandedMask[BitWidth-1] == 0) {
1315 // Right fill the mask of bits for this SUB to demand the most
1316 // significant bit and all those below it.
1317 uint32_t NLZ = DemandedMask.countLeadingZeros();
1318 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1319 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1320 LHSKnownZero, LHSKnownOne, Depth+1))
1322 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1323 LHSKnownZero, LHSKnownOne, Depth+1))
1327 case Instruction::Shl:
1328 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1329 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1330 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1331 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1332 RHSKnownZero, RHSKnownOne, Depth+1))
1334 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1335 "Bits known to be one AND zero?");
1336 RHSKnownZero <<= ShiftAmt;
1337 RHSKnownOne <<= ShiftAmt;
1338 // low bits known zero.
1340 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1343 case Instruction::LShr:
1344 // For a logical shift right
1345 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1346 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1348 // Unsigned shift right.
1349 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1350 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1351 RHSKnownZero, RHSKnownOne, Depth+1))
1353 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1354 "Bits known to be one AND zero?");
1355 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1356 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1358 // Compute the new bits that are at the top now.
1359 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1360 RHSKnownZero |= HighBits; // high bits known zero.
1364 case Instruction::AShr:
1365 // If this is an arithmetic shift right and only the low-bit is set, we can
1366 // always convert this into a logical shr, even if the shift amount is
1367 // variable. The low bit of the shift cannot be an input sign bit unless
1368 // the shift amount is >= the size of the datatype, which is undefined.
1369 if (DemandedMask == 1) {
1370 // Perform the logical shift right.
1371 Value *NewVal = BinaryOperator::createLShr(
1372 I->getOperand(0), I->getOperand(1), I->getName());
1373 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1374 return UpdateValueUsesWith(I, NewVal);
1377 // If the sign bit is the only bit demanded by this ashr, then there is no
1378 // need to do it, the shift doesn't change the high bit.
1379 if (DemandedMask.isSignBit())
1380 return UpdateValueUsesWith(I, I->getOperand(0));
1382 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1383 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1385 // Signed shift right.
1386 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1387 // If any of the "high bits" are demanded, we should set the sign bit as
1389 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1390 DemandedMaskIn.set(BitWidth-1);
1391 if (SimplifyDemandedBits(I->getOperand(0),
1393 RHSKnownZero, RHSKnownOne, Depth+1))
1395 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1396 "Bits known to be one AND zero?");
1397 // Compute the new bits that are at the top now.
1398 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1399 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1400 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1402 // Handle the sign bits.
1403 APInt SignBit(APInt::getSignBit(BitWidth));
1404 // Adjust to where it is now in the mask.
1405 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1407 // If the input sign bit is known to be zero, or if none of the top bits
1408 // are demanded, turn this into an unsigned shift right.
1409 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1410 (HighBits & ~DemandedMask) == HighBits) {
1411 // Perform the logical shift right.
1412 Value *NewVal = BinaryOperator::createLShr(
1413 I->getOperand(0), SA, I->getName());
1414 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1415 return UpdateValueUsesWith(I, NewVal);
1416 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1417 RHSKnownOne |= HighBits;
1423 // If the client is only demanding bits that we know, return the known
1425 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1426 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1431 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1432 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1433 /// actually used by the caller. This method analyzes which elements of the
1434 /// operand are undef and returns that information in UndefElts.
1436 /// If the information about demanded elements can be used to simplify the
1437 /// operation, the operation is simplified, then the resultant value is
1438 /// returned. This returns null if no change was made.
1439 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1440 uint64_t &UndefElts,
1442 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1443 assert(VWidth <= 64 && "Vector too wide to analyze!");
1444 uint64_t EltMask = ~0ULL >> (64-VWidth);
1445 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1446 "Invalid DemandedElts!");
1448 if (isa<UndefValue>(V)) {
1449 // If the entire vector is undefined, just return this info.
1450 UndefElts = EltMask;
1452 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1453 UndefElts = EltMask;
1454 return UndefValue::get(V->getType());
1458 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1459 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1460 Constant *Undef = UndefValue::get(EltTy);
1462 std::vector<Constant*> Elts;
1463 for (unsigned i = 0; i != VWidth; ++i)
1464 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1465 Elts.push_back(Undef);
1466 UndefElts |= (1ULL << i);
1467 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1468 Elts.push_back(Undef);
1469 UndefElts |= (1ULL << i);
1470 } else { // Otherwise, defined.
1471 Elts.push_back(CP->getOperand(i));
1474 // If we changed the constant, return it.
1475 Constant *NewCP = ConstantVector::get(Elts);
1476 return NewCP != CP ? NewCP : 0;
1477 } else if (isa<ConstantAggregateZero>(V)) {
1478 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1480 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1481 Constant *Zero = Constant::getNullValue(EltTy);
1482 Constant *Undef = UndefValue::get(EltTy);
1483 std::vector<Constant*> Elts;
1484 for (unsigned i = 0; i != VWidth; ++i)
1485 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1486 UndefElts = DemandedElts ^ EltMask;
1487 return ConstantVector::get(Elts);
1490 if (!V->hasOneUse()) { // Other users may use these bits.
1491 if (Depth != 0) { // Not at the root.
1492 // TODO: Just compute the UndefElts information recursively.
1496 } else if (Depth == 10) { // Limit search depth.
1500 Instruction *I = dyn_cast<Instruction>(V);
1501 if (!I) return false; // Only analyze instructions.
1503 bool MadeChange = false;
1504 uint64_t UndefElts2;
1506 switch (I->getOpcode()) {
1509 case Instruction::InsertElement: {
1510 // If this is a variable index, we don't know which element it overwrites.
1511 // demand exactly the same input as we produce.
1512 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1514 // Note that we can't propagate undef elt info, because we don't know
1515 // which elt is getting updated.
1516 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1517 UndefElts2, Depth+1);
1518 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1522 // If this is inserting an element that isn't demanded, remove this
1524 unsigned IdxNo = Idx->getZExtValue();
1525 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1526 return AddSoonDeadInstToWorklist(*I, 0);
1528 // Otherwise, the element inserted overwrites whatever was there, so the
1529 // input demanded set is simpler than the output set.
1530 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1531 DemandedElts & ~(1ULL << IdxNo),
1532 UndefElts, Depth+1);
1533 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1535 // The inserted element is defined.
1536 UndefElts |= 1ULL << IdxNo;
1539 case Instruction::BitCast: {
1540 // Vector->vector casts only.
1541 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1543 unsigned InVWidth = VTy->getNumElements();
1544 uint64_t InputDemandedElts = 0;
1547 if (VWidth == InVWidth) {
1548 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1549 // elements as are demanded of us.
1551 InputDemandedElts = DemandedElts;
1552 } else if (VWidth > InVWidth) {
1556 // If there are more elements in the result than there are in the source,
1557 // then an input element is live if any of the corresponding output
1558 // elements are live.
1559 Ratio = VWidth/InVWidth;
1560 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1561 if (DemandedElts & (1ULL << OutIdx))
1562 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1568 // If there are more elements in the source than there are in the result,
1569 // then an input element is live if the corresponding output element is
1571 Ratio = InVWidth/VWidth;
1572 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1573 if (DemandedElts & (1ULL << InIdx/Ratio))
1574 InputDemandedElts |= 1ULL << InIdx;
1577 // div/rem demand all inputs, because they don't want divide by zero.
1578 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1579 UndefElts2, Depth+1);
1581 I->setOperand(0, TmpV);
1585 UndefElts = UndefElts2;
1586 if (VWidth > InVWidth) {
1587 assert(0 && "Unimp");
1588 // If there are more elements in the result than there are in the source,
1589 // then an output element is undef if the corresponding input element is
1591 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1592 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1593 UndefElts |= 1ULL << OutIdx;
1594 } else if (VWidth < InVWidth) {
1595 assert(0 && "Unimp");
1596 // If there are more elements in the source than there are in the result,
1597 // then a result element is undef if all of the corresponding input
1598 // elements are undef.
1599 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1600 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1601 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1602 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1606 case Instruction::And:
1607 case Instruction::Or:
1608 case Instruction::Xor:
1609 case Instruction::Add:
1610 case Instruction::Sub:
1611 case Instruction::Mul:
1612 // div/rem demand all inputs, because they don't want divide by zero.
1613 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1614 UndefElts, Depth+1);
1615 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1616 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1617 UndefElts2, Depth+1);
1618 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1620 // Output elements are undefined if both are undefined. Consider things
1621 // like undef&0. The result is known zero, not undef.
1622 UndefElts &= UndefElts2;
1625 case Instruction::Call: {
1626 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1628 switch (II->getIntrinsicID()) {
1631 // Binary vector operations that work column-wise. A dest element is a
1632 // function of the corresponding input elements from the two inputs.
1633 case Intrinsic::x86_sse_sub_ss:
1634 case Intrinsic::x86_sse_mul_ss:
1635 case Intrinsic::x86_sse_min_ss:
1636 case Intrinsic::x86_sse_max_ss:
1637 case Intrinsic::x86_sse2_sub_sd:
1638 case Intrinsic::x86_sse2_mul_sd:
1639 case Intrinsic::x86_sse2_min_sd:
1640 case Intrinsic::x86_sse2_max_sd:
1641 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1642 UndefElts, Depth+1);
1643 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1644 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1645 UndefElts2, Depth+1);
1646 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1648 // If only the low elt is demanded and this is a scalarizable intrinsic,
1649 // scalarize it now.
1650 if (DemandedElts == 1) {
1651 switch (II->getIntrinsicID()) {
1653 case Intrinsic::x86_sse_sub_ss:
1654 case Intrinsic::x86_sse_mul_ss:
1655 case Intrinsic::x86_sse2_sub_sd:
1656 case Intrinsic::x86_sse2_mul_sd:
1657 // TODO: Lower MIN/MAX/ABS/etc
1658 Value *LHS = II->getOperand(1);
1659 Value *RHS = II->getOperand(2);
1660 // Extract the element as scalars.
1661 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1662 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1664 switch (II->getIntrinsicID()) {
1665 default: assert(0 && "Case stmts out of sync!");
1666 case Intrinsic::x86_sse_sub_ss:
1667 case Intrinsic::x86_sse2_sub_sd:
1668 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1669 II->getName()), *II);
1671 case Intrinsic::x86_sse_mul_ss:
1672 case Intrinsic::x86_sse2_mul_sd:
1673 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1674 II->getName()), *II);
1679 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1681 InsertNewInstBefore(New, *II);
1682 AddSoonDeadInstToWorklist(*II, 0);
1687 // Output elements are undefined if both are undefined. Consider things
1688 // like undef&0. The result is known zero, not undef.
1689 UndefElts &= UndefElts2;
1695 return MadeChange ? I : 0;
1698 /// @returns true if the specified compare predicate is
1699 /// true when both operands are equal...
1700 /// @brief Determine if the icmp Predicate is true when both operands are equal
1701 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1702 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1703 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1704 pred == ICmpInst::ICMP_SLE;
1707 /// @returns true if the specified compare instruction is
1708 /// true when both operands are equal...
1709 /// @brief Determine if the ICmpInst returns true when both operands are equal
1710 static bool isTrueWhenEqual(ICmpInst &ICI) {
1711 return isTrueWhenEqual(ICI.getPredicate());
1714 /// AssociativeOpt - Perform an optimization on an associative operator. This
1715 /// function is designed to check a chain of associative operators for a
1716 /// potential to apply a certain optimization. Since the optimization may be
1717 /// applicable if the expression was reassociated, this checks the chain, then
1718 /// reassociates the expression as necessary to expose the optimization
1719 /// opportunity. This makes use of a special Functor, which must define
1720 /// 'shouldApply' and 'apply' methods.
1722 template<typename Functor>
1723 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1724 unsigned Opcode = Root.getOpcode();
1725 Value *LHS = Root.getOperand(0);
1727 // Quick check, see if the immediate LHS matches...
1728 if (F.shouldApply(LHS))
1729 return F.apply(Root);
1731 // Otherwise, if the LHS is not of the same opcode as the root, return.
1732 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1733 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1734 // Should we apply this transform to the RHS?
1735 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1737 // If not to the RHS, check to see if we should apply to the LHS...
1738 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1739 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1743 // If the functor wants to apply the optimization to the RHS of LHSI,
1744 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1746 BasicBlock *BB = Root.getParent();
1748 // Now all of the instructions are in the current basic block, go ahead
1749 // and perform the reassociation.
1750 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1752 // First move the selected RHS to the LHS of the root...
1753 Root.setOperand(0, LHSI->getOperand(1));
1755 // Make what used to be the LHS of the root be the user of the root...
1756 Value *ExtraOperand = TmpLHSI->getOperand(1);
1757 if (&Root == TmpLHSI) {
1758 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1761 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1762 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1763 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1764 BasicBlock::iterator ARI = &Root; ++ARI;
1765 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1768 // Now propagate the ExtraOperand down the chain of instructions until we
1770 while (TmpLHSI != LHSI) {
1771 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1772 // Move the instruction to immediately before the chain we are
1773 // constructing to avoid breaking dominance properties.
1774 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1775 BB->getInstList().insert(ARI, NextLHSI);
1778 Value *NextOp = NextLHSI->getOperand(1);
1779 NextLHSI->setOperand(1, ExtraOperand);
1781 ExtraOperand = NextOp;
1784 // Now that the instructions are reassociated, have the functor perform
1785 // the transformation...
1786 return F.apply(Root);
1789 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1795 // AddRHS - Implements: X + X --> X << 1
1798 AddRHS(Value *rhs) : RHS(rhs) {}
1799 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1800 Instruction *apply(BinaryOperator &Add) const {
1801 return BinaryOperator::createShl(Add.getOperand(0),
1802 ConstantInt::get(Add.getType(), 1));
1806 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1808 struct AddMaskingAnd {
1810 AddMaskingAnd(Constant *c) : C2(c) {}
1811 bool shouldApply(Value *LHS) const {
1813 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1814 ConstantExpr::getAnd(C1, C2)->isNullValue();
1816 Instruction *apply(BinaryOperator &Add) const {
1817 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1821 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1823 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1824 if (Constant *SOC = dyn_cast<Constant>(SO))
1825 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1827 return IC->InsertNewInstBefore(CastInst::create(
1828 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1831 // Figure out if the constant is the left or the right argument.
1832 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1833 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1835 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1837 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1838 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1841 Value *Op0 = SO, *Op1 = ConstOperand;
1843 std::swap(Op0, Op1);
1845 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1846 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1847 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1848 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1849 SO->getName()+".cmp");
1851 assert(0 && "Unknown binary instruction type!");
1854 return IC->InsertNewInstBefore(New, I);
1857 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1858 // constant as the other operand, try to fold the binary operator into the
1859 // select arguments. This also works for Cast instructions, which obviously do
1860 // not have a second operand.
1861 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1863 // Don't modify shared select instructions
1864 if (!SI->hasOneUse()) return 0;
1865 Value *TV = SI->getOperand(1);
1866 Value *FV = SI->getOperand(2);
1868 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1869 // Bool selects with constant operands can be folded to logical ops.
1870 if (SI->getType() == Type::Int1Ty) return 0;
1872 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1873 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1875 return new SelectInst(SI->getCondition(), SelectTrueVal,
1882 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1883 /// node as operand #0, see if we can fold the instruction into the PHI (which
1884 /// is only possible if all operands to the PHI are constants).
1885 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1886 PHINode *PN = cast<PHINode>(I.getOperand(0));
1887 unsigned NumPHIValues = PN->getNumIncomingValues();
1888 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1890 // Check to see if all of the operands of the PHI are constants. If there is
1891 // one non-constant value, remember the BB it is. If there is more than one
1892 // or if *it* is a PHI, bail out.
1893 BasicBlock *NonConstBB = 0;
1894 for (unsigned i = 0; i != NumPHIValues; ++i)
1895 if (!isa<Constant>(PN->getIncomingValue(i))) {
1896 if (NonConstBB) return 0; // More than one non-const value.
1897 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1898 NonConstBB = PN->getIncomingBlock(i);
1900 // If the incoming non-constant value is in I's block, we have an infinite
1902 if (NonConstBB == I.getParent())
1906 // If there is exactly one non-constant value, we can insert a copy of the
1907 // operation in that block. However, if this is a critical edge, we would be
1908 // inserting the computation one some other paths (e.g. inside a loop). Only
1909 // do this if the pred block is unconditionally branching into the phi block.
1911 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1912 if (!BI || !BI->isUnconditional()) return 0;
1915 // Okay, we can do the transformation: create the new PHI node.
1916 PHINode *NewPN = new PHINode(I.getType(), "");
1917 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1918 InsertNewInstBefore(NewPN, *PN);
1919 NewPN->takeName(PN);
1921 // Next, add all of the operands to the PHI.
1922 if (I.getNumOperands() == 2) {
1923 Constant *C = cast<Constant>(I.getOperand(1));
1924 for (unsigned i = 0; i != NumPHIValues; ++i) {
1926 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1927 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1928 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1930 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1932 assert(PN->getIncomingBlock(i) == NonConstBB);
1933 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1934 InV = BinaryOperator::create(BO->getOpcode(),
1935 PN->getIncomingValue(i), C, "phitmp",
1936 NonConstBB->getTerminator());
1937 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1938 InV = CmpInst::create(CI->getOpcode(),
1940 PN->getIncomingValue(i), C, "phitmp",
1941 NonConstBB->getTerminator());
1943 assert(0 && "Unknown binop!");
1945 AddToWorkList(cast<Instruction>(InV));
1947 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1950 CastInst *CI = cast<CastInst>(&I);
1951 const Type *RetTy = CI->getType();
1952 for (unsigned i = 0; i != NumPHIValues; ++i) {
1954 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1955 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1957 assert(PN->getIncomingBlock(i) == NonConstBB);
1958 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1959 I.getType(), "phitmp",
1960 NonConstBB->getTerminator());
1961 AddToWorkList(cast<Instruction>(InV));
1963 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1966 return ReplaceInstUsesWith(I, NewPN);
1970 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1971 /// value is never equal to -0.0.
1973 /// Note that this function will need to be revisited when we support nondefault
1976 static bool CannotBeNegativeZero(const Value *V) {
1977 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1978 return !CFP->getValueAPF().isNegZero();
1980 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1981 if (const Instruction *I = dyn_cast<Instruction>(V)) {
1982 if (I->getOpcode() == Instruction::Add &&
1983 isa<ConstantFP>(I->getOperand(1)) &&
1984 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1987 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1988 if (II->getIntrinsicID() == Intrinsic::sqrt)
1989 return CannotBeNegativeZero(II->getOperand(1));
1991 if (const CallInst *CI = dyn_cast<CallInst>(I))
1992 if (const Function *F = CI->getCalledFunction()) {
1993 if (F->isDeclaration()) {
1994 switch (F->getNameLen()) {
1995 case 3: // abs(x) != -0.0
1996 if (!strcmp(F->getNameStart(), "abs")) return true;
1998 case 4: // abs[lf](x) != -0.0
1999 if (!strcmp(F->getNameStart(), "absf")) return true;
2000 if (!strcmp(F->getNameStart(), "absl")) return true;
2011 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2012 bool Changed = SimplifyCommutative(I);
2013 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2015 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2016 // X + undef -> undef
2017 if (isa<UndefValue>(RHS))
2018 return ReplaceInstUsesWith(I, RHS);
2021 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2022 if (RHSC->isNullValue())
2023 return ReplaceInstUsesWith(I, LHS);
2024 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2025 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2026 (I.getType())->getValueAPF()))
2027 return ReplaceInstUsesWith(I, LHS);
2030 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2031 // X + (signbit) --> X ^ signbit
2032 const APInt& Val = CI->getValue();
2033 uint32_t BitWidth = Val.getBitWidth();
2034 if (Val == APInt::getSignBit(BitWidth))
2035 return BinaryOperator::createXor(LHS, RHS);
2037 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2038 // (X & 254)+1 -> (X&254)|1
2039 if (!isa<VectorType>(I.getType())) {
2040 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2041 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2042 KnownZero, KnownOne))
2047 if (isa<PHINode>(LHS))
2048 if (Instruction *NV = FoldOpIntoPhi(I))
2051 ConstantInt *XorRHS = 0;
2053 if (isa<ConstantInt>(RHSC) &&
2054 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2055 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2056 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2058 uint32_t Size = TySizeBits / 2;
2059 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2060 APInt CFF80Val(-C0080Val);
2062 if (TySizeBits > Size) {
2063 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2064 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2065 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2066 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2067 // This is a sign extend if the top bits are known zero.
2068 if (!MaskedValueIsZero(XorLHS,
2069 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2070 Size = 0; // Not a sign ext, but can't be any others either.
2075 C0080Val = APIntOps::lshr(C0080Val, Size);
2076 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2077 } while (Size >= 1);
2079 // FIXME: This shouldn't be necessary. When the backends can handle types
2080 // with funny bit widths then this whole cascade of if statements should
2081 // be removed. It is just here to get the size of the "middle" type back
2082 // up to something that the back ends can handle.
2083 const Type *MiddleType = 0;
2086 case 32: MiddleType = Type::Int32Ty; break;
2087 case 16: MiddleType = Type::Int16Ty; break;
2088 case 8: MiddleType = Type::Int8Ty; break;
2091 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2092 InsertNewInstBefore(NewTrunc, I);
2093 return new SExtInst(NewTrunc, I.getType(), I.getName());
2099 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2100 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2102 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2103 if (RHSI->getOpcode() == Instruction::Sub)
2104 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2105 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2107 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2108 if (LHSI->getOpcode() == Instruction::Sub)
2109 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2110 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2115 // -A + -B --> -(A + B)
2116 if (Value *LHSV = dyn_castNegVal(LHS)) {
2117 if (LHS->getType()->isIntOrIntVector()) {
2118 if (Value *RHSV = dyn_castNegVal(RHS)) {
2119 Instruction *NewAdd = BinaryOperator::createAdd(LHSV, RHSV, "sum");
2120 InsertNewInstBefore(NewAdd, I);
2121 return BinaryOperator::createNeg(NewAdd);
2125 return BinaryOperator::createSub(RHS, LHSV);
2129 if (!isa<Constant>(RHS))
2130 if (Value *V = dyn_castNegVal(RHS))
2131 return BinaryOperator::createSub(LHS, V);
2135 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2136 if (X == RHS) // X*C + X --> X * (C+1)
2137 return BinaryOperator::createMul(RHS, AddOne(C2));
2139 // X*C1 + X*C2 --> X * (C1+C2)
2141 if (X == dyn_castFoldableMul(RHS, C1))
2142 return BinaryOperator::createMul(X, Add(C1, C2));
2145 // X + X*C --> X * (C+1)
2146 if (dyn_castFoldableMul(RHS, C2) == LHS)
2147 return BinaryOperator::createMul(LHS, AddOne(C2));
2149 // X + ~X --> -1 since ~X = -X-1
2150 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2151 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2154 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2155 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2156 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2159 // W*X + Y*Z --> W * (X+Z) iff W == Y
2160 if (I.getType()->isIntOrIntVector()) {
2161 Value *W, *X, *Y, *Z;
2162 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2163 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2167 } else if (Y == X) {
2169 } else if (X == Z) {
2176 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, Z,
2177 LHS->getName()), I);
2178 return BinaryOperator::createMul(W, NewAdd);
2183 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2185 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2186 return BinaryOperator::createSub(SubOne(CRHS), X);
2188 // (X & FF00) + xx00 -> (X+xx00) & FF00
2189 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2190 Constant *Anded = And(CRHS, C2);
2191 if (Anded == CRHS) {
2192 // See if all bits from the first bit set in the Add RHS up are included
2193 // in the mask. First, get the rightmost bit.
2194 const APInt& AddRHSV = CRHS->getValue();
2196 // Form a mask of all bits from the lowest bit added through the top.
2197 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2199 // See if the and mask includes all of these bits.
2200 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2202 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2203 // Okay, the xform is safe. Insert the new add pronto.
2204 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2205 LHS->getName()), I);
2206 return BinaryOperator::createAnd(NewAdd, C2);
2211 // Try to fold constant add into select arguments.
2212 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2213 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2217 // add (cast *A to intptrtype) B ->
2218 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2220 CastInst *CI = dyn_cast<CastInst>(LHS);
2223 CI = dyn_cast<CastInst>(RHS);
2226 if (CI && CI->getType()->isSized() &&
2227 (CI->getType()->getPrimitiveSizeInBits() ==
2228 TD->getIntPtrType()->getPrimitiveSizeInBits())
2229 && isa<PointerType>(CI->getOperand(0)->getType())) {
2231 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2232 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2233 PointerType::get(Type::Int8Ty, AS), I);
2234 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2235 return new PtrToIntInst(I2, CI->getType());
2239 // add (select X 0 (sub n A)) A --> select X A n
2241 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2244 SI = dyn_cast<SelectInst>(RHS);
2247 if (SI && SI->hasOneUse()) {
2248 Value *TV = SI->getTrueValue();
2249 Value *FV = SI->getFalseValue();
2252 // Can we fold the add into the argument of the select?
2253 // We check both true and false select arguments for a matching subtract.
2254 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2255 A == Other) // Fold the add into the true select value.
2256 return new SelectInst(SI->getCondition(), N, A);
2257 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2258 A == Other) // Fold the add into the false select value.
2259 return new SelectInst(SI->getCondition(), A, N);
2263 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2264 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2265 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2266 return ReplaceInstUsesWith(I, LHS);
2268 return Changed ? &I : 0;
2271 // isSignBit - Return true if the value represented by the constant only has the
2272 // highest order bit set.
2273 static bool isSignBit(ConstantInt *CI) {
2274 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2275 return CI->getValue() == APInt::getSignBit(NumBits);
2278 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2279 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2281 if (Op0 == Op1) // sub X, X -> 0
2282 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2284 // If this is a 'B = x-(-A)', change to B = x+A...
2285 if (Value *V = dyn_castNegVal(Op1))
2286 return BinaryOperator::createAdd(Op0, V);
2288 if (isa<UndefValue>(Op0))
2289 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2290 if (isa<UndefValue>(Op1))
2291 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2293 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2294 // Replace (-1 - A) with (~A)...
2295 if (C->isAllOnesValue())
2296 return BinaryOperator::createNot(Op1);
2298 // C - ~X == X + (1+C)
2300 if (match(Op1, m_Not(m_Value(X))))
2301 return BinaryOperator::createAdd(X, AddOne(C));
2303 // -(X >>u 31) -> (X >>s 31)
2304 // -(X >>s 31) -> (X >>u 31)
2306 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2307 if (SI->getOpcode() == Instruction::LShr) {
2308 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2309 // Check to see if we are shifting out everything but the sign bit.
2310 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2311 SI->getType()->getPrimitiveSizeInBits()-1) {
2312 // Ok, the transformation is safe. Insert AShr.
2313 return BinaryOperator::create(Instruction::AShr,
2314 SI->getOperand(0), CU, SI->getName());
2318 else if (SI->getOpcode() == Instruction::AShr) {
2319 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2320 // Check to see if we are shifting out everything but the sign bit.
2321 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2322 SI->getType()->getPrimitiveSizeInBits()-1) {
2323 // Ok, the transformation is safe. Insert LShr.
2324 return BinaryOperator::createLShr(
2325 SI->getOperand(0), CU, SI->getName());
2331 // Try to fold constant sub into select arguments.
2332 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2333 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2336 if (isa<PHINode>(Op0))
2337 if (Instruction *NV = FoldOpIntoPhi(I))
2341 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2342 if (Op1I->getOpcode() == Instruction::Add &&
2343 !Op0->getType()->isFPOrFPVector()) {
2344 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2345 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2346 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2347 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2348 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2349 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2350 // C1-(X+C2) --> (C1-C2)-X
2351 return BinaryOperator::createSub(Subtract(CI1, CI2),
2352 Op1I->getOperand(0));
2356 if (Op1I->hasOneUse()) {
2357 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2358 // is not used by anyone else...
2360 if (Op1I->getOpcode() == Instruction::Sub &&
2361 !Op1I->getType()->isFPOrFPVector()) {
2362 // Swap the two operands of the subexpr...
2363 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2364 Op1I->setOperand(0, IIOp1);
2365 Op1I->setOperand(1, IIOp0);
2367 // Create the new top level add instruction...
2368 return BinaryOperator::createAdd(Op0, Op1);
2371 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2373 if (Op1I->getOpcode() == Instruction::And &&
2374 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2375 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2378 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2379 return BinaryOperator::createAnd(Op0, NewNot);
2382 // 0 - (X sdiv C) -> (X sdiv -C)
2383 if (Op1I->getOpcode() == Instruction::SDiv)
2384 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2386 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2387 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2388 ConstantExpr::getNeg(DivRHS));
2390 // X - X*C --> X * (1-C)
2391 ConstantInt *C2 = 0;
2392 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2393 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2394 return BinaryOperator::createMul(Op0, CP1);
2397 // X - ((X / Y) * Y) --> X % Y
2398 if (Op1I->getOpcode() == Instruction::Mul)
2399 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2400 if (Op0 == I->getOperand(0) &&
2401 Op1I->getOperand(1) == I->getOperand(1)) {
2402 if (I->getOpcode() == Instruction::SDiv)
2403 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2404 if (I->getOpcode() == Instruction::UDiv)
2405 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2410 if (!Op0->getType()->isFPOrFPVector())
2411 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2412 if (Op0I->getOpcode() == Instruction::Add) {
2413 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2414 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2415 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2416 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2417 } else if (Op0I->getOpcode() == Instruction::Sub) {
2418 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2419 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2423 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2424 if (X == Op1) // X*C - X --> X * (C-1)
2425 return BinaryOperator::createMul(Op1, SubOne(C1));
2427 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2428 if (X == dyn_castFoldableMul(Op1, C2))
2429 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2434 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2435 /// comparison only checks the sign bit. If it only checks the sign bit, set
2436 /// TrueIfSigned if the result of the comparison is true when the input value is
2438 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2439 bool &TrueIfSigned) {
2441 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2442 TrueIfSigned = true;
2443 return RHS->isZero();
2444 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2445 TrueIfSigned = true;
2446 return RHS->isAllOnesValue();
2447 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2448 TrueIfSigned = false;
2449 return RHS->isAllOnesValue();
2450 case ICmpInst::ICMP_UGT:
2451 // True if LHS u> RHS and RHS == high-bit-mask - 1
2452 TrueIfSigned = true;
2453 return RHS->getValue() ==
2454 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2455 case ICmpInst::ICMP_UGE:
2456 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2457 TrueIfSigned = true;
2458 return RHS->getValue() ==
2459 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2465 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2466 bool Changed = SimplifyCommutative(I);
2467 Value *Op0 = I.getOperand(0);
2469 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2470 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2472 // Simplify mul instructions with a constant RHS...
2473 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2474 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2476 // ((X << C1)*C2) == (X * (C2 << C1))
2477 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2478 if (SI->getOpcode() == Instruction::Shl)
2479 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2480 return BinaryOperator::createMul(SI->getOperand(0),
2481 ConstantExpr::getShl(CI, ShOp));
2484 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2485 if (CI->equalsInt(1)) // X * 1 == X
2486 return ReplaceInstUsesWith(I, Op0);
2487 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2488 return BinaryOperator::createNeg(Op0, I.getName());
2490 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2491 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2492 return BinaryOperator::createShl(Op0,
2493 ConstantInt::get(Op0->getType(), Val.logBase2()));
2495 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2496 if (Op1F->isNullValue())
2497 return ReplaceInstUsesWith(I, Op1);
2499 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2500 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2501 // We need a better interface for long double here.
2502 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2503 if (Op1F->isExactlyValue(1.0))
2504 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2507 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2508 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2509 isa<ConstantInt>(Op0I->getOperand(1))) {
2510 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2511 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2513 InsertNewInstBefore(Add, I);
2514 Value *C1C2 = ConstantExpr::getMul(Op1,
2515 cast<Constant>(Op0I->getOperand(1)));
2516 return BinaryOperator::createAdd(Add, C1C2);
2520 // Try to fold constant mul into select arguments.
2521 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2522 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2525 if (isa<PHINode>(Op0))
2526 if (Instruction *NV = FoldOpIntoPhi(I))
2530 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2531 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2532 return BinaryOperator::createMul(Op0v, Op1v);
2534 // If one of the operands of the multiply is a cast from a boolean value, then
2535 // we know the bool is either zero or one, so this is a 'masking' multiply.
2536 // See if we can simplify things based on how the boolean was originally
2538 CastInst *BoolCast = 0;
2539 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2540 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2543 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2544 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2547 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2548 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2549 const Type *SCOpTy = SCIOp0->getType();
2552 // If the icmp is true iff the sign bit of X is set, then convert this
2553 // multiply into a shift/and combination.
2554 if (isa<ConstantInt>(SCIOp1) &&
2555 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2557 // Shift the X value right to turn it into "all signbits".
2558 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2559 SCOpTy->getPrimitiveSizeInBits()-1);
2561 InsertNewInstBefore(
2562 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2563 BoolCast->getOperand(0)->getName()+
2566 // If the multiply type is not the same as the source type, sign extend
2567 // or truncate to the multiply type.
2568 if (I.getType() != V->getType()) {
2569 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2570 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2571 Instruction::CastOps opcode =
2572 (SrcBits == DstBits ? Instruction::BitCast :
2573 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2574 V = InsertCastBefore(opcode, V, I.getType(), I);
2577 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2578 return BinaryOperator::createAnd(V, OtherOp);
2583 return Changed ? &I : 0;
2586 /// This function implements the transforms on div instructions that work
2587 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2588 /// used by the visitors to those instructions.
2589 /// @brief Transforms common to all three div instructions
2590 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2591 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2593 // undef / X -> 0 for integer.
2594 // undef / X -> undef for FP (the undef could be a snan).
2595 if (isa<UndefValue>(Op0)) {
2596 if (Op0->getType()->isFPOrFPVector())
2597 return ReplaceInstUsesWith(I, Op0);
2598 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2601 // X / undef -> undef
2602 if (isa<UndefValue>(Op1))
2603 return ReplaceInstUsesWith(I, Op1);
2605 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2606 // This does not apply for fdiv.
2607 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2608 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2609 // the same basic block, then we replace the select with Y, and the
2610 // condition of the select with false (if the cond value is in the same BB).
2611 // If the select has uses other than the div, this allows them to be
2612 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2613 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2614 if (ST->isNullValue()) {
2615 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2616 if (CondI && CondI->getParent() == I.getParent())
2617 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2618 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2619 I.setOperand(1, SI->getOperand(2));
2621 UpdateValueUsesWith(SI, SI->getOperand(2));
2625 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2626 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2627 if (ST->isNullValue()) {
2628 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2629 if (CondI && CondI->getParent() == I.getParent())
2630 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2631 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2632 I.setOperand(1, SI->getOperand(1));
2634 UpdateValueUsesWith(SI, SI->getOperand(1));
2642 /// This function implements the transforms common to both integer division
2643 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2644 /// division instructions.
2645 /// @brief Common integer divide transforms
2646 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2647 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2649 if (Instruction *Common = commonDivTransforms(I))
2652 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2654 if (RHS->equalsInt(1))
2655 return ReplaceInstUsesWith(I, Op0);
2657 // (X / C1) / C2 -> X / (C1*C2)
2658 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2659 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2660 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2661 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2662 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2664 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2665 Multiply(RHS, LHSRHS));
2668 if (!RHS->isZero()) { // avoid X udiv 0
2669 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2670 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2672 if (isa<PHINode>(Op0))
2673 if (Instruction *NV = FoldOpIntoPhi(I))
2678 // 0 / X == 0, we don't need to preserve faults!
2679 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2680 if (LHS->equalsInt(0))
2681 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2686 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2687 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2689 // Handle the integer div common cases
2690 if (Instruction *Common = commonIDivTransforms(I))
2693 // X udiv C^2 -> X >> C
2694 // Check to see if this is an unsigned division with an exact power of 2,
2695 // if so, convert to a right shift.
2696 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2697 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2698 return BinaryOperator::createLShr(Op0,
2699 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2702 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2703 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2704 if (RHSI->getOpcode() == Instruction::Shl &&
2705 isa<ConstantInt>(RHSI->getOperand(0))) {
2706 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2707 if (C1.isPowerOf2()) {
2708 Value *N = RHSI->getOperand(1);
2709 const Type *NTy = N->getType();
2710 if (uint32_t C2 = C1.logBase2()) {
2711 Constant *C2V = ConstantInt::get(NTy, C2);
2712 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2714 return BinaryOperator::createLShr(Op0, N);
2719 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2720 // where C1&C2 are powers of two.
2721 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2722 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2723 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2724 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2725 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2726 // Compute the shift amounts
2727 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2728 // Construct the "on true" case of the select
2729 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2730 Instruction *TSI = BinaryOperator::createLShr(
2731 Op0, TC, SI->getName()+".t");
2732 TSI = InsertNewInstBefore(TSI, I);
2734 // Construct the "on false" case of the select
2735 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2736 Instruction *FSI = BinaryOperator::createLShr(
2737 Op0, FC, SI->getName()+".f");
2738 FSI = InsertNewInstBefore(FSI, I);
2740 // construct the select instruction and return it.
2741 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2747 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2748 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2750 // Handle the integer div common cases
2751 if (Instruction *Common = commonIDivTransforms(I))
2754 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2756 if (RHS->isAllOnesValue())
2757 return BinaryOperator::createNeg(Op0);
2760 if (Value *LHSNeg = dyn_castNegVal(Op0))
2761 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2764 // If the sign bits of both operands are zero (i.e. we can prove they are
2765 // unsigned inputs), turn this into a udiv.
2766 if (I.getType()->isInteger()) {
2767 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2768 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2769 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2770 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2777 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2778 return commonDivTransforms(I);
2781 /// GetFactor - If we can prove that the specified value is at least a multiple
2782 /// of some factor, return that factor.
2783 static Constant *GetFactor(Value *V) {
2784 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2787 // Unless we can be tricky, we know this is a multiple of 1.
2788 Constant *Result = ConstantInt::get(V->getType(), 1);
2790 Instruction *I = dyn_cast<Instruction>(V);
2791 if (!I) return Result;
2793 if (I->getOpcode() == Instruction::Mul) {
2794 // Handle multiplies by a constant, etc.
2795 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2796 GetFactor(I->getOperand(1)));
2797 } else if (I->getOpcode() == Instruction::Shl) {
2798 // (X<<C) -> X * (1 << C)
2799 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2800 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2801 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2803 } else if (I->getOpcode() == Instruction::And) {
2804 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2805 // X & 0xFFF0 is known to be a multiple of 16.
2806 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2807 if (Zeros != V->getType()->getPrimitiveSizeInBits())// don't shift by "32"
2808 return ConstantExpr::getShl(Result,
2809 ConstantInt::get(Result->getType(), Zeros));
2811 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2812 // Only handle int->int casts.
2813 if (!CI->isIntegerCast())
2815 Value *Op = CI->getOperand(0);
2816 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2821 /// This function implements the transforms on rem instructions that work
2822 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2823 /// is used by the visitors to those instructions.
2824 /// @brief Transforms common to all three rem instructions
2825 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2826 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2828 // 0 % X == 0 for integer, we don't need to preserve faults!
2829 if (Constant *LHS = dyn_cast<Constant>(Op0))
2830 if (LHS->isNullValue())
2831 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2833 if (isa<UndefValue>(Op0)) { // undef % X -> 0
2834 if (I.getType()->isFPOrFPVector())
2835 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
2836 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2838 if (isa<UndefValue>(Op1))
2839 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2841 // Handle cases involving: rem X, (select Cond, Y, Z)
2842 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2843 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2844 // the same basic block, then we replace the select with Y, and the
2845 // condition of the select with false (if the cond value is in the same
2846 // BB). If the select has uses other than the div, this allows them to be
2848 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2849 if (ST->isNullValue()) {
2850 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2851 if (CondI && CondI->getParent() == I.getParent())
2852 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2853 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2854 I.setOperand(1, SI->getOperand(2));
2856 UpdateValueUsesWith(SI, SI->getOperand(2));
2859 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2860 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2861 if (ST->isNullValue()) {
2862 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2863 if (CondI && CondI->getParent() == I.getParent())
2864 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2865 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2866 I.setOperand(1, SI->getOperand(1));
2868 UpdateValueUsesWith(SI, SI->getOperand(1));
2876 /// This function implements the transforms common to both integer remainder
2877 /// instructions (urem and srem). It is called by the visitors to those integer
2878 /// remainder instructions.
2879 /// @brief Common integer remainder transforms
2880 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2881 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2883 if (Instruction *common = commonRemTransforms(I))
2886 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2887 // X % 0 == undef, we don't need to preserve faults!
2888 if (RHS->equalsInt(0))
2889 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2891 if (RHS->equalsInt(1)) // X % 1 == 0
2892 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2894 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2895 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2896 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2898 } else if (isa<PHINode>(Op0I)) {
2899 if (Instruction *NV = FoldOpIntoPhi(I))
2902 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2903 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2904 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2911 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2912 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2914 if (Instruction *common = commonIRemTransforms(I))
2917 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2918 // X urem C^2 -> X and C
2919 // Check to see if this is an unsigned remainder with an exact power of 2,
2920 // if so, convert to a bitwise and.
2921 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2922 if (C->getValue().isPowerOf2())
2923 return BinaryOperator::createAnd(Op0, SubOne(C));
2926 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2927 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2928 if (RHSI->getOpcode() == Instruction::Shl &&
2929 isa<ConstantInt>(RHSI->getOperand(0))) {
2930 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2931 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2932 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2934 return BinaryOperator::createAnd(Op0, Add);
2939 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2940 // where C1&C2 are powers of two.
2941 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2942 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2943 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2944 // STO == 0 and SFO == 0 handled above.
2945 if ((STO->getValue().isPowerOf2()) &&
2946 (SFO->getValue().isPowerOf2())) {
2947 Value *TrueAnd = InsertNewInstBefore(
2948 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2949 Value *FalseAnd = InsertNewInstBefore(
2950 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2951 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2959 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2960 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2962 // Handle the integer rem common cases
2963 if (Instruction *common = commonIRemTransforms(I))
2966 if (Value *RHSNeg = dyn_castNegVal(Op1))
2967 if (!isa<ConstantInt>(RHSNeg) ||
2968 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2970 AddUsesToWorkList(I);
2971 I.setOperand(1, RHSNeg);
2975 // If the sign bits of both operands are zero (i.e. we can prove they are
2976 // unsigned inputs), turn this into a urem.
2977 if (I.getType()->isInteger()) {
2978 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2979 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2980 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2981 return BinaryOperator::createURem(Op0, Op1, I.getName());
2988 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2989 return commonRemTransforms(I);
2992 // isMaxValueMinusOne - return true if this is Max-1
2993 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2994 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2996 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2997 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
3000 // isMinValuePlusOne - return true if this is Min+1
3001 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
3003 return C->getValue() == 1; // unsigned
3005 // Calculate 1111111111000000000000
3006 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3007 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
3010 // isOneBitSet - Return true if there is exactly one bit set in the specified
3012 static bool isOneBitSet(const ConstantInt *CI) {
3013 return CI->getValue().isPowerOf2();
3016 // isHighOnes - Return true if the constant is of the form 1+0+.
3017 // This is the same as lowones(~X).
3018 static bool isHighOnes(const ConstantInt *CI) {
3019 return (~CI->getValue() + 1).isPowerOf2();
3022 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3023 /// are carefully arranged to allow folding of expressions such as:
3025 /// (A < B) | (A > B) --> (A != B)
3027 /// Note that this is only valid if the first and second predicates have the
3028 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3030 /// Three bits are used to represent the condition, as follows:
3035 /// <=> Value Definition
3036 /// 000 0 Always false
3043 /// 111 7 Always true
3045 static unsigned getICmpCode(const ICmpInst *ICI) {
3046 switch (ICI->getPredicate()) {
3048 case ICmpInst::ICMP_UGT: return 1; // 001
3049 case ICmpInst::ICMP_SGT: return 1; // 001
3050 case ICmpInst::ICMP_EQ: return 2; // 010
3051 case ICmpInst::ICMP_UGE: return 3; // 011
3052 case ICmpInst::ICMP_SGE: return 3; // 011
3053 case ICmpInst::ICMP_ULT: return 4; // 100
3054 case ICmpInst::ICMP_SLT: return 4; // 100
3055 case ICmpInst::ICMP_NE: return 5; // 101
3056 case ICmpInst::ICMP_ULE: return 6; // 110
3057 case ICmpInst::ICMP_SLE: return 6; // 110
3060 assert(0 && "Invalid ICmp predicate!");
3065 /// getICmpValue - This is the complement of getICmpCode, which turns an
3066 /// opcode and two operands into either a constant true or false, or a brand
3067 /// new ICmp instruction. The sign is passed in to determine which kind
3068 /// of predicate to use in new icmp instructions.
3069 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3071 default: assert(0 && "Illegal ICmp code!");
3072 case 0: return ConstantInt::getFalse();
3075 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3077 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3078 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3081 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3083 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3086 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3088 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3089 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3092 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3094 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3095 case 7: return ConstantInt::getTrue();
3099 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3100 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3101 (ICmpInst::isSignedPredicate(p1) &&
3102 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3103 (ICmpInst::isSignedPredicate(p2) &&
3104 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3108 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3109 struct FoldICmpLogical {
3112 ICmpInst::Predicate pred;
3113 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3114 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3115 pred(ICI->getPredicate()) {}
3116 bool shouldApply(Value *V) const {
3117 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3118 if (PredicatesFoldable(pred, ICI->getPredicate()))
3119 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
3120 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
3123 Instruction *apply(Instruction &Log) const {
3124 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3125 if (ICI->getOperand(0) != LHS) {
3126 assert(ICI->getOperand(1) == LHS);
3127 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3130 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3131 unsigned LHSCode = getICmpCode(ICI);
3132 unsigned RHSCode = getICmpCode(RHSICI);
3134 switch (Log.getOpcode()) {
3135 case Instruction::And: Code = LHSCode & RHSCode; break;
3136 case Instruction::Or: Code = LHSCode | RHSCode; break;
3137 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3138 default: assert(0 && "Illegal logical opcode!"); return 0;
3141 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3142 ICmpInst::isSignedPredicate(ICI->getPredicate());
3144 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3145 if (Instruction *I = dyn_cast<Instruction>(RV))
3147 // Otherwise, it's a constant boolean value...
3148 return IC.ReplaceInstUsesWith(Log, RV);
3151 } // end anonymous namespace
3153 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3154 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3155 // guaranteed to be a binary operator.
3156 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3158 ConstantInt *AndRHS,
3159 BinaryOperator &TheAnd) {
3160 Value *X = Op->getOperand(0);
3161 Constant *Together = 0;
3163 Together = And(AndRHS, OpRHS);
3165 switch (Op->getOpcode()) {
3166 case Instruction::Xor:
3167 if (Op->hasOneUse()) {
3168 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3169 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3170 InsertNewInstBefore(And, TheAnd);
3172 return BinaryOperator::createXor(And, Together);
3175 case Instruction::Or:
3176 if (Together == AndRHS) // (X | C) & C --> C
3177 return ReplaceInstUsesWith(TheAnd, AndRHS);
3179 if (Op->hasOneUse() && Together != OpRHS) {
3180 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3181 Instruction *Or = BinaryOperator::createOr(X, Together);
3182 InsertNewInstBefore(Or, TheAnd);
3184 return BinaryOperator::createAnd(Or, AndRHS);
3187 case Instruction::Add:
3188 if (Op->hasOneUse()) {
3189 // Adding a one to a single bit bit-field should be turned into an XOR
3190 // of the bit. First thing to check is to see if this AND is with a
3191 // single bit constant.
3192 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3194 // If there is only one bit set...
3195 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3196 // Ok, at this point, we know that we are masking the result of the
3197 // ADD down to exactly one bit. If the constant we are adding has
3198 // no bits set below this bit, then we can eliminate the ADD.
3199 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3201 // Check to see if any bits below the one bit set in AndRHSV are set.
3202 if ((AddRHS & (AndRHSV-1)) == 0) {
3203 // If not, the only thing that can effect the output of the AND is
3204 // the bit specified by AndRHSV. If that bit is set, the effect of
3205 // the XOR is to toggle the bit. If it is clear, then the ADD has
3207 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3208 TheAnd.setOperand(0, X);
3211 // Pull the XOR out of the AND.
3212 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3213 InsertNewInstBefore(NewAnd, TheAnd);
3214 NewAnd->takeName(Op);
3215 return BinaryOperator::createXor(NewAnd, AndRHS);
3222 case Instruction::Shl: {
3223 // We know that the AND will not produce any of the bits shifted in, so if
3224 // the anded constant includes them, clear them now!
3226 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3227 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3228 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3229 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3231 if (CI->getValue() == ShlMask) {
3232 // Masking out bits that the shift already masks
3233 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3234 } else if (CI != AndRHS) { // Reducing bits set in and.
3235 TheAnd.setOperand(1, CI);
3240 case Instruction::LShr:
3242 // We know that the AND will not produce any of the bits shifted in, so if
3243 // the anded constant includes them, clear them now! This only applies to
3244 // unsigned shifts, because a signed shr may bring in set bits!
3246 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3247 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3248 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3249 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3251 if (CI->getValue() == ShrMask) {
3252 // Masking out bits that the shift already masks.
3253 return ReplaceInstUsesWith(TheAnd, Op);
3254 } else if (CI != AndRHS) {
3255 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3260 case Instruction::AShr:
3262 // See if this is shifting in some sign extension, then masking it out
3264 if (Op->hasOneUse()) {
3265 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3266 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3267 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3268 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3269 if (C == AndRHS) { // Masking out bits shifted in.
3270 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3271 // Make the argument unsigned.
3272 Value *ShVal = Op->getOperand(0);
3273 ShVal = InsertNewInstBefore(
3274 BinaryOperator::createLShr(ShVal, OpRHS,
3275 Op->getName()), TheAnd);
3276 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3285 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3286 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3287 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3288 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3289 /// insert new instructions.
3290 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3291 bool isSigned, bool Inside,
3293 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3294 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3295 "Lo is not <= Hi in range emission code!");
3298 if (Lo == Hi) // Trivially false.
3299 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3301 // V >= Min && V < Hi --> V < Hi
3302 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3303 ICmpInst::Predicate pred = (isSigned ?
3304 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3305 return new ICmpInst(pred, V, Hi);
3308 // Emit V-Lo <u Hi-Lo
3309 Constant *NegLo = ConstantExpr::getNeg(Lo);
3310 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3311 InsertNewInstBefore(Add, IB);
3312 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3313 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3316 if (Lo == Hi) // Trivially true.
3317 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3319 // V < Min || V >= Hi -> V > Hi-1
3320 Hi = SubOne(cast<ConstantInt>(Hi));
3321 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3322 ICmpInst::Predicate pred = (isSigned ?
3323 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3324 return new ICmpInst(pred, V, Hi);
3327 // Emit V-Lo >u Hi-1-Lo
3328 // Note that Hi has already had one subtracted from it, above.
3329 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3330 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3331 InsertNewInstBefore(Add, IB);
3332 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3333 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3336 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3337 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3338 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3339 // not, since all 1s are not contiguous.
3340 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3341 const APInt& V = Val->getValue();
3342 uint32_t BitWidth = Val->getType()->getBitWidth();
3343 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3345 // look for the first zero bit after the run of ones
3346 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3347 // look for the first non-zero bit
3348 ME = V.getActiveBits();
3352 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3353 /// where isSub determines whether the operator is a sub. If we can fold one of
3354 /// the following xforms:
3356 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3357 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3358 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3360 /// return (A +/- B).
3362 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3363 ConstantInt *Mask, bool isSub,
3365 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3366 if (!LHSI || LHSI->getNumOperands() != 2 ||
3367 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3369 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3371 switch (LHSI->getOpcode()) {
3373 case Instruction::And:
3374 if (And(N, Mask) == Mask) {
3375 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3376 if ((Mask->getValue().countLeadingZeros() +
3377 Mask->getValue().countPopulation()) ==
3378 Mask->getValue().getBitWidth())
3381 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3382 // part, we don't need any explicit masks to take them out of A. If that
3383 // is all N is, ignore it.
3384 uint32_t MB = 0, ME = 0;
3385 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3386 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3387 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3388 if (MaskedValueIsZero(RHS, Mask))
3393 case Instruction::Or:
3394 case Instruction::Xor:
3395 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3396 if ((Mask->getValue().countLeadingZeros() +
3397 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3398 && And(N, Mask)->isZero())
3405 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3407 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3408 return InsertNewInstBefore(New, I);
3411 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3412 bool Changed = SimplifyCommutative(I);
3413 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3415 if (isa<UndefValue>(Op1)) // X & undef -> 0
3416 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3420 return ReplaceInstUsesWith(I, Op1);
3422 // See if we can simplify any instructions used by the instruction whose sole
3423 // purpose is to compute bits we don't care about.
3424 if (!isa<VectorType>(I.getType())) {
3425 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3426 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3427 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3428 KnownZero, KnownOne))
3431 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3432 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3433 return ReplaceInstUsesWith(I, I.getOperand(0));
3434 } else if (isa<ConstantAggregateZero>(Op1)) {
3435 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3439 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3440 const APInt& AndRHSMask = AndRHS->getValue();
3441 APInt NotAndRHS(~AndRHSMask);
3443 // Optimize a variety of ((val OP C1) & C2) combinations...
3444 if (isa<BinaryOperator>(Op0)) {
3445 Instruction *Op0I = cast<Instruction>(Op0);
3446 Value *Op0LHS = Op0I->getOperand(0);
3447 Value *Op0RHS = Op0I->getOperand(1);
3448 switch (Op0I->getOpcode()) {
3449 case Instruction::Xor:
3450 case Instruction::Or:
3451 // If the mask is only needed on one incoming arm, push it up.
3452 if (Op0I->hasOneUse()) {
3453 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3454 // Not masking anything out for the LHS, move to RHS.
3455 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3456 Op0RHS->getName()+".masked");
3457 InsertNewInstBefore(NewRHS, I);
3458 return BinaryOperator::create(
3459 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3461 if (!isa<Constant>(Op0RHS) &&
3462 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3463 // Not masking anything out for the RHS, move to LHS.
3464 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3465 Op0LHS->getName()+".masked");
3466 InsertNewInstBefore(NewLHS, I);
3467 return BinaryOperator::create(
3468 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3473 case Instruction::Add:
3474 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3475 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3476 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3477 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3478 return BinaryOperator::createAnd(V, AndRHS);
3479 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3480 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3483 case Instruction::Sub:
3484 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3485 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3486 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3487 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3488 return BinaryOperator::createAnd(V, AndRHS);
3492 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3493 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3495 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3496 // If this is an integer truncation or change from signed-to-unsigned, and
3497 // if the source is an and/or with immediate, transform it. This
3498 // frequently occurs for bitfield accesses.
3499 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3500 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3501 CastOp->getNumOperands() == 2)
3502 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3503 if (CastOp->getOpcode() == Instruction::And) {
3504 // Change: and (cast (and X, C1) to T), C2
3505 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3506 // This will fold the two constants together, which may allow
3507 // other simplifications.
3508 Instruction *NewCast = CastInst::createTruncOrBitCast(
3509 CastOp->getOperand(0), I.getType(),
3510 CastOp->getName()+".shrunk");
3511 NewCast = InsertNewInstBefore(NewCast, I);
3512 // trunc_or_bitcast(C1)&C2
3513 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3514 C3 = ConstantExpr::getAnd(C3, AndRHS);
3515 return BinaryOperator::createAnd(NewCast, C3);
3516 } else if (CastOp->getOpcode() == Instruction::Or) {
3517 // Change: and (cast (or X, C1) to T), C2
3518 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3519 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3520 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3521 return ReplaceInstUsesWith(I, AndRHS);
3526 // Try to fold constant and into select arguments.
3527 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3528 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3530 if (isa<PHINode>(Op0))
3531 if (Instruction *NV = FoldOpIntoPhi(I))
3535 Value *Op0NotVal = dyn_castNotVal(Op0);
3536 Value *Op1NotVal = dyn_castNotVal(Op1);
3538 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3539 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3541 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3542 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3543 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3544 I.getName()+".demorgan");
3545 InsertNewInstBefore(Or, I);
3546 return BinaryOperator::createNot(Or);
3550 Value *A = 0, *B = 0, *C = 0, *D = 0;
3551 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3552 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3553 return ReplaceInstUsesWith(I, Op1);
3555 // (A|B) & ~(A&B) -> A^B
3556 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3557 if ((A == C && B == D) || (A == D && B == C))
3558 return BinaryOperator::createXor(A, B);
3562 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3563 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3564 return ReplaceInstUsesWith(I, Op0);
3566 // ~(A&B) & (A|B) -> A^B
3567 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3568 if ((A == C && B == D) || (A == D && B == C))
3569 return BinaryOperator::createXor(A, B);
3573 if (Op0->hasOneUse() &&
3574 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3575 if (A == Op1) { // (A^B)&A -> A&(A^B)
3576 I.swapOperands(); // Simplify below
3577 std::swap(Op0, Op1);
3578 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3579 cast<BinaryOperator>(Op0)->swapOperands();
3580 I.swapOperands(); // Simplify below
3581 std::swap(Op0, Op1);
3584 if (Op1->hasOneUse() &&
3585 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3586 if (B == Op0) { // B&(A^B) -> B&(B^A)
3587 cast<BinaryOperator>(Op1)->swapOperands();
3590 if (A == Op0) { // A&(A^B) -> A & ~B
3591 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3592 InsertNewInstBefore(NotB, I);
3593 return BinaryOperator::createAnd(A, NotB);
3598 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3599 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3600 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3603 Value *LHSVal, *RHSVal;
3604 ConstantInt *LHSCst, *RHSCst;
3605 ICmpInst::Predicate LHSCC, RHSCC;
3606 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3607 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3608 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3609 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3610 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3611 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3612 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3613 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3615 // Don't try to fold ICMP_SLT + ICMP_ULT.
3616 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3617 ICmpInst::isSignedPredicate(LHSCC) ==
3618 ICmpInst::isSignedPredicate(RHSCC))) {
3619 // Ensure that the larger constant is on the RHS.
3620 ICmpInst::Predicate GT;
3621 if (ICmpInst::isSignedPredicate(LHSCC) ||
3622 (ICmpInst::isEquality(LHSCC) &&
3623 ICmpInst::isSignedPredicate(RHSCC)))
3624 GT = ICmpInst::ICMP_SGT;
3626 GT = ICmpInst::ICMP_UGT;
3628 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3629 ICmpInst *LHS = cast<ICmpInst>(Op0);
3630 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3631 std::swap(LHS, RHS);
3632 std::swap(LHSCst, RHSCst);
3633 std::swap(LHSCC, RHSCC);
3636 // At this point, we know we have have two icmp instructions
3637 // comparing a value against two constants and and'ing the result
3638 // together. Because of the above check, we know that we only have
3639 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3640 // (from the FoldICmpLogical check above), that the two constants
3641 // are not equal and that the larger constant is on the RHS
3642 assert(LHSCst != RHSCst && "Compares not folded above?");
3645 default: assert(0 && "Unknown integer condition code!");
3646 case ICmpInst::ICMP_EQ:
3648 default: assert(0 && "Unknown integer condition code!");
3649 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3650 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3651 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3652 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3653 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3654 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3655 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3656 return ReplaceInstUsesWith(I, LHS);
3658 case ICmpInst::ICMP_NE:
3660 default: assert(0 && "Unknown integer condition code!");
3661 case ICmpInst::ICMP_ULT:
3662 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3663 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3664 break; // (X != 13 & X u< 15) -> no change
3665 case ICmpInst::ICMP_SLT:
3666 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3667 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3668 break; // (X != 13 & X s< 15) -> no change
3669 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3670 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3671 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3672 return ReplaceInstUsesWith(I, RHS);
3673 case ICmpInst::ICMP_NE:
3674 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3675 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3676 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3677 LHSVal->getName()+".off");
3678 InsertNewInstBefore(Add, I);
3679 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3680 ConstantInt::get(Add->getType(), 1));
3682 break; // (X != 13 & X != 15) -> no change
3685 case ICmpInst::ICMP_ULT:
3687 default: assert(0 && "Unknown integer condition code!");
3688 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3689 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3690 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3691 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3693 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3694 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3695 return ReplaceInstUsesWith(I, LHS);
3696 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3700 case ICmpInst::ICMP_SLT:
3702 default: assert(0 && "Unknown integer condition code!");
3703 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3704 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3705 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3706 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3708 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3709 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3710 return ReplaceInstUsesWith(I, LHS);
3711 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3715 case ICmpInst::ICMP_UGT:
3717 default: assert(0 && "Unknown integer condition code!");
3718 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3719 return ReplaceInstUsesWith(I, LHS);
3720 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3721 return ReplaceInstUsesWith(I, RHS);
3722 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3724 case ICmpInst::ICMP_NE:
3725 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3726 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3727 break; // (X u> 13 & X != 15) -> no change
3728 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3729 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3731 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3735 case ICmpInst::ICMP_SGT:
3737 default: assert(0 && "Unknown integer condition code!");
3738 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3739 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3740 return ReplaceInstUsesWith(I, RHS);
3741 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3743 case ICmpInst::ICMP_NE:
3744 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3745 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3746 break; // (X s> 13 & X != 15) -> no change
3747 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3748 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3750 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3758 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3759 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3760 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3761 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3762 const Type *SrcTy = Op0C->getOperand(0)->getType();
3763 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3764 // Only do this if the casts both really cause code to be generated.
3765 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3767 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3769 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3770 Op1C->getOperand(0),
3772 InsertNewInstBefore(NewOp, I);
3773 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3777 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3778 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3779 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3780 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3781 SI0->getOperand(1) == SI1->getOperand(1) &&
3782 (SI0->hasOneUse() || SI1->hasOneUse())) {
3783 Instruction *NewOp =
3784 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3786 SI0->getName()), I);
3787 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3788 SI1->getOperand(1));
3792 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3793 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3794 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3795 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3796 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3797 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3798 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3799 // If either of the constants are nans, then the whole thing returns
3801 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3802 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3803 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3804 RHS->getOperand(0));
3809 return Changed ? &I : 0;
3812 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3813 /// in the result. If it does, and if the specified byte hasn't been filled in
3814 /// yet, fill it in and return false.
3815 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3816 Instruction *I = dyn_cast<Instruction>(V);
3817 if (I == 0) return true;
3819 // If this is an or instruction, it is an inner node of the bswap.
3820 if (I->getOpcode() == Instruction::Or)
3821 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3822 CollectBSwapParts(I->getOperand(1), ByteValues);
3824 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3825 // If this is a shift by a constant int, and it is "24", then its operand
3826 // defines a byte. We only handle unsigned types here.
3827 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3828 // Not shifting the entire input by N-1 bytes?
3829 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3830 8*(ByteValues.size()-1))
3834 if (I->getOpcode() == Instruction::Shl) {
3835 // X << 24 defines the top byte with the lowest of the input bytes.
3836 DestNo = ByteValues.size()-1;
3838 // X >>u 24 defines the low byte with the highest of the input bytes.
3842 // If the destination byte value is already defined, the values are or'd
3843 // together, which isn't a bswap (unless it's an or of the same bits).
3844 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3846 ByteValues[DestNo] = I->getOperand(0);
3850 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3852 Value *Shift = 0, *ShiftLHS = 0;
3853 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3854 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3855 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3857 Instruction *SI = cast<Instruction>(Shift);
3859 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3860 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3861 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3864 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3866 if (AndAmt->getValue().getActiveBits() > 64)
3868 uint64_t AndAmtVal = AndAmt->getZExtValue();
3869 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3870 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3872 // Unknown mask for bswap.
3873 if (DestByte == ByteValues.size()) return true;
3875 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3877 if (SI->getOpcode() == Instruction::Shl)
3878 SrcByte = DestByte - ShiftBytes;
3880 SrcByte = DestByte + ShiftBytes;
3882 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3883 if (SrcByte != ByteValues.size()-DestByte-1)
3886 // If the destination byte value is already defined, the values are or'd
3887 // together, which isn't a bswap (unless it's an or of the same bits).
3888 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3890 ByteValues[DestByte] = SI->getOperand(0);
3894 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3895 /// If so, insert the new bswap intrinsic and return it.
3896 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3897 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3898 if (!ITy || ITy->getBitWidth() % 16)
3899 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3901 /// ByteValues - For each byte of the result, we keep track of which value
3902 /// defines each byte.
3903 SmallVector<Value*, 8> ByteValues;
3904 ByteValues.resize(ITy->getBitWidth()/8);
3906 // Try to find all the pieces corresponding to the bswap.
3907 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3908 CollectBSwapParts(I.getOperand(1), ByteValues))
3911 // Check to see if all of the bytes come from the same value.
3912 Value *V = ByteValues[0];
3913 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3915 // Check to make sure that all of the bytes come from the same value.
3916 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3917 if (ByteValues[i] != V)
3919 const Type *Tys[] = { ITy };
3920 Module *M = I.getParent()->getParent()->getParent();
3921 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3922 return new CallInst(F, V);
3926 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3927 bool Changed = SimplifyCommutative(I);
3928 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3930 if (isa<UndefValue>(Op1)) // X | undef -> -1
3931 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3935 return ReplaceInstUsesWith(I, Op0);
3937 // See if we can simplify any instructions used by the instruction whose sole
3938 // purpose is to compute bits we don't care about.
3939 if (!isa<VectorType>(I.getType())) {
3940 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3941 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3942 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3943 KnownZero, KnownOne))
3945 } else if (isa<ConstantAggregateZero>(Op1)) {
3946 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3947 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3948 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3949 return ReplaceInstUsesWith(I, I.getOperand(1));
3955 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3956 ConstantInt *C1 = 0; Value *X = 0;
3957 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3958 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3959 Instruction *Or = BinaryOperator::createOr(X, RHS);
3960 InsertNewInstBefore(Or, I);
3962 return BinaryOperator::createAnd(Or,
3963 ConstantInt::get(RHS->getValue() | C1->getValue()));
3966 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3967 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3968 Instruction *Or = BinaryOperator::createOr(X, RHS);
3969 InsertNewInstBefore(Or, I);
3971 return BinaryOperator::createXor(Or,
3972 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3975 // Try to fold constant and into select arguments.
3976 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3977 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3979 if (isa<PHINode>(Op0))
3980 if (Instruction *NV = FoldOpIntoPhi(I))
3984 Value *A = 0, *B = 0;
3985 ConstantInt *C1 = 0, *C2 = 0;
3987 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3988 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3989 return ReplaceInstUsesWith(I, Op1);
3990 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3991 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3992 return ReplaceInstUsesWith(I, Op0);
3994 // (A | B) | C and A | (B | C) -> bswap if possible.
3995 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3996 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3997 match(Op1, m_Or(m_Value(), m_Value())) ||
3998 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3999 match(Op1, m_Shift(m_Value(), m_Value())))) {
4000 if (Instruction *BSwap = MatchBSwap(I))
4004 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4005 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4006 MaskedValueIsZero(Op1, C1->getValue())) {
4007 Instruction *NOr = BinaryOperator::createOr(A, Op1);
4008 InsertNewInstBefore(NOr, I);
4010 return BinaryOperator::createXor(NOr, C1);
4013 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4014 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4015 MaskedValueIsZero(Op0, C1->getValue())) {
4016 Instruction *NOr = BinaryOperator::createOr(A, Op0);
4017 InsertNewInstBefore(NOr, I);
4019 return BinaryOperator::createXor(NOr, C1);
4023 Value *C = 0, *D = 0;
4024 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4025 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4026 Value *V1 = 0, *V2 = 0, *V3 = 0;
4027 C1 = dyn_cast<ConstantInt>(C);
4028 C2 = dyn_cast<ConstantInt>(D);
4029 if (C1 && C2) { // (A & C1)|(B & C2)
4030 // If we have: ((V + N) & C1) | (V & C2)
4031 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4032 // replace with V+N.
4033 if (C1->getValue() == ~C2->getValue()) {
4034 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4035 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4036 // Add commutes, try both ways.
4037 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4038 return ReplaceInstUsesWith(I, A);
4039 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4040 return ReplaceInstUsesWith(I, A);
4042 // Or commutes, try both ways.
4043 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4044 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4045 // Add commutes, try both ways.
4046 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4047 return ReplaceInstUsesWith(I, B);
4048 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4049 return ReplaceInstUsesWith(I, B);
4052 V1 = 0; V2 = 0; V3 = 0;
4055 // Check to see if we have any common things being and'ed. If so, find the
4056 // terms for V1 & (V2|V3).
4057 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4058 if (A == B) // (A & C)|(A & D) == A & (C|D)
4059 V1 = A, V2 = C, V3 = D;
4060 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4061 V1 = A, V2 = B, V3 = C;
4062 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4063 V1 = C, V2 = A, V3 = D;
4064 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4065 V1 = C, V2 = A, V3 = B;
4069 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
4070 return BinaryOperator::createAnd(V1, Or);
4075 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4076 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4077 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4078 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4079 SI0->getOperand(1) == SI1->getOperand(1) &&
4080 (SI0->hasOneUse() || SI1->hasOneUse())) {
4081 Instruction *NewOp =
4082 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4084 SI0->getName()), I);
4085 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4086 SI1->getOperand(1));
4090 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4091 if (A == Op1) // ~A | A == -1
4092 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4096 // Note, A is still live here!
4097 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4099 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4101 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4102 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4103 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4104 I.getName()+".demorgan"), I);
4105 return BinaryOperator::createNot(And);
4109 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4110 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4111 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4114 Value *LHSVal, *RHSVal;
4115 ConstantInt *LHSCst, *RHSCst;
4116 ICmpInst::Predicate LHSCC, RHSCC;
4117 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4118 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4119 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4120 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4121 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4122 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4123 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4124 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4125 // We can't fold (ugt x, C) | (sgt x, C2).
4126 PredicatesFoldable(LHSCC, RHSCC)) {
4127 // Ensure that the larger constant is on the RHS.
4128 ICmpInst *LHS = cast<ICmpInst>(Op0);
4130 if (ICmpInst::isSignedPredicate(LHSCC))
4131 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4133 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4136 std::swap(LHS, RHS);
4137 std::swap(LHSCst, RHSCst);
4138 std::swap(LHSCC, RHSCC);
4141 // At this point, we know we have have two icmp instructions
4142 // comparing a value against two constants and or'ing the result
4143 // together. Because of the above check, we know that we only have
4144 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4145 // FoldICmpLogical check above), that the two constants are not
4147 assert(LHSCst != RHSCst && "Compares not folded above?");
4150 default: assert(0 && "Unknown integer condition code!");
4151 case ICmpInst::ICMP_EQ:
4153 default: assert(0 && "Unknown integer condition code!");
4154 case ICmpInst::ICMP_EQ:
4155 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4156 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4157 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4158 LHSVal->getName()+".off");
4159 InsertNewInstBefore(Add, I);
4160 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4161 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4163 break; // (X == 13 | X == 15) -> no change
4164 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4165 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4167 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4168 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4169 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4170 return ReplaceInstUsesWith(I, RHS);
4173 case ICmpInst::ICMP_NE:
4175 default: assert(0 && "Unknown integer condition code!");
4176 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4177 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4178 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4179 return ReplaceInstUsesWith(I, LHS);
4180 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4181 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4182 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4183 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4186 case ICmpInst::ICMP_ULT:
4188 default: assert(0 && "Unknown integer condition code!");
4189 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4191 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4192 // If RHSCst is [us]MAXINT, it is always false. Not handling
4193 // this can cause overflow.
4194 if (RHSCst->isMaxValue(false))
4195 return ReplaceInstUsesWith(I, LHS);
4196 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4198 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4200 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4201 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4202 return ReplaceInstUsesWith(I, RHS);
4203 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4207 case ICmpInst::ICMP_SLT:
4209 default: assert(0 && "Unknown integer condition code!");
4210 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4212 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4213 // If RHSCst is [us]MAXINT, it is always false. Not handling
4214 // this can cause overflow.
4215 if (RHSCst->isMaxValue(true))
4216 return ReplaceInstUsesWith(I, LHS);
4217 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4219 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4221 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4222 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4223 return ReplaceInstUsesWith(I, RHS);
4224 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4228 case ICmpInst::ICMP_UGT:
4230 default: assert(0 && "Unknown integer condition code!");
4231 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4232 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4233 return ReplaceInstUsesWith(I, LHS);
4234 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4236 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4237 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4238 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4239 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4243 case ICmpInst::ICMP_SGT:
4245 default: assert(0 && "Unknown integer condition code!");
4246 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4247 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4248 return ReplaceInstUsesWith(I, LHS);
4249 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4251 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4252 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4253 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4254 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4262 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4263 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4264 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4265 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4266 const Type *SrcTy = Op0C->getOperand(0)->getType();
4267 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4268 // Only do this if the casts both really cause code to be generated.
4269 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4271 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4273 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4274 Op1C->getOperand(0),
4276 InsertNewInstBefore(NewOp, I);
4277 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4283 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4284 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4285 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4286 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4287 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4288 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4289 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4290 // If either of the constants are nans, then the whole thing returns
4292 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4293 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4295 // Otherwise, no need to compare the two constants, compare the
4297 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4298 RHS->getOperand(0));
4303 return Changed ? &I : 0;
4306 // XorSelf - Implements: X ^ X --> 0
4309 XorSelf(Value *rhs) : RHS(rhs) {}
4310 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4311 Instruction *apply(BinaryOperator &Xor) const {
4317 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4318 bool Changed = SimplifyCommutative(I);
4319 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4321 if (isa<UndefValue>(Op1))
4322 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4324 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4325 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4326 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4327 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4330 // See if we can simplify any instructions used by the instruction whose sole
4331 // purpose is to compute bits we don't care about.
4332 if (!isa<VectorType>(I.getType())) {
4333 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4334 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4335 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4336 KnownZero, KnownOne))
4338 } else if (isa<ConstantAggregateZero>(Op1)) {
4339 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4342 // Is this a ~ operation?
4343 if (Value *NotOp = dyn_castNotVal(&I)) {
4344 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4345 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4346 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4347 if (Op0I->getOpcode() == Instruction::And ||
4348 Op0I->getOpcode() == Instruction::Or) {
4349 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4350 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4352 BinaryOperator::createNot(Op0I->getOperand(1),
4353 Op0I->getOperand(1)->getName()+".not");
4354 InsertNewInstBefore(NotY, I);
4355 if (Op0I->getOpcode() == Instruction::And)
4356 return BinaryOperator::createOr(Op0NotVal, NotY);
4358 return BinaryOperator::createAnd(Op0NotVal, NotY);
4365 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4366 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4367 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4368 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4369 return new ICmpInst(ICI->getInversePredicate(),
4370 ICI->getOperand(0), ICI->getOperand(1));
4372 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4373 return new FCmpInst(FCI->getInversePredicate(),
4374 FCI->getOperand(0), FCI->getOperand(1));
4377 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4378 // ~(c-X) == X-c-1 == X+(-c-1)
4379 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4380 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4381 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4382 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4383 ConstantInt::get(I.getType(), 1));
4384 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4387 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4388 if (Op0I->getOpcode() == Instruction::Add) {
4389 // ~(X-c) --> (-c-1)-X
4390 if (RHS->isAllOnesValue()) {
4391 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4392 return BinaryOperator::createSub(
4393 ConstantExpr::getSub(NegOp0CI,
4394 ConstantInt::get(I.getType(), 1)),
4395 Op0I->getOperand(0));
4396 } else if (RHS->getValue().isSignBit()) {
4397 // (X + C) ^ signbit -> (X + C + signbit)
4398 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4399 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4402 } else if (Op0I->getOpcode() == Instruction::Or) {
4403 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4404 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4405 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4406 // Anything in both C1 and C2 is known to be zero, remove it from
4408 Constant *CommonBits = And(Op0CI, RHS);
4409 NewRHS = ConstantExpr::getAnd(NewRHS,
4410 ConstantExpr::getNot(CommonBits));
4411 AddToWorkList(Op0I);
4412 I.setOperand(0, Op0I->getOperand(0));
4413 I.setOperand(1, NewRHS);
4419 // Try to fold constant and into select arguments.
4420 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4421 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4423 if (isa<PHINode>(Op0))
4424 if (Instruction *NV = FoldOpIntoPhi(I))
4428 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4430 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4432 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4434 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4437 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4440 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4441 if (A == Op0) { // B^(B|A) == (A|B)^B
4442 Op1I->swapOperands();
4444 std::swap(Op0, Op1);
4445 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4446 I.swapOperands(); // Simplified below.
4447 std::swap(Op0, Op1);
4449 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4450 if (Op0 == A) // A^(A^B) == B
4451 return ReplaceInstUsesWith(I, B);
4452 else if (Op0 == B) // A^(B^A) == B
4453 return ReplaceInstUsesWith(I, A);
4454 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4455 if (A == Op0) { // A^(A&B) -> A^(B&A)
4456 Op1I->swapOperands();
4459 if (B == Op0) { // A^(B&A) -> (B&A)^A
4460 I.swapOperands(); // Simplified below.
4461 std::swap(Op0, Op1);
4466 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4469 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4470 if (A == Op1) // (B|A)^B == (A|B)^B
4472 if (B == Op1) { // (A|B)^B == A & ~B
4474 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4475 return BinaryOperator::createAnd(A, NotB);
4477 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4478 if (Op1 == A) // (A^B)^A == B
4479 return ReplaceInstUsesWith(I, B);
4480 else if (Op1 == B) // (B^A)^A == B
4481 return ReplaceInstUsesWith(I, A);
4482 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4483 if (A == Op1) // (A&B)^A -> (B&A)^A
4485 if (B == Op1 && // (B&A)^A == ~B & A
4486 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4488 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4489 return BinaryOperator::createAnd(N, Op1);
4494 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4495 if (Op0I && Op1I && Op0I->isShift() &&
4496 Op0I->getOpcode() == Op1I->getOpcode() &&
4497 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4498 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4499 Instruction *NewOp =
4500 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4501 Op1I->getOperand(0),
4502 Op0I->getName()), I);
4503 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4504 Op1I->getOperand(1));
4508 Value *A, *B, *C, *D;
4509 // (A & B)^(A | B) -> A ^ B
4510 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4511 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4512 if ((A == C && B == D) || (A == D && B == C))
4513 return BinaryOperator::createXor(A, B);
4515 // (A | B)^(A & B) -> A ^ B
4516 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4517 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4518 if ((A == C && B == D) || (A == D && B == C))
4519 return BinaryOperator::createXor(A, B);
4523 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4524 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4525 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4526 // (X & Y)^(X & Y) -> (Y^Z) & X
4527 Value *X = 0, *Y = 0, *Z = 0;
4529 X = A, Y = B, Z = D;
4531 X = A, Y = B, Z = C;
4533 X = B, Y = A, Z = D;
4535 X = B, Y = A, Z = C;
4538 Instruction *NewOp =
4539 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4540 return BinaryOperator::createAnd(NewOp, X);
4545 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4546 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4547 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4550 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4551 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4552 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4553 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4554 const Type *SrcTy = Op0C->getOperand(0)->getType();
4555 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4556 // Only do this if the casts both really cause code to be generated.
4557 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4559 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4561 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4562 Op1C->getOperand(0),
4564 InsertNewInstBefore(NewOp, I);
4565 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4569 return Changed ? &I : 0;
4572 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4573 /// overflowed for this type.
4574 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4575 ConstantInt *In2, bool IsSigned = false) {
4576 Result = cast<ConstantInt>(Add(In1, In2));
4579 if (In2->getValue().isNegative())
4580 return Result->getValue().sgt(In1->getValue());
4582 return Result->getValue().slt(In1->getValue());
4584 return Result->getValue().ult(In1->getValue());
4587 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4588 /// code necessary to compute the offset from the base pointer (without adding
4589 /// in the base pointer). Return the result as a signed integer of intptr size.
4590 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4591 TargetData &TD = IC.getTargetData();
4592 gep_type_iterator GTI = gep_type_begin(GEP);
4593 const Type *IntPtrTy = TD.getIntPtrType();
4594 Value *Result = Constant::getNullValue(IntPtrTy);
4596 // Build a mask for high order bits.
4597 unsigned IntPtrWidth = TD.getPointerSize()*8;
4598 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4600 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4601 Value *Op = GEP->getOperand(i);
4602 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4603 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4604 if (OpC->isZero()) continue;
4606 // Handle a struct index, which adds its field offset to the pointer.
4607 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4608 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4610 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4611 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4613 Result = IC.InsertNewInstBefore(
4614 BinaryOperator::createAdd(Result,
4615 ConstantInt::get(IntPtrTy, Size),
4616 GEP->getName()+".offs"), I);
4620 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4621 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4622 Scale = ConstantExpr::getMul(OC, Scale);
4623 if (Constant *RC = dyn_cast<Constant>(Result))
4624 Result = ConstantExpr::getAdd(RC, Scale);
4626 // Emit an add instruction.
4627 Result = IC.InsertNewInstBefore(
4628 BinaryOperator::createAdd(Result, Scale,
4629 GEP->getName()+".offs"), I);
4633 // Convert to correct type.
4634 if (Op->getType() != IntPtrTy) {
4635 if (Constant *OpC = dyn_cast<Constant>(Op))
4636 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4638 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4639 Op->getName()+".c"), I);
4642 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4643 if (Constant *OpC = dyn_cast<Constant>(Op))
4644 Op = ConstantExpr::getMul(OpC, Scale);
4645 else // We'll let instcombine(mul) convert this to a shl if possible.
4646 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4647 GEP->getName()+".idx"), I);
4650 // Emit an add instruction.
4651 if (isa<Constant>(Op) && isa<Constant>(Result))
4652 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4653 cast<Constant>(Result));
4655 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4656 GEP->getName()+".offs"), I);
4661 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4662 /// else. At this point we know that the GEP is on the LHS of the comparison.
4663 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4664 ICmpInst::Predicate Cond,
4666 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4668 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4669 if (isa<PointerType>(CI->getOperand(0)->getType()))
4670 RHS = CI->getOperand(0);
4672 Value *PtrBase = GEPLHS->getOperand(0);
4673 if (PtrBase == RHS) {
4674 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4675 // This transformation is valid because we know pointers can't overflow.
4676 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4677 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4678 Constant::getNullValue(Offset->getType()));
4679 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4680 // If the base pointers are different, but the indices are the same, just
4681 // compare the base pointer.
4682 if (PtrBase != GEPRHS->getOperand(0)) {
4683 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4684 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4685 GEPRHS->getOperand(0)->getType();
4687 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4688 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4689 IndicesTheSame = false;
4693 // If all indices are the same, just compare the base pointers.
4695 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4696 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4698 // Otherwise, the base pointers are different and the indices are
4699 // different, bail out.
4703 // If one of the GEPs has all zero indices, recurse.
4704 bool AllZeros = true;
4705 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4706 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4707 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4712 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4713 ICmpInst::getSwappedPredicate(Cond), I);
4715 // If the other GEP has all zero indices, recurse.
4717 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4718 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4719 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4724 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4726 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4727 // If the GEPs only differ by one index, compare it.
4728 unsigned NumDifferences = 0; // Keep track of # differences.
4729 unsigned DiffOperand = 0; // The operand that differs.
4730 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4731 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4732 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4733 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4734 // Irreconcilable differences.
4738 if (NumDifferences++) break;
4743 if (NumDifferences == 0) // SAME GEP?
4744 return ReplaceInstUsesWith(I, // No comparison is needed here.
4745 ConstantInt::get(Type::Int1Ty,
4746 isTrueWhenEqual(Cond)));
4748 else if (NumDifferences == 1) {
4749 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4750 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4751 // Make sure we do a signed comparison here.
4752 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4756 // Only lower this if the icmp is the only user of the GEP or if we expect
4757 // the result to fold to a constant!
4758 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4759 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4760 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4761 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4762 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4763 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4769 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4770 bool Changed = SimplifyCompare(I);
4771 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4773 // Fold trivial predicates.
4774 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4775 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4776 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4777 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4779 // Simplify 'fcmp pred X, X'
4781 switch (I.getPredicate()) {
4782 default: assert(0 && "Unknown predicate!");
4783 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4784 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4785 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4786 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4787 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4788 case FCmpInst::FCMP_OLT: // True if ordered and less than
4789 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4790 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4792 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4793 case FCmpInst::FCMP_ULT: // True if unordered or less than
4794 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4795 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4796 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4797 I.setPredicate(FCmpInst::FCMP_UNO);
4798 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4801 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4802 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4803 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4804 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4805 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4806 I.setPredicate(FCmpInst::FCMP_ORD);
4807 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4812 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4813 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4815 // Handle fcmp with constant RHS
4816 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4817 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4818 switch (LHSI->getOpcode()) {
4819 case Instruction::PHI:
4820 if (Instruction *NV = FoldOpIntoPhi(I))
4823 case Instruction::Select:
4824 // If either operand of the select is a constant, we can fold the
4825 // comparison into the select arms, which will cause one to be
4826 // constant folded and the select turned into a bitwise or.
4827 Value *Op1 = 0, *Op2 = 0;
4828 if (LHSI->hasOneUse()) {
4829 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4830 // Fold the known value into the constant operand.
4831 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4832 // Insert a new FCmp of the other select operand.
4833 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4834 LHSI->getOperand(2), RHSC,
4836 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4837 // Fold the known value into the constant operand.
4838 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4839 // Insert a new FCmp of the other select operand.
4840 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4841 LHSI->getOperand(1), RHSC,
4847 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4852 return Changed ? &I : 0;
4855 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4856 bool Changed = SimplifyCompare(I);
4857 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4858 const Type *Ty = Op0->getType();
4862 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4863 isTrueWhenEqual(I)));
4865 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4866 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4868 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4869 // addresses never equal each other! We already know that Op0 != Op1.
4870 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4871 isa<ConstantPointerNull>(Op0)) &&
4872 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4873 isa<ConstantPointerNull>(Op1)))
4874 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4875 !isTrueWhenEqual(I)));
4877 // icmp's with boolean values can always be turned into bitwise operations
4878 if (Ty == Type::Int1Ty) {
4879 switch (I.getPredicate()) {
4880 default: assert(0 && "Invalid icmp instruction!");
4881 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4882 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4883 InsertNewInstBefore(Xor, I);
4884 return BinaryOperator::createNot(Xor);
4886 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4887 return BinaryOperator::createXor(Op0, Op1);
4889 case ICmpInst::ICMP_UGT:
4890 case ICmpInst::ICMP_SGT:
4891 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4893 case ICmpInst::ICMP_ULT:
4894 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4895 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4896 InsertNewInstBefore(Not, I);
4897 return BinaryOperator::createAnd(Not, Op1);
4899 case ICmpInst::ICMP_UGE:
4900 case ICmpInst::ICMP_SGE:
4901 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4903 case ICmpInst::ICMP_ULE:
4904 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4905 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4906 InsertNewInstBefore(Not, I);
4907 return BinaryOperator::createOr(Not, Op1);
4912 // See if we are doing a comparison between a constant and an instruction that
4913 // can be folded into the comparison.
4914 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4917 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
4918 if (I.isEquality() && CI->isNullValue() &&
4919 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
4920 // (icmp cond A B) if cond is equality
4921 return new ICmpInst(I.getPredicate(), A, B);
4924 switch (I.getPredicate()) {
4926 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4927 if (CI->isMinValue(false))
4928 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4929 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4930 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4931 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4932 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4933 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4934 if (CI->isMinValue(true))
4935 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4936 ConstantInt::getAllOnesValue(Op0->getType()));
4940 case ICmpInst::ICMP_SLT:
4941 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4942 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4943 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4944 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4945 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4946 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4949 case ICmpInst::ICMP_UGT:
4950 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4951 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4952 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4953 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4954 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4955 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4957 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4958 if (CI->isMaxValue(true))
4959 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4960 ConstantInt::getNullValue(Op0->getType()));
4963 case ICmpInst::ICMP_SGT:
4964 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4965 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4966 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4967 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4968 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4969 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4972 case ICmpInst::ICMP_ULE:
4973 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4974 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4975 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4976 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4977 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4978 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4981 case ICmpInst::ICMP_SLE:
4982 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4983 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4984 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4985 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4986 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4987 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4990 case ICmpInst::ICMP_UGE:
4991 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4992 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4993 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4994 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4995 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4996 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4999 case ICmpInst::ICMP_SGE:
5000 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5001 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5002 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5003 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5004 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5005 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5009 // If we still have a icmp le or icmp ge instruction, turn it into the
5010 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5011 // already been handled above, this requires little checking.
5013 switch (I.getPredicate()) {
5015 case ICmpInst::ICMP_ULE:
5016 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5017 case ICmpInst::ICMP_SLE:
5018 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5019 case ICmpInst::ICMP_UGE:
5020 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5021 case ICmpInst::ICMP_SGE:
5022 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5025 // See if we can fold the comparison based on bits known to be zero or one
5026 // in the input. If this comparison is a normal comparison, it demands all
5027 // bits, if it is a sign bit comparison, it only demands the sign bit.
5030 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5032 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5033 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5034 if (SimplifyDemandedBits(Op0,
5035 isSignBit ? APInt::getSignBit(BitWidth)
5036 : APInt::getAllOnesValue(BitWidth),
5037 KnownZero, KnownOne, 0))
5040 // Given the known and unknown bits, compute a range that the LHS could be
5042 if ((KnownOne | KnownZero) != 0) {
5043 // Compute the Min, Max and RHS values based on the known bits. For the
5044 // EQ and NE we use unsigned values.
5045 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5046 const APInt& RHSVal = CI->getValue();
5047 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5048 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5051 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5054 switch (I.getPredicate()) { // LE/GE have been folded already.
5055 default: assert(0 && "Unknown icmp opcode!");
5056 case ICmpInst::ICMP_EQ:
5057 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5058 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5060 case ICmpInst::ICMP_NE:
5061 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5062 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5064 case ICmpInst::ICMP_ULT:
5065 if (Max.ult(RHSVal))
5066 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5067 if (Min.uge(RHSVal))
5068 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5070 case ICmpInst::ICMP_UGT:
5071 if (Min.ugt(RHSVal))
5072 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5073 if (Max.ule(RHSVal))
5074 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5076 case ICmpInst::ICMP_SLT:
5077 if (Max.slt(RHSVal))
5078 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5079 if (Min.sgt(RHSVal))
5080 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5082 case ICmpInst::ICMP_SGT:
5083 if (Min.sgt(RHSVal))
5084 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5085 if (Max.sle(RHSVal))
5086 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5091 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5092 // instruction, see if that instruction also has constants so that the
5093 // instruction can be folded into the icmp
5094 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5095 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5099 // Handle icmp with constant (but not simple integer constant) RHS
5100 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5101 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5102 switch (LHSI->getOpcode()) {
5103 case Instruction::GetElementPtr:
5104 if (RHSC->isNullValue()) {
5105 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5106 bool isAllZeros = true;
5107 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5108 if (!isa<Constant>(LHSI->getOperand(i)) ||
5109 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5114 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5115 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5119 case Instruction::PHI:
5120 if (Instruction *NV = FoldOpIntoPhi(I))
5123 case Instruction::Select: {
5124 // If either operand of the select is a constant, we can fold the
5125 // comparison into the select arms, which will cause one to be
5126 // constant folded and the select turned into a bitwise or.
5127 Value *Op1 = 0, *Op2 = 0;
5128 if (LHSI->hasOneUse()) {
5129 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5130 // Fold the known value into the constant operand.
5131 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5132 // Insert a new ICmp of the other select operand.
5133 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5134 LHSI->getOperand(2), RHSC,
5136 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5137 // Fold the known value into the constant operand.
5138 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5139 // Insert a new ICmp of the other select operand.
5140 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5141 LHSI->getOperand(1), RHSC,
5147 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5150 case Instruction::Malloc:
5151 // If we have (malloc != null), and if the malloc has a single use, we
5152 // can assume it is successful and remove the malloc.
5153 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5154 AddToWorkList(LHSI);
5155 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5156 !isTrueWhenEqual(I)));
5162 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5163 if (User *GEP = dyn_castGetElementPtr(Op0))
5164 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5166 if (User *GEP = dyn_castGetElementPtr(Op1))
5167 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5168 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5171 // Test to see if the operands of the icmp are casted versions of other
5172 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5174 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5175 if (isa<PointerType>(Op0->getType()) &&
5176 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5177 // We keep moving the cast from the left operand over to the right
5178 // operand, where it can often be eliminated completely.
5179 Op0 = CI->getOperand(0);
5181 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5182 // so eliminate it as well.
5183 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5184 Op1 = CI2->getOperand(0);
5186 // If Op1 is a constant, we can fold the cast into the constant.
5187 if (Op0->getType() != Op1->getType())
5188 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5189 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5191 // Otherwise, cast the RHS right before the icmp
5192 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5194 return new ICmpInst(I.getPredicate(), Op0, Op1);
5198 if (isa<CastInst>(Op0)) {
5199 // Handle the special case of: icmp (cast bool to X), <cst>
5200 // This comes up when you have code like
5203 // For generality, we handle any zero-extension of any operand comparison
5204 // with a constant or another cast from the same type.
5205 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5206 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5210 if (I.isEquality()) {
5211 Value *A, *B, *C, *D;
5212 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5213 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5214 Value *OtherVal = A == Op1 ? B : A;
5215 return new ICmpInst(I.getPredicate(), OtherVal,
5216 Constant::getNullValue(A->getType()));
5219 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5220 // A^c1 == C^c2 --> A == C^(c1^c2)
5221 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5222 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5223 if (Op1->hasOneUse()) {
5224 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5225 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5226 return new ICmpInst(I.getPredicate(), A,
5227 InsertNewInstBefore(Xor, I));
5230 // A^B == A^D -> B == D
5231 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5232 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5233 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5234 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5238 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5239 (A == Op0 || B == Op0)) {
5240 // A == (A^B) -> B == 0
5241 Value *OtherVal = A == Op0 ? B : A;
5242 return new ICmpInst(I.getPredicate(), OtherVal,
5243 Constant::getNullValue(A->getType()));
5245 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5246 // (A-B) == A -> B == 0
5247 return new ICmpInst(I.getPredicate(), B,
5248 Constant::getNullValue(B->getType()));
5250 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5251 // A == (A-B) -> B == 0
5252 return new ICmpInst(I.getPredicate(), B,
5253 Constant::getNullValue(B->getType()));
5256 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5257 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5258 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5259 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5260 Value *X = 0, *Y = 0, *Z = 0;
5263 X = B; Y = D; Z = A;
5264 } else if (A == D) {
5265 X = B; Y = C; Z = A;
5266 } else if (B == C) {
5267 X = A; Y = D; Z = B;
5268 } else if (B == D) {
5269 X = A; Y = C; Z = B;
5272 if (X) { // Build (X^Y) & Z
5273 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5274 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5275 I.setOperand(0, Op1);
5276 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5281 return Changed ? &I : 0;
5285 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5286 /// and CmpRHS are both known to be integer constants.
5287 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5288 ConstantInt *DivRHS) {
5289 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5290 const APInt &CmpRHSV = CmpRHS->getValue();
5292 // FIXME: If the operand types don't match the type of the divide
5293 // then don't attempt this transform. The code below doesn't have the
5294 // logic to deal with a signed divide and an unsigned compare (and
5295 // vice versa). This is because (x /s C1) <s C2 produces different
5296 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5297 // (x /u C1) <u C2. Simply casting the operands and result won't
5298 // work. :( The if statement below tests that condition and bails
5300 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5301 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5303 if (DivRHS->isZero())
5304 return 0; // The ProdOV computation fails on divide by zero.
5306 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5307 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5308 // C2 (CI). By solving for X we can turn this into a range check
5309 // instead of computing a divide.
5310 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5312 // Determine if the product overflows by seeing if the product is
5313 // not equal to the divide. Make sure we do the same kind of divide
5314 // as in the LHS instruction that we're folding.
5315 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5316 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5318 // Get the ICmp opcode
5319 ICmpInst::Predicate Pred = ICI.getPredicate();
5321 // Figure out the interval that is being checked. For example, a comparison
5322 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5323 // Compute this interval based on the constants involved and the signedness of
5324 // the compare/divide. This computes a half-open interval, keeping track of
5325 // whether either value in the interval overflows. After analysis each
5326 // overflow variable is set to 0 if it's corresponding bound variable is valid
5327 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5328 int LoOverflow = 0, HiOverflow = 0;
5329 ConstantInt *LoBound = 0, *HiBound = 0;
5332 if (!DivIsSigned) { // udiv
5333 // e.g. X/5 op 3 --> [15, 20)
5335 HiOverflow = LoOverflow = ProdOV;
5337 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5338 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5339 if (CmpRHSV == 0) { // (X / pos) op 0
5340 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5341 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5343 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5344 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5345 HiOverflow = LoOverflow = ProdOV;
5347 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5348 } else { // (X / pos) op neg
5349 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5350 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5351 LoOverflow = AddWithOverflow(LoBound, Prod,
5352 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5353 HiBound = AddOne(Prod);
5354 HiOverflow = ProdOV ? -1 : 0;
5356 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5357 if (CmpRHSV == 0) { // (X / neg) op 0
5358 // e.g. X/-5 op 0 --> [-4, 5)
5359 LoBound = AddOne(DivRHS);
5360 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5361 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5362 HiOverflow = 1; // [INTMIN+1, overflow)
5363 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5365 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5366 // e.g. X/-5 op 3 --> [-19, -14)
5367 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5369 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5370 HiBound = AddOne(Prod);
5371 } else { // (X / neg) op neg
5372 // e.g. X/-5 op -3 --> [15, 20)
5374 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5375 HiBound = Subtract(Prod, DivRHS);
5378 // Dividing by a negative swaps the condition. LT <-> GT
5379 Pred = ICmpInst::getSwappedPredicate(Pred);
5382 Value *X = DivI->getOperand(0);
5384 default: assert(0 && "Unhandled icmp opcode!");
5385 case ICmpInst::ICMP_EQ:
5386 if (LoOverflow && HiOverflow)
5387 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5388 else if (HiOverflow)
5389 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5390 ICmpInst::ICMP_UGE, X, LoBound);
5391 else if (LoOverflow)
5392 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5393 ICmpInst::ICMP_ULT, X, HiBound);
5395 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5396 case ICmpInst::ICMP_NE:
5397 if (LoOverflow && HiOverflow)
5398 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5399 else if (HiOverflow)
5400 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5401 ICmpInst::ICMP_ULT, X, LoBound);
5402 else if (LoOverflow)
5403 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5404 ICmpInst::ICMP_UGE, X, HiBound);
5406 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5407 case ICmpInst::ICMP_ULT:
5408 case ICmpInst::ICMP_SLT:
5409 if (LoOverflow == +1) // Low bound is greater than input range.
5410 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5411 if (LoOverflow == -1) // Low bound is less than input range.
5412 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5413 return new ICmpInst(Pred, X, LoBound);
5414 case ICmpInst::ICMP_UGT:
5415 case ICmpInst::ICMP_SGT:
5416 if (HiOverflow == +1) // High bound greater than input range.
5417 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5418 else if (HiOverflow == -1) // High bound less than input range.
5419 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5420 if (Pred == ICmpInst::ICMP_UGT)
5421 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5423 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5428 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5430 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5433 const APInt &RHSV = RHS->getValue();
5435 switch (LHSI->getOpcode()) {
5436 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5437 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5438 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5440 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5441 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5442 Value *CompareVal = LHSI->getOperand(0);
5444 // If the sign bit of the XorCST is not set, there is no change to
5445 // the operation, just stop using the Xor.
5446 if (!XorCST->getValue().isNegative()) {
5447 ICI.setOperand(0, CompareVal);
5448 AddToWorkList(LHSI);
5452 // Was the old condition true if the operand is positive?
5453 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5455 // If so, the new one isn't.
5456 isTrueIfPositive ^= true;
5458 if (isTrueIfPositive)
5459 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5461 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5465 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5466 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5467 LHSI->getOperand(0)->hasOneUse()) {
5468 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5470 // If the LHS is an AND of a truncating cast, we can widen the
5471 // and/compare to be the input width without changing the value
5472 // produced, eliminating a cast.
5473 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5474 // We can do this transformation if either the AND constant does not
5475 // have its sign bit set or if it is an equality comparison.
5476 // Extending a relational comparison when we're checking the sign
5477 // bit would not work.
5478 if (Cast->hasOneUse() &&
5479 (ICI.isEquality() || AndCST->getValue().isNonNegative() &&
5480 RHSV.isNonNegative())) {
5482 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5483 APInt NewCST = AndCST->getValue();
5484 NewCST.zext(BitWidth);
5486 NewCI.zext(BitWidth);
5487 Instruction *NewAnd =
5488 BinaryOperator::createAnd(Cast->getOperand(0),
5489 ConstantInt::get(NewCST),LHSI->getName());
5490 InsertNewInstBefore(NewAnd, ICI);
5491 return new ICmpInst(ICI.getPredicate(), NewAnd,
5492 ConstantInt::get(NewCI));
5496 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5497 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5498 // happens a LOT in code produced by the C front-end, for bitfield
5500 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5501 if (Shift && !Shift->isShift())
5505 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5506 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5507 const Type *AndTy = AndCST->getType(); // Type of the and.
5509 // We can fold this as long as we can't shift unknown bits
5510 // into the mask. This can only happen with signed shift
5511 // rights, as they sign-extend.
5513 bool CanFold = Shift->isLogicalShift();
5515 // To test for the bad case of the signed shr, see if any
5516 // of the bits shifted in could be tested after the mask.
5517 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5518 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5520 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5521 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5522 AndCST->getValue()) == 0)
5528 if (Shift->getOpcode() == Instruction::Shl)
5529 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5531 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5533 // Check to see if we are shifting out any of the bits being
5535 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5536 // If we shifted bits out, the fold is not going to work out.
5537 // As a special case, check to see if this means that the
5538 // result is always true or false now.
5539 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5540 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5541 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5542 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5544 ICI.setOperand(1, NewCst);
5545 Constant *NewAndCST;
5546 if (Shift->getOpcode() == Instruction::Shl)
5547 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5549 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5550 LHSI->setOperand(1, NewAndCST);
5551 LHSI->setOperand(0, Shift->getOperand(0));
5552 AddToWorkList(Shift); // Shift is dead.
5553 AddUsesToWorkList(ICI);
5559 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5560 // preferable because it allows the C<<Y expression to be hoisted out
5561 // of a loop if Y is invariant and X is not.
5562 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5563 ICI.isEquality() && !Shift->isArithmeticShift() &&
5564 isa<Instruction>(Shift->getOperand(0))) {
5567 if (Shift->getOpcode() == Instruction::LShr) {
5568 NS = BinaryOperator::createShl(AndCST,
5569 Shift->getOperand(1), "tmp");
5571 // Insert a logical shift.
5572 NS = BinaryOperator::createLShr(AndCST,
5573 Shift->getOperand(1), "tmp");
5575 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5577 // Compute X & (C << Y).
5578 Instruction *NewAnd =
5579 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5580 InsertNewInstBefore(NewAnd, ICI);
5582 ICI.setOperand(0, NewAnd);
5588 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5589 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5592 uint32_t TypeBits = RHSV.getBitWidth();
5594 // Check that the shift amount is in range. If not, don't perform
5595 // undefined shifts. When the shift is visited it will be
5597 if (ShAmt->uge(TypeBits))
5600 if (ICI.isEquality()) {
5601 // If we are comparing against bits always shifted out, the
5602 // comparison cannot succeed.
5604 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5605 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5606 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5607 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5608 return ReplaceInstUsesWith(ICI, Cst);
5611 if (LHSI->hasOneUse()) {
5612 // Otherwise strength reduce the shift into an and.
5613 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5615 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5618 BinaryOperator::createAnd(LHSI->getOperand(0),
5619 Mask, LHSI->getName()+".mask");
5620 Value *And = InsertNewInstBefore(AndI, ICI);
5621 return new ICmpInst(ICI.getPredicate(), And,
5622 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5626 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5627 bool TrueIfSigned = false;
5628 if (LHSI->hasOneUse() &&
5629 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5630 // (X << 31) <s 0 --> (X&1) != 0
5631 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5632 (TypeBits-ShAmt->getZExtValue()-1));
5634 BinaryOperator::createAnd(LHSI->getOperand(0),
5635 Mask, LHSI->getName()+".mask");
5636 Value *And = InsertNewInstBefore(AndI, ICI);
5638 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5639 And, Constant::getNullValue(And->getType()));
5644 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5645 case Instruction::AShr: {
5646 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5649 if (ICI.isEquality()) {
5650 // Check that the shift amount is in range. If not, don't perform
5651 // undefined shifts. When the shift is visited it will be
5653 uint32_t TypeBits = RHSV.getBitWidth();
5654 if (ShAmt->uge(TypeBits))
5656 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5658 // If we are comparing against bits always shifted out, the
5659 // comparison cannot succeed.
5660 APInt Comp = RHSV << ShAmtVal;
5661 if (LHSI->getOpcode() == Instruction::LShr)
5662 Comp = Comp.lshr(ShAmtVal);
5664 Comp = Comp.ashr(ShAmtVal);
5666 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5667 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5668 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5669 return ReplaceInstUsesWith(ICI, Cst);
5672 if (LHSI->hasOneUse() || RHSV == 0) {
5673 // Otherwise strength reduce the shift into an and.
5674 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5675 Constant *Mask = ConstantInt::get(Val);
5678 BinaryOperator::createAnd(LHSI->getOperand(0),
5679 Mask, LHSI->getName()+".mask");
5680 Value *And = InsertNewInstBefore(AndI, ICI);
5681 return new ICmpInst(ICI.getPredicate(), And,
5682 ConstantExpr::getShl(RHS, ShAmt));
5688 case Instruction::SDiv:
5689 case Instruction::UDiv:
5690 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5691 // Fold this div into the comparison, producing a range check.
5692 // Determine, based on the divide type, what the range is being
5693 // checked. If there is an overflow on the low or high side, remember
5694 // it, otherwise compute the range [low, hi) bounding the new value.
5695 // See: InsertRangeTest above for the kinds of replacements possible.
5696 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5697 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5702 case Instruction::Add:
5703 // Fold: icmp pred (add, X, C1), C2
5705 if (!ICI.isEquality()) {
5706 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5708 const APInt &LHSV = LHSC->getValue();
5710 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
5713 if (ICI.isSignedPredicate()) {
5714 if (CR.getLower().isSignBit()) {
5715 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
5716 ConstantInt::get(CR.getUpper()));
5717 } else if (CR.getUpper().isSignBit()) {
5718 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
5719 ConstantInt::get(CR.getLower()));
5722 if (CR.getLower().isMinValue()) {
5723 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
5724 ConstantInt::get(CR.getUpper()));
5725 } else if (CR.getUpper().isMinValue()) {
5726 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
5727 ConstantInt::get(CR.getLower()));
5734 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5735 if (ICI.isEquality()) {
5736 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5738 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5739 // the second operand is a constant, simplify a bit.
5740 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5741 switch (BO->getOpcode()) {
5742 case Instruction::SRem:
5743 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5744 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5745 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5746 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5747 Instruction *NewRem =
5748 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5750 InsertNewInstBefore(NewRem, ICI);
5751 return new ICmpInst(ICI.getPredicate(), NewRem,
5752 Constant::getNullValue(BO->getType()));
5756 case Instruction::Add:
5757 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5758 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5759 if (BO->hasOneUse())
5760 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5761 Subtract(RHS, BOp1C));
5762 } else if (RHSV == 0) {
5763 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5764 // efficiently invertible, or if the add has just this one use.
5765 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5767 if (Value *NegVal = dyn_castNegVal(BOp1))
5768 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5769 else if (Value *NegVal = dyn_castNegVal(BOp0))
5770 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5771 else if (BO->hasOneUse()) {
5772 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5773 InsertNewInstBefore(Neg, ICI);
5775 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5779 case Instruction::Xor:
5780 // For the xor case, we can xor two constants together, eliminating
5781 // the explicit xor.
5782 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5783 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5784 ConstantExpr::getXor(RHS, BOC));
5787 case Instruction::Sub:
5788 // Replace (([sub|xor] A, B) != 0) with (A != B)
5790 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5794 case Instruction::Or:
5795 // If bits are being or'd in that are not present in the constant we
5796 // are comparing against, then the comparison could never succeed!
5797 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5798 Constant *NotCI = ConstantExpr::getNot(RHS);
5799 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5800 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5805 case Instruction::And:
5806 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5807 // If bits are being compared against that are and'd out, then the
5808 // comparison can never succeed!
5809 if ((RHSV & ~BOC->getValue()) != 0)
5810 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5813 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5814 if (RHS == BOC && RHSV.isPowerOf2())
5815 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5816 ICmpInst::ICMP_NE, LHSI,
5817 Constant::getNullValue(RHS->getType()));
5819 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5820 if (isSignBit(BOC)) {
5821 Value *X = BO->getOperand(0);
5822 Constant *Zero = Constant::getNullValue(X->getType());
5823 ICmpInst::Predicate pred = isICMP_NE ?
5824 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5825 return new ICmpInst(pred, X, Zero);
5828 // ((X & ~7) == 0) --> X < 8
5829 if (RHSV == 0 && isHighOnes(BOC)) {
5830 Value *X = BO->getOperand(0);
5831 Constant *NegX = ConstantExpr::getNeg(BOC);
5832 ICmpInst::Predicate pred = isICMP_NE ?
5833 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5834 return new ICmpInst(pred, X, NegX);
5839 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5840 // Handle icmp {eq|ne} <intrinsic>, intcst.
5841 if (II->getIntrinsicID() == Intrinsic::bswap) {
5843 ICI.setOperand(0, II->getOperand(1));
5844 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5848 } else { // Not a ICMP_EQ/ICMP_NE
5849 // If the LHS is a cast from an integral value of the same size,
5850 // then since we know the RHS is a constant, try to simlify.
5851 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5852 Value *CastOp = Cast->getOperand(0);
5853 const Type *SrcTy = CastOp->getType();
5854 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5855 if (SrcTy->isInteger() &&
5856 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5857 // If this is an unsigned comparison, try to make the comparison use
5858 // smaller constant values.
5859 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5860 // X u< 128 => X s> -1
5861 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5862 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5863 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5864 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5865 // X u> 127 => X s< 0
5866 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5867 Constant::getNullValue(SrcTy));
5875 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5876 /// We only handle extending casts so far.
5878 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5879 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5880 Value *LHSCIOp = LHSCI->getOperand(0);
5881 const Type *SrcTy = LHSCIOp->getType();
5882 const Type *DestTy = LHSCI->getType();
5885 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5886 // integer type is the same size as the pointer type.
5887 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5888 getTargetData().getPointerSizeInBits() ==
5889 cast<IntegerType>(DestTy)->getBitWidth()) {
5891 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5892 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5893 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5894 RHSOp = RHSC->getOperand(0);
5895 // If the pointer types don't match, insert a bitcast.
5896 if (LHSCIOp->getType() != RHSOp->getType())
5897 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
5901 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5904 // The code below only handles extension cast instructions, so far.
5906 if (LHSCI->getOpcode() != Instruction::ZExt &&
5907 LHSCI->getOpcode() != Instruction::SExt)
5910 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5911 bool isSignedCmp = ICI.isSignedPredicate();
5913 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5914 // Not an extension from the same type?
5915 RHSCIOp = CI->getOperand(0);
5916 if (RHSCIOp->getType() != LHSCIOp->getType())
5919 // If the signedness of the two casts doesn't agree (i.e. one is a sext
5920 // and the other is a zext), then we can't handle this.
5921 if (CI->getOpcode() != LHSCI->getOpcode())
5924 // Deal with equality cases early.
5925 if (ICI.isEquality())
5926 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5928 // A signed comparison of sign extended values simplifies into a
5929 // signed comparison.
5930 if (isSignedCmp && isSignedExt)
5931 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5933 // The other three cases all fold into an unsigned comparison.
5934 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
5937 // If we aren't dealing with a constant on the RHS, exit early
5938 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5942 // Compute the constant that would happen if we truncated to SrcTy then
5943 // reextended to DestTy.
5944 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5945 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5947 // If the re-extended constant didn't change...
5949 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5950 // For example, we might have:
5951 // %A = sext short %X to uint
5952 // %B = icmp ugt uint %A, 1330
5953 // It is incorrect to transform this into
5954 // %B = icmp ugt short %X, 1330
5955 // because %A may have negative value.
5957 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5958 // OR operation is EQ/NE.
5959 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5960 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5965 // The re-extended constant changed so the constant cannot be represented
5966 // in the shorter type. Consequently, we cannot emit a simple comparison.
5968 // First, handle some easy cases. We know the result cannot be equal at this
5969 // point so handle the ICI.isEquality() cases
5970 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5971 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5972 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5973 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5975 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5976 // should have been folded away previously and not enter in here.
5979 // We're performing a signed comparison.
5980 if (cast<ConstantInt>(CI)->getValue().isNegative())
5981 Result = ConstantInt::getFalse(); // X < (small) --> false
5983 Result = ConstantInt::getTrue(); // X < (large) --> true
5985 // We're performing an unsigned comparison.
5987 // We're performing an unsigned comp with a sign extended value.
5988 // This is true if the input is >= 0. [aka >s -1]
5989 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5990 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5991 NegOne, ICI.getName()), ICI);
5993 // Unsigned extend & unsigned compare -> always true.
5994 Result = ConstantInt::getTrue();
5998 // Finally, return the value computed.
5999 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6000 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6001 return ReplaceInstUsesWith(ICI, Result);
6003 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6004 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6005 "ICmp should be folded!");
6006 if (Constant *CI = dyn_cast<Constant>(Result))
6007 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6009 return BinaryOperator::createNot(Result);
6013 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6014 return commonShiftTransforms(I);
6017 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6018 return commonShiftTransforms(I);
6021 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6022 if (Instruction *R = commonShiftTransforms(I))
6025 Value *Op0 = I.getOperand(0);
6027 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6028 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6029 if (CSI->isAllOnesValue())
6030 return ReplaceInstUsesWith(I, CSI);
6032 // See if we can turn a signed shr into an unsigned shr.
6033 if (MaskedValueIsZero(Op0,
6034 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6035 return BinaryOperator::createLShr(Op0, I.getOperand(1));
6040 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6041 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6042 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6044 // shl X, 0 == X and shr X, 0 == X
6045 // shl 0, X == 0 and shr 0, X == 0
6046 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6047 Op0 == Constant::getNullValue(Op0->getType()))
6048 return ReplaceInstUsesWith(I, Op0);
6050 if (isa<UndefValue>(Op0)) {
6051 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6052 return ReplaceInstUsesWith(I, Op0);
6053 else // undef << X -> 0, undef >>u X -> 0
6054 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6056 if (isa<UndefValue>(Op1)) {
6057 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6058 return ReplaceInstUsesWith(I, Op0);
6059 else // X << undef, X >>u undef -> 0
6060 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6063 // Try to fold constant and into select arguments.
6064 if (isa<Constant>(Op0))
6065 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6066 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6069 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6070 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6075 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6076 BinaryOperator &I) {
6077 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6079 // See if we can simplify any instructions used by the instruction whose sole
6080 // purpose is to compute bits we don't care about.
6081 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6082 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6083 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6084 KnownZero, KnownOne))
6087 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6088 // of a signed value.
6090 if (Op1->uge(TypeBits)) {
6091 if (I.getOpcode() != Instruction::AShr)
6092 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6094 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6099 // ((X*C1) << C2) == (X * (C1 << C2))
6100 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6101 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6102 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6103 return BinaryOperator::createMul(BO->getOperand(0),
6104 ConstantExpr::getShl(BOOp, Op1));
6106 // Try to fold constant and into select arguments.
6107 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6108 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6110 if (isa<PHINode>(Op0))
6111 if (Instruction *NV = FoldOpIntoPhi(I))
6114 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6115 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6116 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6117 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6118 // place. Don't try to do this transformation in this case. Also, we
6119 // require that the input operand is a shift-by-constant so that we have
6120 // confidence that the shifts will get folded together. We could do this
6121 // xform in more cases, but it is unlikely to be profitable.
6122 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6123 isa<ConstantInt>(TrOp->getOperand(1))) {
6124 // Okay, we'll do this xform. Make the shift of shift.
6125 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6126 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6128 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6130 // For logical shifts, the truncation has the effect of making the high
6131 // part of the register be zeros. Emulate this by inserting an AND to
6132 // clear the top bits as needed. This 'and' will usually be zapped by
6133 // other xforms later if dead.
6134 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6135 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6136 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6138 // The mask we constructed says what the trunc would do if occurring
6139 // between the shifts. We want to know the effect *after* the second
6140 // shift. We know that it is a logical shift by a constant, so adjust the
6141 // mask as appropriate.
6142 if (I.getOpcode() == Instruction::Shl)
6143 MaskV <<= Op1->getZExtValue();
6145 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6146 MaskV = MaskV.lshr(Op1->getZExtValue());
6149 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6151 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6153 // Return the value truncated to the interesting size.
6154 return new TruncInst(And, I.getType());
6158 if (Op0->hasOneUse()) {
6159 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6160 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6163 switch (Op0BO->getOpcode()) {
6165 case Instruction::Add:
6166 case Instruction::And:
6167 case Instruction::Or:
6168 case Instruction::Xor: {
6169 // These operators commute.
6170 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6171 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6172 match(Op0BO->getOperand(1),
6173 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6174 Instruction *YS = BinaryOperator::createShl(
6175 Op0BO->getOperand(0), Op1,
6177 InsertNewInstBefore(YS, I); // (Y << C)
6179 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6180 Op0BO->getOperand(1)->getName());
6181 InsertNewInstBefore(X, I); // (X + (Y << C))
6182 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6183 return BinaryOperator::createAnd(X, ConstantInt::get(
6184 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6187 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6188 Value *Op0BOOp1 = Op0BO->getOperand(1);
6189 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6191 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6192 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6194 Instruction *YS = BinaryOperator::createShl(
6195 Op0BO->getOperand(0), Op1,
6197 InsertNewInstBefore(YS, I); // (Y << C)
6199 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6200 V1->getName()+".mask");
6201 InsertNewInstBefore(XM, I); // X & (CC << C)
6203 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6208 case Instruction::Sub: {
6209 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6210 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6211 match(Op0BO->getOperand(0),
6212 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6213 Instruction *YS = BinaryOperator::createShl(
6214 Op0BO->getOperand(1), Op1,
6216 InsertNewInstBefore(YS, I); // (Y << C)
6218 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6219 Op0BO->getOperand(0)->getName());
6220 InsertNewInstBefore(X, I); // (X + (Y << C))
6221 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6222 return BinaryOperator::createAnd(X, ConstantInt::get(
6223 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6226 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6227 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6228 match(Op0BO->getOperand(0),
6229 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6230 m_ConstantInt(CC))) && V2 == Op1 &&
6231 cast<BinaryOperator>(Op0BO->getOperand(0))
6232 ->getOperand(0)->hasOneUse()) {
6233 Instruction *YS = BinaryOperator::createShl(
6234 Op0BO->getOperand(1), Op1,
6236 InsertNewInstBefore(YS, I); // (Y << C)
6238 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6239 V1->getName()+".mask");
6240 InsertNewInstBefore(XM, I); // X & (CC << C)
6242 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6250 // If the operand is an bitwise operator with a constant RHS, and the
6251 // shift is the only use, we can pull it out of the shift.
6252 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6253 bool isValid = true; // Valid only for And, Or, Xor
6254 bool highBitSet = false; // Transform if high bit of constant set?
6256 switch (Op0BO->getOpcode()) {
6257 default: isValid = false; break; // Do not perform transform!
6258 case Instruction::Add:
6259 isValid = isLeftShift;
6261 case Instruction::Or:
6262 case Instruction::Xor:
6265 case Instruction::And:
6270 // If this is a signed shift right, and the high bit is modified
6271 // by the logical operation, do not perform the transformation.
6272 // The highBitSet boolean indicates the value of the high bit of
6273 // the constant which would cause it to be modified for this
6276 if (isValid && I.getOpcode() == Instruction::AShr)
6277 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6280 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6282 Instruction *NewShift =
6283 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6284 InsertNewInstBefore(NewShift, I);
6285 NewShift->takeName(Op0BO);
6287 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6294 // Find out if this is a shift of a shift by a constant.
6295 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6296 if (ShiftOp && !ShiftOp->isShift())
6299 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6300 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6301 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6302 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6303 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6304 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6305 Value *X = ShiftOp->getOperand(0);
6307 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6308 if (AmtSum > TypeBits)
6311 const IntegerType *Ty = cast<IntegerType>(I.getType());
6313 // Check for (X << c1) << c2 and (X >> c1) >> c2
6314 if (I.getOpcode() == ShiftOp->getOpcode()) {
6315 return BinaryOperator::create(I.getOpcode(), X,
6316 ConstantInt::get(Ty, AmtSum));
6317 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6318 I.getOpcode() == Instruction::AShr) {
6319 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6320 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6321 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6322 I.getOpcode() == Instruction::LShr) {
6323 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6324 Instruction *Shift =
6325 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6326 InsertNewInstBefore(Shift, I);
6328 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6329 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6332 // Okay, if we get here, one shift must be left, and the other shift must be
6333 // right. See if the amounts are equal.
6334 if (ShiftAmt1 == ShiftAmt2) {
6335 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6336 if (I.getOpcode() == Instruction::Shl) {
6337 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6338 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6340 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6341 if (I.getOpcode() == Instruction::LShr) {
6342 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6343 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6345 // We can simplify ((X << C) >>s C) into a trunc + sext.
6346 // NOTE: we could do this for any C, but that would make 'unusual' integer
6347 // types. For now, just stick to ones well-supported by the code
6349 const Type *SExtType = 0;
6350 switch (Ty->getBitWidth() - ShiftAmt1) {
6357 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6362 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6363 InsertNewInstBefore(NewTrunc, I);
6364 return new SExtInst(NewTrunc, Ty);
6366 // Otherwise, we can't handle it yet.
6367 } else if (ShiftAmt1 < ShiftAmt2) {
6368 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6370 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6371 if (I.getOpcode() == Instruction::Shl) {
6372 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6373 ShiftOp->getOpcode() == Instruction::AShr);
6374 Instruction *Shift =
6375 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6376 InsertNewInstBefore(Shift, I);
6378 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6379 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6382 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6383 if (I.getOpcode() == Instruction::LShr) {
6384 assert(ShiftOp->getOpcode() == Instruction::Shl);
6385 Instruction *Shift =
6386 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6387 InsertNewInstBefore(Shift, I);
6389 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6390 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6393 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6395 assert(ShiftAmt2 < ShiftAmt1);
6396 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6398 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6399 if (I.getOpcode() == Instruction::Shl) {
6400 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6401 ShiftOp->getOpcode() == Instruction::AShr);
6402 Instruction *Shift =
6403 BinaryOperator::create(ShiftOp->getOpcode(), X,
6404 ConstantInt::get(Ty, ShiftDiff));
6405 InsertNewInstBefore(Shift, I);
6407 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6408 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6411 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6412 if (I.getOpcode() == Instruction::LShr) {
6413 assert(ShiftOp->getOpcode() == Instruction::Shl);
6414 Instruction *Shift =
6415 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6416 InsertNewInstBefore(Shift, I);
6418 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6419 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6422 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6429 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6430 /// expression. If so, decompose it, returning some value X, such that Val is
6433 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6435 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6436 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6437 Offset = CI->getZExtValue();
6439 return ConstantInt::get(Type::Int32Ty, 0);
6440 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6441 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6442 if (I->getOpcode() == Instruction::Shl) {
6443 // This is a value scaled by '1 << the shift amt'.
6444 Scale = 1U << RHS->getZExtValue();
6446 return I->getOperand(0);
6447 } else if (I->getOpcode() == Instruction::Mul) {
6448 // This value is scaled by 'RHS'.
6449 Scale = RHS->getZExtValue();
6451 return I->getOperand(0);
6452 } else if (I->getOpcode() == Instruction::Add) {
6453 // We have X+C. Check to see if we really have (X*C2)+C1,
6454 // where C1 is divisible by C2.
6457 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6458 Offset += RHS->getZExtValue();
6465 // Otherwise, we can't look past this.
6472 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6473 /// try to eliminate the cast by moving the type information into the alloc.
6474 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6475 AllocationInst &AI) {
6476 const PointerType *PTy = cast<PointerType>(CI.getType());
6478 // Remove any uses of AI that are dead.
6479 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6481 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6482 Instruction *User = cast<Instruction>(*UI++);
6483 if (isInstructionTriviallyDead(User)) {
6484 while (UI != E && *UI == User)
6485 ++UI; // If this instruction uses AI more than once, don't break UI.
6488 DOUT << "IC: DCE: " << *User;
6489 EraseInstFromFunction(*User);
6493 // Get the type really allocated and the type casted to.
6494 const Type *AllocElTy = AI.getAllocatedType();
6495 const Type *CastElTy = PTy->getElementType();
6496 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6498 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6499 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6500 if (CastElTyAlign < AllocElTyAlign) return 0;
6502 // If the allocation has multiple uses, only promote it if we are strictly
6503 // increasing the alignment of the resultant allocation. If we keep it the
6504 // same, we open the door to infinite loops of various kinds.
6505 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6507 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6508 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6509 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6511 // See if we can satisfy the modulus by pulling a scale out of the array
6513 unsigned ArraySizeScale;
6515 Value *NumElements = // See if the array size is a decomposable linear expr.
6516 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6518 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6520 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6521 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6523 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6528 // If the allocation size is constant, form a constant mul expression
6529 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6530 if (isa<ConstantInt>(NumElements))
6531 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6532 // otherwise multiply the amount and the number of elements
6533 else if (Scale != 1) {
6534 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6535 Amt = InsertNewInstBefore(Tmp, AI);
6539 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6540 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6541 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6542 Amt = InsertNewInstBefore(Tmp, AI);
6545 AllocationInst *New;
6546 if (isa<MallocInst>(AI))
6547 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6549 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6550 InsertNewInstBefore(New, AI);
6553 // If the allocation has multiple uses, insert a cast and change all things
6554 // that used it to use the new cast. This will also hack on CI, but it will
6556 if (!AI.hasOneUse()) {
6557 AddUsesToWorkList(AI);
6558 // New is the allocation instruction, pointer typed. AI is the original
6559 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6560 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6561 InsertNewInstBefore(NewCast, AI);
6562 AI.replaceAllUsesWith(NewCast);
6564 return ReplaceInstUsesWith(CI, New);
6567 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6568 /// and return it as type Ty without inserting any new casts and without
6569 /// changing the computed value. This is used by code that tries to decide
6570 /// whether promoting or shrinking integer operations to wider or smaller types
6571 /// will allow us to eliminate a truncate or extend.
6573 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6574 /// extension operation if Ty is larger.
6575 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6576 unsigned CastOpc, int &NumCastsRemoved) {
6577 // We can always evaluate constants in another type.
6578 if (isa<ConstantInt>(V))
6581 Instruction *I = dyn_cast<Instruction>(V);
6582 if (!I) return false;
6584 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6586 // If this is an extension or truncate, we can often eliminate it.
6587 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6588 // If this is a cast from the destination type, we can trivially eliminate
6589 // it, and this will remove a cast overall.
6590 if (I->getOperand(0)->getType() == Ty) {
6591 // If the first operand is itself a cast, and is eliminable, do not count
6592 // this as an eliminable cast. We would prefer to eliminate those two
6594 if (!isa<CastInst>(I->getOperand(0)))
6600 // We can't extend or shrink something that has multiple uses: doing so would
6601 // require duplicating the instruction in general, which isn't profitable.
6602 if (!I->hasOneUse()) return false;
6604 switch (I->getOpcode()) {
6605 case Instruction::Add:
6606 case Instruction::Sub:
6607 case Instruction::And:
6608 case Instruction::Or:
6609 case Instruction::Xor:
6610 // These operators can all arbitrarily be extended or truncated.
6611 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6613 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6616 case Instruction::Mul:
6617 // A multiply can be truncated by truncating its operands.
6618 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6619 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6621 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6624 case Instruction::Shl:
6625 // If we are truncating the result of this SHL, and if it's a shift of a
6626 // constant amount, we can always perform a SHL in a smaller type.
6627 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6628 uint32_t BitWidth = Ty->getBitWidth();
6629 if (BitWidth < OrigTy->getBitWidth() &&
6630 CI->getLimitedValue(BitWidth) < BitWidth)
6631 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6635 case Instruction::LShr:
6636 // If this is a truncate of a logical shr, we can truncate it to a smaller
6637 // lshr iff we know that the bits we would otherwise be shifting in are
6639 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6640 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6641 uint32_t BitWidth = Ty->getBitWidth();
6642 if (BitWidth < OrigBitWidth &&
6643 MaskedValueIsZero(I->getOperand(0),
6644 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6645 CI->getLimitedValue(BitWidth) < BitWidth) {
6646 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6651 case Instruction::ZExt:
6652 case Instruction::SExt:
6653 case Instruction::Trunc:
6654 // If this is the same kind of case as our original (e.g. zext+zext), we
6655 // can safely replace it. Note that replacing it does not reduce the number
6656 // of casts in the input.
6657 if (I->getOpcode() == CastOpc)
6662 // TODO: Can handle more cases here.
6669 /// EvaluateInDifferentType - Given an expression that
6670 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6671 /// evaluate the expression.
6672 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6674 if (Constant *C = dyn_cast<Constant>(V))
6675 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6677 // Otherwise, it must be an instruction.
6678 Instruction *I = cast<Instruction>(V);
6679 Instruction *Res = 0;
6680 switch (I->getOpcode()) {
6681 case Instruction::Add:
6682 case Instruction::Sub:
6683 case Instruction::Mul:
6684 case Instruction::And:
6685 case Instruction::Or:
6686 case Instruction::Xor:
6687 case Instruction::AShr:
6688 case Instruction::LShr:
6689 case Instruction::Shl: {
6690 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6691 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6692 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6693 LHS, RHS, I->getName());
6696 case Instruction::Trunc:
6697 case Instruction::ZExt:
6698 case Instruction::SExt:
6699 // If the source type of the cast is the type we're trying for then we can
6700 // just return the source. There's no need to insert it because it is not
6702 if (I->getOperand(0)->getType() == Ty)
6703 return I->getOperand(0);
6705 // Otherwise, must be the same type of case, so just reinsert a new one.
6706 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6710 // TODO: Can handle more cases here.
6711 assert(0 && "Unreachable!");
6715 return InsertNewInstBefore(Res, *I);
6718 /// @brief Implement the transforms common to all CastInst visitors.
6719 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6720 Value *Src = CI.getOperand(0);
6722 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6723 // eliminate it now.
6724 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6725 if (Instruction::CastOps opc =
6726 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6727 // The first cast (CSrc) is eliminable so we need to fix up or replace
6728 // the second cast (CI). CSrc will then have a good chance of being dead.
6729 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6733 // If we are casting a select then fold the cast into the select
6734 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6735 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6738 // If we are casting a PHI then fold the cast into the PHI
6739 if (isa<PHINode>(Src))
6740 if (Instruction *NV = FoldOpIntoPhi(CI))
6746 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6747 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6748 Value *Src = CI.getOperand(0);
6750 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6751 // If casting the result of a getelementptr instruction with no offset, turn
6752 // this into a cast of the original pointer!
6753 if (GEP->hasAllZeroIndices()) {
6754 // Changing the cast operand is usually not a good idea but it is safe
6755 // here because the pointer operand is being replaced with another
6756 // pointer operand so the opcode doesn't need to change.
6758 CI.setOperand(0, GEP->getOperand(0));
6762 // If the GEP has a single use, and the base pointer is a bitcast, and the
6763 // GEP computes a constant offset, see if we can convert these three
6764 // instructions into fewer. This typically happens with unions and other
6765 // non-type-safe code.
6766 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6767 if (GEP->hasAllConstantIndices()) {
6768 // We are guaranteed to get a constant from EmitGEPOffset.
6769 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6770 int64_t Offset = OffsetV->getSExtValue();
6772 // Get the base pointer input of the bitcast, and the type it points to.
6773 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6774 const Type *GEPIdxTy =
6775 cast<PointerType>(OrigBase->getType())->getElementType();
6776 if (GEPIdxTy->isSized()) {
6777 SmallVector<Value*, 8> NewIndices;
6779 // Start with the index over the outer type. Note that the type size
6780 // might be zero (even if the offset isn't zero) if the indexed type
6781 // is something like [0 x {int, int}]
6782 const Type *IntPtrTy = TD->getIntPtrType();
6783 int64_t FirstIdx = 0;
6784 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6785 FirstIdx = Offset/TySize;
6788 // Handle silly modulus not returning values values [0..TySize).
6792 assert(Offset >= 0);
6794 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6797 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6799 // Index into the types. If we fail, set OrigBase to null.
6801 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6802 const StructLayout *SL = TD->getStructLayout(STy);
6803 if (Offset < (int64_t)SL->getSizeInBytes()) {
6804 unsigned Elt = SL->getElementContainingOffset(Offset);
6805 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6807 Offset -= SL->getElementOffset(Elt);
6808 GEPIdxTy = STy->getElementType(Elt);
6810 // Otherwise, we can't index into this, bail out.
6814 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6815 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6816 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6817 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6820 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6822 GEPIdxTy = STy->getElementType();
6824 // Otherwise, we can't index into this, bail out.
6830 // If we were able to index down into an element, create the GEP
6831 // and bitcast the result. This eliminates one bitcast, potentially
6833 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6835 NewIndices.end(), "");
6836 InsertNewInstBefore(NGEP, CI);
6837 NGEP->takeName(GEP);
6839 if (isa<BitCastInst>(CI))
6840 return new BitCastInst(NGEP, CI.getType());
6841 assert(isa<PtrToIntInst>(CI));
6842 return new PtrToIntInst(NGEP, CI.getType());
6849 return commonCastTransforms(CI);
6854 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6855 /// integer types. This function implements the common transforms for all those
6857 /// @brief Implement the transforms common to CastInst with integer operands
6858 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6859 if (Instruction *Result = commonCastTransforms(CI))
6862 Value *Src = CI.getOperand(0);
6863 const Type *SrcTy = Src->getType();
6864 const Type *DestTy = CI.getType();
6865 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6866 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6868 // See if we can simplify any instructions used by the LHS whose sole
6869 // purpose is to compute bits we don't care about.
6870 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6871 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6872 KnownZero, KnownOne))
6875 // If the source isn't an instruction or has more than one use then we
6876 // can't do anything more.
6877 Instruction *SrcI = dyn_cast<Instruction>(Src);
6878 if (!SrcI || !Src->hasOneUse())
6881 // Attempt to propagate the cast into the instruction for int->int casts.
6882 int NumCastsRemoved = 0;
6883 if (!isa<BitCastInst>(CI) &&
6884 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6885 CI.getOpcode(), NumCastsRemoved)) {
6886 // If this cast is a truncate, evaluting in a different type always
6887 // eliminates the cast, so it is always a win. If this is a zero-extension,
6888 // we need to do an AND to maintain the clear top-part of the computation,
6889 // so we require that the input have eliminated at least one cast. If this
6890 // is a sign extension, we insert two new casts (to do the extension) so we
6891 // require that two casts have been eliminated.
6893 switch (CI.getOpcode()) {
6895 // All the others use floating point so we shouldn't actually
6896 // get here because of the check above.
6897 assert(0 && "Unknown cast type");
6898 case Instruction::Trunc:
6901 case Instruction::ZExt:
6902 DoXForm = NumCastsRemoved >= 1;
6904 case Instruction::SExt:
6905 DoXForm = NumCastsRemoved >= 2;
6910 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6911 CI.getOpcode() == Instruction::SExt);
6912 assert(Res->getType() == DestTy);
6913 switch (CI.getOpcode()) {
6914 default: assert(0 && "Unknown cast type!");
6915 case Instruction::Trunc:
6916 case Instruction::BitCast:
6917 // Just replace this cast with the result.
6918 return ReplaceInstUsesWith(CI, Res);
6919 case Instruction::ZExt: {
6920 // We need to emit an AND to clear the high bits.
6921 assert(SrcBitSize < DestBitSize && "Not a zext?");
6922 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6924 return BinaryOperator::createAnd(Res, C);
6926 case Instruction::SExt:
6927 // We need to emit a cast to truncate, then a cast to sext.
6928 return CastInst::create(Instruction::SExt,
6929 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6935 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6936 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6938 switch (SrcI->getOpcode()) {
6939 case Instruction::Add:
6940 case Instruction::Mul:
6941 case Instruction::And:
6942 case Instruction::Or:
6943 case Instruction::Xor:
6944 // If we are discarding information, rewrite.
6945 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6946 // Don't insert two casts if they cannot be eliminated. We allow
6947 // two casts to be inserted if the sizes are the same. This could
6948 // only be converting signedness, which is a noop.
6949 if (DestBitSize == SrcBitSize ||
6950 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6951 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6952 Instruction::CastOps opcode = CI.getOpcode();
6953 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6954 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6955 return BinaryOperator::create(
6956 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6960 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6961 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6962 SrcI->getOpcode() == Instruction::Xor &&
6963 Op1 == ConstantInt::getTrue() &&
6964 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6965 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6966 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6969 case Instruction::SDiv:
6970 case Instruction::UDiv:
6971 case Instruction::SRem:
6972 case Instruction::URem:
6973 // If we are just changing the sign, rewrite.
6974 if (DestBitSize == SrcBitSize) {
6975 // Don't insert two casts if they cannot be eliminated. We allow
6976 // two casts to be inserted if the sizes are the same. This could
6977 // only be converting signedness, which is a noop.
6978 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6979 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6980 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6982 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6984 return BinaryOperator::create(
6985 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6990 case Instruction::Shl:
6991 // Allow changing the sign of the source operand. Do not allow
6992 // changing the size of the shift, UNLESS the shift amount is a
6993 // constant. We must not change variable sized shifts to a smaller
6994 // size, because it is undefined to shift more bits out than exist
6996 if (DestBitSize == SrcBitSize ||
6997 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6998 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6999 Instruction::BitCast : Instruction::Trunc);
7000 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7001 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7002 return BinaryOperator::createShl(Op0c, Op1c);
7005 case Instruction::AShr:
7006 // If this is a signed shr, and if all bits shifted in are about to be
7007 // truncated off, turn it into an unsigned shr to allow greater
7009 if (DestBitSize < SrcBitSize &&
7010 isa<ConstantInt>(Op1)) {
7011 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7012 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7013 // Insert the new logical shift right.
7014 return BinaryOperator::createLShr(Op0, Op1);
7022 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7023 if (Instruction *Result = commonIntCastTransforms(CI))
7026 Value *Src = CI.getOperand(0);
7027 const Type *Ty = CI.getType();
7028 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7029 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7031 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7032 switch (SrcI->getOpcode()) {
7034 case Instruction::LShr:
7035 // We can shrink lshr to something smaller if we know the bits shifted in
7036 // are already zeros.
7037 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7038 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7040 // Get a mask for the bits shifting in.
7041 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7042 Value* SrcIOp0 = SrcI->getOperand(0);
7043 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7044 if (ShAmt >= DestBitWidth) // All zeros.
7045 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7047 // Okay, we can shrink this. Truncate the input, then return a new
7049 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7050 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7052 return BinaryOperator::createLShr(V1, V2);
7054 } else { // This is a variable shr.
7056 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7057 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7058 // loop-invariant and CSE'd.
7059 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7060 Value *One = ConstantInt::get(SrcI->getType(), 1);
7062 Value *V = InsertNewInstBefore(
7063 BinaryOperator::createShl(One, SrcI->getOperand(1),
7065 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7066 SrcI->getOperand(0),
7068 Value *Zero = Constant::getNullValue(V->getType());
7069 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7079 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7080 // If one of the common conversion will work ..
7081 if (Instruction *Result = commonIntCastTransforms(CI))
7084 Value *Src = CI.getOperand(0);
7086 // If this is a cast of a cast
7087 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7088 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7089 // types and if the sizes are just right we can convert this into a logical
7090 // 'and' which will be much cheaper than the pair of casts.
7091 if (isa<TruncInst>(CSrc)) {
7092 // Get the sizes of the types involved
7093 Value *A = CSrc->getOperand(0);
7094 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7095 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7096 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7097 // If we're actually extending zero bits and the trunc is a no-op
7098 if (MidSize < DstSize && SrcSize == DstSize) {
7099 // Replace both of the casts with an And of the type mask.
7100 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7101 Constant *AndConst = ConstantInt::get(AndValue);
7103 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7104 // Unfortunately, if the type changed, we need to cast it back.
7105 if (And->getType() != CI.getType()) {
7106 And->setName(CSrc->getName()+".mask");
7107 InsertNewInstBefore(And, CI);
7108 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7115 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7116 // If we are just checking for a icmp eq of a single bit and zext'ing it
7117 // to an integer, then shift the bit to the appropriate place and then
7118 // cast to integer to avoid the comparison.
7119 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7120 const APInt &Op1CV = Op1C->getValue();
7122 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7123 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7124 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7125 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7126 Value *In = ICI->getOperand(0);
7127 Value *Sh = ConstantInt::get(In->getType(),
7128 In->getType()->getPrimitiveSizeInBits()-1);
7129 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7130 In->getName()+".lobit"),
7132 if (In->getType() != CI.getType())
7133 In = CastInst::createIntegerCast(In, CI.getType(),
7134 false/*ZExt*/, "tmp", &CI);
7136 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7137 Constant *One = ConstantInt::get(In->getType(), 1);
7138 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7139 In->getName()+".not"),
7143 return ReplaceInstUsesWith(CI, In);
7148 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7149 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7150 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7151 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7152 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7153 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7154 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7155 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7156 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7157 // This only works for EQ and NE
7158 ICI->isEquality()) {
7159 // If Op1C some other power of two, convert:
7160 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7161 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7162 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7163 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7165 APInt KnownZeroMask(~KnownZero);
7166 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7167 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7168 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7169 // (X&4) == 2 --> false
7170 // (X&4) != 2 --> true
7171 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7172 Res = ConstantExpr::getZExt(Res, CI.getType());
7173 return ReplaceInstUsesWith(CI, Res);
7176 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7177 Value *In = ICI->getOperand(0);
7179 // Perform a logical shr by shiftamt.
7180 // Insert the shift to put the result in the low bit.
7181 In = InsertNewInstBefore(
7182 BinaryOperator::createLShr(In,
7183 ConstantInt::get(In->getType(), ShiftAmt),
7184 In->getName()+".lobit"), CI);
7187 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7188 Constant *One = ConstantInt::get(In->getType(), 1);
7189 In = BinaryOperator::createXor(In, One, "tmp");
7190 InsertNewInstBefore(cast<Instruction>(In), CI);
7193 if (CI.getType() == In->getType())
7194 return ReplaceInstUsesWith(CI, In);
7196 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7204 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7205 if (Instruction *I = commonIntCastTransforms(CI))
7208 Value *Src = CI.getOperand(0);
7210 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7211 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7212 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7213 // If we are just checking for a icmp eq of a single bit and zext'ing it
7214 // to an integer, then shift the bit to the appropriate place and then
7215 // cast to integer to avoid the comparison.
7216 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7217 const APInt &Op1CV = Op1C->getValue();
7219 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7220 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7221 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7222 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7223 Value *In = ICI->getOperand(0);
7224 Value *Sh = ConstantInt::get(In->getType(),
7225 In->getType()->getPrimitiveSizeInBits()-1);
7226 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7227 In->getName()+".lobit"),
7229 if (In->getType() != CI.getType())
7230 In = CastInst::createIntegerCast(In, CI.getType(),
7231 true/*SExt*/, "tmp", &CI);
7233 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7234 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7235 In->getName()+".not"), CI);
7237 return ReplaceInstUsesWith(CI, In);
7245 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7246 /// in the specified FP type without changing its value.
7247 static Constant *FitsInFPType(ConstantFP *CFP, const Type *FPTy,
7248 const fltSemantics &Sem) {
7249 APFloat F = CFP->getValueAPF();
7250 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7251 return ConstantFP::get(FPTy, F);
7255 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7256 /// through it until we get the source value.
7257 static Value *LookThroughFPExtensions(Value *V) {
7258 if (Instruction *I = dyn_cast<Instruction>(V))
7259 if (I->getOpcode() == Instruction::FPExt)
7260 return LookThroughFPExtensions(I->getOperand(0));
7262 // If this value is a constant, return the constant in the smallest FP type
7263 // that can accurately represent it. This allows us to turn
7264 // (float)((double)X+2.0) into x+2.0f.
7265 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7266 if (CFP->getType() == Type::PPC_FP128Ty)
7267 return V; // No constant folding of this.
7268 // See if the value can be truncated to float and then reextended.
7269 if (Value *V = FitsInFPType(CFP, Type::FloatTy, APFloat::IEEEsingle))
7271 if (CFP->getType() == Type::DoubleTy)
7272 return V; // Won't shrink.
7273 if (Value *V = FitsInFPType(CFP, Type::DoubleTy, APFloat::IEEEdouble))
7275 // Don't try to shrink to various long double types.
7281 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7282 if (Instruction *I = commonCastTransforms(CI))
7285 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7286 // smaller than the destination type, we can eliminate the truncate by doing
7287 // the add as the smaller type. This applies to add/sub/mul/div as well as
7288 // many builtins (sqrt, etc).
7289 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7290 if (OpI && OpI->hasOneUse()) {
7291 switch (OpI->getOpcode()) {
7293 case Instruction::Add:
7294 case Instruction::Sub:
7295 case Instruction::Mul:
7296 case Instruction::FDiv:
7297 case Instruction::FRem:
7298 const Type *SrcTy = OpI->getType();
7299 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7300 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7301 if (LHSTrunc->getType() != SrcTy &&
7302 RHSTrunc->getType() != SrcTy) {
7303 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7304 // If the source types were both smaller than the destination type of
7305 // the cast, do this xform.
7306 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7307 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7308 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7310 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7312 return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7321 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7322 return commonCastTransforms(CI);
7325 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7326 return commonCastTransforms(CI);
7329 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7330 return commonCastTransforms(CI);
7333 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7334 return commonCastTransforms(CI);
7337 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7338 return commonCastTransforms(CI);
7341 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7342 return commonPointerCastTransforms(CI);
7345 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7346 if (Instruction *I = commonCastTransforms(CI))
7349 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7350 if (!DestPointee->isSized()) return 0;
7352 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7355 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7356 m_ConstantInt(Cst)))) {
7357 // If the source and destination operands have the same type, see if this
7358 // is a single-index GEP.
7359 if (X->getType() == CI.getType()) {
7360 // Get the size of the pointee type.
7361 uint64_t Size = TD->getABITypeSizeInBits(DestPointee);
7363 // Convert the constant to intptr type.
7364 APInt Offset = Cst->getValue();
7365 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7367 // If Offset is evenly divisible by Size, we can do this xform.
7368 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7369 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7370 return new GetElementPtrInst(X, ConstantInt::get(Offset));
7373 // TODO: Could handle other cases, e.g. where add is indexing into field of
7375 } else if (CI.getOperand(0)->hasOneUse() &&
7376 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7377 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7378 // "inttoptr+GEP" instead of "add+intptr".
7380 // Get the size of the pointee type.
7381 uint64_t Size = TD->getABITypeSize(DestPointee);
7383 // Convert the constant to intptr type.
7384 APInt Offset = Cst->getValue();
7385 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7387 // If Offset is evenly divisible by Size, we can do this xform.
7388 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7389 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7391 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7393 return new GetElementPtrInst(P, ConstantInt::get(Offset), "tmp");
7399 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7400 // If the operands are integer typed then apply the integer transforms,
7401 // otherwise just apply the common ones.
7402 Value *Src = CI.getOperand(0);
7403 const Type *SrcTy = Src->getType();
7404 const Type *DestTy = CI.getType();
7406 if (SrcTy->isInteger() && DestTy->isInteger()) {
7407 if (Instruction *Result = commonIntCastTransforms(CI))
7409 } else if (isa<PointerType>(SrcTy)) {
7410 if (Instruction *I = commonPointerCastTransforms(CI))
7413 if (Instruction *Result = commonCastTransforms(CI))
7418 // Get rid of casts from one type to the same type. These are useless and can
7419 // be replaced by the operand.
7420 if (DestTy == Src->getType())
7421 return ReplaceInstUsesWith(CI, Src);
7423 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7424 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7425 const Type *DstElTy = DstPTy->getElementType();
7426 const Type *SrcElTy = SrcPTy->getElementType();
7428 // If we are casting a malloc or alloca to a pointer to a type of the same
7429 // size, rewrite the allocation instruction to allocate the "right" type.
7430 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7431 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7434 // If the source and destination are pointers, and this cast is equivalent
7435 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7436 // This can enhance SROA and other transforms that want type-safe pointers.
7437 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7438 unsigned NumZeros = 0;
7439 while (SrcElTy != DstElTy &&
7440 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7441 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7442 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7446 // If we found a path from the src to dest, create the getelementptr now.
7447 if (SrcElTy == DstElTy) {
7448 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7449 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7450 ((Instruction*) NULL));
7454 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7455 if (SVI->hasOneUse()) {
7456 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7457 // a bitconvert to a vector with the same # elts.
7458 if (isa<VectorType>(DestTy) &&
7459 cast<VectorType>(DestTy)->getNumElements() ==
7460 SVI->getType()->getNumElements()) {
7462 // If either of the operands is a cast from CI.getType(), then
7463 // evaluating the shuffle in the casted destination's type will allow
7464 // us to eliminate at least one cast.
7465 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7466 Tmp->getOperand(0)->getType() == DestTy) ||
7467 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7468 Tmp->getOperand(0)->getType() == DestTy)) {
7469 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7470 SVI->getOperand(0), DestTy, &CI);
7471 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7472 SVI->getOperand(1), DestTy, &CI);
7473 // Return a new shuffle vector. Use the same element ID's, as we
7474 // know the vector types match #elts.
7475 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7483 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7485 /// %D = select %cond, %C, %A
7487 /// %C = select %cond, %B, 0
7490 /// Assuming that the specified instruction is an operand to the select, return
7491 /// a bitmask indicating which operands of this instruction are foldable if they
7492 /// equal the other incoming value of the select.
7494 static unsigned GetSelectFoldableOperands(Instruction *I) {
7495 switch (I->getOpcode()) {
7496 case Instruction::Add:
7497 case Instruction::Mul:
7498 case Instruction::And:
7499 case Instruction::Or:
7500 case Instruction::Xor:
7501 return 3; // Can fold through either operand.
7502 case Instruction::Sub: // Can only fold on the amount subtracted.
7503 case Instruction::Shl: // Can only fold on the shift amount.
7504 case Instruction::LShr:
7505 case Instruction::AShr:
7508 return 0; // Cannot fold
7512 /// GetSelectFoldableConstant - For the same transformation as the previous
7513 /// function, return the identity constant that goes into the select.
7514 static Constant *GetSelectFoldableConstant(Instruction *I) {
7515 switch (I->getOpcode()) {
7516 default: assert(0 && "This cannot happen!"); abort();
7517 case Instruction::Add:
7518 case Instruction::Sub:
7519 case Instruction::Or:
7520 case Instruction::Xor:
7521 case Instruction::Shl:
7522 case Instruction::LShr:
7523 case Instruction::AShr:
7524 return Constant::getNullValue(I->getType());
7525 case Instruction::And:
7526 return Constant::getAllOnesValue(I->getType());
7527 case Instruction::Mul:
7528 return ConstantInt::get(I->getType(), 1);
7532 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7533 /// have the same opcode and only one use each. Try to simplify this.
7534 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7536 if (TI->getNumOperands() == 1) {
7537 // If this is a non-volatile load or a cast from the same type,
7540 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7543 return 0; // unknown unary op.
7546 // Fold this by inserting a select from the input values.
7547 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7548 FI->getOperand(0), SI.getName()+".v");
7549 InsertNewInstBefore(NewSI, SI);
7550 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7554 // Only handle binary operators here.
7555 if (!isa<BinaryOperator>(TI))
7558 // Figure out if the operations have any operands in common.
7559 Value *MatchOp, *OtherOpT, *OtherOpF;
7561 if (TI->getOperand(0) == FI->getOperand(0)) {
7562 MatchOp = TI->getOperand(0);
7563 OtherOpT = TI->getOperand(1);
7564 OtherOpF = FI->getOperand(1);
7565 MatchIsOpZero = true;
7566 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7567 MatchOp = TI->getOperand(1);
7568 OtherOpT = TI->getOperand(0);
7569 OtherOpF = FI->getOperand(0);
7570 MatchIsOpZero = false;
7571 } else if (!TI->isCommutative()) {
7573 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7574 MatchOp = TI->getOperand(0);
7575 OtherOpT = TI->getOperand(1);
7576 OtherOpF = FI->getOperand(0);
7577 MatchIsOpZero = true;
7578 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7579 MatchOp = TI->getOperand(1);
7580 OtherOpT = TI->getOperand(0);
7581 OtherOpF = FI->getOperand(1);
7582 MatchIsOpZero = true;
7587 // If we reach here, they do have operations in common.
7588 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7589 OtherOpF, SI.getName()+".v");
7590 InsertNewInstBefore(NewSI, SI);
7592 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7594 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7596 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7598 assert(0 && "Shouldn't get here");
7602 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7603 Value *CondVal = SI.getCondition();
7604 Value *TrueVal = SI.getTrueValue();
7605 Value *FalseVal = SI.getFalseValue();
7607 // select true, X, Y -> X
7608 // select false, X, Y -> Y
7609 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7610 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7612 // select C, X, X -> X
7613 if (TrueVal == FalseVal)
7614 return ReplaceInstUsesWith(SI, TrueVal);
7616 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7617 return ReplaceInstUsesWith(SI, FalseVal);
7618 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7619 return ReplaceInstUsesWith(SI, TrueVal);
7620 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7621 if (isa<Constant>(TrueVal))
7622 return ReplaceInstUsesWith(SI, TrueVal);
7624 return ReplaceInstUsesWith(SI, FalseVal);
7627 if (SI.getType() == Type::Int1Ty) {
7628 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7629 if (C->getZExtValue()) {
7630 // Change: A = select B, true, C --> A = or B, C
7631 return BinaryOperator::createOr(CondVal, FalseVal);
7633 // Change: A = select B, false, C --> A = and !B, C
7635 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7636 "not."+CondVal->getName()), SI);
7637 return BinaryOperator::createAnd(NotCond, FalseVal);
7639 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7640 if (C->getZExtValue() == false) {
7641 // Change: A = select B, C, false --> A = and B, C
7642 return BinaryOperator::createAnd(CondVal, TrueVal);
7644 // Change: A = select B, C, true --> A = or !B, C
7646 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7647 "not."+CondVal->getName()), SI);
7648 return BinaryOperator::createOr(NotCond, TrueVal);
7652 // select a, b, a -> a&b
7653 // select a, a, b -> a|b
7654 if (CondVal == TrueVal)
7655 return BinaryOperator::createOr(CondVal, FalseVal);
7656 else if (CondVal == FalseVal)
7657 return BinaryOperator::createAnd(CondVal, TrueVal);
7660 // Selecting between two integer constants?
7661 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7662 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7663 // select C, 1, 0 -> zext C to int
7664 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7665 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7666 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7667 // select C, 0, 1 -> zext !C to int
7669 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7670 "not."+CondVal->getName()), SI);
7671 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7674 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7676 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7678 // (x <s 0) ? -1 : 0 -> ashr x, 31
7679 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7680 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7681 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7682 // The comparison constant and the result are not neccessarily the
7683 // same width. Make an all-ones value by inserting a AShr.
7684 Value *X = IC->getOperand(0);
7685 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7686 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7687 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7689 InsertNewInstBefore(SRA, SI);
7691 // Finally, convert to the type of the select RHS. We figure out
7692 // if this requires a SExt, Trunc or BitCast based on the sizes.
7693 Instruction::CastOps opc = Instruction::BitCast;
7694 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7695 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7696 if (SRASize < SISize)
7697 opc = Instruction::SExt;
7698 else if (SRASize > SISize)
7699 opc = Instruction::Trunc;
7700 return CastInst::create(opc, SRA, SI.getType());
7705 // If one of the constants is zero (we know they can't both be) and we
7706 // have an icmp instruction with zero, and we have an 'and' with the
7707 // non-constant value, eliminate this whole mess. This corresponds to
7708 // cases like this: ((X & 27) ? 27 : 0)
7709 if (TrueValC->isZero() || FalseValC->isZero())
7710 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7711 cast<Constant>(IC->getOperand(1))->isNullValue())
7712 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7713 if (ICA->getOpcode() == Instruction::And &&
7714 isa<ConstantInt>(ICA->getOperand(1)) &&
7715 (ICA->getOperand(1) == TrueValC ||
7716 ICA->getOperand(1) == FalseValC) &&
7717 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7718 // Okay, now we know that everything is set up, we just don't
7719 // know whether we have a icmp_ne or icmp_eq and whether the
7720 // true or false val is the zero.
7721 bool ShouldNotVal = !TrueValC->isZero();
7722 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7725 V = InsertNewInstBefore(BinaryOperator::create(
7726 Instruction::Xor, V, ICA->getOperand(1)), SI);
7727 return ReplaceInstUsesWith(SI, V);
7732 // See if we are selecting two values based on a comparison of the two values.
7733 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7734 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7735 // Transform (X == Y) ? X : Y -> Y
7736 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7737 // This is not safe in general for floating point:
7738 // consider X== -0, Y== +0.
7739 // It becomes safe if either operand is a nonzero constant.
7740 ConstantFP *CFPt, *CFPf;
7741 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7742 !CFPt->getValueAPF().isZero()) ||
7743 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7744 !CFPf->getValueAPF().isZero()))
7745 return ReplaceInstUsesWith(SI, FalseVal);
7747 // Transform (X != Y) ? X : Y -> X
7748 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7749 return ReplaceInstUsesWith(SI, TrueVal);
7750 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7752 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7753 // Transform (X == Y) ? Y : X -> X
7754 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7755 // This is not safe in general for floating point:
7756 // consider X== -0, Y== +0.
7757 // It becomes safe if either operand is a nonzero constant.
7758 ConstantFP *CFPt, *CFPf;
7759 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7760 !CFPt->getValueAPF().isZero()) ||
7761 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7762 !CFPf->getValueAPF().isZero()))
7763 return ReplaceInstUsesWith(SI, FalseVal);
7765 // Transform (X != Y) ? Y : X -> Y
7766 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7767 return ReplaceInstUsesWith(SI, TrueVal);
7768 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7772 // See if we are selecting two values based on a comparison of the two values.
7773 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7774 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7775 // Transform (X == Y) ? X : Y -> Y
7776 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7777 return ReplaceInstUsesWith(SI, FalseVal);
7778 // Transform (X != Y) ? X : Y -> X
7779 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7780 return ReplaceInstUsesWith(SI, TrueVal);
7781 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7783 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7784 // Transform (X == Y) ? Y : X -> X
7785 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7786 return ReplaceInstUsesWith(SI, FalseVal);
7787 // Transform (X != Y) ? Y : X -> Y
7788 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7789 return ReplaceInstUsesWith(SI, TrueVal);
7790 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7794 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7795 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7796 if (TI->hasOneUse() && FI->hasOneUse()) {
7797 Instruction *AddOp = 0, *SubOp = 0;
7799 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7800 if (TI->getOpcode() == FI->getOpcode())
7801 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7804 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7805 // even legal for FP.
7806 if (TI->getOpcode() == Instruction::Sub &&
7807 FI->getOpcode() == Instruction::Add) {
7808 AddOp = FI; SubOp = TI;
7809 } else if (FI->getOpcode() == Instruction::Sub &&
7810 TI->getOpcode() == Instruction::Add) {
7811 AddOp = TI; SubOp = FI;
7815 Value *OtherAddOp = 0;
7816 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7817 OtherAddOp = AddOp->getOperand(1);
7818 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7819 OtherAddOp = AddOp->getOperand(0);
7823 // So at this point we know we have (Y -> OtherAddOp):
7824 // select C, (add X, Y), (sub X, Z)
7825 Value *NegVal; // Compute -Z
7826 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7827 NegVal = ConstantExpr::getNeg(C);
7829 NegVal = InsertNewInstBefore(
7830 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7833 Value *NewTrueOp = OtherAddOp;
7834 Value *NewFalseOp = NegVal;
7836 std::swap(NewTrueOp, NewFalseOp);
7837 Instruction *NewSel =
7838 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7840 NewSel = InsertNewInstBefore(NewSel, SI);
7841 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7846 // See if we can fold the select into one of our operands.
7847 if (SI.getType()->isInteger()) {
7848 // See the comment above GetSelectFoldableOperands for a description of the
7849 // transformation we are doing here.
7850 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7851 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7852 !isa<Constant>(FalseVal))
7853 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7854 unsigned OpToFold = 0;
7855 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7857 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7862 Constant *C = GetSelectFoldableConstant(TVI);
7863 Instruction *NewSel =
7864 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7865 InsertNewInstBefore(NewSel, SI);
7866 NewSel->takeName(TVI);
7867 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7868 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7870 assert(0 && "Unknown instruction!!");
7875 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7876 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7877 !isa<Constant>(TrueVal))
7878 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7879 unsigned OpToFold = 0;
7880 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7882 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7887 Constant *C = GetSelectFoldableConstant(FVI);
7888 Instruction *NewSel =
7889 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7890 InsertNewInstBefore(NewSel, SI);
7891 NewSel->takeName(FVI);
7892 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7893 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7895 assert(0 && "Unknown instruction!!");
7900 if (BinaryOperator::isNot(CondVal)) {
7901 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7902 SI.setOperand(1, FalseVal);
7903 SI.setOperand(2, TrueVal);
7910 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7911 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7912 /// and it is more than the alignment of the ultimate object, see if we can
7913 /// increase the alignment of the ultimate object, making this check succeed.
7914 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7915 unsigned PrefAlign = 0) {
7916 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7917 unsigned Align = GV->getAlignment();
7918 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7919 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7921 // If there is a large requested alignment and we can, bump up the alignment
7923 if (PrefAlign > Align && GV->hasInitializer()) {
7924 GV->setAlignment(PrefAlign);
7928 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7929 unsigned Align = AI->getAlignment();
7930 if (Align == 0 && TD) {
7931 if (isa<AllocaInst>(AI))
7932 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7933 else if (isa<MallocInst>(AI)) {
7934 // Malloc returns maximally aligned memory.
7935 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7938 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7941 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7945 // If there is a requested alignment and if this is an alloca, round up. We
7946 // don't do this for malloc, because some systems can't respect the request.
7947 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7948 AI->setAlignment(PrefAlign);
7952 } else if (isa<BitCastInst>(V) ||
7953 (isa<ConstantExpr>(V) &&
7954 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7955 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7957 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7958 // If all indexes are zero, it is just the alignment of the base pointer.
7959 bool AllZeroOperands = true;
7960 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7961 if (!isa<Constant>(GEPI->getOperand(i)) ||
7962 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7963 AllZeroOperands = false;
7967 if (AllZeroOperands) {
7968 // Treat this like a bitcast.
7969 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7972 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7973 if (BaseAlignment == 0) return 0;
7975 // Otherwise, if the base alignment is >= the alignment we expect for the
7976 // base pointer type, then we know that the resultant pointer is aligned at
7977 // least as much as its type requires.
7980 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7981 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7982 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7983 if (Align <= BaseAlignment) {
7984 const Type *GEPTy = GEPI->getType();
7985 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7986 Align = std::min(Align, (unsigned)
7987 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7995 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
7996 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7997 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7998 unsigned MinAlign = std::min(DstAlign, SrcAlign);
7999 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8001 if (CopyAlign < MinAlign) {
8002 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8006 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8008 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8009 if (MemOpLength == 0) return 0;
8011 // Source and destination pointer types are always "i8*" for intrinsic. See
8012 // if the size is something we can handle with a single primitive load/store.
8013 // A single load+store correctly handles overlapping memory in the memmove
8015 unsigned Size = MemOpLength->getZExtValue();
8016 if (Size == 0 || Size > 8 || (Size&(Size-1)))
8017 return 0; // If not 1/2/4/8 bytes, exit.
8019 // Use an integer load+store unless we can find something better.
8020 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8022 // Memcpy forces the use of i8* for the source and destination. That means
8023 // that if you're using memcpy to move one double around, you'll get a cast
8024 // from double* to i8*. We'd much rather use a double load+store rather than
8025 // an i64 load+store, here because this improves the odds that the source or
8026 // dest address will be promotable. See if we can find a better type than the
8027 // integer datatype.
8028 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8029 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8030 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8031 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8032 // down through these levels if so.
8033 while (!SrcETy->isFirstClassType()) {
8034 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8035 if (STy->getNumElements() == 1)
8036 SrcETy = STy->getElementType(0);
8039 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8040 if (ATy->getNumElements() == 1)
8041 SrcETy = ATy->getElementType();
8048 if (SrcETy->isFirstClassType())
8049 NewPtrTy = PointerType::getUnqual(SrcETy);
8054 // If the memcpy/memmove provides better alignment info than we can
8056 SrcAlign = std::max(SrcAlign, CopyAlign);
8057 DstAlign = std::max(DstAlign, CopyAlign);
8059 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8060 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8061 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8062 InsertNewInstBefore(L, *MI);
8063 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8065 // Set the size of the copy to 0, it will be deleted on the next iteration.
8066 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8070 /// visitCallInst - CallInst simplification. This mostly only handles folding
8071 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8072 /// the heavy lifting.
8074 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8075 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8076 if (!II) return visitCallSite(&CI);
8078 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8080 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8081 bool Changed = false;
8083 // memmove/cpy/set of zero bytes is a noop.
8084 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8085 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8087 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8088 if (CI->getZExtValue() == 1) {
8089 // Replace the instruction with just byte operations. We would
8090 // transform other cases to loads/stores, but we don't know if
8091 // alignment is sufficient.
8095 // If we have a memmove and the source operation is a constant global,
8096 // then the source and dest pointers can't alias, so we can change this
8097 // into a call to memcpy.
8098 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8099 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8100 if (GVSrc->isConstant()) {
8101 Module *M = CI.getParent()->getParent()->getParent();
8102 Intrinsic::ID MemCpyID;
8103 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8104 MemCpyID = Intrinsic::memcpy_i32;
8106 MemCpyID = Intrinsic::memcpy_i64;
8107 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8112 // If we can determine a pointer alignment that is bigger than currently
8113 // set, update the alignment.
8114 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8115 if (Instruction *I = SimplifyMemTransfer(MI))
8117 } else if (isa<MemSetInst>(MI)) {
8118 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
8119 if (MI->getAlignment()->getZExtValue() < Alignment) {
8120 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8125 if (Changed) return II;
8127 switch (II->getIntrinsicID()) {
8129 case Intrinsic::ppc_altivec_lvx:
8130 case Intrinsic::ppc_altivec_lvxl:
8131 case Intrinsic::x86_sse_loadu_ps:
8132 case Intrinsic::x86_sse2_loadu_pd:
8133 case Intrinsic::x86_sse2_loadu_dq:
8134 // Turn PPC lvx -> load if the pointer is known aligned.
8135 // Turn X86 loadups -> load if the pointer is known aligned.
8136 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8137 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8138 PointerType::getUnqual(II->getType()),
8140 return new LoadInst(Ptr);
8143 case Intrinsic::ppc_altivec_stvx:
8144 case Intrinsic::ppc_altivec_stvxl:
8145 // Turn stvx -> store if the pointer is known aligned.
8146 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
8147 const Type *OpPtrTy =
8148 PointerType::getUnqual(II->getOperand(1)->getType());
8149 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8150 return new StoreInst(II->getOperand(1), Ptr);
8153 case Intrinsic::x86_sse_storeu_ps:
8154 case Intrinsic::x86_sse2_storeu_pd:
8155 case Intrinsic::x86_sse2_storeu_dq:
8156 case Intrinsic::x86_sse2_storel_dq:
8157 // Turn X86 storeu -> store if the pointer is known aligned.
8158 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8159 const Type *OpPtrTy =
8160 PointerType::getUnqual(II->getOperand(2)->getType());
8161 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8162 return new StoreInst(II->getOperand(2), Ptr);
8166 case Intrinsic::x86_sse_cvttss2si: {
8167 // These intrinsics only demands the 0th element of its input vector. If
8168 // we can simplify the input based on that, do so now.
8170 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8172 II->setOperand(1, V);
8178 case Intrinsic::ppc_altivec_vperm:
8179 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8180 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8181 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8183 // Check that all of the elements are integer constants or undefs.
8184 bool AllEltsOk = true;
8185 for (unsigned i = 0; i != 16; ++i) {
8186 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8187 !isa<UndefValue>(Mask->getOperand(i))) {
8194 // Cast the input vectors to byte vectors.
8195 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8196 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8197 Value *Result = UndefValue::get(Op0->getType());
8199 // Only extract each element once.
8200 Value *ExtractedElts[32];
8201 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8203 for (unsigned i = 0; i != 16; ++i) {
8204 if (isa<UndefValue>(Mask->getOperand(i)))
8206 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8207 Idx &= 31; // Match the hardware behavior.
8209 if (ExtractedElts[Idx] == 0) {
8211 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8212 InsertNewInstBefore(Elt, CI);
8213 ExtractedElts[Idx] = Elt;
8216 // Insert this value into the result vector.
8217 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
8218 InsertNewInstBefore(cast<Instruction>(Result), CI);
8220 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8225 case Intrinsic::stackrestore: {
8226 // If the save is right next to the restore, remove the restore. This can
8227 // happen when variable allocas are DCE'd.
8228 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8229 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8230 BasicBlock::iterator BI = SS;
8232 return EraseInstFromFunction(CI);
8236 // Scan down this block to see if there is another stack restore in the
8237 // same block without an intervening call/alloca.
8238 BasicBlock::iterator BI = II;
8239 TerminatorInst *TI = II->getParent()->getTerminator();
8240 bool CannotRemove = false;
8241 for (++BI; &*BI != TI; ++BI) {
8242 if (isa<AllocaInst>(BI)) {
8243 CannotRemove = true;
8246 if (isa<CallInst>(BI)) {
8247 if (!isa<IntrinsicInst>(BI)) {
8248 CannotRemove = true;
8251 // If there is a stackrestore below this one, remove this one.
8252 return EraseInstFromFunction(CI);
8256 // If the stack restore is in a return/unwind block and if there are no
8257 // allocas or calls between the restore and the return, nuke the restore.
8258 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8259 return EraseInstFromFunction(CI);
8265 return visitCallSite(II);
8268 // InvokeInst simplification
8270 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8271 return visitCallSite(&II);
8274 // visitCallSite - Improvements for call and invoke instructions.
8276 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8277 bool Changed = false;
8279 // If the callee is a constexpr cast of a function, attempt to move the cast
8280 // to the arguments of the call/invoke.
8281 if (transformConstExprCastCall(CS)) return 0;
8283 Value *Callee = CS.getCalledValue();
8285 if (Function *CalleeF = dyn_cast<Function>(Callee))
8286 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8287 Instruction *OldCall = CS.getInstruction();
8288 // If the call and callee calling conventions don't match, this call must
8289 // be unreachable, as the call is undefined.
8290 new StoreInst(ConstantInt::getTrue(),
8291 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8293 if (!OldCall->use_empty())
8294 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8295 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8296 return EraseInstFromFunction(*OldCall);
8300 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8301 // This instruction is not reachable, just remove it. We insert a store to
8302 // undef so that we know that this code is not reachable, despite the fact
8303 // that we can't modify the CFG here.
8304 new StoreInst(ConstantInt::getTrue(),
8305 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8306 CS.getInstruction());
8308 if (!CS.getInstruction()->use_empty())
8309 CS.getInstruction()->
8310 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8312 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8313 // Don't break the CFG, insert a dummy cond branch.
8314 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8315 ConstantInt::getTrue(), II);
8317 return EraseInstFromFunction(*CS.getInstruction());
8320 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8321 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8322 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8323 return transformCallThroughTrampoline(CS);
8325 const PointerType *PTy = cast<PointerType>(Callee->getType());
8326 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8327 if (FTy->isVarArg()) {
8328 // See if we can optimize any arguments passed through the varargs area of
8330 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8331 E = CS.arg_end(); I != E; ++I)
8332 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8333 // If this cast does not effect the value passed through the varargs
8334 // area, we can eliminate the use of the cast.
8335 Value *Op = CI->getOperand(0);
8336 if (CI->isLosslessCast()) {
8343 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8344 // Inline asm calls cannot throw - mark them 'nounwind'.
8345 CS.setDoesNotThrow();
8349 return Changed ? CS.getInstruction() : 0;
8352 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8353 // attempt to move the cast to the arguments of the call/invoke.
8355 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8356 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8357 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8358 if (CE->getOpcode() != Instruction::BitCast ||
8359 !isa<Function>(CE->getOperand(0)))
8361 Function *Callee = cast<Function>(CE->getOperand(0));
8362 Instruction *Caller = CS.getInstruction();
8363 const ParamAttrsList* CallerPAL = CS.getParamAttrs();
8365 // Okay, this is a cast from a function to a different type. Unless doing so
8366 // would cause a type conversion of one of our arguments, change this call to
8367 // be a direct call with arguments casted to the appropriate types.
8369 const FunctionType *FT = Callee->getFunctionType();
8370 const Type *OldRetTy = Caller->getType();
8372 // Check to see if we are changing the return type...
8373 if (OldRetTy != FT->getReturnType()) {
8374 if (Callee->isDeclaration() && !Caller->use_empty() &&
8375 // Conversion is ok if changing from pointer to int of same size.
8376 !(isa<PointerType>(FT->getReturnType()) &&
8377 TD->getIntPtrType() == OldRetTy))
8378 return false; // Cannot transform this return value.
8380 if (!Caller->use_empty() &&
8381 // void -> non-void is handled specially
8382 FT->getReturnType() != Type::VoidTy &&
8383 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8384 return false; // Cannot transform this return value.
8386 if (CallerPAL && !Caller->use_empty()) {
8387 ParameterAttributes RAttrs = CallerPAL->getParamAttrs(0);
8388 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8389 return false; // Attribute not compatible with transformed value.
8392 // If the callsite is an invoke instruction, and the return value is used by
8393 // a PHI node in a successor, we cannot change the return type of the call
8394 // because there is no place to put the cast instruction (without breaking
8395 // the critical edge). Bail out in this case.
8396 if (!Caller->use_empty())
8397 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8398 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8400 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8401 if (PN->getParent() == II->getNormalDest() ||
8402 PN->getParent() == II->getUnwindDest())
8406 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8407 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8409 CallSite::arg_iterator AI = CS.arg_begin();
8410 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8411 const Type *ParamTy = FT->getParamType(i);
8412 const Type *ActTy = (*AI)->getType();
8414 if (!CastInst::isCastable(ActTy, ParamTy))
8415 return false; // Cannot transform this parameter value.
8418 ParameterAttributes PAttrs = CallerPAL->getParamAttrs(i + 1);
8419 if (PAttrs & ParamAttr::typeIncompatible(ParamTy))
8420 return false; // Attribute not compatible with transformed value.
8423 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8424 // Some conversions are safe even if we do not have a body.
8425 // Either we can cast directly, or we can upconvert the argument
8426 bool isConvertible = ActTy == ParamTy ||
8427 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8428 (ParamTy->isInteger() && ActTy->isInteger() &&
8429 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8430 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8431 && c->getValue().isStrictlyPositive());
8432 if (Callee->isDeclaration() && !isConvertible) return false;
8435 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8436 Callee->isDeclaration())
8437 return false; // Do not delete arguments unless we have a function body...
8439 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && CallerPAL)
8440 // In this case we have more arguments than the new function type, but we
8441 // won't be dropping them. Check that these extra arguments have attributes
8442 // that are compatible with being a vararg call argument.
8443 for (unsigned i = CallerPAL->size(); i; --i) {
8444 if (CallerPAL->getParamIndex(i - 1) <= FT->getNumParams())
8446 ParameterAttributes PAttrs = CallerPAL->getParamAttrsAtIndex(i - 1);
8447 if (PAttrs & ParamAttr::VarArgsIncompatible)
8451 // Okay, we decided that this is a safe thing to do: go ahead and start
8452 // inserting cast instructions as necessary...
8453 std::vector<Value*> Args;
8454 Args.reserve(NumActualArgs);
8455 ParamAttrsVector attrVec;
8456 attrVec.reserve(NumCommonArgs);
8458 // Get any return attributes.
8459 ParameterAttributes RAttrs = CallerPAL ? CallerPAL->getParamAttrs(0) :
8462 // If the return value is not being used, the type may not be compatible
8463 // with the existing attributes. Wipe out any problematic attributes.
8464 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
8466 // Add the new return attributes.
8468 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8470 AI = CS.arg_begin();
8471 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8472 const Type *ParamTy = FT->getParamType(i);
8473 if ((*AI)->getType() == ParamTy) {
8474 Args.push_back(*AI);
8476 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8477 false, ParamTy, false);
8478 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8479 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8482 // Add any parameter attributes.
8483 ParameterAttributes PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) :
8486 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8489 // If the function takes more arguments than the call was taking, add them
8491 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8492 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8494 // If we are removing arguments to the function, emit an obnoxious warning...
8495 if (FT->getNumParams() < NumActualArgs)
8496 if (!FT->isVarArg()) {
8497 cerr << "WARNING: While resolving call to function '"
8498 << Callee->getName() << "' arguments were dropped!\n";
8500 // Add all of the arguments in their promoted form to the arg list...
8501 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8502 const Type *PTy = getPromotedType((*AI)->getType());
8503 if (PTy != (*AI)->getType()) {
8504 // Must promote to pass through va_arg area!
8505 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8507 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8508 InsertNewInstBefore(Cast, *Caller);
8509 Args.push_back(Cast);
8511 Args.push_back(*AI);
8514 // Add any parameter attributes.
8515 ParameterAttributes PAttrs = CallerPAL ?
8516 CallerPAL->getParamAttrs(i + 1) :
8519 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8523 if (FT->getReturnType() == Type::VoidTy)
8524 Caller->setName(""); // Void type should not have a name.
8526 const ParamAttrsList* NewCallerPAL = ParamAttrsList::get(attrVec);
8529 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8530 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8531 Args.begin(), Args.end(), Caller->getName(), Caller);
8532 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8533 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
8535 NC = new CallInst(Callee, Args.begin(), Args.end(),
8536 Caller->getName(), Caller);
8537 CallInst *CI = cast<CallInst>(Caller);
8538 if (CI->isTailCall())
8539 cast<CallInst>(NC)->setTailCall();
8540 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8541 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
8544 // Insert a cast of the return type as necessary.
8546 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8547 if (NV->getType() != Type::VoidTy) {
8548 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8550 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8552 // If this is an invoke instruction, we should insert it after the first
8553 // non-phi, instruction in the normal successor block.
8554 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8555 BasicBlock::iterator I = II->getNormalDest()->begin();
8556 while (isa<PHINode>(I)) ++I;
8557 InsertNewInstBefore(NC, *I);
8559 // Otherwise, it's a call, just insert cast right after the call instr
8560 InsertNewInstBefore(NC, *Caller);
8562 AddUsersToWorkList(*Caller);
8564 NV = UndefValue::get(Caller->getType());
8568 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8569 Caller->replaceAllUsesWith(NV);
8570 Caller->eraseFromParent();
8571 RemoveFromWorkList(Caller);
8575 // transformCallThroughTrampoline - Turn a call to a function created by the
8576 // init_trampoline intrinsic into a direct call to the underlying function.
8578 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8579 Value *Callee = CS.getCalledValue();
8580 const PointerType *PTy = cast<PointerType>(Callee->getType());
8581 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8582 const ParamAttrsList *Attrs = CS.getParamAttrs();
8584 // If the call already has the 'nest' attribute somewhere then give up -
8585 // otherwise 'nest' would occur twice after splicing in the chain.
8586 if (Attrs && Attrs->hasAttrSomewhere(ParamAttr::Nest))
8589 IntrinsicInst *Tramp =
8590 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8593 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8594 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8595 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8597 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8598 unsigned NestIdx = 1;
8599 const Type *NestTy = 0;
8600 ParameterAttributes NestAttr = ParamAttr::None;
8602 // Look for a parameter marked with the 'nest' attribute.
8603 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8604 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8605 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8606 // Record the parameter type and any other attributes.
8608 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8613 Instruction *Caller = CS.getInstruction();
8614 std::vector<Value*> NewArgs;
8615 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8617 ParamAttrsVector NewAttrs;
8618 NewAttrs.reserve(Attrs ? Attrs->size() + 1 : 1);
8620 // Insert the nest argument into the call argument list, which may
8621 // mean appending it. Likewise for attributes.
8623 // Add any function result attributes.
8624 ParameterAttributes Attr = Attrs ? Attrs->getParamAttrs(0) :
8627 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8631 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8633 if (Idx == NestIdx) {
8634 // Add the chain argument and attributes.
8635 Value *NestVal = Tramp->getOperand(3);
8636 if (NestVal->getType() != NestTy)
8637 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8638 NewArgs.push_back(NestVal);
8639 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8645 // Add the original argument and attributes.
8646 NewArgs.push_back(*I);
8647 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8650 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8656 // The trampoline may have been bitcast to a bogus type (FTy).
8657 // Handle this by synthesizing a new function type, equal to FTy
8658 // with the chain parameter inserted.
8660 std::vector<const Type*> NewTypes;
8661 NewTypes.reserve(FTy->getNumParams()+1);
8663 // Insert the chain's type into the list of parameter types, which may
8664 // mean appending it.
8667 FunctionType::param_iterator I = FTy->param_begin(),
8668 E = FTy->param_end();
8672 // Add the chain's type.
8673 NewTypes.push_back(NestTy);
8678 // Add the original type.
8679 NewTypes.push_back(*I);
8685 // Replace the trampoline call with a direct call. Let the generic
8686 // code sort out any function type mismatches.
8687 FunctionType *NewFTy =
8688 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8689 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8690 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8691 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8693 Instruction *NewCaller;
8694 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8695 NewCaller = new InvokeInst(NewCallee,
8696 II->getNormalDest(), II->getUnwindDest(),
8697 NewArgs.begin(), NewArgs.end(),
8698 Caller->getName(), Caller);
8699 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8700 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8702 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8703 Caller->getName(), Caller);
8704 if (cast<CallInst>(Caller)->isTailCall())
8705 cast<CallInst>(NewCaller)->setTailCall();
8706 cast<CallInst>(NewCaller)->
8707 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8708 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8710 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8711 Caller->replaceAllUsesWith(NewCaller);
8712 Caller->eraseFromParent();
8713 RemoveFromWorkList(Caller);
8718 // Replace the trampoline call with a direct call. Since there is no 'nest'
8719 // parameter, there is no need to adjust the argument list. Let the generic
8720 // code sort out any function type mismatches.
8721 Constant *NewCallee =
8722 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8723 CS.setCalledFunction(NewCallee);
8724 return CS.getInstruction();
8727 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8728 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8729 /// and a single binop.
8730 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8731 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8732 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8733 isa<CmpInst>(FirstInst));
8734 unsigned Opc = FirstInst->getOpcode();
8735 Value *LHSVal = FirstInst->getOperand(0);
8736 Value *RHSVal = FirstInst->getOperand(1);
8738 const Type *LHSType = LHSVal->getType();
8739 const Type *RHSType = RHSVal->getType();
8741 // Scan to see if all operands are the same opcode, all have one use, and all
8742 // kill their operands (i.e. the operands have one use).
8743 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8744 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8745 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8746 // Verify type of the LHS matches so we don't fold cmp's of different
8747 // types or GEP's with different index types.
8748 I->getOperand(0)->getType() != LHSType ||
8749 I->getOperand(1)->getType() != RHSType)
8752 // If they are CmpInst instructions, check their predicates
8753 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8754 if (cast<CmpInst>(I)->getPredicate() !=
8755 cast<CmpInst>(FirstInst)->getPredicate())
8758 // Keep track of which operand needs a phi node.
8759 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8760 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8763 // Otherwise, this is safe to transform, determine if it is profitable.
8765 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8766 // Indexes are often folded into load/store instructions, so we don't want to
8767 // hide them behind a phi.
8768 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8771 Value *InLHS = FirstInst->getOperand(0);
8772 Value *InRHS = FirstInst->getOperand(1);
8773 PHINode *NewLHS = 0, *NewRHS = 0;
8775 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8776 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8777 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8778 InsertNewInstBefore(NewLHS, PN);
8783 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8784 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8785 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8786 InsertNewInstBefore(NewRHS, PN);
8790 // Add all operands to the new PHIs.
8791 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8793 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8794 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8797 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8798 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8802 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8803 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8804 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8805 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8808 assert(isa<GetElementPtrInst>(FirstInst));
8809 return new GetElementPtrInst(LHSVal, RHSVal);
8813 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8814 /// of the block that defines it. This means that it must be obvious the value
8815 /// of the load is not changed from the point of the load to the end of the
8818 /// Finally, it is safe, but not profitable, to sink a load targetting a
8819 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8821 static bool isSafeToSinkLoad(LoadInst *L) {
8822 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8824 for (++BBI; BBI != E; ++BBI)
8825 if (BBI->mayWriteToMemory())
8828 // Check for non-address taken alloca. If not address-taken already, it isn't
8829 // profitable to do this xform.
8830 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8831 bool isAddressTaken = false;
8832 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8834 if (isa<LoadInst>(UI)) continue;
8835 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8836 // If storing TO the alloca, then the address isn't taken.
8837 if (SI->getOperand(1) == AI) continue;
8839 isAddressTaken = true;
8843 if (!isAddressTaken)
8851 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8852 // operator and they all are only used by the PHI, PHI together their
8853 // inputs, and do the operation once, to the result of the PHI.
8854 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8855 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8857 // Scan the instruction, looking for input operations that can be folded away.
8858 // If all input operands to the phi are the same instruction (e.g. a cast from
8859 // the same type or "+42") we can pull the operation through the PHI, reducing
8860 // code size and simplifying code.
8861 Constant *ConstantOp = 0;
8862 const Type *CastSrcTy = 0;
8863 bool isVolatile = false;
8864 if (isa<CastInst>(FirstInst)) {
8865 CastSrcTy = FirstInst->getOperand(0)->getType();
8866 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8867 // Can fold binop, compare or shift here if the RHS is a constant,
8868 // otherwise call FoldPHIArgBinOpIntoPHI.
8869 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8870 if (ConstantOp == 0)
8871 return FoldPHIArgBinOpIntoPHI(PN);
8872 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8873 isVolatile = LI->isVolatile();
8874 // We can't sink the load if the loaded value could be modified between the
8875 // load and the PHI.
8876 if (LI->getParent() != PN.getIncomingBlock(0) ||
8877 !isSafeToSinkLoad(LI))
8879 } else if (isa<GetElementPtrInst>(FirstInst)) {
8880 if (FirstInst->getNumOperands() == 2)
8881 return FoldPHIArgBinOpIntoPHI(PN);
8882 // Can't handle general GEPs yet.
8885 return 0; // Cannot fold this operation.
8888 // Check to see if all arguments are the same operation.
8889 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8890 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8891 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8892 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8895 if (I->getOperand(0)->getType() != CastSrcTy)
8896 return 0; // Cast operation must match.
8897 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8898 // We can't sink the load if the loaded value could be modified between
8899 // the load and the PHI.
8900 if (LI->isVolatile() != isVolatile ||
8901 LI->getParent() != PN.getIncomingBlock(i) ||
8902 !isSafeToSinkLoad(LI))
8904 } else if (I->getOperand(1) != ConstantOp) {
8909 // Okay, they are all the same operation. Create a new PHI node of the
8910 // correct type, and PHI together all of the LHS's of the instructions.
8911 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8912 PN.getName()+".in");
8913 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8915 Value *InVal = FirstInst->getOperand(0);
8916 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8918 // Add all operands to the new PHI.
8919 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8920 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8921 if (NewInVal != InVal)
8923 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8928 // The new PHI unions all of the same values together. This is really
8929 // common, so we handle it intelligently here for compile-time speed.
8933 InsertNewInstBefore(NewPN, PN);
8937 // Insert and return the new operation.
8938 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8939 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8940 else if (isa<LoadInst>(FirstInst))
8941 return new LoadInst(PhiVal, "", isVolatile);
8942 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8943 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8944 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8945 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8946 PhiVal, ConstantOp);
8948 assert(0 && "Unknown operation");
8952 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8954 static bool DeadPHICycle(PHINode *PN,
8955 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8956 if (PN->use_empty()) return true;
8957 if (!PN->hasOneUse()) return false;
8959 // Remember this node, and if we find the cycle, return.
8960 if (!PotentiallyDeadPHIs.insert(PN))
8963 // Don't scan crazily complex things.
8964 if (PotentiallyDeadPHIs.size() == 16)
8967 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8968 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8973 /// PHIsEqualValue - Return true if this phi node is always equal to
8974 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8975 /// z = some value; x = phi (y, z); y = phi (x, z)
8976 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8977 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8978 // See if we already saw this PHI node.
8979 if (!ValueEqualPHIs.insert(PN))
8982 // Don't scan crazily complex things.
8983 if (ValueEqualPHIs.size() == 16)
8986 // Scan the operands to see if they are either phi nodes or are equal to
8988 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8989 Value *Op = PN->getIncomingValue(i);
8990 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8991 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8993 } else if (Op != NonPhiInVal)
9001 // PHINode simplification
9003 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9004 // If LCSSA is around, don't mess with Phi nodes
9005 if (MustPreserveLCSSA) return 0;
9007 if (Value *V = PN.hasConstantValue())
9008 return ReplaceInstUsesWith(PN, V);
9010 // If all PHI operands are the same operation, pull them through the PHI,
9011 // reducing code size.
9012 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9013 PN.getIncomingValue(0)->hasOneUse())
9014 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9017 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9018 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9019 // PHI)... break the cycle.
9020 if (PN.hasOneUse()) {
9021 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9022 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9023 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9024 PotentiallyDeadPHIs.insert(&PN);
9025 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9026 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9029 // If this phi has a single use, and if that use just computes a value for
9030 // the next iteration of a loop, delete the phi. This occurs with unused
9031 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9032 // common case here is good because the only other things that catch this
9033 // are induction variable analysis (sometimes) and ADCE, which is only run
9035 if (PHIUser->hasOneUse() &&
9036 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9037 PHIUser->use_back() == &PN) {
9038 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9042 // We sometimes end up with phi cycles that non-obviously end up being the
9043 // same value, for example:
9044 // z = some value; x = phi (y, z); y = phi (x, z)
9045 // where the phi nodes don't necessarily need to be in the same block. Do a
9046 // quick check to see if the PHI node only contains a single non-phi value, if
9047 // so, scan to see if the phi cycle is actually equal to that value.
9049 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9050 // Scan for the first non-phi operand.
9051 while (InValNo != NumOperandVals &&
9052 isa<PHINode>(PN.getIncomingValue(InValNo)))
9055 if (InValNo != NumOperandVals) {
9056 Value *NonPhiInVal = PN.getOperand(InValNo);
9058 // Scan the rest of the operands to see if there are any conflicts, if so
9059 // there is no need to recursively scan other phis.
9060 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9061 Value *OpVal = PN.getIncomingValue(InValNo);
9062 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9066 // If we scanned over all operands, then we have one unique value plus
9067 // phi values. Scan PHI nodes to see if they all merge in each other or
9069 if (InValNo == NumOperandVals) {
9070 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9071 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9072 return ReplaceInstUsesWith(PN, NonPhiInVal);
9079 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9080 Instruction *InsertPoint,
9082 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9083 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9084 // We must cast correctly to the pointer type. Ensure that we
9085 // sign extend the integer value if it is smaller as this is
9086 // used for address computation.
9087 Instruction::CastOps opcode =
9088 (VTySize < PtrSize ? Instruction::SExt :
9089 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9090 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9094 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9095 Value *PtrOp = GEP.getOperand(0);
9096 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9097 // If so, eliminate the noop.
9098 if (GEP.getNumOperands() == 1)
9099 return ReplaceInstUsesWith(GEP, PtrOp);
9101 if (isa<UndefValue>(GEP.getOperand(0)))
9102 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9104 bool HasZeroPointerIndex = false;
9105 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9106 HasZeroPointerIndex = C->isNullValue();
9108 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9109 return ReplaceInstUsesWith(GEP, PtrOp);
9111 // Eliminate unneeded casts for indices.
9112 bool MadeChange = false;
9114 gep_type_iterator GTI = gep_type_begin(GEP);
9115 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
9116 if (isa<SequentialType>(*GTI)) {
9117 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
9118 if (CI->getOpcode() == Instruction::ZExt ||
9119 CI->getOpcode() == Instruction::SExt) {
9120 const Type *SrcTy = CI->getOperand(0)->getType();
9121 // We can eliminate a cast from i32 to i64 iff the target
9122 // is a 32-bit pointer target.
9123 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9125 GEP.setOperand(i, CI->getOperand(0));
9129 // If we are using a wider index than needed for this platform, shrink it
9130 // to what we need. If the incoming value needs a cast instruction,
9131 // insert it. This explicit cast can make subsequent optimizations more
9133 Value *Op = GEP.getOperand(i);
9134 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits())
9135 if (Constant *C = dyn_cast<Constant>(Op)) {
9136 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
9139 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9141 GEP.setOperand(i, Op);
9146 if (MadeChange) return &GEP;
9148 // If this GEP instruction doesn't move the pointer, and if the input operand
9149 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9150 // real input to the dest type.
9151 if (GEP.hasAllZeroIndices()) {
9152 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9153 // If the bitcast is of an allocation, and the allocation will be
9154 // converted to match the type of the cast, don't touch this.
9155 if (isa<AllocationInst>(BCI->getOperand(0))) {
9156 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9157 if (Instruction *I = visitBitCast(*BCI)) {
9160 BCI->getParent()->getInstList().insert(BCI, I);
9161 ReplaceInstUsesWith(*BCI, I);
9166 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9170 // Combine Indices - If the source pointer to this getelementptr instruction
9171 // is a getelementptr instruction, combine the indices of the two
9172 // getelementptr instructions into a single instruction.
9174 SmallVector<Value*, 8> SrcGEPOperands;
9175 if (User *Src = dyn_castGetElementPtr(PtrOp))
9176 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9178 if (!SrcGEPOperands.empty()) {
9179 // Note that if our source is a gep chain itself that we wait for that
9180 // chain to be resolved before we perform this transformation. This
9181 // avoids us creating a TON of code in some cases.
9183 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9184 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9185 return 0; // Wait until our source is folded to completion.
9187 SmallVector<Value*, 8> Indices;
9189 // Find out whether the last index in the source GEP is a sequential idx.
9190 bool EndsWithSequential = false;
9191 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9192 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9193 EndsWithSequential = !isa<StructType>(*I);
9195 // Can we combine the two pointer arithmetics offsets?
9196 if (EndsWithSequential) {
9197 // Replace: gep (gep %P, long B), long A, ...
9198 // With: T = long A+B; gep %P, T, ...
9200 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9201 if (SO1 == Constant::getNullValue(SO1->getType())) {
9203 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9206 // If they aren't the same type, convert both to an integer of the
9207 // target's pointer size.
9208 if (SO1->getType() != GO1->getType()) {
9209 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9210 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9211 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9212 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9214 unsigned PS = TD->getPointerSizeInBits();
9215 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9216 // Convert GO1 to SO1's type.
9217 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9219 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9220 // Convert SO1 to GO1's type.
9221 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9223 const Type *PT = TD->getIntPtrType();
9224 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9225 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9229 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9230 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9232 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
9233 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9237 // Recycle the GEP we already have if possible.
9238 if (SrcGEPOperands.size() == 2) {
9239 GEP.setOperand(0, SrcGEPOperands[0]);
9240 GEP.setOperand(1, Sum);
9243 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9244 SrcGEPOperands.end()-1);
9245 Indices.push_back(Sum);
9246 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9248 } else if (isa<Constant>(*GEP.idx_begin()) &&
9249 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9250 SrcGEPOperands.size() != 1) {
9251 // Otherwise we can do the fold if the first index of the GEP is a zero
9252 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9253 SrcGEPOperands.end());
9254 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9257 if (!Indices.empty())
9258 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
9259 Indices.end(), GEP.getName());
9261 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9262 // GEP of global variable. If all of the indices for this GEP are
9263 // constants, we can promote this to a constexpr instead of an instruction.
9265 // Scan for nonconstants...
9266 SmallVector<Constant*, 8> Indices;
9267 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9268 for (; I != E && isa<Constant>(*I); ++I)
9269 Indices.push_back(cast<Constant>(*I));
9271 if (I == E) { // If they are all constants...
9272 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9273 &Indices[0],Indices.size());
9275 // Replace all uses of the GEP with the new constexpr...
9276 return ReplaceInstUsesWith(GEP, CE);
9278 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9279 if (!isa<PointerType>(X->getType())) {
9280 // Not interesting. Source pointer must be a cast from pointer.
9281 } else if (HasZeroPointerIndex) {
9282 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9283 // into : GEP [10 x i8]* X, i32 0, ...
9285 // This occurs when the program declares an array extern like "int X[];"
9287 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9288 const PointerType *XTy = cast<PointerType>(X->getType());
9289 if (const ArrayType *XATy =
9290 dyn_cast<ArrayType>(XTy->getElementType()))
9291 if (const ArrayType *CATy =
9292 dyn_cast<ArrayType>(CPTy->getElementType()))
9293 if (CATy->getElementType() == XATy->getElementType()) {
9294 // At this point, we know that the cast source type is a pointer
9295 // to an array of the same type as the destination pointer
9296 // array. Because the array type is never stepped over (there
9297 // is a leading zero) we can fold the cast into this GEP.
9298 GEP.setOperand(0, X);
9301 } else if (GEP.getNumOperands() == 2) {
9302 // Transform things like:
9303 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9304 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9305 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9306 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9307 if (isa<ArrayType>(SrcElTy) &&
9308 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9309 TD->getABITypeSize(ResElTy)) {
9311 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9312 Idx[1] = GEP.getOperand(1);
9313 Value *V = InsertNewInstBefore(
9314 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
9315 // V and GEP are both pointer types --> BitCast
9316 return new BitCastInst(V, GEP.getType());
9319 // Transform things like:
9320 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9321 // (where tmp = 8*tmp2) into:
9322 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9324 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9325 uint64_t ArrayEltSize =
9326 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9328 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9329 // allow either a mul, shift, or constant here.
9331 ConstantInt *Scale = 0;
9332 if (ArrayEltSize == 1) {
9333 NewIdx = GEP.getOperand(1);
9334 Scale = ConstantInt::get(NewIdx->getType(), 1);
9335 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9336 NewIdx = ConstantInt::get(CI->getType(), 1);
9338 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9339 if (Inst->getOpcode() == Instruction::Shl &&
9340 isa<ConstantInt>(Inst->getOperand(1))) {
9341 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9342 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9343 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9344 NewIdx = Inst->getOperand(0);
9345 } else if (Inst->getOpcode() == Instruction::Mul &&
9346 isa<ConstantInt>(Inst->getOperand(1))) {
9347 Scale = cast<ConstantInt>(Inst->getOperand(1));
9348 NewIdx = Inst->getOperand(0);
9352 // If the index will be to exactly the right offset with the scale taken
9353 // out, perform the transformation. Note, we don't know whether Scale is
9354 // signed or not. We'll use unsigned version of division/modulo
9355 // operation after making sure Scale doesn't have the sign bit set.
9356 if (Scale && Scale->getSExtValue() >= 0LL &&
9357 Scale->getZExtValue() % ArrayEltSize == 0) {
9358 Scale = ConstantInt::get(Scale->getType(),
9359 Scale->getZExtValue() / ArrayEltSize);
9360 if (Scale->getZExtValue() != 1) {
9361 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9363 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9364 NewIdx = InsertNewInstBefore(Sc, GEP);
9367 // Insert the new GEP instruction.
9369 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9371 Instruction *NewGEP =
9372 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
9373 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9374 // The NewGEP must be pointer typed, so must the old one -> BitCast
9375 return new BitCastInst(NewGEP, GEP.getType());
9384 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9385 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9386 if (AI.isArrayAllocation()) // Check C != 1
9387 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9389 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9390 AllocationInst *New = 0;
9392 // Create and insert the replacement instruction...
9393 if (isa<MallocInst>(AI))
9394 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9396 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9397 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9400 InsertNewInstBefore(New, AI);
9402 // Scan to the end of the allocation instructions, to skip over a block of
9403 // allocas if possible...
9405 BasicBlock::iterator It = New;
9406 while (isa<AllocationInst>(*It)) ++It;
9408 // Now that I is pointing to the first non-allocation-inst in the block,
9409 // insert our getelementptr instruction...
9411 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9415 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9416 New->getName()+".sub", It);
9418 // Now make everything use the getelementptr instead of the original
9420 return ReplaceInstUsesWith(AI, V);
9421 } else if (isa<UndefValue>(AI.getArraySize())) {
9422 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9425 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9426 // Note that we only do this for alloca's, because malloc should allocate and
9427 // return a unique pointer, even for a zero byte allocation.
9428 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9429 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9430 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9435 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9436 Value *Op = FI.getOperand(0);
9438 // free undef -> unreachable.
9439 if (isa<UndefValue>(Op)) {
9440 // Insert a new store to null because we cannot modify the CFG here.
9441 new StoreInst(ConstantInt::getTrue(),
9442 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9443 return EraseInstFromFunction(FI);
9446 // If we have 'free null' delete the instruction. This can happen in stl code
9447 // when lots of inlining happens.
9448 if (isa<ConstantPointerNull>(Op))
9449 return EraseInstFromFunction(FI);
9451 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9452 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9453 FI.setOperand(0, CI->getOperand(0));
9457 // Change free (gep X, 0,0,0,0) into free(X)
9458 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9459 if (GEPI->hasAllZeroIndices()) {
9460 AddToWorkList(GEPI);
9461 FI.setOperand(0, GEPI->getOperand(0));
9466 // Change free(malloc) into nothing, if the malloc has a single use.
9467 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9468 if (MI->hasOneUse()) {
9469 EraseInstFromFunction(FI);
9470 return EraseInstFromFunction(*MI);
9477 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9478 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9479 const TargetData *TD) {
9480 User *CI = cast<User>(LI.getOperand(0));
9481 Value *CastOp = CI->getOperand(0);
9483 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9484 // Instead of loading constant c string, use corresponding integer value
9485 // directly if string length is small enough.
9486 const std::string &Str = CE->getOperand(0)->getStringValue();
9488 unsigned len = Str.length();
9489 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9490 unsigned numBits = Ty->getPrimitiveSizeInBits();
9491 // Replace LI with immediate integer store.
9492 if ((numBits >> 3) == len + 1) {
9493 APInt StrVal(numBits, 0);
9494 APInt SingleChar(numBits, 0);
9495 if (TD->isLittleEndian()) {
9496 for (signed i = len-1; i >= 0; i--) {
9497 SingleChar = (uint64_t) Str[i];
9498 StrVal = (StrVal << 8) | SingleChar;
9501 for (unsigned i = 0; i < len; i++) {
9502 SingleChar = (uint64_t) Str[i];
9503 StrVal = (StrVal << 8) | SingleChar;
9505 // Append NULL at the end.
9507 StrVal = (StrVal << 8) | SingleChar;
9509 Value *NL = ConstantInt::get(StrVal);
9510 return IC.ReplaceInstUsesWith(LI, NL);
9515 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9516 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9517 const Type *SrcPTy = SrcTy->getElementType();
9519 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9520 isa<VectorType>(DestPTy)) {
9521 // If the source is an array, the code below will not succeed. Check to
9522 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9524 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9525 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9526 if (ASrcTy->getNumElements() != 0) {
9528 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9529 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9530 SrcTy = cast<PointerType>(CastOp->getType());
9531 SrcPTy = SrcTy->getElementType();
9534 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9535 isa<VectorType>(SrcPTy)) &&
9536 // Do not allow turning this into a load of an integer, which is then
9537 // casted to a pointer, this pessimizes pointer analysis a lot.
9538 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9539 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9540 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9542 // Okay, we are casting from one integer or pointer type to another of
9543 // the same size. Instead of casting the pointer before the load, cast
9544 // the result of the loaded value.
9545 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9547 LI.isVolatile()),LI);
9548 // Now cast the result of the load.
9549 return new BitCastInst(NewLoad, LI.getType());
9556 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9557 /// from this value cannot trap. If it is not obviously safe to load from the
9558 /// specified pointer, we do a quick local scan of the basic block containing
9559 /// ScanFrom, to determine if the address is already accessed.
9560 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9561 // If it is an alloca it is always safe to load from.
9562 if (isa<AllocaInst>(V)) return true;
9564 // If it is a global variable it is mostly safe to load from.
9565 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9566 // Don't try to evaluate aliases. External weak GV can be null.
9567 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9569 // Otherwise, be a little bit agressive by scanning the local block where we
9570 // want to check to see if the pointer is already being loaded or stored
9571 // from/to. If so, the previous load or store would have already trapped,
9572 // so there is no harm doing an extra load (also, CSE will later eliminate
9573 // the load entirely).
9574 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9579 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9580 if (LI->getOperand(0) == V) return true;
9581 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9582 if (SI->getOperand(1) == V) return true;
9588 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9589 /// until we find the underlying object a pointer is referring to or something
9590 /// we don't understand. Note that the returned pointer may be offset from the
9591 /// input, because we ignore GEP indices.
9592 static Value *GetUnderlyingObject(Value *Ptr) {
9594 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9595 if (CE->getOpcode() == Instruction::BitCast ||
9596 CE->getOpcode() == Instruction::GetElementPtr)
9597 Ptr = CE->getOperand(0);
9600 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9601 Ptr = BCI->getOperand(0);
9602 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9603 Ptr = GEP->getOperand(0);
9610 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9611 Value *Op = LI.getOperand(0);
9613 // Attempt to improve the alignment.
9614 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9615 if (KnownAlign > LI.getAlignment())
9616 LI.setAlignment(KnownAlign);
9618 // load (cast X) --> cast (load X) iff safe
9619 if (isa<CastInst>(Op))
9620 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9623 // None of the following transforms are legal for volatile loads.
9624 if (LI.isVolatile()) return 0;
9626 if (&LI.getParent()->front() != &LI) {
9627 BasicBlock::iterator BBI = &LI; --BBI;
9628 // If the instruction immediately before this is a store to the same
9629 // address, do a simple form of store->load forwarding.
9630 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9631 if (SI->getOperand(1) == LI.getOperand(0))
9632 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9633 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9634 if (LIB->getOperand(0) == LI.getOperand(0))
9635 return ReplaceInstUsesWith(LI, LIB);
9638 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9639 const Value *GEPI0 = GEPI->getOperand(0);
9640 // TODO: Consider a target hook for valid address spaces for this xform.
9641 if (isa<ConstantPointerNull>(GEPI0) &&
9642 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
9643 // Insert a new store to null instruction before the load to indicate
9644 // that this code is not reachable. We do this instead of inserting
9645 // an unreachable instruction directly because we cannot modify the
9647 new StoreInst(UndefValue::get(LI.getType()),
9648 Constant::getNullValue(Op->getType()), &LI);
9649 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9653 if (Constant *C = dyn_cast<Constant>(Op)) {
9654 // load null/undef -> undef
9655 // TODO: Consider a target hook for valid address spaces for this xform.
9656 if (isa<UndefValue>(C) || (C->isNullValue() &&
9657 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
9658 // Insert a new store to null instruction before the load to indicate that
9659 // this code is not reachable. We do this instead of inserting an
9660 // unreachable instruction directly because we cannot modify the CFG.
9661 new StoreInst(UndefValue::get(LI.getType()),
9662 Constant::getNullValue(Op->getType()), &LI);
9663 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9666 // Instcombine load (constant global) into the value loaded.
9667 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9668 if (GV->isConstant() && !GV->isDeclaration())
9669 return ReplaceInstUsesWith(LI, GV->getInitializer());
9671 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9672 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9673 if (CE->getOpcode() == Instruction::GetElementPtr) {
9674 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9675 if (GV->isConstant() && !GV->isDeclaration())
9677 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9678 return ReplaceInstUsesWith(LI, V);
9679 if (CE->getOperand(0)->isNullValue()) {
9680 // Insert a new store to null instruction before the load to indicate
9681 // that this code is not reachable. We do this instead of inserting
9682 // an unreachable instruction directly because we cannot modify the
9684 new StoreInst(UndefValue::get(LI.getType()),
9685 Constant::getNullValue(Op->getType()), &LI);
9686 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9689 } else if (CE->isCast()) {
9690 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9695 // If this load comes from anywhere in a constant global, and if the global
9696 // is all undef or zero, we know what it loads.
9697 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9698 if (GV->isConstant() && GV->hasInitializer()) {
9699 if (GV->getInitializer()->isNullValue())
9700 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9701 else if (isa<UndefValue>(GV->getInitializer()))
9702 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9706 if (Op->hasOneUse()) {
9707 // Change select and PHI nodes to select values instead of addresses: this
9708 // helps alias analysis out a lot, allows many others simplifications, and
9709 // exposes redundancy in the code.
9711 // Note that we cannot do the transformation unless we know that the
9712 // introduced loads cannot trap! Something like this is valid as long as
9713 // the condition is always false: load (select bool %C, int* null, int* %G),
9714 // but it would not be valid if we transformed it to load from null
9717 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9718 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9719 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9720 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9721 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9722 SI->getOperand(1)->getName()+".val"), LI);
9723 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9724 SI->getOperand(2)->getName()+".val"), LI);
9725 return new SelectInst(SI->getCondition(), V1, V2);
9728 // load (select (cond, null, P)) -> load P
9729 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9730 if (C->isNullValue()) {
9731 LI.setOperand(0, SI->getOperand(2));
9735 // load (select (cond, P, null)) -> load P
9736 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9737 if (C->isNullValue()) {
9738 LI.setOperand(0, SI->getOperand(1));
9746 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9748 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9749 User *CI = cast<User>(SI.getOperand(1));
9750 Value *CastOp = CI->getOperand(0);
9752 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9753 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9754 const Type *SrcPTy = SrcTy->getElementType();
9756 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9757 // If the source is an array, the code below will not succeed. Check to
9758 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9760 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9761 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9762 if (ASrcTy->getNumElements() != 0) {
9764 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9765 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9766 SrcTy = cast<PointerType>(CastOp->getType());
9767 SrcPTy = SrcTy->getElementType();
9770 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9771 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9772 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9774 // Okay, we are casting from one integer or pointer type to another of
9775 // the same size. Instead of casting the pointer before
9776 // the store, cast the value to be stored.
9778 Value *SIOp0 = SI.getOperand(0);
9779 Instruction::CastOps opcode = Instruction::BitCast;
9780 const Type* CastSrcTy = SIOp0->getType();
9781 const Type* CastDstTy = SrcPTy;
9782 if (isa<PointerType>(CastDstTy)) {
9783 if (CastSrcTy->isInteger())
9784 opcode = Instruction::IntToPtr;
9785 } else if (isa<IntegerType>(CastDstTy)) {
9786 if (isa<PointerType>(SIOp0->getType()))
9787 opcode = Instruction::PtrToInt;
9789 if (Constant *C = dyn_cast<Constant>(SIOp0))
9790 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9792 NewCast = IC.InsertNewInstBefore(
9793 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9795 return new StoreInst(NewCast, CastOp);
9802 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9803 Value *Val = SI.getOperand(0);
9804 Value *Ptr = SI.getOperand(1);
9806 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9807 EraseInstFromFunction(SI);
9812 // If the RHS is an alloca with a single use, zapify the store, making the
9814 if (Ptr->hasOneUse()) {
9815 if (isa<AllocaInst>(Ptr)) {
9816 EraseInstFromFunction(SI);
9821 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9822 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9823 GEP->getOperand(0)->hasOneUse()) {
9824 EraseInstFromFunction(SI);
9830 // Attempt to improve the alignment.
9831 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9832 if (KnownAlign > SI.getAlignment())
9833 SI.setAlignment(KnownAlign);
9835 // Do really simple DSE, to catch cases where there are several consequtive
9836 // stores to the same location, separated by a few arithmetic operations. This
9837 // situation often occurs with bitfield accesses.
9838 BasicBlock::iterator BBI = &SI;
9839 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9843 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9844 // Prev store isn't volatile, and stores to the same location?
9845 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9848 EraseInstFromFunction(*PrevSI);
9854 // If this is a load, we have to stop. However, if the loaded value is from
9855 // the pointer we're loading and is producing the pointer we're storing,
9856 // then *this* store is dead (X = load P; store X -> P).
9857 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9858 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9859 EraseInstFromFunction(SI);
9863 // Otherwise, this is a load from some other location. Stores before it
9868 // Don't skip over loads or things that can modify memory.
9869 if (BBI->mayWriteToMemory())
9874 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9876 // store X, null -> turns into 'unreachable' in SimplifyCFG
9877 if (isa<ConstantPointerNull>(Ptr)) {
9878 if (!isa<UndefValue>(Val)) {
9879 SI.setOperand(0, UndefValue::get(Val->getType()));
9880 if (Instruction *U = dyn_cast<Instruction>(Val))
9881 AddToWorkList(U); // Dropped a use.
9884 return 0; // Do not modify these!
9887 // store undef, Ptr -> noop
9888 if (isa<UndefValue>(Val)) {
9889 EraseInstFromFunction(SI);
9894 // If the pointer destination is a cast, see if we can fold the cast into the
9896 if (isa<CastInst>(Ptr))
9897 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9899 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9901 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9905 // If this store is the last instruction in the basic block, and if the block
9906 // ends with an unconditional branch, try to move it to the successor block.
9908 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9909 if (BI->isUnconditional())
9910 if (SimplifyStoreAtEndOfBlock(SI))
9911 return 0; // xform done!
9916 /// SimplifyStoreAtEndOfBlock - Turn things like:
9917 /// if () { *P = v1; } else { *P = v2 }
9918 /// into a phi node with a store in the successor.
9920 /// Simplify things like:
9921 /// *P = v1; if () { *P = v2; }
9922 /// into a phi node with a store in the successor.
9924 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9925 BasicBlock *StoreBB = SI.getParent();
9927 // Check to see if the successor block has exactly two incoming edges. If
9928 // so, see if the other predecessor contains a store to the same location.
9929 // if so, insert a PHI node (if needed) and move the stores down.
9930 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9932 // Determine whether Dest has exactly two predecessors and, if so, compute
9933 // the other predecessor.
9934 pred_iterator PI = pred_begin(DestBB);
9935 BasicBlock *OtherBB = 0;
9939 if (PI == pred_end(DestBB))
9942 if (*PI != StoreBB) {
9947 if (++PI != pred_end(DestBB))
9951 // Verify that the other block ends in a branch and is not otherwise empty.
9952 BasicBlock::iterator BBI = OtherBB->getTerminator();
9953 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9954 if (!OtherBr || BBI == OtherBB->begin())
9957 // If the other block ends in an unconditional branch, check for the 'if then
9958 // else' case. there is an instruction before the branch.
9959 StoreInst *OtherStore = 0;
9960 if (OtherBr->isUnconditional()) {
9961 // If this isn't a store, or isn't a store to the same location, bail out.
9963 OtherStore = dyn_cast<StoreInst>(BBI);
9964 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9967 // Otherwise, the other block ended with a conditional branch. If one of the
9968 // destinations is StoreBB, then we have the if/then case.
9969 if (OtherBr->getSuccessor(0) != StoreBB &&
9970 OtherBr->getSuccessor(1) != StoreBB)
9973 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9974 // if/then triangle. See if there is a store to the same ptr as SI that
9975 // lives in OtherBB.
9977 // Check to see if we find the matching store.
9978 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9979 if (OtherStore->getOperand(1) != SI.getOperand(1))
9983 // If we find something that may be using the stored value, or if we run
9984 // out of instructions, we can't do the xform.
9985 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9986 BBI == OtherBB->begin())
9990 // In order to eliminate the store in OtherBr, we have to
9991 // make sure nothing reads the stored value in StoreBB.
9992 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9993 // FIXME: This should really be AA driven.
9994 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9999 // Insert a PHI node now if we need it.
10000 Value *MergedVal = OtherStore->getOperand(0);
10001 if (MergedVal != SI.getOperand(0)) {
10002 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
10003 PN->reserveOperandSpace(2);
10004 PN->addIncoming(SI.getOperand(0), SI.getParent());
10005 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10006 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10009 // Advance to a place where it is safe to insert the new store and
10011 BBI = DestBB->begin();
10012 while (isa<PHINode>(BBI)) ++BBI;
10013 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10014 OtherStore->isVolatile()), *BBI);
10016 // Nuke the old stores.
10017 EraseInstFromFunction(SI);
10018 EraseInstFromFunction(*OtherStore);
10024 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10025 // Change br (not X), label True, label False to: br X, label False, True
10027 BasicBlock *TrueDest;
10028 BasicBlock *FalseDest;
10029 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10030 !isa<Constant>(X)) {
10031 // Swap Destinations and condition...
10032 BI.setCondition(X);
10033 BI.setSuccessor(0, FalseDest);
10034 BI.setSuccessor(1, TrueDest);
10038 // Cannonicalize fcmp_one -> fcmp_oeq
10039 FCmpInst::Predicate FPred; Value *Y;
10040 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10041 TrueDest, FalseDest)))
10042 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10043 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10044 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10045 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10046 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10047 NewSCC->takeName(I);
10048 // Swap Destinations and condition...
10049 BI.setCondition(NewSCC);
10050 BI.setSuccessor(0, FalseDest);
10051 BI.setSuccessor(1, TrueDest);
10052 RemoveFromWorkList(I);
10053 I->eraseFromParent();
10054 AddToWorkList(NewSCC);
10058 // Cannonicalize icmp_ne -> icmp_eq
10059 ICmpInst::Predicate IPred;
10060 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10061 TrueDest, FalseDest)))
10062 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10063 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10064 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10065 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10066 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10067 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10068 NewSCC->takeName(I);
10069 // Swap Destinations and condition...
10070 BI.setCondition(NewSCC);
10071 BI.setSuccessor(0, FalseDest);
10072 BI.setSuccessor(1, TrueDest);
10073 RemoveFromWorkList(I);
10074 I->eraseFromParent();;
10075 AddToWorkList(NewSCC);
10082 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10083 Value *Cond = SI.getCondition();
10084 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10085 if (I->getOpcode() == Instruction::Add)
10086 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10087 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10088 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10089 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10091 SI.setOperand(0, I->getOperand(0));
10099 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10100 /// is to leave as a vector operation.
10101 static bool CheapToScalarize(Value *V, bool isConstant) {
10102 if (isa<ConstantAggregateZero>(V))
10104 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10105 if (isConstant) return true;
10106 // If all elts are the same, we can extract.
10107 Constant *Op0 = C->getOperand(0);
10108 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10109 if (C->getOperand(i) != Op0)
10113 Instruction *I = dyn_cast<Instruction>(V);
10114 if (!I) return false;
10116 // Insert element gets simplified to the inserted element or is deleted if
10117 // this is constant idx extract element and its a constant idx insertelt.
10118 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10119 isa<ConstantInt>(I->getOperand(2)))
10121 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10123 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10124 if (BO->hasOneUse() &&
10125 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10126 CheapToScalarize(BO->getOperand(1), isConstant)))
10128 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10129 if (CI->hasOneUse() &&
10130 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10131 CheapToScalarize(CI->getOperand(1), isConstant)))
10137 /// Read and decode a shufflevector mask.
10139 /// It turns undef elements into values that are larger than the number of
10140 /// elements in the input.
10141 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10142 unsigned NElts = SVI->getType()->getNumElements();
10143 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10144 return std::vector<unsigned>(NElts, 0);
10145 if (isa<UndefValue>(SVI->getOperand(2)))
10146 return std::vector<unsigned>(NElts, 2*NElts);
10148 std::vector<unsigned> Result;
10149 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10150 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
10151 if (isa<UndefValue>(CP->getOperand(i)))
10152 Result.push_back(NElts*2); // undef -> 8
10154 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
10158 /// FindScalarElement - Given a vector and an element number, see if the scalar
10159 /// value is already around as a register, for example if it were inserted then
10160 /// extracted from the vector.
10161 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10162 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10163 const VectorType *PTy = cast<VectorType>(V->getType());
10164 unsigned Width = PTy->getNumElements();
10165 if (EltNo >= Width) // Out of range access.
10166 return UndefValue::get(PTy->getElementType());
10168 if (isa<UndefValue>(V))
10169 return UndefValue::get(PTy->getElementType());
10170 else if (isa<ConstantAggregateZero>(V))
10171 return Constant::getNullValue(PTy->getElementType());
10172 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10173 return CP->getOperand(EltNo);
10174 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10175 // If this is an insert to a variable element, we don't know what it is.
10176 if (!isa<ConstantInt>(III->getOperand(2)))
10178 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10180 // If this is an insert to the element we are looking for, return the
10182 if (EltNo == IIElt)
10183 return III->getOperand(1);
10185 // Otherwise, the insertelement doesn't modify the value, recurse on its
10187 return FindScalarElement(III->getOperand(0), EltNo);
10188 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10189 unsigned InEl = getShuffleMask(SVI)[EltNo];
10191 return FindScalarElement(SVI->getOperand(0), InEl);
10192 else if (InEl < Width*2)
10193 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10195 return UndefValue::get(PTy->getElementType());
10198 // Otherwise, we don't know.
10202 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10204 // If vector val is undef, replace extract with scalar undef.
10205 if (isa<UndefValue>(EI.getOperand(0)))
10206 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10208 // If vector val is constant 0, replace extract with scalar 0.
10209 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10210 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10212 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10213 // If vector val is constant with uniform operands, replace EI
10214 // with that operand
10215 Constant *op0 = C->getOperand(0);
10216 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10217 if (C->getOperand(i) != op0) {
10222 return ReplaceInstUsesWith(EI, op0);
10225 // If extracting a specified index from the vector, see if we can recursively
10226 // find a previously computed scalar that was inserted into the vector.
10227 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10228 unsigned IndexVal = IdxC->getZExtValue();
10229 unsigned VectorWidth =
10230 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10232 // If this is extracting an invalid index, turn this into undef, to avoid
10233 // crashing the code below.
10234 if (IndexVal >= VectorWidth)
10235 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10237 // This instruction only demands the single element from the input vector.
10238 // If the input vector has a single use, simplify it based on this use
10240 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10241 uint64_t UndefElts;
10242 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10245 EI.setOperand(0, V);
10250 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10251 return ReplaceInstUsesWith(EI, Elt);
10253 // If the this extractelement is directly using a bitcast from a vector of
10254 // the same number of elements, see if we can find the source element from
10255 // it. In this case, we will end up needing to bitcast the scalars.
10256 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10257 if (const VectorType *VT =
10258 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10259 if (VT->getNumElements() == VectorWidth)
10260 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10261 return new BitCastInst(Elt, EI.getType());
10265 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10266 if (I->hasOneUse()) {
10267 // Push extractelement into predecessor operation if legal and
10268 // profitable to do so
10269 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10270 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10271 if (CheapToScalarize(BO, isConstantElt)) {
10272 ExtractElementInst *newEI0 =
10273 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10274 EI.getName()+".lhs");
10275 ExtractElementInst *newEI1 =
10276 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10277 EI.getName()+".rhs");
10278 InsertNewInstBefore(newEI0, EI);
10279 InsertNewInstBefore(newEI1, EI);
10280 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10282 } else if (isa<LoadInst>(I)) {
10284 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10285 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10286 PointerType::get(EI.getType(), AS),EI);
10287 GetElementPtrInst *GEP =
10288 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
10289 InsertNewInstBefore(GEP, EI);
10290 return new LoadInst(GEP);
10293 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10294 // Extracting the inserted element?
10295 if (IE->getOperand(2) == EI.getOperand(1))
10296 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10297 // If the inserted and extracted elements are constants, they must not
10298 // be the same value, extract from the pre-inserted value instead.
10299 if (isa<Constant>(IE->getOperand(2)) &&
10300 isa<Constant>(EI.getOperand(1))) {
10301 AddUsesToWorkList(EI);
10302 EI.setOperand(0, IE->getOperand(0));
10305 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10306 // If this is extracting an element from a shufflevector, figure out where
10307 // it came from and extract from the appropriate input element instead.
10308 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10309 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10311 if (SrcIdx < SVI->getType()->getNumElements())
10312 Src = SVI->getOperand(0);
10313 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10314 SrcIdx -= SVI->getType()->getNumElements();
10315 Src = SVI->getOperand(1);
10317 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10319 return new ExtractElementInst(Src, SrcIdx);
10326 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10327 /// elements from either LHS or RHS, return the shuffle mask and true.
10328 /// Otherwise, return false.
10329 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10330 std::vector<Constant*> &Mask) {
10331 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10332 "Invalid CollectSingleShuffleElements");
10333 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10335 if (isa<UndefValue>(V)) {
10336 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10338 } else if (V == LHS) {
10339 for (unsigned i = 0; i != NumElts; ++i)
10340 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10342 } else if (V == RHS) {
10343 for (unsigned i = 0; i != NumElts; ++i)
10344 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10346 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10347 // If this is an insert of an extract from some other vector, include it.
10348 Value *VecOp = IEI->getOperand(0);
10349 Value *ScalarOp = IEI->getOperand(1);
10350 Value *IdxOp = IEI->getOperand(2);
10352 if (!isa<ConstantInt>(IdxOp))
10354 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10356 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10357 // Okay, we can handle this if the vector we are insertinting into is
10358 // transitively ok.
10359 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10360 // If so, update the mask to reflect the inserted undef.
10361 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10364 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10365 if (isa<ConstantInt>(EI->getOperand(1)) &&
10366 EI->getOperand(0)->getType() == V->getType()) {
10367 unsigned ExtractedIdx =
10368 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10370 // This must be extracting from either LHS or RHS.
10371 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10372 // Okay, we can handle this if the vector we are insertinting into is
10373 // transitively ok.
10374 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10375 // If so, update the mask to reflect the inserted value.
10376 if (EI->getOperand(0) == LHS) {
10377 Mask[InsertedIdx & (NumElts-1)] =
10378 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10380 assert(EI->getOperand(0) == RHS);
10381 Mask[InsertedIdx & (NumElts-1)] =
10382 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10391 // TODO: Handle shufflevector here!
10396 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10397 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10398 /// that computes V and the LHS value of the shuffle.
10399 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10401 assert(isa<VectorType>(V->getType()) &&
10402 (RHS == 0 || V->getType() == RHS->getType()) &&
10403 "Invalid shuffle!");
10404 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10406 if (isa<UndefValue>(V)) {
10407 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10409 } else if (isa<ConstantAggregateZero>(V)) {
10410 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10412 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10413 // If this is an insert of an extract from some other vector, include it.
10414 Value *VecOp = IEI->getOperand(0);
10415 Value *ScalarOp = IEI->getOperand(1);
10416 Value *IdxOp = IEI->getOperand(2);
10418 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10419 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10420 EI->getOperand(0)->getType() == V->getType()) {
10421 unsigned ExtractedIdx =
10422 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10423 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10425 // Either the extracted from or inserted into vector must be RHSVec,
10426 // otherwise we'd end up with a shuffle of three inputs.
10427 if (EI->getOperand(0) == RHS || RHS == 0) {
10428 RHS = EI->getOperand(0);
10429 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10430 Mask[InsertedIdx & (NumElts-1)] =
10431 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10435 if (VecOp == RHS) {
10436 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10437 // Everything but the extracted element is replaced with the RHS.
10438 for (unsigned i = 0; i != NumElts; ++i) {
10439 if (i != InsertedIdx)
10440 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10445 // If this insertelement is a chain that comes from exactly these two
10446 // vectors, return the vector and the effective shuffle.
10447 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10448 return EI->getOperand(0);
10453 // TODO: Handle shufflevector here!
10455 // Otherwise, can't do anything fancy. Return an identity vector.
10456 for (unsigned i = 0; i != NumElts; ++i)
10457 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10461 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10462 Value *VecOp = IE.getOperand(0);
10463 Value *ScalarOp = IE.getOperand(1);
10464 Value *IdxOp = IE.getOperand(2);
10466 // Inserting an undef or into an undefined place, remove this.
10467 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10468 ReplaceInstUsesWith(IE, VecOp);
10470 // If the inserted element was extracted from some other vector, and if the
10471 // indexes are constant, try to turn this into a shufflevector operation.
10472 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10473 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10474 EI->getOperand(0)->getType() == IE.getType()) {
10475 unsigned NumVectorElts = IE.getType()->getNumElements();
10476 unsigned ExtractedIdx =
10477 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10478 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10480 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10481 return ReplaceInstUsesWith(IE, VecOp);
10483 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10484 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10486 // If we are extracting a value from a vector, then inserting it right
10487 // back into the same place, just use the input vector.
10488 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10489 return ReplaceInstUsesWith(IE, VecOp);
10491 // We could theoretically do this for ANY input. However, doing so could
10492 // turn chains of insertelement instructions into a chain of shufflevector
10493 // instructions, and right now we do not merge shufflevectors. As such,
10494 // only do this in a situation where it is clear that there is benefit.
10495 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10496 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10497 // the values of VecOp, except then one read from EIOp0.
10498 // Build a new shuffle mask.
10499 std::vector<Constant*> Mask;
10500 if (isa<UndefValue>(VecOp))
10501 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10503 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10504 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10507 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10508 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10509 ConstantVector::get(Mask));
10512 // If this insertelement isn't used by some other insertelement, turn it
10513 // (and any insertelements it points to), into one big shuffle.
10514 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10515 std::vector<Constant*> Mask;
10517 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10518 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10519 // We now have a shuffle of LHS, RHS, Mask.
10520 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10529 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10530 Value *LHS = SVI.getOperand(0);
10531 Value *RHS = SVI.getOperand(1);
10532 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10534 bool MadeChange = false;
10536 // Undefined shuffle mask -> undefined value.
10537 if (isa<UndefValue>(SVI.getOperand(2)))
10538 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10540 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10541 // the undef, change them to undefs.
10542 if (isa<UndefValue>(SVI.getOperand(1))) {
10543 // Scan to see if there are any references to the RHS. If so, replace them
10544 // with undef element refs and set MadeChange to true.
10545 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10546 if (Mask[i] >= e && Mask[i] != 2*e) {
10553 // Remap any references to RHS to use LHS.
10554 std::vector<Constant*> Elts;
10555 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10556 if (Mask[i] == 2*e)
10557 Elts.push_back(UndefValue::get(Type::Int32Ty));
10559 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10561 SVI.setOperand(2, ConstantVector::get(Elts));
10565 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10566 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10567 if (LHS == RHS || isa<UndefValue>(LHS)) {
10568 if (isa<UndefValue>(LHS) && LHS == RHS) {
10569 // shuffle(undef,undef,mask) -> undef.
10570 return ReplaceInstUsesWith(SVI, LHS);
10573 // Remap any references to RHS to use LHS.
10574 std::vector<Constant*> Elts;
10575 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10576 if (Mask[i] >= 2*e)
10577 Elts.push_back(UndefValue::get(Type::Int32Ty));
10579 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10580 (Mask[i] < e && isa<UndefValue>(LHS)))
10581 Mask[i] = 2*e; // Turn into undef.
10583 Mask[i] &= (e-1); // Force to LHS.
10584 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10587 SVI.setOperand(0, SVI.getOperand(1));
10588 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10589 SVI.setOperand(2, ConstantVector::get(Elts));
10590 LHS = SVI.getOperand(0);
10591 RHS = SVI.getOperand(1);
10595 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10596 bool isLHSID = true, isRHSID = true;
10598 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10599 if (Mask[i] >= e*2) continue; // Ignore undef values.
10600 // Is this an identity shuffle of the LHS value?
10601 isLHSID &= (Mask[i] == i);
10603 // Is this an identity shuffle of the RHS value?
10604 isRHSID &= (Mask[i]-e == i);
10607 // Eliminate identity shuffles.
10608 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10609 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10611 // If the LHS is a shufflevector itself, see if we can combine it with this
10612 // one without producing an unusual shuffle. Here we are really conservative:
10613 // we are absolutely afraid of producing a shuffle mask not in the input
10614 // program, because the code gen may not be smart enough to turn a merged
10615 // shuffle into two specific shuffles: it may produce worse code. As such,
10616 // we only merge two shuffles if the result is one of the two input shuffle
10617 // masks. In this case, merging the shuffles just removes one instruction,
10618 // which we know is safe. This is good for things like turning:
10619 // (splat(splat)) -> splat.
10620 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10621 if (isa<UndefValue>(RHS)) {
10622 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10624 std::vector<unsigned> NewMask;
10625 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10626 if (Mask[i] >= 2*e)
10627 NewMask.push_back(2*e);
10629 NewMask.push_back(LHSMask[Mask[i]]);
10631 // If the result mask is equal to the src shuffle or this shuffle mask, do
10632 // the replacement.
10633 if (NewMask == LHSMask || NewMask == Mask) {
10634 std::vector<Constant*> Elts;
10635 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10636 if (NewMask[i] >= e*2) {
10637 Elts.push_back(UndefValue::get(Type::Int32Ty));
10639 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10642 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10643 LHSSVI->getOperand(1),
10644 ConstantVector::get(Elts));
10649 return MadeChange ? &SVI : 0;
10655 /// TryToSinkInstruction - Try to move the specified instruction from its
10656 /// current block into the beginning of DestBlock, which can only happen if it's
10657 /// safe to move the instruction past all of the instructions between it and the
10658 /// end of its block.
10659 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10660 assert(I->hasOneUse() && "Invariants didn't hold!");
10662 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10663 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10665 // Do not sink alloca instructions out of the entry block.
10666 if (isa<AllocaInst>(I) && I->getParent() ==
10667 &DestBlock->getParent()->getEntryBlock())
10670 // We can only sink load instructions if there is nothing between the load and
10671 // the end of block that could change the value.
10672 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10673 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10675 if (Scan->mayWriteToMemory())
10679 BasicBlock::iterator InsertPos = DestBlock->begin();
10680 while (isa<PHINode>(InsertPos)) ++InsertPos;
10682 I->moveBefore(InsertPos);
10688 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10689 /// all reachable code to the worklist.
10691 /// This has a couple of tricks to make the code faster and more powerful. In
10692 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10693 /// them to the worklist (this significantly speeds up instcombine on code where
10694 /// many instructions are dead or constant). Additionally, if we find a branch
10695 /// whose condition is a known constant, we only visit the reachable successors.
10697 static void AddReachableCodeToWorklist(BasicBlock *BB,
10698 SmallPtrSet<BasicBlock*, 64> &Visited,
10700 const TargetData *TD) {
10701 std::vector<BasicBlock*> Worklist;
10702 Worklist.push_back(BB);
10704 while (!Worklist.empty()) {
10705 BB = Worklist.back();
10706 Worklist.pop_back();
10708 // We have now visited this block! If we've already been here, ignore it.
10709 if (!Visited.insert(BB)) continue;
10711 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10712 Instruction *Inst = BBI++;
10714 // DCE instruction if trivially dead.
10715 if (isInstructionTriviallyDead(Inst)) {
10717 DOUT << "IC: DCE: " << *Inst;
10718 Inst->eraseFromParent();
10722 // ConstantProp instruction if trivially constant.
10723 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10724 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10725 Inst->replaceAllUsesWith(C);
10727 Inst->eraseFromParent();
10731 IC.AddToWorkList(Inst);
10734 // Recursively visit successors. If this is a branch or switch on a
10735 // constant, only visit the reachable successor.
10736 TerminatorInst *TI = BB->getTerminator();
10737 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10738 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10739 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10740 Worklist.push_back(BI->getSuccessor(!CondVal));
10743 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10744 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10745 // See if this is an explicit destination.
10746 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10747 if (SI->getCaseValue(i) == Cond) {
10748 Worklist.push_back(SI->getSuccessor(i));
10752 // Otherwise it is the default destination.
10753 Worklist.push_back(SI->getSuccessor(0));
10758 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10759 Worklist.push_back(TI->getSuccessor(i));
10763 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10764 bool Changed = false;
10765 TD = &getAnalysis<TargetData>();
10767 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10768 << F.getNameStr() << "\n");
10771 // Do a depth-first traversal of the function, populate the worklist with
10772 // the reachable instructions. Ignore blocks that are not reachable. Keep
10773 // track of which blocks we visit.
10774 SmallPtrSet<BasicBlock*, 64> Visited;
10775 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10777 // Do a quick scan over the function. If we find any blocks that are
10778 // unreachable, remove any instructions inside of them. This prevents
10779 // the instcombine code from having to deal with some bad special cases.
10780 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10781 if (!Visited.count(BB)) {
10782 Instruction *Term = BB->getTerminator();
10783 while (Term != BB->begin()) { // Remove instrs bottom-up
10784 BasicBlock::iterator I = Term; --I;
10786 DOUT << "IC: DCE: " << *I;
10789 if (!I->use_empty())
10790 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10791 I->eraseFromParent();
10796 while (!Worklist.empty()) {
10797 Instruction *I = RemoveOneFromWorkList();
10798 if (I == 0) continue; // skip null values.
10800 // Check to see if we can DCE the instruction.
10801 if (isInstructionTriviallyDead(I)) {
10802 // Add operands to the worklist.
10803 if (I->getNumOperands() < 4)
10804 AddUsesToWorkList(*I);
10807 DOUT << "IC: DCE: " << *I;
10809 I->eraseFromParent();
10810 RemoveFromWorkList(I);
10814 // Instruction isn't dead, see if we can constant propagate it.
10815 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10816 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10818 // Add operands to the worklist.
10819 AddUsesToWorkList(*I);
10820 ReplaceInstUsesWith(*I, C);
10823 I->eraseFromParent();
10824 RemoveFromWorkList(I);
10828 // See if we can trivially sink this instruction to a successor basic block.
10829 if (I->hasOneUse()) {
10830 BasicBlock *BB = I->getParent();
10831 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10832 if (UserParent != BB) {
10833 bool UserIsSuccessor = false;
10834 // See if the user is one of our successors.
10835 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10836 if (*SI == UserParent) {
10837 UserIsSuccessor = true;
10841 // If the user is one of our immediate successors, and if that successor
10842 // only has us as a predecessors (we'd have to split the critical edge
10843 // otherwise), we can keep going.
10844 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10845 next(pred_begin(UserParent)) == pred_end(UserParent))
10846 // Okay, the CFG is simple enough, try to sink this instruction.
10847 Changed |= TryToSinkInstruction(I, UserParent);
10851 // Now that we have an instruction, try combining it to simplify it...
10855 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10856 if (Instruction *Result = visit(*I)) {
10858 // Should we replace the old instruction with a new one?
10860 DOUT << "IC: Old = " << *I
10861 << " New = " << *Result;
10863 // Everything uses the new instruction now.
10864 I->replaceAllUsesWith(Result);
10866 // Push the new instruction and any users onto the worklist.
10867 AddToWorkList(Result);
10868 AddUsersToWorkList(*Result);
10870 // Move the name to the new instruction first.
10871 Result->takeName(I);
10873 // Insert the new instruction into the basic block...
10874 BasicBlock *InstParent = I->getParent();
10875 BasicBlock::iterator InsertPos = I;
10877 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10878 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10881 InstParent->getInstList().insert(InsertPos, Result);
10883 // Make sure that we reprocess all operands now that we reduced their
10885 AddUsesToWorkList(*I);
10887 // Instructions can end up on the worklist more than once. Make sure
10888 // we do not process an instruction that has been deleted.
10889 RemoveFromWorkList(I);
10891 // Erase the old instruction.
10892 InstParent->getInstList().erase(I);
10895 DOUT << "IC: Mod = " << OrigI
10896 << " New = " << *I;
10899 // If the instruction was modified, it's possible that it is now dead.
10900 // if so, remove it.
10901 if (isInstructionTriviallyDead(I)) {
10902 // Make sure we process all operands now that we are reducing their
10904 AddUsesToWorkList(*I);
10906 // Instructions may end up in the worklist more than once. Erase all
10907 // occurrences of this instruction.
10908 RemoveFromWorkList(I);
10909 I->eraseFromParent();
10912 AddUsersToWorkList(*I);
10919 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10921 // Do an explicit clear, this shrinks the map if needed.
10922 WorklistMap.clear();
10927 bool InstCombiner::runOnFunction(Function &F) {
10928 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10930 bool EverMadeChange = false;
10932 // Iterate while there is work to do.
10933 unsigned Iteration = 0;
10934 while (DoOneIteration(F, Iteration++))
10935 EverMadeChange = true;
10936 return EverMadeChange;
10939 FunctionPass *llvm::createInstructionCombiningPass() {
10940 return new InstCombiner();