1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Target/TargetData.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Support/CallSite.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/GetElementPtrTypeIterator.h"
48 #include "llvm/Support/InstVisitor.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/PatternMatch.h"
51 #include "llvm/Support/Compiler.h"
52 #include "llvm/ADT/Statistic.h"
53 #include "llvm/ADT/STLExtras.h"
56 using namespace llvm::PatternMatch;
58 STATISTIC(NumCombined , "Number of insts combined");
59 STATISTIC(NumConstProp, "Number of constant folds");
60 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
61 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
62 STATISTIC(NumSunkInst , "Number of instructions sunk");
65 class VISIBILITY_HIDDEN InstCombiner
66 : public FunctionPass,
67 public InstVisitor<InstCombiner, Instruction*> {
68 // Worklist of all of the instructions that need to be simplified.
69 std::vector<Instruction*> WorkList;
72 /// AddUsersToWorkList - When an instruction is simplified, add all users of
73 /// the instruction to the work lists because they might get more simplified
76 void AddUsersToWorkList(Value &I) {
77 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
79 WorkList.push_back(cast<Instruction>(*UI));
82 /// AddUsesToWorkList - When an instruction is simplified, add operands to
83 /// the work lists because they might get more simplified now.
85 void AddUsesToWorkList(Instruction &I) {
86 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
87 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
88 WorkList.push_back(Op);
91 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
92 /// dead. Add all of its operands to the worklist, turning them into
93 /// undef's to reduce the number of uses of those instructions.
95 /// Return the specified operand before it is turned into an undef.
97 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
98 Value *R = I.getOperand(op);
100 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
101 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
102 WorkList.push_back(Op);
103 // Set the operand to undef to drop the use.
104 I.setOperand(i, UndefValue::get(Op->getType()));
110 // removeFromWorkList - remove all instances of I from the worklist.
111 void removeFromWorkList(Instruction *I);
113 virtual bool runOnFunction(Function &F);
115 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
116 AU.addRequired<TargetData>();
117 AU.addPreservedID(LCSSAID);
118 AU.setPreservesCFG();
121 TargetData &getTargetData() const { return *TD; }
123 // Visitation implementation - Implement instruction combining for different
124 // instruction types. The semantics are as follows:
126 // null - No change was made
127 // I - Change was made, I is still valid, I may be dead though
128 // otherwise - Change was made, replace I with returned instruction
130 Instruction *visitAdd(BinaryOperator &I);
131 Instruction *visitSub(BinaryOperator &I);
132 Instruction *visitMul(BinaryOperator &I);
133 Instruction *visitURem(BinaryOperator &I);
134 Instruction *visitSRem(BinaryOperator &I);
135 Instruction *visitFRem(BinaryOperator &I);
136 Instruction *commonRemTransforms(BinaryOperator &I);
137 Instruction *commonIRemTransforms(BinaryOperator &I);
138 Instruction *commonDivTransforms(BinaryOperator &I);
139 Instruction *commonIDivTransforms(BinaryOperator &I);
140 Instruction *visitUDiv(BinaryOperator &I);
141 Instruction *visitSDiv(BinaryOperator &I);
142 Instruction *visitFDiv(BinaryOperator &I);
143 Instruction *visitAnd(BinaryOperator &I);
144 Instruction *visitOr (BinaryOperator &I);
145 Instruction *visitXor(BinaryOperator &I);
146 Instruction *visitFCmpInst(FCmpInst &I);
147 Instruction *visitICmpInst(ICmpInst &I);
148 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
150 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
151 ICmpInst::Predicate Cond, Instruction &I);
152 Instruction *visitShiftInst(ShiftInst &I);
153 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
155 Instruction *commonCastTransforms(CastInst &CI);
156 Instruction *commonIntCastTransforms(CastInst &CI);
157 Instruction *visitTrunc(CastInst &CI);
158 Instruction *visitZExt(CastInst &CI);
159 Instruction *visitSExt(CastInst &CI);
160 Instruction *visitFPTrunc(CastInst &CI);
161 Instruction *visitFPExt(CastInst &CI);
162 Instruction *visitFPToUI(CastInst &CI);
163 Instruction *visitFPToSI(CastInst &CI);
164 Instruction *visitUIToFP(CastInst &CI);
165 Instruction *visitSIToFP(CastInst &CI);
166 Instruction *visitPtrToInt(CastInst &CI);
167 Instruction *visitIntToPtr(CastInst &CI);
168 Instruction *visitBitCast(CastInst &CI);
169 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
171 Instruction *visitSelectInst(SelectInst &CI);
172 Instruction *visitCallInst(CallInst &CI);
173 Instruction *visitInvokeInst(InvokeInst &II);
174 Instruction *visitPHINode(PHINode &PN);
175 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
176 Instruction *visitAllocationInst(AllocationInst &AI);
177 Instruction *visitFreeInst(FreeInst &FI);
178 Instruction *visitLoadInst(LoadInst &LI);
179 Instruction *visitStoreInst(StoreInst &SI);
180 Instruction *visitBranchInst(BranchInst &BI);
181 Instruction *visitSwitchInst(SwitchInst &SI);
182 Instruction *visitInsertElementInst(InsertElementInst &IE);
183 Instruction *visitExtractElementInst(ExtractElementInst &EI);
184 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
186 // visitInstruction - Specify what to return for unhandled instructions...
187 Instruction *visitInstruction(Instruction &I) { return 0; }
190 Instruction *visitCallSite(CallSite CS);
191 bool transformConstExprCastCall(CallSite CS);
194 // InsertNewInstBefore - insert an instruction New before instruction Old
195 // in the program. Add the new instruction to the worklist.
197 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
198 assert(New && New->getParent() == 0 &&
199 "New instruction already inserted into a basic block!");
200 BasicBlock *BB = Old.getParent();
201 BB->getInstList().insert(&Old, New); // Insert inst
202 WorkList.push_back(New); // Add to worklist
206 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
207 /// This also adds the cast to the worklist. Finally, this returns the
209 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
211 if (V->getType() == Ty) return V;
213 if (Constant *CV = dyn_cast<Constant>(V))
214 return ConstantExpr::getCast(opc, CV, Ty);
216 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
217 WorkList.push_back(C);
221 // ReplaceInstUsesWith - This method is to be used when an instruction is
222 // found to be dead, replacable with another preexisting expression. Here
223 // we add all uses of I to the worklist, replace all uses of I with the new
224 // value, then return I, so that the inst combiner will know that I was
227 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
228 AddUsersToWorkList(I); // Add all modified instrs to worklist
230 I.replaceAllUsesWith(V);
233 // If we are replacing the instruction with itself, this must be in a
234 // segment of unreachable code, so just clobber the instruction.
235 I.replaceAllUsesWith(UndefValue::get(I.getType()));
240 // UpdateValueUsesWith - This method is to be used when an value is
241 // found to be replacable with another preexisting expression or was
242 // updated. Here we add all uses of I to the worklist, replace all uses of
243 // I with the new value (unless the instruction was just updated), then
244 // return true, so that the inst combiner will know that I was modified.
246 bool UpdateValueUsesWith(Value *Old, Value *New) {
247 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
249 Old->replaceAllUsesWith(New);
250 if (Instruction *I = dyn_cast<Instruction>(Old))
251 WorkList.push_back(I);
252 if (Instruction *I = dyn_cast<Instruction>(New))
253 WorkList.push_back(I);
257 // EraseInstFromFunction - When dealing with an instruction that has side
258 // effects or produces a void value, we can't rely on DCE to delete the
259 // instruction. Instead, visit methods should return the value returned by
261 Instruction *EraseInstFromFunction(Instruction &I) {
262 assert(I.use_empty() && "Cannot erase instruction that is used!");
263 AddUsesToWorkList(I);
264 removeFromWorkList(&I);
266 return 0; // Don't do anything with FI
270 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
271 /// InsertBefore instruction. This is specialized a bit to avoid inserting
272 /// casts that are known to not do anything...
274 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
275 Value *V, const Type *DestTy,
276 Instruction *InsertBefore);
278 /// SimplifyCommutative - This performs a few simplifications for
279 /// commutative operators.
280 bool SimplifyCommutative(BinaryOperator &I);
282 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
283 /// most-complex to least-complex order.
284 bool SimplifyCompare(CmpInst &I);
286 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
287 uint64_t &KnownZero, uint64_t &KnownOne,
290 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
291 uint64_t &UndefElts, unsigned Depth = 0);
293 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
294 // PHI node as operand #0, see if we can fold the instruction into the PHI
295 // (which is only possible if all operands to the PHI are constants).
296 Instruction *FoldOpIntoPhi(Instruction &I);
298 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
299 // operator and they all are only used by the PHI, PHI together their
300 // inputs, and do the operation once, to the result of the PHI.
301 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
302 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
305 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
306 ConstantInt *AndRHS, BinaryOperator &TheAnd);
308 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
309 bool isSub, Instruction &I);
310 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
311 bool isSigned, bool Inside, Instruction &IB);
312 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
313 Instruction *MatchBSwap(BinaryOperator &I);
315 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
318 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
321 // getComplexity: Assign a complexity or rank value to LLVM Values...
322 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
323 static unsigned getComplexity(Value *V) {
324 if (isa<Instruction>(V)) {
325 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
329 if (isa<Argument>(V)) return 3;
330 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
333 // isOnlyUse - Return true if this instruction will be deleted if we stop using
335 static bool isOnlyUse(Value *V) {
336 return V->hasOneUse() || isa<Constant>(V);
339 // getPromotedType - Return the specified type promoted as it would be to pass
340 // though a va_arg area...
341 static const Type *getPromotedType(const Type *Ty) {
342 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
343 if (ITy->getBitWidth() < 32)
344 return Type::Int32Ty;
345 } else if (Ty == Type::FloatTy)
346 return Type::DoubleTy;
350 /// getBitCastOperand - If the specified operand is a CastInst or a constant
351 /// expression bitcast, return the operand value, otherwise return null.
352 static Value *getBitCastOperand(Value *V) {
353 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
354 return I->getOperand(0);
355 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
356 if (CE->getOpcode() == Instruction::BitCast)
357 return CE->getOperand(0);
361 /// This function is a wrapper around CastInst::isEliminableCastPair. It
362 /// simply extracts arguments and returns what that function returns.
363 /// @Determine if it is valid to eliminate a Convert pair
364 static Instruction::CastOps
365 isEliminableCastPair(
366 const CastInst *CI, ///< The first cast instruction
367 unsigned opcode, ///< The opcode of the second cast instruction
368 const Type *DstTy, ///< The target type for the second cast instruction
369 TargetData *TD ///< The target data for pointer size
372 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
373 const Type *MidTy = CI->getType(); // B from above
375 // Get the opcodes of the two Cast instructions
376 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
377 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
379 return Instruction::CastOps(
380 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
381 DstTy, TD->getIntPtrType()));
384 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
385 /// in any code being generated. It does not require codegen if V is simple
386 /// enough or if the cast can be folded into other casts.
387 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
388 const Type *Ty, TargetData *TD) {
389 if (V->getType() == Ty || isa<Constant>(V)) return false;
391 // If this is another cast that can be eliminated, it isn't codegen either.
392 if (const CastInst *CI = dyn_cast<CastInst>(V))
393 if (isEliminableCastPair(CI, opcode, Ty, TD))
398 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
399 /// InsertBefore instruction. This is specialized a bit to avoid inserting
400 /// casts that are known to not do anything...
402 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
403 Value *V, const Type *DestTy,
404 Instruction *InsertBefore) {
405 if (V->getType() == DestTy) return V;
406 if (Constant *C = dyn_cast<Constant>(V))
407 return ConstantExpr::getCast(opcode, C, DestTy);
409 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
412 // SimplifyCommutative - This performs a few simplifications for commutative
415 // 1. Order operands such that they are listed from right (least complex) to
416 // left (most complex). This puts constants before unary operators before
419 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
420 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
422 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
423 bool Changed = false;
424 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
425 Changed = !I.swapOperands();
427 if (!I.isAssociative()) return Changed;
428 Instruction::BinaryOps Opcode = I.getOpcode();
429 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
430 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
431 if (isa<Constant>(I.getOperand(1))) {
432 Constant *Folded = ConstantExpr::get(I.getOpcode(),
433 cast<Constant>(I.getOperand(1)),
434 cast<Constant>(Op->getOperand(1)));
435 I.setOperand(0, Op->getOperand(0));
436 I.setOperand(1, Folded);
438 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
439 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
440 isOnlyUse(Op) && isOnlyUse(Op1)) {
441 Constant *C1 = cast<Constant>(Op->getOperand(1));
442 Constant *C2 = cast<Constant>(Op1->getOperand(1));
444 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
445 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
446 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
449 WorkList.push_back(New);
450 I.setOperand(0, New);
451 I.setOperand(1, Folded);
458 /// SimplifyCompare - For a CmpInst this function just orders the operands
459 /// so that theyare listed from right (least complex) to left (most complex).
460 /// This puts constants before unary operators before binary operators.
461 bool InstCombiner::SimplifyCompare(CmpInst &I) {
462 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
465 // Compare instructions are not associative so there's nothing else we can do.
469 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
470 // if the LHS is a constant zero (which is the 'negate' form).
472 static inline Value *dyn_castNegVal(Value *V) {
473 if (BinaryOperator::isNeg(V))
474 return BinaryOperator::getNegArgument(V);
476 // Constants can be considered to be negated values if they can be folded.
477 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
478 return ConstantExpr::getNeg(C);
482 static inline Value *dyn_castNotVal(Value *V) {
483 if (BinaryOperator::isNot(V))
484 return BinaryOperator::getNotArgument(V);
486 // Constants can be considered to be not'ed values...
487 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
488 return ConstantExpr::getNot(C);
492 // dyn_castFoldableMul - If this value is a multiply that can be folded into
493 // other computations (because it has a constant operand), return the
494 // non-constant operand of the multiply, and set CST to point to the multiplier.
495 // Otherwise, return null.
497 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
498 if (V->hasOneUse() && V->getType()->isInteger())
499 if (Instruction *I = dyn_cast<Instruction>(V)) {
500 if (I->getOpcode() == Instruction::Mul)
501 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
502 return I->getOperand(0);
503 if (I->getOpcode() == Instruction::Shl)
504 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
505 // The multiplier is really 1 << CST.
506 Constant *One = ConstantInt::get(V->getType(), 1);
507 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
508 return I->getOperand(0);
514 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
515 /// expression, return it.
516 static User *dyn_castGetElementPtr(Value *V) {
517 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
518 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
519 if (CE->getOpcode() == Instruction::GetElementPtr)
520 return cast<User>(V);
524 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
525 static ConstantInt *AddOne(ConstantInt *C) {
526 return cast<ConstantInt>(ConstantExpr::getAdd(C,
527 ConstantInt::get(C->getType(), 1)));
529 static ConstantInt *SubOne(ConstantInt *C) {
530 return cast<ConstantInt>(ConstantExpr::getSub(C,
531 ConstantInt::get(C->getType(), 1)));
534 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
535 /// known to be either zero or one and return them in the KnownZero/KnownOne
536 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
538 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
539 uint64_t &KnownOne, unsigned Depth = 0) {
540 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
541 // we cannot optimize based on the assumption that it is zero without changing
542 // it to be an explicit zero. If we don't change it to zero, other code could
543 // optimized based on the contradictory assumption that it is non-zero.
544 // Because instcombine aggressively folds operations with undef args anyway,
545 // this won't lose us code quality.
546 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
547 // We know all of the bits for a constant!
548 KnownOne = CI->getZExtValue() & Mask;
549 KnownZero = ~KnownOne & Mask;
553 KnownZero = KnownOne = 0; // Don't know anything.
554 if (Depth == 6 || Mask == 0)
555 return; // Limit search depth.
557 uint64_t KnownZero2, KnownOne2;
558 Instruction *I = dyn_cast<Instruction>(V);
561 Mask &= V->getType()->getIntegralTypeMask();
563 switch (I->getOpcode()) {
564 case Instruction::And:
565 // If either the LHS or the RHS are Zero, the result is zero.
566 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
568 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
569 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
570 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
572 // Output known-1 bits are only known if set in both the LHS & RHS.
573 KnownOne &= KnownOne2;
574 // Output known-0 are known to be clear if zero in either the LHS | RHS.
575 KnownZero |= KnownZero2;
577 case Instruction::Or:
578 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
580 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
581 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
582 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
584 // Output known-0 bits are only known if clear in both the LHS & RHS.
585 KnownZero &= KnownZero2;
586 // Output known-1 are known to be set if set in either the LHS | RHS.
587 KnownOne |= KnownOne2;
589 case Instruction::Xor: {
590 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
591 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
592 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
593 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
595 // Output known-0 bits are known if clear or set in both the LHS & RHS.
596 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
597 // Output known-1 are known to be set if set in only one of the LHS, RHS.
598 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
599 KnownZero = KnownZeroOut;
602 case Instruction::Select:
603 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
604 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
605 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
606 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
608 // Only known if known in both the LHS and RHS.
609 KnownOne &= KnownOne2;
610 KnownZero &= KnownZero2;
612 case Instruction::FPTrunc:
613 case Instruction::FPExt:
614 case Instruction::FPToUI:
615 case Instruction::FPToSI:
616 case Instruction::SIToFP:
617 case Instruction::PtrToInt:
618 case Instruction::UIToFP:
619 case Instruction::IntToPtr:
620 return; // Can't work with floating point or pointers
621 case Instruction::Trunc:
622 // All these have integer operands
623 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
625 case Instruction::BitCast: {
626 const Type *SrcTy = I->getOperand(0)->getType();
627 if (SrcTy->isIntegral()) {
628 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
633 case Instruction::ZExt: {
634 // Compute the bits in the result that are not present in the input.
635 const Type *SrcTy = I->getOperand(0)->getType();
636 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
637 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
639 Mask &= SrcTy->getIntegralTypeMask();
640 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
641 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
642 // The top bits are known to be zero.
643 KnownZero |= NewBits;
646 case Instruction::SExt: {
647 // Compute the bits in the result that are not present in the input.
648 const Type *SrcTy = I->getOperand(0)->getType();
649 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
650 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
652 Mask &= SrcTy->getIntegralTypeMask();
653 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
654 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
656 // If the sign bit of the input is known set or clear, then we know the
657 // top bits of the result.
658 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
659 if (KnownZero & InSignBit) { // Input sign bit known zero
660 KnownZero |= NewBits;
661 KnownOne &= ~NewBits;
662 } else if (KnownOne & InSignBit) { // Input sign bit known set
664 KnownZero &= ~NewBits;
665 } else { // Input sign bit unknown
666 KnownZero &= ~NewBits;
667 KnownOne &= ~NewBits;
671 case Instruction::Shl:
672 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
673 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
674 uint64_t ShiftAmt = SA->getZExtValue();
676 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
677 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
678 KnownZero <<= ShiftAmt;
679 KnownOne <<= ShiftAmt;
680 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
684 case Instruction::LShr:
685 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
686 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
687 // Compute the new bits that are at the top now.
688 uint64_t ShiftAmt = SA->getZExtValue();
689 uint64_t HighBits = (1ULL << ShiftAmt)-1;
690 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
692 // Unsigned shift right.
694 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
695 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
696 KnownZero >>= ShiftAmt;
697 KnownOne >>= ShiftAmt;
698 KnownZero |= HighBits; // high bits known zero.
702 case Instruction::AShr:
703 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
704 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
705 // Compute the new bits that are at the top now.
706 uint64_t ShiftAmt = SA->getZExtValue();
707 uint64_t HighBits = (1ULL << ShiftAmt)-1;
708 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
710 // Signed shift right.
712 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
713 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
714 KnownZero >>= ShiftAmt;
715 KnownOne >>= ShiftAmt;
717 // Handle the sign bits.
718 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
719 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
721 if (KnownZero & SignBit) { // New bits are known zero.
722 KnownZero |= HighBits;
723 } else if (KnownOne & SignBit) { // New bits are known one.
724 KnownOne |= HighBits;
732 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
733 /// this predicate to simplify operations downstream. Mask is known to be zero
734 /// for bits that V cannot have.
735 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
736 uint64_t KnownZero, KnownOne;
737 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
738 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
739 return (KnownZero & Mask) == Mask;
742 /// ShrinkDemandedConstant - Check to see if the specified operand of the
743 /// specified instruction is a constant integer. If so, check to see if there
744 /// are any bits set in the constant that are not demanded. If so, shrink the
745 /// constant and return true.
746 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
748 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
749 if (!OpC) return false;
751 // If there are no bits set that aren't demanded, nothing to do.
752 if ((~Demanded & OpC->getZExtValue()) == 0)
755 // This is producing any bits that are not needed, shrink the RHS.
756 uint64_t Val = Demanded & OpC->getZExtValue();
757 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Val));
761 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
762 // set of known zero and one bits, compute the maximum and minimum values that
763 // could have the specified known zero and known one bits, returning them in
765 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
768 int64_t &Min, int64_t &Max) {
769 uint64_t TypeBits = Ty->getIntegralTypeMask();
770 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
772 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
774 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
775 // bit if it is unknown.
777 Max = KnownOne|UnknownBits;
779 if (SignBit & UnknownBits) { // Sign bit is unknown
784 // Sign extend the min/max values.
785 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
786 Min = (Min << ShAmt) >> ShAmt;
787 Max = (Max << ShAmt) >> ShAmt;
790 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
791 // a set of known zero and one bits, compute the maximum and minimum values that
792 // could have the specified known zero and known one bits, returning them in
794 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
799 uint64_t TypeBits = Ty->getIntegralTypeMask();
800 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
802 // The minimum value is when the unknown bits are all zeros.
804 // The maximum value is when the unknown bits are all ones.
805 Max = KnownOne|UnknownBits;
809 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
810 /// DemandedMask bits of the result of V are ever used downstream. If we can
811 /// use this information to simplify V, do so and return true. Otherwise,
812 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
813 /// the expression (used to simplify the caller). The KnownZero/One bits may
814 /// only be accurate for those bits in the DemandedMask.
815 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
816 uint64_t &KnownZero, uint64_t &KnownOne,
818 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
819 // We know all of the bits for a constant!
820 KnownOne = CI->getZExtValue() & DemandedMask;
821 KnownZero = ~KnownOne & DemandedMask;
825 KnownZero = KnownOne = 0;
826 if (!V->hasOneUse()) { // Other users may use these bits.
827 if (Depth != 0) { // Not at the root.
828 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
829 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
832 // If this is the root being simplified, allow it to have multiple uses,
833 // just set the DemandedMask to all bits.
834 DemandedMask = V->getType()->getIntegralTypeMask();
835 } else if (DemandedMask == 0) { // Not demanding any bits from V.
836 if (V != UndefValue::get(V->getType()))
837 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
839 } else if (Depth == 6) { // Limit search depth.
843 Instruction *I = dyn_cast<Instruction>(V);
844 if (!I) return false; // Only analyze instructions.
846 DemandedMask &= V->getType()->getIntegralTypeMask();
848 uint64_t KnownZero2 = 0, KnownOne2 = 0;
849 switch (I->getOpcode()) {
851 case Instruction::And:
852 // If either the LHS or the RHS are Zero, the result is zero.
853 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
854 KnownZero, KnownOne, Depth+1))
856 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
858 // If something is known zero on the RHS, the bits aren't demanded on the
860 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
861 KnownZero2, KnownOne2, Depth+1))
863 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
865 // If all of the demanded bits are known 1 on one side, return the other.
866 // These bits cannot contribute to the result of the 'and'.
867 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
868 return UpdateValueUsesWith(I, I->getOperand(0));
869 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
870 return UpdateValueUsesWith(I, I->getOperand(1));
872 // If all of the demanded bits in the inputs are known zeros, return zero.
873 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
874 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
876 // If the RHS is a constant, see if we can simplify it.
877 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
878 return UpdateValueUsesWith(I, I);
880 // Output known-1 bits are only known if set in both the LHS & RHS.
881 KnownOne &= KnownOne2;
882 // Output known-0 are known to be clear if zero in either the LHS | RHS.
883 KnownZero |= KnownZero2;
885 case Instruction::Or:
886 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
887 KnownZero, KnownOne, Depth+1))
889 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
890 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
891 KnownZero2, KnownOne2, Depth+1))
893 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
895 // If all of the demanded bits are known zero on one side, return the other.
896 // These bits cannot contribute to the result of the 'or'.
897 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
898 return UpdateValueUsesWith(I, I->getOperand(0));
899 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
900 return UpdateValueUsesWith(I, I->getOperand(1));
902 // If all of the potentially set bits on one side are known to be set on
903 // the other side, just use the 'other' side.
904 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
905 (DemandedMask & (~KnownZero)))
906 return UpdateValueUsesWith(I, I->getOperand(0));
907 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
908 (DemandedMask & (~KnownZero2)))
909 return UpdateValueUsesWith(I, I->getOperand(1));
911 // If the RHS is a constant, see if we can simplify it.
912 if (ShrinkDemandedConstant(I, 1, DemandedMask))
913 return UpdateValueUsesWith(I, I);
915 // Output known-0 bits are only known if clear in both the LHS & RHS.
916 KnownZero &= KnownZero2;
917 // Output known-1 are known to be set if set in either the LHS | RHS.
918 KnownOne |= KnownOne2;
920 case Instruction::Xor: {
921 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
922 KnownZero, KnownOne, Depth+1))
924 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
925 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
926 KnownZero2, KnownOne2, Depth+1))
928 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
930 // If all of the demanded bits are known zero on one side, return the other.
931 // These bits cannot contribute to the result of the 'xor'.
932 if ((DemandedMask & KnownZero) == DemandedMask)
933 return UpdateValueUsesWith(I, I->getOperand(0));
934 if ((DemandedMask & KnownZero2) == DemandedMask)
935 return UpdateValueUsesWith(I, I->getOperand(1));
937 // Output known-0 bits are known if clear or set in both the LHS & RHS.
938 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
939 // Output known-1 are known to be set if set in only one of the LHS, RHS.
940 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
942 // If all of the demanded bits are known to be zero on one side or the
943 // other, turn this into an *inclusive* or.
944 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
945 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
947 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
949 InsertNewInstBefore(Or, *I);
950 return UpdateValueUsesWith(I, Or);
953 // If all of the demanded bits on one side are known, and all of the set
954 // bits on that side are also known to be set on the other side, turn this
955 // into an AND, as we know the bits will be cleared.
956 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
957 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
958 if ((KnownOne & KnownOne2) == KnownOne) {
959 Constant *AndC = ConstantInt::get(I->getType(),
960 ~KnownOne & DemandedMask);
962 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
963 InsertNewInstBefore(And, *I);
964 return UpdateValueUsesWith(I, And);
968 // If the RHS is a constant, see if we can simplify it.
969 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
970 if (ShrinkDemandedConstant(I, 1, DemandedMask))
971 return UpdateValueUsesWith(I, I);
973 KnownZero = KnownZeroOut;
974 KnownOne = KnownOneOut;
977 case Instruction::Select:
978 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
979 KnownZero, KnownOne, Depth+1))
981 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
982 KnownZero2, KnownOne2, Depth+1))
984 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
985 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
987 // If the operands are constants, see if we can simplify them.
988 if (ShrinkDemandedConstant(I, 1, DemandedMask))
989 return UpdateValueUsesWith(I, I);
990 if (ShrinkDemandedConstant(I, 2, DemandedMask))
991 return UpdateValueUsesWith(I, I);
993 // Only known if known in both the LHS and RHS.
994 KnownOne &= KnownOne2;
995 KnownZero &= KnownZero2;
997 case Instruction::Trunc:
998 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
999 KnownZero, KnownOne, Depth+1))
1001 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1003 case Instruction::BitCast:
1004 if (!I->getOperand(0)->getType()->isIntegral())
1007 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1008 KnownZero, KnownOne, Depth+1))
1010 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1012 case Instruction::ZExt: {
1013 // Compute the bits in the result that are not present in the input.
1014 const Type *SrcTy = I->getOperand(0)->getType();
1015 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1016 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1018 DemandedMask &= SrcTy->getIntegralTypeMask();
1019 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1020 KnownZero, KnownOne, Depth+1))
1022 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1023 // The top bits are known to be zero.
1024 KnownZero |= NewBits;
1027 case Instruction::SExt: {
1028 // Compute the bits in the result that are not present in the input.
1029 const Type *SrcTy = I->getOperand(0)->getType();
1030 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1031 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1033 // Get the sign bit for the source type
1034 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1035 int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
1037 // If any of the sign extended bits are demanded, we know that the sign
1039 if (NewBits & DemandedMask)
1040 InputDemandedBits |= InSignBit;
1042 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1043 KnownZero, KnownOne, Depth+1))
1045 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1047 // If the sign bit of the input is known set or clear, then we know the
1048 // top bits of the result.
1050 // If the input sign bit is known zero, or if the NewBits are not demanded
1051 // convert this into a zero extension.
1052 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1053 // Convert to ZExt cast
1054 CastInst *NewCast = CastInst::create(
1055 Instruction::ZExt, I->getOperand(0), I->getType(), I->getName(), I);
1056 return UpdateValueUsesWith(I, NewCast);
1057 } else if (KnownOne & InSignBit) { // Input sign bit known set
1058 KnownOne |= NewBits;
1059 KnownZero &= ~NewBits;
1060 } else { // Input sign bit unknown
1061 KnownZero &= ~NewBits;
1062 KnownOne &= ~NewBits;
1066 case Instruction::Add:
1067 // If there is a constant on the RHS, there are a variety of xformations
1069 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1070 // If null, this should be simplified elsewhere. Some of the xforms here
1071 // won't work if the RHS is zero.
1072 if (RHS->isNullValue())
1075 // Figure out what the input bits are. If the top bits of the and result
1076 // are not demanded, then the add doesn't demand them from its input
1079 // Shift the demanded mask up so that it's at the top of the uint64_t.
1080 unsigned BitWidth = I->getType()->getPrimitiveSizeInBits();
1081 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1083 // If the top bit of the output is demanded, demand everything from the
1084 // input. Otherwise, we demand all the input bits except NLZ top bits.
1085 uint64_t InDemandedBits = ~0ULL >> (64-BitWidth+NLZ);
1087 // Find information about known zero/one bits in the input.
1088 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1089 KnownZero2, KnownOne2, Depth+1))
1092 // If the RHS of the add has bits set that can't affect the input, reduce
1094 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1095 return UpdateValueUsesWith(I, I);
1097 // Avoid excess work.
1098 if (KnownZero2 == 0 && KnownOne2 == 0)
1101 // Turn it into OR if input bits are zero.
1102 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1104 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1106 InsertNewInstBefore(Or, *I);
1107 return UpdateValueUsesWith(I, Or);
1110 // We can say something about the output known-zero and known-one bits,
1111 // depending on potential carries from the input constant and the
1112 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1113 // bits set and the RHS constant is 0x01001, then we know we have a known
1114 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1116 // To compute this, we first compute the potential carry bits. These are
1117 // the bits which may be modified. I'm not aware of a better way to do
1119 uint64_t RHSVal = RHS->getZExtValue();
1121 bool CarryIn = false;
1122 uint64_t CarryBits = 0;
1123 uint64_t CurBit = 1;
1124 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1125 // Record the current carry in.
1126 if (CarryIn) CarryBits |= CurBit;
1130 // This bit has a carry out unless it is "zero + zero" or
1131 // "zero + anything" with no carry in.
1132 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1133 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1134 } else if (!CarryIn &&
1135 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1136 CarryOut = false; // 0 + anything has no carry out if no carry in.
1138 // Otherwise, we have to assume we have a carry out.
1142 // This stage's carry out becomes the next stage's carry-in.
1146 // Now that we know which bits have carries, compute the known-1/0 sets.
1148 // Bits are known one if they are known zero in one operand and one in the
1149 // other, and there is no input carry.
1150 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1152 // Bits are known zero if they are known zero in both operands and there
1153 // is no input carry.
1154 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1157 case Instruction::Shl:
1158 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1159 uint64_t ShiftAmt = SA->getZExtValue();
1160 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1161 KnownZero, KnownOne, Depth+1))
1163 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1164 KnownZero <<= ShiftAmt;
1165 KnownOne <<= ShiftAmt;
1166 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1169 case Instruction::LShr:
1170 // For a logical shift right
1171 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1172 unsigned ShiftAmt = SA->getZExtValue();
1174 // Compute the new bits that are at the top now.
1175 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1176 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1177 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1178 // Unsigned shift right.
1179 if (SimplifyDemandedBits(I->getOperand(0),
1180 (DemandedMask << ShiftAmt) & TypeMask,
1181 KnownZero, KnownOne, Depth+1))
1183 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1184 KnownZero &= TypeMask;
1185 KnownOne &= TypeMask;
1186 KnownZero >>= ShiftAmt;
1187 KnownOne >>= ShiftAmt;
1188 KnownZero |= HighBits; // high bits known zero.
1191 case Instruction::AShr:
1192 // If this is an arithmetic shift right and only the low-bit is set, we can
1193 // always convert this into a logical shr, even if the shift amount is
1194 // variable. The low bit of the shift cannot be an input sign bit unless
1195 // the shift amount is >= the size of the datatype, which is undefined.
1196 if (DemandedMask == 1) {
1197 // Perform the logical shift right.
1198 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1199 I->getOperand(1), I->getName());
1200 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1201 return UpdateValueUsesWith(I, NewVal);
1204 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1205 unsigned ShiftAmt = SA->getZExtValue();
1207 // Compute the new bits that are at the top now.
1208 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1209 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1210 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1211 // Signed shift right.
1212 if (SimplifyDemandedBits(I->getOperand(0),
1213 (DemandedMask << ShiftAmt) & TypeMask,
1214 KnownZero, KnownOne, Depth+1))
1216 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1217 KnownZero &= TypeMask;
1218 KnownOne &= TypeMask;
1219 KnownZero >>= ShiftAmt;
1220 KnownOne >>= ShiftAmt;
1222 // Handle the sign bits.
1223 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1224 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1226 // If the input sign bit is known to be zero, or if none of the top bits
1227 // are demanded, turn this into an unsigned shift right.
1228 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1229 // Perform the logical shift right.
1230 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1232 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1233 return UpdateValueUsesWith(I, NewVal);
1234 } else if (KnownOne & SignBit) { // New bits are known one.
1235 KnownOne |= HighBits;
1241 // If the client is only demanding bits that we know, return the known
1243 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1244 return UpdateValueUsesWith(I, ConstantInt::get(I->getType(), KnownOne));
1249 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1250 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1251 /// actually used by the caller. This method analyzes which elements of the
1252 /// operand are undef and returns that information in UndefElts.
1254 /// If the information about demanded elements can be used to simplify the
1255 /// operation, the operation is simplified, then the resultant value is
1256 /// returned. This returns null if no change was made.
1257 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1258 uint64_t &UndefElts,
1260 unsigned VWidth = cast<PackedType>(V->getType())->getNumElements();
1261 assert(VWidth <= 64 && "Vector too wide to analyze!");
1262 uint64_t EltMask = ~0ULL >> (64-VWidth);
1263 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1264 "Invalid DemandedElts!");
1266 if (isa<UndefValue>(V)) {
1267 // If the entire vector is undefined, just return this info.
1268 UndefElts = EltMask;
1270 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1271 UndefElts = EltMask;
1272 return UndefValue::get(V->getType());
1276 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V)) {
1277 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1278 Constant *Undef = UndefValue::get(EltTy);
1280 std::vector<Constant*> Elts;
1281 for (unsigned i = 0; i != VWidth; ++i)
1282 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1283 Elts.push_back(Undef);
1284 UndefElts |= (1ULL << i);
1285 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1286 Elts.push_back(Undef);
1287 UndefElts |= (1ULL << i);
1288 } else { // Otherwise, defined.
1289 Elts.push_back(CP->getOperand(i));
1292 // If we changed the constant, return it.
1293 Constant *NewCP = ConstantPacked::get(Elts);
1294 return NewCP != CP ? NewCP : 0;
1295 } else if (isa<ConstantAggregateZero>(V)) {
1296 // Simplify the CAZ to a ConstantPacked where the non-demanded elements are
1298 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1299 Constant *Zero = Constant::getNullValue(EltTy);
1300 Constant *Undef = UndefValue::get(EltTy);
1301 std::vector<Constant*> Elts;
1302 for (unsigned i = 0; i != VWidth; ++i)
1303 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1304 UndefElts = DemandedElts ^ EltMask;
1305 return ConstantPacked::get(Elts);
1308 if (!V->hasOneUse()) { // Other users may use these bits.
1309 if (Depth != 0) { // Not at the root.
1310 // TODO: Just compute the UndefElts information recursively.
1314 } else if (Depth == 10) { // Limit search depth.
1318 Instruction *I = dyn_cast<Instruction>(V);
1319 if (!I) return false; // Only analyze instructions.
1321 bool MadeChange = false;
1322 uint64_t UndefElts2;
1324 switch (I->getOpcode()) {
1327 case Instruction::InsertElement: {
1328 // If this is a variable index, we don't know which element it overwrites.
1329 // demand exactly the same input as we produce.
1330 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1332 // Note that we can't propagate undef elt info, because we don't know
1333 // which elt is getting updated.
1334 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1335 UndefElts2, Depth+1);
1336 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1340 // If this is inserting an element that isn't demanded, remove this
1342 unsigned IdxNo = Idx->getZExtValue();
1343 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1344 return AddSoonDeadInstToWorklist(*I, 0);
1346 // Otherwise, the element inserted overwrites whatever was there, so the
1347 // input demanded set is simpler than the output set.
1348 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1349 DemandedElts & ~(1ULL << IdxNo),
1350 UndefElts, Depth+1);
1351 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1353 // The inserted element is defined.
1354 UndefElts |= 1ULL << IdxNo;
1358 case Instruction::And:
1359 case Instruction::Or:
1360 case Instruction::Xor:
1361 case Instruction::Add:
1362 case Instruction::Sub:
1363 case Instruction::Mul:
1364 // div/rem demand all inputs, because they don't want divide by zero.
1365 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1366 UndefElts, Depth+1);
1367 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1368 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1369 UndefElts2, Depth+1);
1370 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1372 // Output elements are undefined if both are undefined. Consider things
1373 // like undef&0. The result is known zero, not undef.
1374 UndefElts &= UndefElts2;
1377 case Instruction::Call: {
1378 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1380 switch (II->getIntrinsicID()) {
1383 // Binary vector operations that work column-wise. A dest element is a
1384 // function of the corresponding input elements from the two inputs.
1385 case Intrinsic::x86_sse_sub_ss:
1386 case Intrinsic::x86_sse_mul_ss:
1387 case Intrinsic::x86_sse_min_ss:
1388 case Intrinsic::x86_sse_max_ss:
1389 case Intrinsic::x86_sse2_sub_sd:
1390 case Intrinsic::x86_sse2_mul_sd:
1391 case Intrinsic::x86_sse2_min_sd:
1392 case Intrinsic::x86_sse2_max_sd:
1393 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1394 UndefElts, Depth+1);
1395 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1396 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1397 UndefElts2, Depth+1);
1398 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1400 // If only the low elt is demanded and this is a scalarizable intrinsic,
1401 // scalarize it now.
1402 if (DemandedElts == 1) {
1403 switch (II->getIntrinsicID()) {
1405 case Intrinsic::x86_sse_sub_ss:
1406 case Intrinsic::x86_sse_mul_ss:
1407 case Intrinsic::x86_sse2_sub_sd:
1408 case Intrinsic::x86_sse2_mul_sd:
1409 // TODO: Lower MIN/MAX/ABS/etc
1410 Value *LHS = II->getOperand(1);
1411 Value *RHS = II->getOperand(2);
1412 // Extract the element as scalars.
1413 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1414 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1416 switch (II->getIntrinsicID()) {
1417 default: assert(0 && "Case stmts out of sync!");
1418 case Intrinsic::x86_sse_sub_ss:
1419 case Intrinsic::x86_sse2_sub_sd:
1420 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1421 II->getName()), *II);
1423 case Intrinsic::x86_sse_mul_ss:
1424 case Intrinsic::x86_sse2_mul_sd:
1425 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1426 II->getName()), *II);
1431 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1433 InsertNewInstBefore(New, *II);
1434 AddSoonDeadInstToWorklist(*II, 0);
1439 // Output elements are undefined if both are undefined. Consider things
1440 // like undef&0. The result is known zero, not undef.
1441 UndefElts &= UndefElts2;
1447 return MadeChange ? I : 0;
1450 /// @returns true if the specified compare instruction is
1451 /// true when both operands are equal...
1452 /// @brief Determine if the ICmpInst returns true if both operands are equal
1453 static bool isTrueWhenEqual(ICmpInst &ICI) {
1454 ICmpInst::Predicate pred = ICI.getPredicate();
1455 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1456 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1457 pred == ICmpInst::ICMP_SLE;
1460 /// @returns true if the specified compare instruction is
1461 /// true when both operands are equal...
1462 /// @brief Determine if the FCmpInst returns true if both operands are equal
1463 static bool isTrueWhenEqual(FCmpInst &FCI) {
1464 FCmpInst::Predicate pred = FCI.getPredicate();
1465 return pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ ||
1466 pred == FCmpInst::FCMP_OGE || pred == FCmpInst::FCMP_UGE ||
1467 pred == FCmpInst::FCMP_OLE || pred == FCmpInst::FCMP_ULE;
1470 /// AssociativeOpt - Perform an optimization on an associative operator. This
1471 /// function is designed to check a chain of associative operators for a
1472 /// potential to apply a certain optimization. Since the optimization may be
1473 /// applicable if the expression was reassociated, this checks the chain, then
1474 /// reassociates the expression as necessary to expose the optimization
1475 /// opportunity. This makes use of a special Functor, which must define
1476 /// 'shouldApply' and 'apply' methods.
1478 template<typename Functor>
1479 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1480 unsigned Opcode = Root.getOpcode();
1481 Value *LHS = Root.getOperand(0);
1483 // Quick check, see if the immediate LHS matches...
1484 if (F.shouldApply(LHS))
1485 return F.apply(Root);
1487 // Otherwise, if the LHS is not of the same opcode as the root, return.
1488 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1489 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1490 // Should we apply this transform to the RHS?
1491 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1493 // If not to the RHS, check to see if we should apply to the LHS...
1494 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1495 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1499 // If the functor wants to apply the optimization to the RHS of LHSI,
1500 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1502 BasicBlock *BB = Root.getParent();
1504 // Now all of the instructions are in the current basic block, go ahead
1505 // and perform the reassociation.
1506 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1508 // First move the selected RHS to the LHS of the root...
1509 Root.setOperand(0, LHSI->getOperand(1));
1511 // Make what used to be the LHS of the root be the user of the root...
1512 Value *ExtraOperand = TmpLHSI->getOperand(1);
1513 if (&Root == TmpLHSI) {
1514 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1517 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1518 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1519 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1520 BasicBlock::iterator ARI = &Root; ++ARI;
1521 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1524 // Now propagate the ExtraOperand down the chain of instructions until we
1526 while (TmpLHSI != LHSI) {
1527 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1528 // Move the instruction to immediately before the chain we are
1529 // constructing to avoid breaking dominance properties.
1530 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1531 BB->getInstList().insert(ARI, NextLHSI);
1534 Value *NextOp = NextLHSI->getOperand(1);
1535 NextLHSI->setOperand(1, ExtraOperand);
1537 ExtraOperand = NextOp;
1540 // Now that the instructions are reassociated, have the functor perform
1541 // the transformation...
1542 return F.apply(Root);
1545 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1551 // AddRHS - Implements: X + X --> X << 1
1554 AddRHS(Value *rhs) : RHS(rhs) {}
1555 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1556 Instruction *apply(BinaryOperator &Add) const {
1557 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
1558 ConstantInt::get(Type::Int8Ty, 1));
1562 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1564 struct AddMaskingAnd {
1566 AddMaskingAnd(Constant *c) : C2(c) {}
1567 bool shouldApply(Value *LHS) const {
1569 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1570 ConstantExpr::getAnd(C1, C2)->isNullValue();
1572 Instruction *apply(BinaryOperator &Add) const {
1573 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1577 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1579 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1580 if (Constant *SOC = dyn_cast<Constant>(SO))
1581 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1583 return IC->InsertNewInstBefore(CastInst::create(
1584 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1587 // Figure out if the constant is the left or the right argument.
1588 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1589 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1591 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1593 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1594 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1597 Value *Op0 = SO, *Op1 = ConstOperand;
1599 std::swap(Op0, Op1);
1601 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1602 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1603 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1604 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1605 SO->getName()+".cmp");
1606 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1607 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
1609 assert(0 && "Unknown binary instruction type!");
1612 return IC->InsertNewInstBefore(New, I);
1615 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1616 // constant as the other operand, try to fold the binary operator into the
1617 // select arguments. This also works for Cast instructions, which obviously do
1618 // not have a second operand.
1619 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1621 // Don't modify shared select instructions
1622 if (!SI->hasOneUse()) return 0;
1623 Value *TV = SI->getOperand(1);
1624 Value *FV = SI->getOperand(2);
1626 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1627 // Bool selects with constant operands can be folded to logical ops.
1628 if (SI->getType() == Type::Int1Ty) return 0;
1630 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1631 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1633 return new SelectInst(SI->getCondition(), SelectTrueVal,
1640 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1641 /// node as operand #0, see if we can fold the instruction into the PHI (which
1642 /// is only possible if all operands to the PHI are constants).
1643 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1644 PHINode *PN = cast<PHINode>(I.getOperand(0));
1645 unsigned NumPHIValues = PN->getNumIncomingValues();
1646 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1648 // Check to see if all of the operands of the PHI are constants. If there is
1649 // one non-constant value, remember the BB it is. If there is more than one
1651 BasicBlock *NonConstBB = 0;
1652 for (unsigned i = 0; i != NumPHIValues; ++i)
1653 if (!isa<Constant>(PN->getIncomingValue(i))) {
1654 if (NonConstBB) return 0; // More than one non-const value.
1655 NonConstBB = PN->getIncomingBlock(i);
1657 // If the incoming non-constant value is in I's block, we have an infinite
1659 if (NonConstBB == I.getParent())
1663 // If there is exactly one non-constant value, we can insert a copy of the
1664 // operation in that block. However, if this is a critical edge, we would be
1665 // inserting the computation one some other paths (e.g. inside a loop). Only
1666 // do this if the pred block is unconditionally branching into the phi block.
1668 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1669 if (!BI || !BI->isUnconditional()) return 0;
1672 // Okay, we can do the transformation: create the new PHI node.
1673 PHINode *NewPN = new PHINode(I.getType(), I.getName());
1675 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1676 InsertNewInstBefore(NewPN, *PN);
1678 // Next, add all of the operands to the PHI.
1679 if (I.getNumOperands() == 2) {
1680 Constant *C = cast<Constant>(I.getOperand(1));
1681 for (unsigned i = 0; i != NumPHIValues; ++i) {
1683 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1684 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1685 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1687 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1689 assert(PN->getIncomingBlock(i) == NonConstBB);
1690 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1691 InV = BinaryOperator::create(BO->getOpcode(),
1692 PN->getIncomingValue(i), C, "phitmp",
1693 NonConstBB->getTerminator());
1694 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1695 InV = CmpInst::create(CI->getOpcode(),
1697 PN->getIncomingValue(i), C, "phitmp",
1698 NonConstBB->getTerminator());
1699 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1700 InV = new ShiftInst(SI->getOpcode(),
1701 PN->getIncomingValue(i), C, "phitmp",
1702 NonConstBB->getTerminator());
1704 assert(0 && "Unknown binop!");
1706 WorkList.push_back(cast<Instruction>(InV));
1708 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1711 CastInst *CI = cast<CastInst>(&I);
1712 const Type *RetTy = CI->getType();
1713 for (unsigned i = 0; i != NumPHIValues; ++i) {
1715 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1716 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1718 assert(PN->getIncomingBlock(i) == NonConstBB);
1719 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1720 I.getType(), "phitmp",
1721 NonConstBB->getTerminator());
1722 WorkList.push_back(cast<Instruction>(InV));
1724 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1727 return ReplaceInstUsesWith(I, NewPN);
1730 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1731 bool Changed = SimplifyCommutative(I);
1732 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1734 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1735 // X + undef -> undef
1736 if (isa<UndefValue>(RHS))
1737 return ReplaceInstUsesWith(I, RHS);
1740 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1741 if (RHSC->isNullValue())
1742 return ReplaceInstUsesWith(I, LHS);
1743 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1744 if (CFP->isExactlyValue(-0.0))
1745 return ReplaceInstUsesWith(I, LHS);
1748 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1749 // X + (signbit) --> X ^ signbit
1750 uint64_t Val = CI->getZExtValue();
1751 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1752 return BinaryOperator::createXor(LHS, RHS);
1754 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1755 // (X & 254)+1 -> (X&254)|1
1756 uint64_t KnownZero, KnownOne;
1757 if (!isa<PackedType>(I.getType()) &&
1758 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
1759 KnownZero, KnownOne))
1763 if (isa<PHINode>(LHS))
1764 if (Instruction *NV = FoldOpIntoPhi(I))
1767 ConstantInt *XorRHS = 0;
1769 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1770 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1771 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1772 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1774 uint64_t C0080Val = 1ULL << 31;
1775 int64_t CFF80Val = -C0080Val;
1778 if (TySizeBits > Size) {
1780 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1781 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1782 if (RHSSExt == CFF80Val) {
1783 if (XorRHS->getZExtValue() == C0080Val)
1785 } else if (RHSZExt == C0080Val) {
1786 if (XorRHS->getSExtValue() == CFF80Val)
1790 // This is a sign extend if the top bits are known zero.
1791 uint64_t Mask = ~0ULL;
1792 Mask <<= 64-(TySizeBits-Size);
1793 Mask &= XorLHS->getType()->getIntegralTypeMask();
1794 if (!MaskedValueIsZero(XorLHS, Mask))
1795 Size = 0; // Not a sign ext, but can't be any others either.
1802 } while (Size >= 8);
1805 const Type *MiddleType = 0;
1808 case 32: MiddleType = Type::Int32Ty; break;
1809 case 16: MiddleType = Type::Int16Ty; break;
1810 case 8: MiddleType = Type::Int8Ty; break;
1813 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1814 InsertNewInstBefore(NewTrunc, I);
1815 return new SExtInst(NewTrunc, I.getType());
1821 if (I.getType()->isInteger()) {
1822 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1824 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1825 if (RHSI->getOpcode() == Instruction::Sub)
1826 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1827 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1829 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1830 if (LHSI->getOpcode() == Instruction::Sub)
1831 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1832 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1837 if (Value *V = dyn_castNegVal(LHS))
1838 return BinaryOperator::createSub(RHS, V);
1841 if (!isa<Constant>(RHS))
1842 if (Value *V = dyn_castNegVal(RHS))
1843 return BinaryOperator::createSub(LHS, V);
1847 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1848 if (X == RHS) // X*C + X --> X * (C+1)
1849 return BinaryOperator::createMul(RHS, AddOne(C2));
1851 // X*C1 + X*C2 --> X * (C1+C2)
1853 if (X == dyn_castFoldableMul(RHS, C1))
1854 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1857 // X + X*C --> X * (C+1)
1858 if (dyn_castFoldableMul(RHS, C2) == LHS)
1859 return BinaryOperator::createMul(LHS, AddOne(C2));
1861 // X + ~X --> -1 since ~X = -X-1
1862 if (dyn_castNotVal(LHS) == RHS ||
1863 dyn_castNotVal(RHS) == LHS)
1864 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
1867 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1868 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1869 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
1872 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1874 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1875 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1876 return BinaryOperator::createSub(C, X);
1879 // (X & FF00) + xx00 -> (X+xx00) & FF00
1880 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1881 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1882 if (Anded == CRHS) {
1883 // See if all bits from the first bit set in the Add RHS up are included
1884 // in the mask. First, get the rightmost bit.
1885 uint64_t AddRHSV = CRHS->getZExtValue();
1887 // Form a mask of all bits from the lowest bit added through the top.
1888 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1889 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
1891 // See if the and mask includes all of these bits.
1892 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1894 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1895 // Okay, the xform is safe. Insert the new add pronto.
1896 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1897 LHS->getName()), I);
1898 return BinaryOperator::createAnd(NewAdd, C2);
1903 // Try to fold constant add into select arguments.
1904 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1905 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1909 // add (cast *A to intptrtype) B ->
1910 // cast (GEP (cast *A to sbyte*) B) ->
1913 CastInst *CI = dyn_cast<CastInst>(LHS);
1916 CI = dyn_cast<CastInst>(RHS);
1919 if (CI && CI->getType()->isSized() &&
1920 (CI->getType()->getPrimitiveSizeInBits() ==
1921 TD->getIntPtrType()->getPrimitiveSizeInBits())
1922 && isa<PointerType>(CI->getOperand(0)->getType())) {
1923 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
1924 PointerType::get(Type::Int8Ty), I);
1925 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1926 return new PtrToIntInst(I2, CI->getType());
1930 return Changed ? &I : 0;
1933 // isSignBit - Return true if the value represented by the constant only has the
1934 // highest order bit set.
1935 static bool isSignBit(ConstantInt *CI) {
1936 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1937 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1940 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1942 static Value *RemoveNoopCast(Value *V) {
1943 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1944 const Type *CTy = CI->getType();
1945 const Type *OpTy = CI->getOperand(0)->getType();
1946 if (CTy->isInteger() && OpTy->isInteger()) {
1947 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1948 return RemoveNoopCast(CI->getOperand(0));
1949 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1950 return RemoveNoopCast(CI->getOperand(0));
1955 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1956 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1958 if (Op0 == Op1) // sub X, X -> 0
1959 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1961 // If this is a 'B = x-(-A)', change to B = x+A...
1962 if (Value *V = dyn_castNegVal(Op1))
1963 return BinaryOperator::createAdd(Op0, V);
1965 if (isa<UndefValue>(Op0))
1966 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1967 if (isa<UndefValue>(Op1))
1968 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1970 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1971 // Replace (-1 - A) with (~A)...
1972 if (C->isAllOnesValue())
1973 return BinaryOperator::createNot(Op1);
1975 // C - ~X == X + (1+C)
1977 if (match(Op1, m_Not(m_Value(X))))
1978 return BinaryOperator::createAdd(X,
1979 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1980 // -((uint)X >> 31) -> ((int)X >> 31)
1981 // -((int)X >> 31) -> ((uint)X >> 31)
1982 if (C->isNullValue()) {
1983 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1984 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1985 if (SI->getOpcode() == Instruction::LShr) {
1986 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1987 // Check to see if we are shifting out everything but the sign bit.
1988 if (CU->getZExtValue() ==
1989 SI->getType()->getPrimitiveSizeInBits()-1) {
1990 // Ok, the transformation is safe. Insert AShr.
1991 // FIXME: Once integer types are signless, this cast should be
1993 Value *ShiftOp = SI->getOperand(0);
1994 return new ShiftInst(Instruction::AShr, ShiftOp, CU,
1999 else if (SI->getOpcode() == Instruction::AShr) {
2000 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2001 // Check to see if we are shifting out everything but the sign bit.
2002 if (CU->getZExtValue() ==
2003 SI->getType()->getPrimitiveSizeInBits()-1) {
2005 // Ok, the transformation is safe. Insert LShr.
2006 return new ShiftInst(Instruction::LShr, SI->getOperand(0), CU,
2013 // Try to fold constant sub into select arguments.
2014 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2015 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2018 if (isa<PHINode>(Op0))
2019 if (Instruction *NV = FoldOpIntoPhi(I))
2023 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2024 if (Op1I->getOpcode() == Instruction::Add &&
2025 !Op0->getType()->isFPOrFPVector()) {
2026 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2027 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2028 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2029 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2030 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2031 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2032 // C1-(X+C2) --> (C1-C2)-X
2033 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2034 Op1I->getOperand(0));
2038 if (Op1I->hasOneUse()) {
2039 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2040 // is not used by anyone else...
2042 if (Op1I->getOpcode() == Instruction::Sub &&
2043 !Op1I->getType()->isFPOrFPVector()) {
2044 // Swap the two operands of the subexpr...
2045 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2046 Op1I->setOperand(0, IIOp1);
2047 Op1I->setOperand(1, IIOp0);
2049 // Create the new top level add instruction...
2050 return BinaryOperator::createAdd(Op0, Op1);
2053 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2055 if (Op1I->getOpcode() == Instruction::And &&
2056 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2057 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2060 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2061 return BinaryOperator::createAnd(Op0, NewNot);
2064 // 0 - (X sdiv C) -> (X sdiv -C)
2065 if (Op1I->getOpcode() == Instruction::SDiv)
2066 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2067 if (CSI->isNullValue())
2068 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2069 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2070 ConstantExpr::getNeg(DivRHS));
2072 // X - X*C --> X * (1-C)
2073 ConstantInt *C2 = 0;
2074 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2076 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2077 return BinaryOperator::createMul(Op0, CP1);
2082 if (!Op0->getType()->isFPOrFPVector())
2083 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2084 if (Op0I->getOpcode() == Instruction::Add) {
2085 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2086 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2087 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2088 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2089 } else if (Op0I->getOpcode() == Instruction::Sub) {
2090 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2091 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2095 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2096 if (X == Op1) { // X*C - X --> X * (C-1)
2097 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2098 return BinaryOperator::createMul(Op1, CP1);
2101 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2102 if (X == dyn_castFoldableMul(Op1, C2))
2103 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2108 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2109 /// really just returns true if the most significant (sign) bit is set.
2110 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2112 case ICmpInst::ICMP_SLT:
2113 // True if LHS s< RHS and RHS == 0
2114 return RHS->isNullValue();
2115 case ICmpInst::ICMP_SLE:
2116 // True if LHS s<= RHS and RHS == -1
2117 return RHS->isAllOnesValue();
2118 case ICmpInst::ICMP_UGE:
2119 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2120 return RHS->getZExtValue() == (1ULL <<
2121 (RHS->getType()->getPrimitiveSizeInBits()-1));
2122 case ICmpInst::ICMP_UGT:
2123 // True if LHS u> RHS and RHS == high-bit-mask - 1
2124 return RHS->getZExtValue() ==
2125 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2131 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2132 bool Changed = SimplifyCommutative(I);
2133 Value *Op0 = I.getOperand(0);
2135 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2136 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2138 // Simplify mul instructions with a constant RHS...
2139 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2140 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2142 // ((X << C1)*C2) == (X * (C2 << C1))
2143 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
2144 if (SI->getOpcode() == Instruction::Shl)
2145 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2146 return BinaryOperator::createMul(SI->getOperand(0),
2147 ConstantExpr::getShl(CI, ShOp));
2149 if (CI->isNullValue())
2150 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2151 if (CI->equalsInt(1)) // X * 1 == X
2152 return ReplaceInstUsesWith(I, Op0);
2153 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2154 return BinaryOperator::createNeg(Op0, I.getName());
2156 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2157 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2158 uint64_t C = Log2_64(Val);
2159 return new ShiftInst(Instruction::Shl, Op0,
2160 ConstantInt::get(Type::Int8Ty, C));
2162 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2163 if (Op1F->isNullValue())
2164 return ReplaceInstUsesWith(I, Op1);
2166 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2167 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2168 if (Op1F->getValue() == 1.0)
2169 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2172 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2173 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2174 isa<ConstantInt>(Op0I->getOperand(1))) {
2175 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2176 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2178 InsertNewInstBefore(Add, I);
2179 Value *C1C2 = ConstantExpr::getMul(Op1,
2180 cast<Constant>(Op0I->getOperand(1)));
2181 return BinaryOperator::createAdd(Add, C1C2);
2185 // Try to fold constant mul into select arguments.
2186 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2187 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2190 if (isa<PHINode>(Op0))
2191 if (Instruction *NV = FoldOpIntoPhi(I))
2195 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2196 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2197 return BinaryOperator::createMul(Op0v, Op1v);
2199 // If one of the operands of the multiply is a cast from a boolean value, then
2200 // we know the bool is either zero or one, so this is a 'masking' multiply.
2201 // See if we can simplify things based on how the boolean was originally
2203 CastInst *BoolCast = 0;
2204 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2205 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2208 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2209 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2212 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2213 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2214 const Type *SCOpTy = SCIOp0->getType();
2216 // If the icmp is true iff the sign bit of X is set, then convert this
2217 // multiply into a shift/and combination.
2218 if (isa<ConstantInt>(SCIOp1) &&
2219 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2220 // Shift the X value right to turn it into "all signbits".
2221 Constant *Amt = ConstantInt::get(Type::Int8Ty,
2222 SCOpTy->getPrimitiveSizeInBits()-1);
2224 InsertNewInstBefore(new ShiftInst(Instruction::AShr, SCIOp0, Amt,
2225 BoolCast->getOperand(0)->getName()+
2228 // If the multiply type is not the same as the source type, sign extend
2229 // or truncate to the multiply type.
2230 if (I.getType() != V->getType()) {
2231 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2232 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2233 Instruction::CastOps opcode =
2234 (SrcBits == DstBits ? Instruction::BitCast :
2235 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2236 V = InsertCastBefore(opcode, V, I.getType(), I);
2239 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2240 return BinaryOperator::createAnd(V, OtherOp);
2245 return Changed ? &I : 0;
2248 /// This function implements the transforms on div instructions that work
2249 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2250 /// used by the visitors to those instructions.
2251 /// @brief Transforms common to all three div instructions
2252 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2253 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2256 if (isa<UndefValue>(Op0))
2257 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2259 // X / undef -> undef
2260 if (isa<UndefValue>(Op1))
2261 return ReplaceInstUsesWith(I, Op1);
2263 // Handle cases involving: div X, (select Cond, Y, Z)
2264 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2265 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2266 // same basic block, then we replace the select with Y, and the condition
2267 // of the select with false (if the cond value is in the same BB). If the
2268 // select has uses other than the div, this allows them to be simplified
2269 // also. Note that div X, Y is just as good as div X, 0 (undef)
2270 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2271 if (ST->isNullValue()) {
2272 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2273 if (CondI && CondI->getParent() == I.getParent())
2274 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2275 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2276 I.setOperand(1, SI->getOperand(2));
2278 UpdateValueUsesWith(SI, SI->getOperand(2));
2282 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2283 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2284 if (ST->isNullValue()) {
2285 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2286 if (CondI && CondI->getParent() == I.getParent())
2287 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2288 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2289 I.setOperand(1, SI->getOperand(1));
2291 UpdateValueUsesWith(SI, SI->getOperand(1));
2299 /// This function implements the transforms common to both integer division
2300 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2301 /// division instructions.
2302 /// @brief Common integer divide transforms
2303 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2304 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2306 if (Instruction *Common = commonDivTransforms(I))
2309 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2311 if (RHS->equalsInt(1))
2312 return ReplaceInstUsesWith(I, Op0);
2314 // (X / C1) / C2 -> X / (C1*C2)
2315 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2316 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2317 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2318 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2319 ConstantExpr::getMul(RHS, LHSRHS));
2322 if (!RHS->isNullValue()) { // avoid X udiv 0
2323 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2324 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2326 if (isa<PHINode>(Op0))
2327 if (Instruction *NV = FoldOpIntoPhi(I))
2332 // 0 / X == 0, we don't need to preserve faults!
2333 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2334 if (LHS->equalsInt(0))
2335 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2340 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2341 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2343 // Handle the integer div common cases
2344 if (Instruction *Common = commonIDivTransforms(I))
2347 // X udiv C^2 -> X >> C
2348 // Check to see if this is an unsigned division with an exact power of 2,
2349 // if so, convert to a right shift.
2350 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2351 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2352 if (isPowerOf2_64(Val)) {
2353 uint64_t ShiftAmt = Log2_64(Val);
2354 return new ShiftInst(Instruction::LShr, Op0,
2355 ConstantInt::get(Type::Int8Ty, ShiftAmt));
2359 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2360 if (ShiftInst *RHSI = dyn_cast<ShiftInst>(I.getOperand(1))) {
2361 if (RHSI->getOpcode() == Instruction::Shl &&
2362 isa<ConstantInt>(RHSI->getOperand(0))) {
2363 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2364 if (isPowerOf2_64(C1)) {
2365 Value *N = RHSI->getOperand(1);
2366 const Type *NTy = N->getType();
2367 if (uint64_t C2 = Log2_64(C1)) {
2368 Constant *C2V = ConstantInt::get(NTy, C2);
2369 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2371 return new ShiftInst(Instruction::LShr, Op0, N);
2376 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2377 // where C1&C2 are powers of two.
2378 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2379 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2380 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2381 if (!STO->isNullValue() && !STO->isNullValue()) {
2382 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2383 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2384 // Compute the shift amounts
2385 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2386 // Construct the "on true" case of the select
2387 Constant *TC = ConstantInt::get(Type::Int8Ty, TSA);
2389 new ShiftInst(Instruction::LShr, Op0, TC, SI->getName()+".t");
2390 TSI = InsertNewInstBefore(TSI, I);
2392 // Construct the "on false" case of the select
2393 Constant *FC = ConstantInt::get(Type::Int8Ty, FSA);
2395 new ShiftInst(Instruction::LShr, Op0, FC, SI->getName()+".f");
2396 FSI = InsertNewInstBefore(FSI, I);
2398 // construct the select instruction and return it.
2399 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2406 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2407 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2409 // Handle the integer div common cases
2410 if (Instruction *Common = commonIDivTransforms(I))
2413 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2415 if (RHS->isAllOnesValue())
2416 return BinaryOperator::createNeg(Op0);
2419 if (Value *LHSNeg = dyn_castNegVal(Op0))
2420 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2423 // If the sign bits of both operands are zero (i.e. we can prove they are
2424 // unsigned inputs), turn this into a udiv.
2425 if (I.getType()->isInteger()) {
2426 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2427 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2428 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2435 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2436 return commonDivTransforms(I);
2439 /// GetFactor - If we can prove that the specified value is at least a multiple
2440 /// of some factor, return that factor.
2441 static Constant *GetFactor(Value *V) {
2442 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2445 // Unless we can be tricky, we know this is a multiple of 1.
2446 Constant *Result = ConstantInt::get(V->getType(), 1);
2448 Instruction *I = dyn_cast<Instruction>(V);
2449 if (!I) return Result;
2451 if (I->getOpcode() == Instruction::Mul) {
2452 // Handle multiplies by a constant, etc.
2453 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2454 GetFactor(I->getOperand(1)));
2455 } else if (I->getOpcode() == Instruction::Shl) {
2456 // (X<<C) -> X * (1 << C)
2457 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2458 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2459 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2461 } else if (I->getOpcode() == Instruction::And) {
2462 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2463 // X & 0xFFF0 is known to be a multiple of 16.
2464 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2465 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2466 return ConstantExpr::getShl(Result,
2467 ConstantInt::get(Type::Int8Ty, Zeros));
2469 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2470 // Only handle int->int casts.
2471 if (!CI->isIntegerCast())
2473 Value *Op = CI->getOperand(0);
2474 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2479 /// This function implements the transforms on rem instructions that work
2480 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2481 /// is used by the visitors to those instructions.
2482 /// @brief Transforms common to all three rem instructions
2483 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2484 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2486 // 0 % X == 0, we don't need to preserve faults!
2487 if (Constant *LHS = dyn_cast<Constant>(Op0))
2488 if (LHS->isNullValue())
2489 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2491 if (isa<UndefValue>(Op0)) // undef % X -> 0
2492 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2493 if (isa<UndefValue>(Op1))
2494 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2496 // Handle cases involving: rem X, (select Cond, Y, Z)
2497 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2498 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2499 // the same basic block, then we replace the select with Y, and the
2500 // condition of the select with false (if the cond value is in the same
2501 // BB). If the select has uses other than the div, this allows them to be
2503 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2504 if (ST->isNullValue()) {
2505 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2506 if (CondI && CondI->getParent() == I.getParent())
2507 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2508 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2509 I.setOperand(1, SI->getOperand(2));
2511 UpdateValueUsesWith(SI, SI->getOperand(2));
2514 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2515 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2516 if (ST->isNullValue()) {
2517 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2518 if (CondI && CondI->getParent() == I.getParent())
2519 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2520 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2521 I.setOperand(1, SI->getOperand(1));
2523 UpdateValueUsesWith(SI, SI->getOperand(1));
2531 /// This function implements the transforms common to both integer remainder
2532 /// instructions (urem and srem). It is called by the visitors to those integer
2533 /// remainder instructions.
2534 /// @brief Common integer remainder transforms
2535 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2536 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2538 if (Instruction *common = commonRemTransforms(I))
2541 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2542 // X % 0 == undef, we don't need to preserve faults!
2543 if (RHS->equalsInt(0))
2544 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2546 if (RHS->equalsInt(1)) // X % 1 == 0
2547 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2549 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2550 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2551 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2553 } else if (isa<PHINode>(Op0I)) {
2554 if (Instruction *NV = FoldOpIntoPhi(I))
2557 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2558 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2559 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2566 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2567 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2569 if (Instruction *common = commonIRemTransforms(I))
2572 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2573 // X urem C^2 -> X and C
2574 // Check to see if this is an unsigned remainder with an exact power of 2,
2575 // if so, convert to a bitwise and.
2576 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2577 if (isPowerOf2_64(C->getZExtValue()))
2578 return BinaryOperator::createAnd(Op0, SubOne(C));
2581 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2582 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2583 if (RHSI->getOpcode() == Instruction::Shl &&
2584 isa<ConstantInt>(RHSI->getOperand(0))) {
2585 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2586 if (isPowerOf2_64(C1)) {
2587 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2588 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2590 return BinaryOperator::createAnd(Op0, Add);
2595 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2596 // where C1&C2 are powers of two.
2597 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2598 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2599 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2600 // STO == 0 and SFO == 0 handled above.
2601 if (isPowerOf2_64(STO->getZExtValue()) &&
2602 isPowerOf2_64(SFO->getZExtValue())) {
2603 Value *TrueAnd = InsertNewInstBefore(
2604 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2605 Value *FalseAnd = InsertNewInstBefore(
2606 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2607 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2615 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2616 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2618 if (Instruction *common = commonIRemTransforms(I))
2621 if (Value *RHSNeg = dyn_castNegVal(Op1))
2622 if (!isa<ConstantInt>(RHSNeg) ||
2623 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2625 AddUsesToWorkList(I);
2626 I.setOperand(1, RHSNeg);
2630 // If the top bits of both operands are zero (i.e. we can prove they are
2631 // unsigned inputs), turn this into a urem.
2632 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2633 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2634 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2635 return BinaryOperator::createURem(Op0, Op1, I.getName());
2641 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2642 return commonRemTransforms(I);
2645 // isMaxValueMinusOne - return true if this is Max-1
2646 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2648 // Calculate 0111111111..11111
2649 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2650 int64_t Val = INT64_MAX; // All ones
2651 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2652 return C->getSExtValue() == Val-1;
2654 return C->getZExtValue() == C->getType()->getIntegralTypeMask()-1;
2657 // isMinValuePlusOne - return true if this is Min+1
2658 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2660 // Calculate 1111111111000000000000
2661 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2662 int64_t Val = -1; // All ones
2663 Val <<= TypeBits-1; // Shift over to the right spot
2664 return C->getSExtValue() == Val+1;
2666 return C->getZExtValue() == 1; // unsigned
2669 // isOneBitSet - Return true if there is exactly one bit set in the specified
2671 static bool isOneBitSet(const ConstantInt *CI) {
2672 uint64_t V = CI->getZExtValue();
2673 return V && (V & (V-1)) == 0;
2676 #if 0 // Currently unused
2677 // isLowOnes - Return true if the constant is of the form 0+1+.
2678 static bool isLowOnes(const ConstantInt *CI) {
2679 uint64_t V = CI->getZExtValue();
2681 // There won't be bits set in parts that the type doesn't contain.
2682 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2684 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2685 return U && V && (U & V) == 0;
2689 // isHighOnes - Return true if the constant is of the form 1+0+.
2690 // This is the same as lowones(~X).
2691 static bool isHighOnes(const ConstantInt *CI) {
2692 uint64_t V = ~CI->getZExtValue();
2693 if (~V == 0) return false; // 0's does not match "1+"
2695 // There won't be bits set in parts that the type doesn't contain.
2696 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2698 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2699 return U && V && (U & V) == 0;
2702 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2703 /// are carefully arranged to allow folding of expressions such as:
2705 /// (A < B) | (A > B) --> (A != B)
2707 /// Note that this is only valid if the first and second predicates have the
2708 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2710 /// Three bits are used to represent the condition, as follows:
2715 /// <=> Value Definition
2716 /// 000 0 Always false
2723 /// 111 7 Always true
2725 static unsigned getICmpCode(const ICmpInst *ICI) {
2726 switch (ICI->getPredicate()) {
2728 case ICmpInst::ICMP_UGT: return 1; // 001
2729 case ICmpInst::ICMP_SGT: return 1; // 001
2730 case ICmpInst::ICMP_EQ: return 2; // 010
2731 case ICmpInst::ICMP_UGE: return 3; // 011
2732 case ICmpInst::ICMP_SGE: return 3; // 011
2733 case ICmpInst::ICMP_ULT: return 4; // 100
2734 case ICmpInst::ICMP_SLT: return 4; // 100
2735 case ICmpInst::ICMP_NE: return 5; // 101
2736 case ICmpInst::ICMP_ULE: return 6; // 110
2737 case ICmpInst::ICMP_SLE: return 6; // 110
2740 assert(0 && "Invalid ICmp predicate!");
2745 /// getICmpValue - This is the complement of getICmpCode, which turns an
2746 /// opcode and two operands into either a constant true or false, or a brand
2747 /// new /// ICmp instruction. The sign is passed in to determine which kind
2748 /// of predicate to use in new icmp instructions.
2749 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2751 default: assert(0 && "Illegal ICmp code!");
2752 case 0: return ConstantInt::getFalse();
2755 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2757 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2758 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2761 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2763 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2766 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2768 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2769 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2772 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2774 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2775 case 7: return ConstantInt::getTrue();
2779 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2780 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2781 (ICmpInst::isSignedPredicate(p1) &&
2782 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2783 (ICmpInst::isSignedPredicate(p2) &&
2784 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2788 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2789 struct FoldICmpLogical {
2792 ICmpInst::Predicate pred;
2793 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2794 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2795 pred(ICI->getPredicate()) {}
2796 bool shouldApply(Value *V) const {
2797 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2798 if (PredicatesFoldable(pred, ICI->getPredicate()))
2799 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2800 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2803 Instruction *apply(Instruction &Log) const {
2804 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2805 if (ICI->getOperand(0) != LHS) {
2806 assert(ICI->getOperand(1) == LHS);
2807 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2810 unsigned LHSCode = getICmpCode(ICI);
2811 unsigned RHSCode = getICmpCode(cast<ICmpInst>(Log.getOperand(1)));
2813 switch (Log.getOpcode()) {
2814 case Instruction::And: Code = LHSCode & RHSCode; break;
2815 case Instruction::Or: Code = LHSCode | RHSCode; break;
2816 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2817 default: assert(0 && "Illegal logical opcode!"); return 0;
2820 Value *RV = getICmpValue(ICmpInst::isSignedPredicate(pred), Code, LHS, RHS);
2821 if (Instruction *I = dyn_cast<Instruction>(RV))
2823 // Otherwise, it's a constant boolean value...
2824 return IC.ReplaceInstUsesWith(Log, RV);
2827 } // end anonymous namespace
2829 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2830 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2831 // guaranteed to be either a shift instruction or a binary operator.
2832 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2834 ConstantInt *AndRHS,
2835 BinaryOperator &TheAnd) {
2836 Value *X = Op->getOperand(0);
2837 Constant *Together = 0;
2838 if (!isa<ShiftInst>(Op))
2839 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2841 switch (Op->getOpcode()) {
2842 case Instruction::Xor:
2843 if (Op->hasOneUse()) {
2844 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2845 std::string OpName = Op->getName(); Op->setName("");
2846 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2847 InsertNewInstBefore(And, TheAnd);
2848 return BinaryOperator::createXor(And, Together);
2851 case Instruction::Or:
2852 if (Together == AndRHS) // (X | C) & C --> C
2853 return ReplaceInstUsesWith(TheAnd, AndRHS);
2855 if (Op->hasOneUse() && Together != OpRHS) {
2856 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2857 std::string Op0Name = Op->getName(); Op->setName("");
2858 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2859 InsertNewInstBefore(Or, TheAnd);
2860 return BinaryOperator::createAnd(Or, AndRHS);
2863 case Instruction::Add:
2864 if (Op->hasOneUse()) {
2865 // Adding a one to a single bit bit-field should be turned into an XOR
2866 // of the bit. First thing to check is to see if this AND is with a
2867 // single bit constant.
2868 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2870 // Clear bits that are not part of the constant.
2871 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
2873 // If there is only one bit set...
2874 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2875 // Ok, at this point, we know that we are masking the result of the
2876 // ADD down to exactly one bit. If the constant we are adding has
2877 // no bits set below this bit, then we can eliminate the ADD.
2878 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2880 // Check to see if any bits below the one bit set in AndRHSV are set.
2881 if ((AddRHS & (AndRHSV-1)) == 0) {
2882 // If not, the only thing that can effect the output of the AND is
2883 // the bit specified by AndRHSV. If that bit is set, the effect of
2884 // the XOR is to toggle the bit. If it is clear, then the ADD has
2886 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2887 TheAnd.setOperand(0, X);
2890 std::string Name = Op->getName(); Op->setName("");
2891 // Pull the XOR out of the AND.
2892 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2893 InsertNewInstBefore(NewAnd, TheAnd);
2894 return BinaryOperator::createXor(NewAnd, AndRHS);
2901 case Instruction::Shl: {
2902 // We know that the AND will not produce any of the bits shifted in, so if
2903 // the anded constant includes them, clear them now!
2905 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2906 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2907 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2909 if (CI == ShlMask) { // Masking out bits that the shift already masks
2910 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2911 } else if (CI != AndRHS) { // Reducing bits set in and.
2912 TheAnd.setOperand(1, CI);
2917 case Instruction::LShr:
2919 // We know that the AND will not produce any of the bits shifted in, so if
2920 // the anded constant includes them, clear them now! This only applies to
2921 // unsigned shifts, because a signed shr may bring in set bits!
2923 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2924 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2925 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2927 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2928 return ReplaceInstUsesWith(TheAnd, Op);
2929 } else if (CI != AndRHS) {
2930 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2935 case Instruction::AShr:
2937 // See if this is shifting in some sign extension, then masking it out
2939 if (Op->hasOneUse()) {
2940 Constant *AllOne = ConstantInt::getAllOnesValue(AndRHS->getType());
2941 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2942 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
2943 if (C == AndRHS) { // Masking out bits shifted in.
2944 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2945 // Make the argument unsigned.
2946 Value *ShVal = Op->getOperand(0);
2947 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::LShr, ShVal,
2948 OpRHS, Op->getName()), TheAnd);
2949 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
2958 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2959 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2960 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
2961 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
2962 /// insert new instructions.
2963 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2964 bool isSigned, bool Inside,
2966 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
2967 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
2968 "Lo is not <= Hi in range emission code!");
2971 if (Lo == Hi) // Trivially false.
2972 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
2974 // V >= Min && V < Hi --> V < Hi
2975 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2976 ICmpInst::Predicate pred = (isSigned ?
2977 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
2978 return new ICmpInst(pred, V, Hi);
2981 // Emit V-Lo <u Hi-Lo
2982 Constant *NegLo = ConstantExpr::getNeg(Lo);
2983 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2984 InsertNewInstBefore(Add, IB);
2985 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
2986 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
2989 if (Lo == Hi) // Trivially true.
2990 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
2992 // V < Min || V >= Hi ->'V > Hi-1'
2993 Hi = SubOne(cast<ConstantInt>(Hi));
2994 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
2995 ICmpInst::Predicate pred = (isSigned ?
2996 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
2997 return new ICmpInst(pred, V, Hi);
3000 // Emit V-Lo > Hi-1-Lo
3001 Constant *NegLo = ConstantExpr::getNeg(Lo);
3002 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3003 InsertNewInstBefore(Add, IB);
3004 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3005 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3008 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3009 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3010 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3011 // not, since all 1s are not contiguous.
3012 static bool isRunOfOnes(ConstantInt *Val, unsigned &MB, unsigned &ME) {
3013 uint64_t V = Val->getZExtValue();
3014 if (!isShiftedMask_64(V)) return false;
3016 // look for the first zero bit after the run of ones
3017 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
3018 // look for the first non-zero bit
3019 ME = 64-CountLeadingZeros_64(V);
3025 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3026 /// where isSub determines whether the operator is a sub. If we can fold one of
3027 /// the following xforms:
3029 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3030 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3031 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3033 /// return (A +/- B).
3035 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3036 ConstantInt *Mask, bool isSub,
3038 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3039 if (!LHSI || LHSI->getNumOperands() != 2 ||
3040 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3042 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3044 switch (LHSI->getOpcode()) {
3046 case Instruction::And:
3047 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3048 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3049 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3052 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3053 // part, we don't need any explicit masks to take them out of A. If that
3054 // is all N is, ignore it.
3056 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3057 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
3059 if (MaskedValueIsZero(RHS, Mask))
3064 case Instruction::Or:
3065 case Instruction::Xor:
3066 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3067 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3068 ConstantExpr::getAnd(N, Mask)->isNullValue())
3075 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3077 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3078 return InsertNewInstBefore(New, I);
3081 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3082 bool Changed = SimplifyCommutative(I);
3083 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3085 if (isa<UndefValue>(Op1)) // X & undef -> 0
3086 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3090 return ReplaceInstUsesWith(I, Op1);
3092 // See if we can simplify any instructions used by the instruction whose sole
3093 // purpose is to compute bits we don't care about.
3094 uint64_t KnownZero, KnownOne;
3095 if (!isa<PackedType>(I.getType()) &&
3096 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3097 KnownZero, KnownOne))
3100 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3101 uint64_t AndRHSMask = AndRHS->getZExtValue();
3102 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
3103 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3105 // Optimize a variety of ((val OP C1) & C2) combinations...
3106 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
3107 Instruction *Op0I = cast<Instruction>(Op0);
3108 Value *Op0LHS = Op0I->getOperand(0);
3109 Value *Op0RHS = Op0I->getOperand(1);
3110 switch (Op0I->getOpcode()) {
3111 case Instruction::Xor:
3112 case Instruction::Or:
3113 // If the mask is only needed on one incoming arm, push it up.
3114 if (Op0I->hasOneUse()) {
3115 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3116 // Not masking anything out for the LHS, move to RHS.
3117 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3118 Op0RHS->getName()+".masked");
3119 InsertNewInstBefore(NewRHS, I);
3120 return BinaryOperator::create(
3121 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3123 if (!isa<Constant>(Op0RHS) &&
3124 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3125 // Not masking anything out for the RHS, move to LHS.
3126 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3127 Op0LHS->getName()+".masked");
3128 InsertNewInstBefore(NewLHS, I);
3129 return BinaryOperator::create(
3130 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3135 case Instruction::Add:
3136 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3137 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3138 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3139 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3140 return BinaryOperator::createAnd(V, AndRHS);
3141 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3142 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3145 case Instruction::Sub:
3146 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3147 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3148 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3149 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3150 return BinaryOperator::createAnd(V, AndRHS);
3154 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3155 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3157 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3158 // If this is an integer truncation or change from signed-to-unsigned, and
3159 // if the source is an and/or with immediate, transform it. This
3160 // frequently occurs for bitfield accesses.
3161 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3162 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3163 CastOp->getNumOperands() == 2)
3164 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3165 if (CastOp->getOpcode() == Instruction::And) {
3166 // Change: and (cast (and X, C1) to T), C2
3167 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3168 // This will fold the two constants together, which may allow
3169 // other simplifications.
3170 Instruction *NewCast = CastInst::createTruncOrBitCast(
3171 CastOp->getOperand(0), I.getType(),
3172 CastOp->getName()+".shrunk");
3173 NewCast = InsertNewInstBefore(NewCast, I);
3174 // trunc_or_bitcast(C1)&C2
3175 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3176 C3 = ConstantExpr::getAnd(C3, AndRHS);
3177 return BinaryOperator::createAnd(NewCast, C3);
3178 } else if (CastOp->getOpcode() == Instruction::Or) {
3179 // Change: and (cast (or X, C1) to T), C2
3180 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3181 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3182 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3183 return ReplaceInstUsesWith(I, AndRHS);
3188 // Try to fold constant and into select arguments.
3189 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3190 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3192 if (isa<PHINode>(Op0))
3193 if (Instruction *NV = FoldOpIntoPhi(I))
3197 Value *Op0NotVal = dyn_castNotVal(Op0);
3198 Value *Op1NotVal = dyn_castNotVal(Op1);
3200 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3201 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3203 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3204 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3205 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3206 I.getName()+".demorgan");
3207 InsertNewInstBefore(Or, I);
3208 return BinaryOperator::createNot(Or);
3212 Value *A = 0, *B = 0;
3213 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3214 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3215 return ReplaceInstUsesWith(I, Op1);
3216 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3217 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3218 return ReplaceInstUsesWith(I, Op0);
3220 if (Op0->hasOneUse() &&
3221 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3222 if (A == Op1) { // (A^B)&A -> A&(A^B)
3223 I.swapOperands(); // Simplify below
3224 std::swap(Op0, Op1);
3225 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3226 cast<BinaryOperator>(Op0)->swapOperands();
3227 I.swapOperands(); // Simplify below
3228 std::swap(Op0, Op1);
3231 if (Op1->hasOneUse() &&
3232 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3233 if (B == Op0) { // B&(A^B) -> B&(B^A)
3234 cast<BinaryOperator>(Op1)->swapOperands();
3237 if (A == Op0) { // A&(A^B) -> A & ~B
3238 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3239 InsertNewInstBefore(NotB, I);
3240 return BinaryOperator::createAnd(A, NotB);
3245 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3246 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3247 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3250 Value *LHSVal, *RHSVal;
3251 ConstantInt *LHSCst, *RHSCst;
3252 ICmpInst::Predicate LHSCC, RHSCC;
3253 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3254 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3255 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3256 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3257 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3258 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3259 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3260 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3261 // Ensure that the larger constant is on the RHS.
3262 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3263 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3264 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3265 ICmpInst *LHS = cast<ICmpInst>(Op0);
3266 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3267 std::swap(LHS, RHS);
3268 std::swap(LHSCst, RHSCst);
3269 std::swap(LHSCC, RHSCC);
3272 // At this point, we know we have have two icmp instructions
3273 // comparing a value against two constants and and'ing the result
3274 // together. Because of the above check, we know that we only have
3275 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3276 // (from the FoldICmpLogical check above), that the two constants
3277 // are not equal and that the larger constant is on the RHS
3278 assert(LHSCst != RHSCst && "Compares not folded above?");
3281 default: assert(0 && "Unknown integer condition code!");
3282 case ICmpInst::ICMP_EQ:
3284 default: assert(0 && "Unknown integer condition code!");
3285 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3286 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3287 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3288 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3289 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3290 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3291 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3292 return ReplaceInstUsesWith(I, LHS);
3294 case ICmpInst::ICMP_NE:
3296 default: assert(0 && "Unknown integer condition code!");
3297 case ICmpInst::ICMP_ULT:
3298 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3299 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3300 break; // (X != 13 & X u< 15) -> no change
3301 case ICmpInst::ICMP_SLT:
3302 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3303 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3304 break; // (X != 13 & X s< 15) -> no change
3305 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3306 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3307 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3308 return ReplaceInstUsesWith(I, RHS);
3309 case ICmpInst::ICMP_NE:
3310 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3311 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3312 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3313 LHSVal->getName()+".off");
3314 InsertNewInstBefore(Add, I);
3315 return new ICmpInst(ICmpInst::ICMP_UGT, Add, AddCST);
3317 break; // (X != 13 & X != 15) -> no change
3320 case ICmpInst::ICMP_ULT:
3322 default: assert(0 && "Unknown integer condition code!");
3323 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3324 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3325 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3326 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3328 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3329 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3330 return ReplaceInstUsesWith(I, LHS);
3331 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3335 case ICmpInst::ICMP_SLT:
3337 default: assert(0 && "Unknown integer condition code!");
3338 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3339 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3340 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3341 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3343 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3344 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3345 return ReplaceInstUsesWith(I, LHS);
3346 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3350 case ICmpInst::ICMP_UGT:
3352 default: assert(0 && "Unknown integer condition code!");
3353 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3354 return ReplaceInstUsesWith(I, LHS);
3355 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3356 return ReplaceInstUsesWith(I, RHS);
3357 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3359 case ICmpInst::ICMP_NE:
3360 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3361 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3362 break; // (X u> 13 & X != 15) -> no change
3363 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3364 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3366 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3370 case ICmpInst::ICMP_SGT:
3372 default: assert(0 && "Unknown integer condition code!");
3373 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3374 return ReplaceInstUsesWith(I, LHS);
3375 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3376 return ReplaceInstUsesWith(I, RHS);
3377 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3379 case ICmpInst::ICMP_NE:
3380 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3381 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3382 break; // (X s> 13 & X != 15) -> no change
3383 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3384 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3386 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3394 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3395 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3396 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3397 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3398 const Type *SrcTy = Op0C->getOperand(0)->getType();
3399 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3400 // Only do this if the casts both really cause code to be generated.
3401 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3403 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3405 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3406 Op1C->getOperand(0),
3408 InsertNewInstBefore(NewOp, I);
3409 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3413 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3414 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3415 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3416 if (SI0->getOpcode() == SI1->getOpcode() &&
3417 SI0->getOperand(1) == SI1->getOperand(1) &&
3418 (SI0->hasOneUse() || SI1->hasOneUse())) {
3419 Instruction *NewOp =
3420 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3422 SI0->getName()), I);
3423 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3427 return Changed ? &I : 0;
3430 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3431 /// in the result. If it does, and if the specified byte hasn't been filled in
3432 /// yet, fill it in and return false.
3433 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
3434 Instruction *I = dyn_cast<Instruction>(V);
3435 if (I == 0) return true;
3437 // If this is an or instruction, it is an inner node of the bswap.
3438 if (I->getOpcode() == Instruction::Or)
3439 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3440 CollectBSwapParts(I->getOperand(1), ByteValues);
3442 // If this is a shift by a constant int, and it is "24", then its operand
3443 // defines a byte. We only handle unsigned types here.
3444 if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
3445 // Not shifting the entire input by N-1 bytes?
3446 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3447 8*(ByteValues.size()-1))
3451 if (I->getOpcode() == Instruction::Shl) {
3452 // X << 24 defines the top byte with the lowest of the input bytes.
3453 DestNo = ByteValues.size()-1;
3455 // X >>u 24 defines the low byte with the highest of the input bytes.
3459 // If the destination byte value is already defined, the values are or'd
3460 // together, which isn't a bswap (unless it's an or of the same bits).
3461 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3463 ByteValues[DestNo] = I->getOperand(0);
3467 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3469 Value *Shift = 0, *ShiftLHS = 0;
3470 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3471 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3472 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3474 Instruction *SI = cast<Instruction>(Shift);
3476 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3477 if (ShiftAmt->getZExtValue() & 7 ||
3478 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3481 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3483 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3484 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3486 // Unknown mask for bswap.
3487 if (DestByte == ByteValues.size()) return true;
3489 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3491 if (SI->getOpcode() == Instruction::Shl)
3492 SrcByte = DestByte - ShiftBytes;
3494 SrcByte = DestByte + ShiftBytes;
3496 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3497 if (SrcByte != ByteValues.size()-DestByte-1)
3500 // If the destination byte value is already defined, the values are or'd
3501 // together, which isn't a bswap (unless it's an or of the same bits).
3502 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3504 ByteValues[DestByte] = SI->getOperand(0);
3508 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3509 /// If so, insert the new bswap intrinsic and return it.
3510 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3511 // We can only handle bswap of unsigned integers, and cannot bswap one byte.
3512 if (I.getType() == Type::Int8Ty)
3515 /// ByteValues - For each byte of the result, we keep track of which value
3516 /// defines each byte.
3517 std::vector<Value*> ByteValues;
3518 ByteValues.resize(TD->getTypeSize(I.getType()));
3520 // Try to find all the pieces corresponding to the bswap.
3521 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3522 CollectBSwapParts(I.getOperand(1), ByteValues))
3525 // Check to see if all of the bytes come from the same value.
3526 Value *V = ByteValues[0];
3527 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3529 // Check to make sure that all of the bytes come from the same value.
3530 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3531 if (ByteValues[i] != V)
3534 // If they do then *success* we can turn this into a bswap. Figure out what
3535 // bswap to make it into.
3536 Module *M = I.getParent()->getParent()->getParent();
3537 const char *FnName = 0;
3538 if (I.getType() == Type::Int16Ty)
3539 FnName = "llvm.bswap.i16";
3540 else if (I.getType() == Type::Int32Ty)
3541 FnName = "llvm.bswap.i32";
3542 else if (I.getType() == Type::Int64Ty)
3543 FnName = "llvm.bswap.i64";
3545 assert(0 && "Unknown integer type!");
3546 Constant *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3547 return new CallInst(F, V);
3551 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3552 bool Changed = SimplifyCommutative(I);
3553 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3555 if (isa<UndefValue>(Op1))
3556 return ReplaceInstUsesWith(I, // X | undef -> -1
3557 ConstantInt::getAllOnesValue(I.getType()));
3561 return ReplaceInstUsesWith(I, Op0);
3563 // See if we can simplify any instructions used by the instruction whose sole
3564 // purpose is to compute bits we don't care about.
3565 uint64_t KnownZero, KnownOne;
3566 if (!isa<PackedType>(I.getType()) &&
3567 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3568 KnownZero, KnownOne))
3572 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3573 ConstantInt *C1 = 0; Value *X = 0;
3574 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3575 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3576 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
3578 InsertNewInstBefore(Or, I);
3579 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3582 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3583 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3584 std::string Op0Name = Op0->getName(); Op0->setName("");
3585 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
3586 InsertNewInstBefore(Or, I);
3587 return BinaryOperator::createXor(Or,
3588 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3591 // Try to fold constant and into select arguments.
3592 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3593 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3595 if (isa<PHINode>(Op0))
3596 if (Instruction *NV = FoldOpIntoPhi(I))
3600 Value *A = 0, *B = 0;
3601 ConstantInt *C1 = 0, *C2 = 0;
3603 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3604 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3605 return ReplaceInstUsesWith(I, Op1);
3606 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3607 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3608 return ReplaceInstUsesWith(I, Op0);
3610 // (A | B) | C and A | (B | C) -> bswap if possible.
3611 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3612 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3613 match(Op1, m_Or(m_Value(), m_Value())) ||
3614 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3615 match(Op1, m_Shift(m_Value(), m_Value())))) {
3616 if (Instruction *BSwap = MatchBSwap(I))
3620 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3621 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3622 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3623 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
3625 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3628 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3629 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3630 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3631 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
3633 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3636 // (A & C1)|(B & C2)
3637 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3638 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3640 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3641 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3644 // If we have: ((V + N) & C1) | (V & C2)
3645 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3646 // replace with V+N.
3647 if (C1 == ConstantExpr::getNot(C2)) {
3648 Value *V1 = 0, *V2 = 0;
3649 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3650 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3651 // Add commutes, try both ways.
3652 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3653 return ReplaceInstUsesWith(I, A);
3654 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3655 return ReplaceInstUsesWith(I, A);
3657 // Or commutes, try both ways.
3658 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3659 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3660 // Add commutes, try both ways.
3661 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3662 return ReplaceInstUsesWith(I, B);
3663 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3664 return ReplaceInstUsesWith(I, B);
3669 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3670 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3671 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3672 if (SI0->getOpcode() == SI1->getOpcode() &&
3673 SI0->getOperand(1) == SI1->getOperand(1) &&
3674 (SI0->hasOneUse() || SI1->hasOneUse())) {
3675 Instruction *NewOp =
3676 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3678 SI0->getName()), I);
3679 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3683 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3684 if (A == Op1) // ~A | A == -1
3685 return ReplaceInstUsesWith(I,
3686 ConstantInt::getAllOnesValue(I.getType()));
3690 // Note, A is still live here!
3691 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3693 return ReplaceInstUsesWith(I,
3694 ConstantInt::getAllOnesValue(I.getType()));
3696 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3697 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3698 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3699 I.getName()+".demorgan"), I);
3700 return BinaryOperator::createNot(And);
3704 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3705 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3706 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3709 Value *LHSVal, *RHSVal;
3710 ConstantInt *LHSCst, *RHSCst;
3711 ICmpInst::Predicate LHSCC, RHSCC;
3712 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3713 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3714 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3715 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3716 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3717 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3718 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3719 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3720 // Ensure that the larger constant is on the RHS.
3721 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3722 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3723 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3724 ICmpInst *LHS = cast<ICmpInst>(Op0);
3725 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3726 std::swap(LHS, RHS);
3727 std::swap(LHSCst, RHSCst);
3728 std::swap(LHSCC, RHSCC);
3731 // At this point, we know we have have two icmp instructions
3732 // comparing a value against two constants and or'ing the result
3733 // together. Because of the above check, we know that we only have
3734 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3735 // FoldICmpLogical check above), that the two constants are not
3737 assert(LHSCst != RHSCst && "Compares not folded above?");
3740 default: assert(0 && "Unknown integer condition code!");
3741 case ICmpInst::ICMP_EQ:
3743 default: assert(0 && "Unknown integer condition code!");
3744 case ICmpInst::ICMP_EQ:
3745 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3746 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3747 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3748 LHSVal->getName()+".off");
3749 InsertNewInstBefore(Add, I);
3750 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3751 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3753 break; // (X == 13 | X == 15) -> no change
3754 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3755 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3757 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3758 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3759 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3760 return ReplaceInstUsesWith(I, RHS);
3763 case ICmpInst::ICMP_NE:
3765 default: assert(0 && "Unknown integer condition code!");
3766 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3767 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3768 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3769 return ReplaceInstUsesWith(I, LHS);
3770 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3771 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3772 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3773 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3776 case ICmpInst::ICMP_ULT:
3778 default: assert(0 && "Unknown integer condition code!");
3779 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3781 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3782 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3784 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3786 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3787 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3788 return ReplaceInstUsesWith(I, RHS);
3789 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3793 case ICmpInst::ICMP_SLT:
3795 default: assert(0 && "Unknown integer condition code!");
3796 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3798 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3799 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3801 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3803 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3804 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3805 return ReplaceInstUsesWith(I, RHS);
3806 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3810 case ICmpInst::ICMP_UGT:
3812 default: assert(0 && "Unknown integer condition code!");
3813 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3814 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3815 return ReplaceInstUsesWith(I, LHS);
3816 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3818 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3819 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3820 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3821 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3825 case ICmpInst::ICMP_SGT:
3827 default: assert(0 && "Unknown integer condition code!");
3828 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3829 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3830 return ReplaceInstUsesWith(I, LHS);
3831 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3833 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3834 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3835 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3836 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3844 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3845 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3846 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3847 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3848 const Type *SrcTy = Op0C->getOperand(0)->getType();
3849 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3850 // Only do this if the casts both really cause code to be generated.
3851 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3853 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3855 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3856 Op1C->getOperand(0),
3858 InsertNewInstBefore(NewOp, I);
3859 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3864 return Changed ? &I : 0;
3867 // XorSelf - Implements: X ^ X --> 0
3870 XorSelf(Value *rhs) : RHS(rhs) {}
3871 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3872 Instruction *apply(BinaryOperator &Xor) const {
3878 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3879 bool Changed = SimplifyCommutative(I);
3880 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3882 if (isa<UndefValue>(Op1))
3883 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3885 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3886 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3887 assert(Result == &I && "AssociativeOpt didn't work?");
3888 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3891 // See if we can simplify any instructions used by the instruction whose sole
3892 // purpose is to compute bits we don't care about.
3893 uint64_t KnownZero, KnownOne;
3894 if (!isa<PackedType>(I.getType()) &&
3895 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3896 KnownZero, KnownOne))
3899 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3900 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
3901 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3902 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
3903 return new ICmpInst(ICI->getInversePredicate(),
3904 ICI->getOperand(0), ICI->getOperand(1));
3906 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3907 // ~(c-X) == X-c-1 == X+(-c-1)
3908 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3909 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3910 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3911 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3912 ConstantInt::get(I.getType(), 1));
3913 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3916 // ~(~X & Y) --> (X | ~Y)
3917 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3918 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3919 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3921 BinaryOperator::createNot(Op0I->getOperand(1),
3922 Op0I->getOperand(1)->getName()+".not");
3923 InsertNewInstBefore(NotY, I);
3924 return BinaryOperator::createOr(Op0NotVal, NotY);
3928 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3929 if (Op0I->getOpcode() == Instruction::Add) {
3930 // ~(X-c) --> (-c-1)-X
3931 if (RHS->isAllOnesValue()) {
3932 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3933 return BinaryOperator::createSub(
3934 ConstantExpr::getSub(NegOp0CI,
3935 ConstantInt::get(I.getType(), 1)),
3936 Op0I->getOperand(0));
3938 } else if (Op0I->getOpcode() == Instruction::Or) {
3939 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3940 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3941 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3942 // Anything in both C1 and C2 is known to be zero, remove it from
3944 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3945 NewRHS = ConstantExpr::getAnd(NewRHS,
3946 ConstantExpr::getNot(CommonBits));
3947 WorkList.push_back(Op0I);
3948 I.setOperand(0, Op0I->getOperand(0));
3949 I.setOperand(1, NewRHS);
3955 // Try to fold constant and into select arguments.
3956 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3957 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3959 if (isa<PHINode>(Op0))
3960 if (Instruction *NV = FoldOpIntoPhi(I))
3964 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3966 return ReplaceInstUsesWith(I,
3967 ConstantInt::getAllOnesValue(I.getType()));
3969 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3971 return ReplaceInstUsesWith(I,
3972 ConstantInt::getAllOnesValue(I.getType()));
3974 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3975 if (Op1I->getOpcode() == Instruction::Or) {
3976 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3977 Op1I->swapOperands();
3979 std::swap(Op0, Op1);
3980 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3981 I.swapOperands(); // Simplified below.
3982 std::swap(Op0, Op1);
3984 } else if (Op1I->getOpcode() == Instruction::Xor) {
3985 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3986 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3987 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3988 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3989 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3990 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3991 Op1I->swapOperands();
3992 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3993 I.swapOperands(); // Simplified below.
3994 std::swap(Op0, Op1);
3998 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3999 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
4000 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
4001 Op0I->swapOperands();
4002 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
4003 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
4004 InsertNewInstBefore(NotB, I);
4005 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
4007 } else if (Op0I->getOpcode() == Instruction::Xor) {
4008 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
4009 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
4010 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
4011 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
4012 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
4013 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
4014 Op0I->swapOperands();
4015 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
4016 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4017 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
4018 InsertNewInstBefore(N, I);
4019 return BinaryOperator::createAnd(N, Op1);
4023 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4024 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4025 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4028 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4029 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4030 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4031 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4032 const Type *SrcTy = Op0C->getOperand(0)->getType();
4033 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
4034 // Only do this if the casts both really cause code to be generated.
4035 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4037 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4039 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4040 Op1C->getOperand(0),
4042 InsertNewInstBefore(NewOp, I);
4043 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4047 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4048 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
4049 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
4050 if (SI0->getOpcode() == SI1->getOpcode() &&
4051 SI0->getOperand(1) == SI1->getOperand(1) &&
4052 (SI0->hasOneUse() || SI1->hasOneUse())) {
4053 Instruction *NewOp =
4054 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
4056 SI0->getName()), I);
4057 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
4061 return Changed ? &I : 0;
4064 static bool isPositive(ConstantInt *C) {
4065 return C->getSExtValue() >= 0;
4068 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4069 /// overflowed for this type.
4070 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4072 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4074 return cast<ConstantInt>(Result)->getZExtValue() <
4075 cast<ConstantInt>(In1)->getZExtValue();
4078 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4079 /// code necessary to compute the offset from the base pointer (without adding
4080 /// in the base pointer). Return the result as a signed integer of intptr size.
4081 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4082 TargetData &TD = IC.getTargetData();
4083 gep_type_iterator GTI = gep_type_begin(GEP);
4084 const Type *IntPtrTy = TD.getIntPtrType();
4085 Value *Result = Constant::getNullValue(IntPtrTy);
4087 // Build a mask for high order bits.
4088 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4090 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4091 Value *Op = GEP->getOperand(i);
4092 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4093 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4094 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4095 if (!OpC->isNullValue()) {
4096 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4097 Scale = ConstantExpr::getMul(OpC, Scale);
4098 if (Constant *RC = dyn_cast<Constant>(Result))
4099 Result = ConstantExpr::getAdd(RC, Scale);
4101 // Emit an add instruction.
4102 Result = IC.InsertNewInstBefore(
4103 BinaryOperator::createAdd(Result, Scale,
4104 GEP->getName()+".offs"), I);
4108 // Convert to correct type.
4109 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4110 Op->getName()+".c"), I);
4112 // We'll let instcombine(mul) convert this to a shl if possible.
4113 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4114 GEP->getName()+".idx"), I);
4116 // Emit an add instruction.
4117 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4118 GEP->getName()+".offs"), I);
4124 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4125 /// else. At this point we know that the GEP is on the LHS of the comparison.
4126 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4127 ICmpInst::Predicate Cond,
4129 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4131 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4132 if (isa<PointerType>(CI->getOperand(0)->getType()))
4133 RHS = CI->getOperand(0);
4135 Value *PtrBase = GEPLHS->getOperand(0);
4136 if (PtrBase == RHS) {
4137 // As an optimization, we don't actually have to compute the actual value of
4138 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4139 // each index is zero or not.
4140 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4141 Instruction *InVal = 0;
4142 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4143 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4145 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4146 if (isa<UndefValue>(C)) // undef index -> undef.
4147 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4148 if (C->isNullValue())
4150 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4151 EmitIt = false; // This is indexing into a zero sized array?
4152 } else if (isa<ConstantInt>(C))
4153 return ReplaceInstUsesWith(I, // No comparison is needed here.
4154 ConstantInt::get(Type::Int1Ty,
4155 Cond == ICmpInst::ICMP_NE));
4160 new ICmpInst(Cond, GEPLHS->getOperand(i),
4161 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4165 InVal = InsertNewInstBefore(InVal, I);
4166 InsertNewInstBefore(Comp, I);
4167 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4168 InVal = BinaryOperator::createOr(InVal, Comp);
4169 else // True if all are equal
4170 InVal = BinaryOperator::createAnd(InVal, Comp);
4178 // No comparison is needed here, all indexes = 0
4179 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4180 Cond == ICmpInst::ICMP_EQ));
4183 // Only lower this if the icmp is the only user of the GEP or if we expect
4184 // the result to fold to a constant!
4185 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4186 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4187 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4188 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4189 Constant::getNullValue(Offset->getType()));
4191 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4192 // If the base pointers are different, but the indices are the same, just
4193 // compare the base pointer.
4194 if (PtrBase != GEPRHS->getOperand(0)) {
4195 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4196 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4197 GEPRHS->getOperand(0)->getType();
4199 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4200 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4201 IndicesTheSame = false;
4205 // If all indices are the same, just compare the base pointers.
4207 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4208 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4210 // Otherwise, the base pointers are different and the indices are
4211 // different, bail out.
4215 // If one of the GEPs has all zero indices, recurse.
4216 bool AllZeros = true;
4217 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4218 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4219 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4224 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4225 ICmpInst::getSwappedPredicate(Cond), I);
4227 // If the other GEP has all zero indices, recurse.
4229 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4230 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4231 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4236 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4238 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4239 // If the GEPs only differ by one index, compare it.
4240 unsigned NumDifferences = 0; // Keep track of # differences.
4241 unsigned DiffOperand = 0; // The operand that differs.
4242 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4243 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4244 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4245 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4246 // Irreconcilable differences.
4250 if (NumDifferences++) break;
4255 if (NumDifferences == 0) // SAME GEP?
4256 return ReplaceInstUsesWith(I, // No comparison is needed here.
4257 ConstantInt::get(Type::Int1Ty,
4258 Cond == ICmpInst::ICMP_EQ));
4259 else if (NumDifferences == 1) {
4260 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4261 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4262 // Make sure we do a signed comparison here.
4263 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4267 // Only lower this if the icmp is the only user of the GEP or if we expect
4268 // the result to fold to a constant!
4269 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4270 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4271 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4272 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4273 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4274 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4280 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4281 bool Changed = SimplifyCompare(I);
4282 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4286 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4287 isTrueWhenEqual(I)));
4289 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4290 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4292 // Handle fcmp with constant RHS
4293 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4294 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4295 switch (LHSI->getOpcode()) {
4296 case Instruction::PHI:
4297 if (Instruction *NV = FoldOpIntoPhi(I))
4300 case Instruction::Select:
4301 // If either operand of the select is a constant, we can fold the
4302 // comparison into the select arms, which will cause one to be
4303 // constant folded and the select turned into a bitwise or.
4304 Value *Op1 = 0, *Op2 = 0;
4305 if (LHSI->hasOneUse()) {
4306 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4307 // Fold the known value into the constant operand.
4308 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4309 // Insert a new FCmp of the other select operand.
4310 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4311 LHSI->getOperand(2), RHSC,
4313 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4314 // Fold the known value into the constant operand.
4315 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4316 // Insert a new FCmp of the other select operand.
4317 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4318 LHSI->getOperand(1), RHSC,
4324 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4329 return Changed ? &I : 0;
4332 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4333 bool Changed = SimplifyCompare(I);
4334 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4335 const Type *Ty = Op0->getType();
4339 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4340 isTrueWhenEqual(I)));
4342 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4343 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4345 // icmp of GlobalValues can never equal each other as long as they aren't
4346 // external weak linkage type.
4347 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4348 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4349 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4350 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4351 !isTrueWhenEqual(I)));
4353 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4354 // addresses never equal each other! We already know that Op0 != Op1.
4355 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4356 isa<ConstantPointerNull>(Op0)) &&
4357 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4358 isa<ConstantPointerNull>(Op1)))
4359 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4360 !isTrueWhenEqual(I)));
4362 // icmp's with boolean values can always be turned into bitwise operations
4363 if (Ty == Type::Int1Ty) {
4364 switch (I.getPredicate()) {
4365 default: assert(0 && "Invalid icmp instruction!");
4366 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4367 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4368 InsertNewInstBefore(Xor, I);
4369 return BinaryOperator::createNot(Xor);
4371 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4372 return BinaryOperator::createXor(Op0, Op1);
4374 case ICmpInst::ICMP_UGT:
4375 case ICmpInst::ICMP_SGT:
4376 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4378 case ICmpInst::ICMP_ULT:
4379 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4380 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4381 InsertNewInstBefore(Not, I);
4382 return BinaryOperator::createAnd(Not, Op1);
4384 case ICmpInst::ICMP_UGE:
4385 case ICmpInst::ICMP_SGE:
4386 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4388 case ICmpInst::ICMP_ULE:
4389 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4390 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4391 InsertNewInstBefore(Not, I);
4392 return BinaryOperator::createOr(Not, Op1);
4397 // See if we are doing a comparison between a constant and an instruction that
4398 // can be folded into the comparison.
4399 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4400 switch (I.getPredicate()) {
4402 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4403 if (CI->isMinValue(false))
4404 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4405 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4406 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4407 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4408 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4411 case ICmpInst::ICMP_SLT:
4412 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4413 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4414 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4415 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4416 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4417 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4420 case ICmpInst::ICMP_UGT:
4421 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4422 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4423 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4424 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4425 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4426 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4429 case ICmpInst::ICMP_SGT:
4430 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4431 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4432 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4433 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4434 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4435 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4438 case ICmpInst::ICMP_ULE:
4439 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4440 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4441 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4442 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4443 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4444 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4447 case ICmpInst::ICMP_SLE:
4448 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4449 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4450 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4451 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4452 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4453 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4456 case ICmpInst::ICMP_UGE:
4457 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4458 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4459 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4460 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4461 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4462 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4465 case ICmpInst::ICMP_SGE:
4466 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4467 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4468 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4469 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4470 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4471 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4475 // If we still have a icmp le or icmp ge instruction, turn it into the
4476 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4477 // already been handled above, this requires little checking.
4479 if (I.getPredicate() == ICmpInst::ICMP_ULE)
4480 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4481 if (I.getPredicate() == ICmpInst::ICMP_SLE)
4482 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4483 if (I.getPredicate() == ICmpInst::ICMP_UGE)
4484 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4485 if (I.getPredicate() == ICmpInst::ICMP_SGE)
4486 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4488 // See if we can fold the comparison based on bits known to be zero or one
4490 uint64_t KnownZero, KnownOne;
4491 if (SimplifyDemandedBits(Op0, Ty->getIntegralTypeMask(),
4492 KnownZero, KnownOne, 0))
4495 // Given the known and unknown bits, compute a range that the LHS could be
4497 if (KnownOne | KnownZero) {
4498 // Compute the Min, Max and RHS values based on the known bits. For the
4499 // EQ and NE we use unsigned values.
4500 uint64_t UMin = 0, UMax = 0, URHSVal = 0;
4501 int64_t SMin = 0, SMax = 0, SRHSVal = 0;
4502 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4503 SRHSVal = CI->getSExtValue();
4504 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
4507 URHSVal = CI->getZExtValue();
4508 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
4511 switch (I.getPredicate()) { // LE/GE have been folded already.
4512 default: assert(0 && "Unknown icmp opcode!");
4513 case ICmpInst::ICMP_EQ:
4514 if (UMax < URHSVal || UMin > URHSVal)
4515 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4517 case ICmpInst::ICMP_NE:
4518 if (UMax < URHSVal || UMin > URHSVal)
4519 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4521 case ICmpInst::ICMP_ULT:
4523 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4525 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4527 case ICmpInst::ICMP_UGT:
4529 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4531 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4533 case ICmpInst::ICMP_SLT:
4535 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4537 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4539 case ICmpInst::ICMP_SGT:
4541 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4543 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4548 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4549 // instruction, see if that instruction also has constants so that the
4550 // instruction can be folded into the icmp
4551 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4552 switch (LHSI->getOpcode()) {
4553 case Instruction::And:
4554 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4555 LHSI->getOperand(0)->hasOneUse()) {
4556 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4558 // If the LHS is an AND of a truncating cast, we can widen the
4559 // and/compare to be the input width without changing the value
4560 // produced, eliminating a cast.
4561 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4562 // We can do this transformation if either the AND constant does not
4563 // have its sign bit set or if it is an equality comparison.
4564 // Extending a relational comparison when we're checking the sign
4565 // bit would not work.
4566 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4568 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4569 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4570 ConstantInt *NewCST;
4572 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4573 AndCST->getZExtValue());
4574 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4575 CI->getZExtValue());
4576 Instruction *NewAnd =
4577 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4579 InsertNewInstBefore(NewAnd, I);
4580 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4584 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4585 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4586 // happens a LOT in code produced by the C front-end, for bitfield
4588 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
4590 // Check to see if there is a noop-cast between the shift and the and.
4592 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
4593 if (CI->getOpcode() == Instruction::BitCast)
4594 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
4598 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4599 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4600 const Type *AndTy = AndCST->getType(); // Type of the and.
4602 // We can fold this as long as we can't shift unknown bits
4603 // into the mask. This can only happen with signed shift
4604 // rights, as they sign-extend.
4606 bool CanFold = Shift->isLogicalShift();
4608 // To test for the bad case of the signed shr, see if any
4609 // of the bits shifted in could be tested after the mask.
4610 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4611 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4613 Constant *OShAmt = ConstantInt::get(Type::Int8Ty, ShAmtVal);
4615 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4617 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4623 if (Shift->getOpcode() == Instruction::Shl)
4624 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4626 NewCst = ConstantExpr::getShl(CI, ShAmt);
4628 // Check to see if we are shifting out any of the bits being
4630 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4631 // If we shifted bits out, the fold is not going to work out.
4632 // As a special case, check to see if this means that the
4633 // result is always true or false now.
4634 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4635 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4636 if (I.getPredicate() == ICmpInst::ICMP_NE)
4637 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4639 I.setOperand(1, NewCst);
4640 Constant *NewAndCST;
4641 if (Shift->getOpcode() == Instruction::Shl)
4642 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4644 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4645 LHSI->setOperand(1, NewAndCST);
4646 LHSI->setOperand(0, Shift->getOperand(0));
4647 WorkList.push_back(Shift); // Shift is dead.
4648 AddUsesToWorkList(I);
4654 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4655 // preferable because it allows the C<<Y expression to be hoisted out
4656 // of a loop if Y is invariant and X is not.
4657 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4658 I.isEquality() && !Shift->isArithmeticShift() &&
4659 isa<Instruction>(Shift->getOperand(0))) {
4662 if (Shift->getOpcode() == Instruction::LShr) {
4663 NS = new ShiftInst(Instruction::Shl, AndCST, Shift->getOperand(1),
4666 // Insert a logical shift.
4667 NS = new ShiftInst(Instruction::LShr, AndCST,
4668 Shift->getOperand(1), "tmp");
4670 InsertNewInstBefore(cast<Instruction>(NS), I);
4672 // Compute X & (C << Y).
4673 Instruction *NewAnd = BinaryOperator::createAnd(
4674 Shift->getOperand(0), NS, LHSI->getName());
4675 InsertNewInstBefore(NewAnd, I);
4677 I.setOperand(0, NewAnd);
4683 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4684 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4685 if (I.isEquality()) {
4686 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4688 // Check that the shift amount is in range. If not, don't perform
4689 // undefined shifts. When the shift is visited it will be
4691 if (ShAmt->getZExtValue() >= TypeBits)
4694 // If we are comparing against bits always shifted out, the
4695 // comparison cannot succeed.
4697 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4698 if (Comp != CI) {// Comparing against a bit that we know is zero.
4699 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4700 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4701 return ReplaceInstUsesWith(I, Cst);
4704 if (LHSI->hasOneUse()) {
4705 // Otherwise strength reduce the shift into an and.
4706 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4707 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4708 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4711 BinaryOperator::createAnd(LHSI->getOperand(0),
4712 Mask, LHSI->getName()+".mask");
4713 Value *And = InsertNewInstBefore(AndI, I);
4714 return new ICmpInst(I.getPredicate(), And,
4715 ConstantExpr::getLShr(CI, ShAmt));
4721 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4722 case Instruction::AShr:
4723 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4724 if (I.isEquality()) {
4725 // Check that the shift amount is in range. If not, don't perform
4726 // undefined shifts. When the shift is visited it will be
4728 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4729 if (ShAmt->getZExtValue() >= TypeBits)
4732 // If we are comparing against bits always shifted out, the
4733 // comparison cannot succeed.
4735 if (LHSI->getOpcode() == Instruction::LShr)
4736 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4739 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4742 if (Comp != CI) {// Comparing against a bit that we know is zero.
4743 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4744 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
4745 return ReplaceInstUsesWith(I, Cst);
4748 if (LHSI->hasOneUse() || CI->isNullValue()) {
4749 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4751 // Otherwise strength reduce the shift into an and.
4752 uint64_t Val = ~0ULL; // All ones.
4753 Val <<= ShAmtVal; // Shift over to the right spot.
4754 Val &= ~0ULL >> (64-TypeBits);
4755 Constant *Mask = ConstantInt::get(CI->getType(), Val);
4758 BinaryOperator::createAnd(LHSI->getOperand(0),
4759 Mask, LHSI->getName()+".mask");
4760 Value *And = InsertNewInstBefore(AndI, I);
4761 return new ICmpInst(I.getPredicate(), And,
4762 ConstantExpr::getShl(CI, ShAmt));
4768 case Instruction::SDiv:
4769 case Instruction::UDiv:
4770 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4771 // Fold this div into the comparison, producing a range check.
4772 // Determine, based on the divide type, what the range is being
4773 // checked. If there is an overflow on the low or high side, remember
4774 // it, otherwise compute the range [low, hi) bounding the new value.
4775 // See: InsertRangeTest above for the kinds of replacements possible.
4776 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4777 // FIXME: If the operand types don't match the type of the divide
4778 // then don't attempt this transform. The code below doesn't have the
4779 // logic to deal with a signed divide and an unsigned compare (and
4780 // vice versa). This is because (x /s C1) <s C2 produces different
4781 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4782 // (x /u C1) <u C2. Simply casting the operands and result won't
4783 // work. :( The if statement below tests that condition and bails
4785 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4786 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4789 // Initialize the variables that will indicate the nature of the
4791 bool LoOverflow = false, HiOverflow = false;
4792 ConstantInt *LoBound = 0, *HiBound = 0;
4794 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4795 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4796 // C2 (CI). By solving for X we can turn this into a range check
4797 // instead of computing a divide.
4799 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4801 // Determine if the product overflows by seeing if the product is
4802 // not equal to the divide. Make sure we do the same kind of divide
4803 // as in the LHS instruction that we're folding.
4804 bool ProdOV = !DivRHS->isNullValue() &&
4805 (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4806 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4808 // Get the ICmp opcode
4809 ICmpInst::Predicate predicate = I.getPredicate();
4811 if (DivRHS->isNullValue()) {
4812 // Don't hack on divide by zeros!
4813 } else if (!DivIsSigned) { // udiv
4815 LoOverflow = ProdOV;
4816 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4817 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4818 if (CI->isNullValue()) { // (X / pos) op 0
4820 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4822 } else if (isPositive(CI)) { // (X / pos) op pos
4824 LoOverflow = ProdOV;
4825 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4826 } else { // (X / pos) op neg
4827 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4828 LoOverflow = AddWithOverflow(LoBound, Prod,
4829 cast<ConstantInt>(DivRHSH));
4831 HiOverflow = ProdOV;
4833 } else { // Divisor is < 0.
4834 if (CI->isNullValue()) { // (X / neg) op 0
4835 LoBound = AddOne(DivRHS);
4836 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4837 if (HiBound == DivRHS)
4838 LoBound = 0; // - INTMIN = INTMIN
4839 } else if (isPositive(CI)) { // (X / neg) op pos
4840 HiOverflow = LoOverflow = ProdOV;
4842 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4843 HiBound = AddOne(Prod);
4844 } else { // (X / neg) op neg
4846 LoOverflow = HiOverflow = ProdOV;
4847 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4850 // Dividing by a negate swaps the condition.
4851 predicate = ICmpInst::getSwappedPredicate(predicate);
4855 Value *X = LHSI->getOperand(0);
4856 switch (predicate) {
4857 default: assert(0 && "Unhandled icmp opcode!");
4858 case ICmpInst::ICMP_EQ:
4859 if (LoOverflow && HiOverflow)
4860 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4861 else if (HiOverflow)
4862 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4863 ICmpInst::ICMP_UGE, X, LoBound);
4864 else if (LoOverflow)
4865 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4866 ICmpInst::ICMP_ULT, X, HiBound);
4868 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4870 case ICmpInst::ICMP_NE:
4871 if (LoOverflow && HiOverflow)
4872 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4873 else if (HiOverflow)
4874 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4875 ICmpInst::ICMP_ULT, X, LoBound);
4876 else if (LoOverflow)
4877 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4878 ICmpInst::ICMP_UGE, X, HiBound);
4880 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4882 case ICmpInst::ICMP_ULT:
4883 case ICmpInst::ICMP_SLT:
4885 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4886 return new ICmpInst(predicate, X, LoBound);
4887 case ICmpInst::ICMP_UGT:
4888 case ICmpInst::ICMP_SGT:
4890 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4891 if (predicate == ICmpInst::ICMP_UGT)
4892 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
4894 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
4901 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
4902 if (I.isEquality()) {
4903 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4905 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4906 // the second operand is a constant, simplify a bit.
4907 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4908 switch (BO->getOpcode()) {
4909 case Instruction::SRem:
4910 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4911 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4913 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4914 if (V > 1 && isPowerOf2_64(V)) {
4915 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4916 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4917 return new ICmpInst(I.getPredicate(), NewRem,
4918 Constant::getNullValue(BO->getType()));
4922 case Instruction::Add:
4923 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4924 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4925 if (BO->hasOneUse())
4926 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4927 ConstantExpr::getSub(CI, BOp1C));
4928 } else if (CI->isNullValue()) {
4929 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4930 // efficiently invertible, or if the add has just this one use.
4931 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4933 if (Value *NegVal = dyn_castNegVal(BOp1))
4934 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
4935 else if (Value *NegVal = dyn_castNegVal(BOp0))
4936 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
4937 else if (BO->hasOneUse()) {
4938 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
4940 InsertNewInstBefore(Neg, I);
4941 return new ICmpInst(I.getPredicate(), BOp0, Neg);
4945 case Instruction::Xor:
4946 // For the xor case, we can xor two constants together, eliminating
4947 // the explicit xor.
4948 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4949 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4950 ConstantExpr::getXor(CI, BOC));
4953 case Instruction::Sub:
4954 // Replace (([sub|xor] A, B) != 0) with (A != B)
4955 if (CI->isNullValue())
4956 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4960 case Instruction::Or:
4961 // If bits are being or'd in that are not present in the constant we
4962 // are comparing against, then the comparison could never succeed!
4963 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4964 Constant *NotCI = ConstantExpr::getNot(CI);
4965 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4966 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4971 case Instruction::And:
4972 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4973 // If bits are being compared against that are and'd out, then the
4974 // comparison can never succeed!
4975 if (!ConstantExpr::getAnd(CI,
4976 ConstantExpr::getNot(BOC))->isNullValue())
4977 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4980 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4981 if (CI == BOC && isOneBitSet(CI))
4982 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
4983 ICmpInst::ICMP_NE, Op0,
4984 Constant::getNullValue(CI->getType()));
4986 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
4987 if (isSignBit(BOC)) {
4988 Value *X = BO->getOperand(0);
4989 Constant *Zero = Constant::getNullValue(X->getType());
4990 ICmpInst::Predicate pred = isICMP_NE ?
4991 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
4992 return new ICmpInst(pred, X, Zero);
4995 // ((X & ~7) == 0) --> X < 8
4996 if (CI->isNullValue() && isHighOnes(BOC)) {
4997 Value *X = BO->getOperand(0);
4998 Constant *NegX = ConstantExpr::getNeg(BOC);
4999 ICmpInst::Predicate pred = isICMP_NE ?
5000 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5001 return new ICmpInst(pred, X, NegX);
5007 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5008 // Handle set{eq|ne} <intrinsic>, intcst.
5009 switch (II->getIntrinsicID()) {
5011 case Intrinsic::bswap_i16:
5012 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5013 WorkList.push_back(II); // Dead?
5014 I.setOperand(0, II->getOperand(1));
5015 I.setOperand(1, ConstantInt::get(Type::Int16Ty,
5016 ByteSwap_16(CI->getZExtValue())));
5018 case Intrinsic::bswap_i32:
5019 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5020 WorkList.push_back(II); // Dead?
5021 I.setOperand(0, II->getOperand(1));
5022 I.setOperand(1, ConstantInt::get(Type::Int32Ty,
5023 ByteSwap_32(CI->getZExtValue())));
5025 case Intrinsic::bswap_i64:
5026 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5027 WorkList.push_back(II); // Dead?
5028 I.setOperand(0, II->getOperand(1));
5029 I.setOperand(1, ConstantInt::get(Type::Int64Ty,
5030 ByteSwap_64(CI->getZExtValue())));
5034 } else { // Not a ICMP_EQ/ICMP_NE
5035 // If the LHS is a cast from an integral value of the same size, then
5036 // since we know the RHS is a constant, try to simlify.
5037 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5038 Value *CastOp = Cast->getOperand(0);
5039 const Type *SrcTy = CastOp->getType();
5040 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5041 if (SrcTy->isInteger() &&
5042 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5043 // If this is an unsigned comparison, try to make the comparison use
5044 // smaller constant values.
5045 switch (I.getPredicate()) {
5047 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5048 ConstantInt *CUI = cast<ConstantInt>(CI);
5049 if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
5050 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5051 ConstantInt::get(SrcTy, -1));
5054 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5055 ConstantInt *CUI = cast<ConstantInt>(CI);
5056 if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
5057 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5058 Constant::getNullValue(SrcTy));
5068 // Handle icmp with constant RHS
5069 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5070 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5071 switch (LHSI->getOpcode()) {
5072 case Instruction::GetElementPtr:
5073 if (RHSC->isNullValue()) {
5074 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5075 bool isAllZeros = true;
5076 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5077 if (!isa<Constant>(LHSI->getOperand(i)) ||
5078 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5083 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5084 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5088 case Instruction::PHI:
5089 if (Instruction *NV = FoldOpIntoPhi(I))
5092 case Instruction::Select:
5093 // If either operand of the select is a constant, we can fold the
5094 // comparison into the select arms, which will cause one to be
5095 // constant folded and the select turned into a bitwise or.
5096 Value *Op1 = 0, *Op2 = 0;
5097 if (LHSI->hasOneUse()) {
5098 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5099 // Fold the known value into the constant operand.
5100 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5101 // Insert a new ICmp of the other select operand.
5102 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5103 LHSI->getOperand(2), RHSC,
5105 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5106 // Fold the known value into the constant operand.
5107 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5108 // Insert a new ICmp of the other select operand.
5109 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5110 LHSI->getOperand(1), RHSC,
5116 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5121 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5122 if (User *GEP = dyn_castGetElementPtr(Op0))
5123 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5125 if (User *GEP = dyn_castGetElementPtr(Op1))
5126 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5127 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5130 // Test to see if the operands of the icmp are casted versions of other
5131 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5133 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5134 if (isa<PointerType>(Op0->getType()) &&
5135 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5136 // We keep moving the cast from the left operand over to the right
5137 // operand, where it can often be eliminated completely.
5138 Op0 = CI->getOperand(0);
5140 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5141 // so eliminate it as well.
5142 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5143 Op1 = CI2->getOperand(0);
5145 // If Op1 is a constant, we can fold the cast into the constant.
5146 if (Op0->getType() != Op1->getType())
5147 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5148 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5150 // Otherwise, cast the RHS right before the icmp
5151 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5153 return new ICmpInst(I.getPredicate(), Op0, Op1);
5157 if (isa<CastInst>(Op0)) {
5158 // Handle the special case of: icmp (cast bool to X), <cst>
5159 // This comes up when you have code like
5162 // For generality, we handle any zero-extension of any operand comparison
5163 // with a constant or another cast from the same type.
5164 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5165 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5169 if (I.isEquality()) {
5170 Value *A, *B, *C, *D;
5171 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5172 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5173 Value *OtherVal = A == Op1 ? B : A;
5174 return new ICmpInst(I.getPredicate(), OtherVal,
5175 Constant::getNullValue(A->getType()));
5178 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5179 // A^c1 == C^c2 --> A == C^(c1^c2)
5180 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5181 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5182 if (Op1->hasOneUse()) {
5183 Constant *NC = ConstantExpr::getXor(C1, C2);
5184 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5185 return new ICmpInst(I.getPredicate(), A,
5186 InsertNewInstBefore(Xor, I));
5189 // A^B == A^D -> B == D
5190 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5191 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5192 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5193 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5197 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5198 (A == Op0 || B == Op0)) {
5199 // A == (A^B) -> B == 0
5200 Value *OtherVal = A == Op0 ? B : A;
5201 return new ICmpInst(I.getPredicate(), OtherVal,
5202 Constant::getNullValue(A->getType()));
5204 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5205 // (A-B) == A -> B == 0
5206 return new ICmpInst(I.getPredicate(), B,
5207 Constant::getNullValue(B->getType()));
5209 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5210 // A == (A-B) -> B == 0
5211 return new ICmpInst(I.getPredicate(), B,
5212 Constant::getNullValue(B->getType()));
5215 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5216 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5217 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5218 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5219 Value *X = 0, *Y = 0, *Z = 0;
5222 X = B; Y = D; Z = A;
5223 } else if (A == D) {
5224 X = B; Y = C; Z = A;
5225 } else if (B == C) {
5226 X = A; Y = D; Z = B;
5227 } else if (B == D) {
5228 X = A; Y = C; Z = B;
5231 if (X) { // Build (X^Y) & Z
5232 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5233 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5234 I.setOperand(0, Op1);
5235 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5240 return Changed ? &I : 0;
5243 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5244 // We only handle extending casts so far.
5246 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5247 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5248 Value *LHSCIOp = LHSCI->getOperand(0);
5249 const Type *SrcTy = LHSCIOp->getType();
5250 const Type *DestTy = LHSCI->getType();
5253 // We only handle extension cast instructions, so far. Enforce this.
5254 if (LHSCI->getOpcode() != Instruction::ZExt &&
5255 LHSCI->getOpcode() != Instruction::SExt)
5258 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5259 bool isSignedCmp = ICI.isSignedPredicate();
5261 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5262 // Not an extension from the same type?
5263 RHSCIOp = CI->getOperand(0);
5264 if (RHSCIOp->getType() != LHSCIOp->getType())
5267 // Okay, just insert a compare of the reduced operands now!
5268 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5271 // If we aren't dealing with a constant on the RHS, exit early
5272 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5276 // Compute the constant that would happen if we truncated to SrcTy then
5277 // reextended to DestTy.
5278 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5279 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5281 // If the re-extended constant didn't change...
5283 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5284 // For example, we might have:
5285 // %A = sext short %X to uint
5286 // %B = icmp ugt uint %A, 1330
5287 // It is incorrect to transform this into
5288 // %B = icmp ugt short %X, 1330
5289 // because %A may have negative value.
5291 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5292 // OR operation is EQ/NE.
5293 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5294 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5299 // The re-extended constant changed so the constant cannot be represented
5300 // in the shorter type. Consequently, we cannot emit a simple comparison.
5302 // First, handle some easy cases. We know the result cannot be equal at this
5303 // point so handle the ICI.isEquality() cases
5304 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5305 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5306 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5307 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5309 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5310 // should have been folded away previously and not enter in here.
5313 // We're performing a signed comparison.
5314 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5315 Result = ConstantInt::getFalse(); // X < (small) --> false
5317 Result = ConstantInt::getTrue(); // X < (large) --> true
5319 // We're performing an unsigned comparison.
5321 // We're performing an unsigned comp with a sign extended value.
5322 // This is true if the input is >= 0. [aka >s -1]
5323 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5324 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5325 NegOne, ICI.getName()), ICI);
5327 // Unsigned extend & unsigned compare -> always true.
5328 Result = ConstantInt::getTrue();
5332 // Finally, return the value computed.
5333 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5334 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5335 return ReplaceInstUsesWith(ICI, Result);
5337 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5338 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5339 "ICmp should be folded!");
5340 if (Constant *CI = dyn_cast<Constant>(Result))
5341 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5343 return BinaryOperator::createNot(Result);
5347 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
5348 assert(I.getOperand(1)->getType() == Type::Int8Ty);
5349 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5351 // shl X, 0 == X and shr X, 0 == X
5352 // shl 0, X == 0 and shr 0, X == 0
5353 if (Op1 == Constant::getNullValue(Type::Int8Ty) ||
5354 Op0 == Constant::getNullValue(Op0->getType()))
5355 return ReplaceInstUsesWith(I, Op0);
5357 if (isa<UndefValue>(Op0)) {
5358 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5359 return ReplaceInstUsesWith(I, Op0);
5360 else // undef << X -> 0, undef >>u X -> 0
5361 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5363 if (isa<UndefValue>(Op1)) {
5364 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5365 return ReplaceInstUsesWith(I, Op0);
5366 else // X << undef, X >>u undef -> 0
5367 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5370 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5371 if (I.getOpcode() == Instruction::AShr)
5372 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5373 if (CSI->isAllOnesValue())
5374 return ReplaceInstUsesWith(I, CSI);
5376 // Try to fold constant and into select arguments.
5377 if (isa<Constant>(Op0))
5378 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5379 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5382 // See if we can turn a signed shr into an unsigned shr.
5383 if (I.isArithmeticShift()) {
5384 if (MaskedValueIsZero(Op0,
5385 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5386 return new ShiftInst(Instruction::LShr, Op0, Op1, I.getName());
5390 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5391 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5396 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5398 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5399 bool isSignedShift = I.getOpcode() == Instruction::AShr;
5400 bool isUnsignedShift = !isSignedShift;
5402 // See if we can simplify any instructions used by the instruction whose sole
5403 // purpose is to compute bits we don't care about.
5404 uint64_t KnownZero, KnownOne;
5405 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
5406 KnownZero, KnownOne))
5409 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5410 // of a signed value.
5412 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5413 if (Op1->getZExtValue() >= TypeBits) {
5414 if (isUnsignedShift || isLeftShift)
5415 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5417 I.setOperand(1, ConstantInt::get(Type::Int8Ty, TypeBits-1));
5422 // ((X*C1) << C2) == (X * (C1 << C2))
5423 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5424 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5425 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5426 return BinaryOperator::createMul(BO->getOperand(0),
5427 ConstantExpr::getShl(BOOp, Op1));
5429 // Try to fold constant and into select arguments.
5430 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5431 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5433 if (isa<PHINode>(Op0))
5434 if (Instruction *NV = FoldOpIntoPhi(I))
5437 if (Op0->hasOneUse()) {
5438 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5439 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5442 switch (Op0BO->getOpcode()) {
5444 case Instruction::Add:
5445 case Instruction::And:
5446 case Instruction::Or:
5447 case Instruction::Xor:
5448 // These operators commute.
5449 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5450 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5451 match(Op0BO->getOperand(1),
5452 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5453 Instruction *YS = new ShiftInst(Instruction::Shl,
5454 Op0BO->getOperand(0), Op1,
5456 InsertNewInstBefore(YS, I); // (Y << C)
5458 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5459 Op0BO->getOperand(1)->getName());
5460 InsertNewInstBefore(X, I); // (X + (Y << C))
5461 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5462 C2 = ConstantExpr::getShl(C2, Op1);
5463 return BinaryOperator::createAnd(X, C2);
5466 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5467 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5468 match(Op0BO->getOperand(1),
5469 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5470 m_ConstantInt(CC))) && V2 == Op1 &&
5471 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
5472 Instruction *YS = new ShiftInst(Instruction::Shl,
5473 Op0BO->getOperand(0), Op1,
5475 InsertNewInstBefore(YS, I); // (Y << C)
5477 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5478 V1->getName()+".mask");
5479 InsertNewInstBefore(XM, I); // X & (CC << C)
5481 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5485 case Instruction::Sub:
5486 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5487 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5488 match(Op0BO->getOperand(0),
5489 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5490 Instruction *YS = new ShiftInst(Instruction::Shl,
5491 Op0BO->getOperand(1), Op1,
5493 InsertNewInstBefore(YS, I); // (Y << C)
5495 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5496 Op0BO->getOperand(0)->getName());
5497 InsertNewInstBefore(X, I); // (X + (Y << C))
5498 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5499 C2 = ConstantExpr::getShl(C2, Op1);
5500 return BinaryOperator::createAnd(X, C2);
5503 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5504 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5505 match(Op0BO->getOperand(0),
5506 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5507 m_ConstantInt(CC))) && V2 == Op1 &&
5508 cast<BinaryOperator>(Op0BO->getOperand(0))
5509 ->getOperand(0)->hasOneUse()) {
5510 Instruction *YS = new ShiftInst(Instruction::Shl,
5511 Op0BO->getOperand(1), Op1,
5513 InsertNewInstBefore(YS, I); // (Y << C)
5515 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5516 V1->getName()+".mask");
5517 InsertNewInstBefore(XM, I); // X & (CC << C)
5519 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5526 // If the operand is an bitwise operator with a constant RHS, and the
5527 // shift is the only use, we can pull it out of the shift.
5528 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5529 bool isValid = true; // Valid only for And, Or, Xor
5530 bool highBitSet = false; // Transform if high bit of constant set?
5532 switch (Op0BO->getOpcode()) {
5533 default: isValid = false; break; // Do not perform transform!
5534 case Instruction::Add:
5535 isValid = isLeftShift;
5537 case Instruction::Or:
5538 case Instruction::Xor:
5541 case Instruction::And:
5546 // If this is a signed shift right, and the high bit is modified
5547 // by the logical operation, do not perform the transformation.
5548 // The highBitSet boolean indicates the value of the high bit of
5549 // the constant which would cause it to be modified for this
5552 if (isValid && !isLeftShift && isSignedShift) {
5553 uint64_t Val = Op0C->getZExtValue();
5554 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5558 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5560 Instruction *NewShift =
5561 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
5564 InsertNewInstBefore(NewShift, I);
5566 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5573 // Find out if this is a shift of a shift by a constant.
5574 ShiftInst *ShiftOp = 0;
5575 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
5577 else if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5578 // If this is a noop-integer cast of a shift instruction, use the shift.
5579 if (isa<ShiftInst>(CI->getOperand(0))) {
5580 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
5584 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5585 // Find the operands and properties of the input shift. Note that the
5586 // signedness of the input shift may differ from the current shift if there
5587 // is a noop cast between the two.
5588 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
5589 bool isShiftOfSignedShift = ShiftOp->getOpcode() == Instruction::AShr;
5590 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
5592 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5594 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5595 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5597 // Check for (A << c1) << c2 and (A >> c1) >> c2.
5598 if (isLeftShift == isShiftOfLeftShift) {
5599 // Do not fold these shifts if the first one is signed and the second one
5600 // is unsigned and this is a right shift. Further, don't do any folding
5602 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
5605 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5606 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
5607 Amt = Op0->getType()->getPrimitiveSizeInBits();
5609 Value *Op = ShiftOp->getOperand(0);
5610 ShiftInst *ShiftResult = new ShiftInst(I.getOpcode(), Op,
5611 ConstantInt::get(Type::Int8Ty, Amt));
5612 if (I.getType() == ShiftResult->getType())
5614 InsertNewInstBefore(ShiftResult, I);
5615 return CastInst::create(Instruction::BitCast, ShiftResult, I.getType());
5618 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
5619 // signed types, we can only support the (A >> c1) << c2 configuration,
5620 // because it can not turn an arbitrary bit of A into a sign bit.
5621 if (isUnsignedShift || isLeftShift) {
5622 // Calculate bitmask for what gets shifted off the edge.
5623 Constant *C = ConstantInt::getAllOnesValue(I.getType());
5625 C = ConstantExpr::getShl(C, ShiftAmt1C);
5627 C = ConstantExpr::getLShr(C, ShiftAmt1C);
5629 Value *Op = ShiftOp->getOperand(0);
5632 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
5633 InsertNewInstBefore(Mask, I);
5635 // Figure out what flavor of shift we should use...
5636 if (ShiftAmt1 == ShiftAmt2) {
5637 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
5638 } else if (ShiftAmt1 < ShiftAmt2) {
5639 return new ShiftInst(I.getOpcode(), Mask,
5640 ConstantInt::get(Type::Int8Ty, ShiftAmt2-ShiftAmt1));
5641 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
5642 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
5643 return new ShiftInst(Instruction::LShr, Mask,
5644 ConstantInt::get(Type::Int8Ty, ShiftAmt1-ShiftAmt2));
5646 return new ShiftInst(ShiftOp->getOpcode(), Mask,
5647 ConstantInt::get(Type::Int8Ty, ShiftAmt1-ShiftAmt2));
5650 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
5651 Instruction *Shift =
5652 new ShiftInst(ShiftOp->getOpcode(), Mask,
5653 ConstantInt::get(Type::Int8Ty, ShiftAmt1-ShiftAmt2));
5654 InsertNewInstBefore(Shift, I);
5656 C = ConstantInt::getAllOnesValue(Shift->getType());
5657 C = ConstantExpr::getShl(C, Op1);
5658 return BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
5661 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
5662 // this case, C1 == C2 and C1 is 8, 16, or 32.
5663 if (ShiftAmt1 == ShiftAmt2) {
5664 const Type *SExtType = 0;
5665 switch (Op0->getType()->getPrimitiveSizeInBits() - ShiftAmt1) {
5666 case 8 : SExtType = Type::Int8Ty; break;
5667 case 16: SExtType = Type::Int16Ty; break;
5668 case 32: SExtType = Type::Int32Ty; break;
5672 Instruction *NewTrunc =
5673 new TruncInst(ShiftOp->getOperand(0), SExtType, "sext");
5674 InsertNewInstBefore(NewTrunc, I);
5675 return new SExtInst(NewTrunc, I.getType());
5684 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5685 /// expression. If so, decompose it, returning some value X, such that Val is
5688 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5690 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
5691 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5692 Offset = CI->getZExtValue();
5694 return ConstantInt::get(Type::Int32Ty, 0);
5695 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5696 if (I->getNumOperands() == 2) {
5697 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5698 if (I->getOpcode() == Instruction::Shl) {
5699 // This is a value scaled by '1 << the shift amt'.
5700 Scale = 1U << CUI->getZExtValue();
5702 return I->getOperand(0);
5703 } else if (I->getOpcode() == Instruction::Mul) {
5704 // This value is scaled by 'CUI'.
5705 Scale = CUI->getZExtValue();
5707 return I->getOperand(0);
5708 } else if (I->getOpcode() == Instruction::Add) {
5709 // We have X+C. Check to see if we really have (X*C2)+C1,
5710 // where C1 is divisible by C2.
5713 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5714 Offset += CUI->getZExtValue();
5715 if (SubScale > 1 && (Offset % SubScale == 0)) {
5724 // Otherwise, we can't look past this.
5731 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5732 /// try to eliminate the cast by moving the type information into the alloc.
5733 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5734 AllocationInst &AI) {
5735 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5736 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5738 // Remove any uses of AI that are dead.
5739 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5740 std::vector<Instruction*> DeadUsers;
5741 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5742 Instruction *User = cast<Instruction>(*UI++);
5743 if (isInstructionTriviallyDead(User)) {
5744 while (UI != E && *UI == User)
5745 ++UI; // If this instruction uses AI more than once, don't break UI.
5747 // Add operands to the worklist.
5748 AddUsesToWorkList(*User);
5750 DOUT << "IC: DCE: " << *User;
5752 User->eraseFromParent();
5753 removeFromWorkList(User);
5757 // Get the type really allocated and the type casted to.
5758 const Type *AllocElTy = AI.getAllocatedType();
5759 const Type *CastElTy = PTy->getElementType();
5760 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5762 unsigned AllocElTyAlign = TD->getTypeAlignment(AllocElTy);
5763 unsigned CastElTyAlign = TD->getTypeAlignment(CastElTy);
5764 if (CastElTyAlign < AllocElTyAlign) return 0;
5766 // If the allocation has multiple uses, only promote it if we are strictly
5767 // increasing the alignment of the resultant allocation. If we keep it the
5768 // same, we open the door to infinite loops of various kinds.
5769 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5771 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5772 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5773 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5775 // See if we can satisfy the modulus by pulling a scale out of the array
5777 unsigned ArraySizeScale, ArrayOffset;
5778 Value *NumElements = // See if the array size is a decomposable linear expr.
5779 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5781 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5783 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5784 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5786 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5791 // If the allocation size is constant, form a constant mul expression
5792 Amt = ConstantInt::get(Type::Int32Ty, Scale);
5793 if (isa<ConstantInt>(NumElements))
5794 Amt = ConstantExpr::getMul(
5795 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5796 // otherwise multiply the amount and the number of elements
5797 else if (Scale != 1) {
5798 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5799 Amt = InsertNewInstBefore(Tmp, AI);
5803 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5804 Value *Off = ConstantInt::get(Type::Int32Ty, Offset);
5805 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5806 Amt = InsertNewInstBefore(Tmp, AI);
5809 std::string Name = AI.getName(); AI.setName("");
5810 AllocationInst *New;
5811 if (isa<MallocInst>(AI))
5812 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
5814 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
5815 InsertNewInstBefore(New, AI);
5817 // If the allocation has multiple uses, insert a cast and change all things
5818 // that used it to use the new cast. This will also hack on CI, but it will
5820 if (!AI.hasOneUse()) {
5821 AddUsesToWorkList(AI);
5822 // New is the allocation instruction, pointer typed. AI is the original
5823 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5824 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5825 InsertNewInstBefore(NewCast, AI);
5826 AI.replaceAllUsesWith(NewCast);
5828 return ReplaceInstUsesWith(CI, New);
5831 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5832 /// and return it without inserting any new casts. This is used by code that
5833 /// tries to decide whether promoting or shrinking integer operations to wider
5834 /// or smaller types will allow us to eliminate a truncate or extend.
5835 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5836 int &NumCastsRemoved) {
5837 if (isa<Constant>(V)) return true;
5839 Instruction *I = dyn_cast<Instruction>(V);
5840 if (!I || !I->hasOneUse()) return false;
5842 switch (I->getOpcode()) {
5843 case Instruction::And:
5844 case Instruction::Or:
5845 case Instruction::Xor:
5846 // These operators can all arbitrarily be extended or truncated.
5847 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5848 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5849 case Instruction::AShr:
5850 case Instruction::LShr:
5851 case Instruction::Shl:
5852 // If this is just a bitcast changing the sign of the operation, we can
5853 // convert if the operand can be converted.
5854 if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
5855 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5857 case Instruction::Trunc:
5858 case Instruction::ZExt:
5859 case Instruction::SExt:
5860 case Instruction::BitCast:
5861 // If this is a cast from the destination type, we can trivially eliminate
5862 // it, and this will remove a cast overall.
5863 if (I->getOperand(0)->getType() == Ty) {
5864 // If the first operand is itself a cast, and is eliminable, do not count
5865 // this as an eliminable cast. We would prefer to eliminate those two
5867 if (isa<CastInst>(I->getOperand(0)))
5875 // TODO: Can handle more cases here.
5882 /// EvaluateInDifferentType - Given an expression that
5883 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5884 /// evaluate the expression.
5885 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
5887 if (Constant *C = dyn_cast<Constant>(V))
5888 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
5890 // Otherwise, it must be an instruction.
5891 Instruction *I = cast<Instruction>(V);
5892 Instruction *Res = 0;
5893 switch (I->getOpcode()) {
5894 case Instruction::And:
5895 case Instruction::Or:
5896 case Instruction::Xor: {
5897 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5898 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
5899 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5900 LHS, RHS, I->getName());
5903 case Instruction::AShr:
5904 case Instruction::LShr:
5905 case Instruction::Shl: {
5906 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5907 Res = new ShiftInst((Instruction::OtherOps)I->getOpcode(), LHS,
5908 I->getOperand(1), I->getName());
5911 case Instruction::Trunc:
5912 case Instruction::ZExt:
5913 case Instruction::SExt:
5914 case Instruction::BitCast:
5915 // If the source type of the cast is the type we're trying for then we can
5916 // just return the source. There's no need to insert it because its not new.
5917 if (I->getOperand(0)->getType() == Ty)
5918 return I->getOperand(0);
5920 // Some other kind of cast, which shouldn't happen, so just ..
5923 // TODO: Can handle more cases here.
5924 assert(0 && "Unreachable!");
5928 return InsertNewInstBefore(Res, *I);
5931 /// @brief Implement the transforms common to all CastInst visitors.
5932 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
5933 Value *Src = CI.getOperand(0);
5935 // Casting undef to anything results in undef so might as just replace it and
5936 // get rid of the cast.
5937 if (isa<UndefValue>(Src)) // cast undef -> undef
5938 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5940 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
5941 // eliminate it now.
5942 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5943 if (Instruction::CastOps opc =
5944 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
5945 // The first cast (CSrc) is eliminable so we need to fix up or replace
5946 // the second cast (CI). CSrc will then have a good chance of being dead.
5947 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
5951 // If casting the result of a getelementptr instruction with no offset, turn
5952 // this into a cast of the original pointer!
5954 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
5955 bool AllZeroOperands = true;
5956 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
5957 if (!isa<Constant>(GEP->getOperand(i)) ||
5958 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
5959 AllZeroOperands = false;
5962 if (AllZeroOperands) {
5963 // Changing the cast operand is usually not a good idea but it is safe
5964 // here because the pointer operand is being replaced with another
5965 // pointer operand so the opcode doesn't need to change.
5966 CI.setOperand(0, GEP->getOperand(0));
5971 // If we are casting a malloc or alloca to a pointer to a type of the same
5972 // size, rewrite the allocation instruction to allocate the "right" type.
5973 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
5974 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
5977 // If we are casting a select then fold the cast into the select
5978 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
5979 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
5982 // If we are casting a PHI then fold the cast into the PHI
5983 if (isa<PHINode>(Src))
5984 if (Instruction *NV = FoldOpIntoPhi(CI))
5990 /// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
5991 /// integers. This function implements the common transforms for all those
5993 /// @brief Implement the transforms common to CastInst with integer operands
5994 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
5995 if (Instruction *Result = commonCastTransforms(CI))
5998 Value *Src = CI.getOperand(0);
5999 const Type *SrcTy = Src->getType();
6000 const Type *DestTy = CI.getType();
6001 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6002 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6004 // See if we can simplify any instructions used by the LHS whose sole
6005 // purpose is to compute bits we don't care about.
6006 uint64_t KnownZero = 0, KnownOne = 0;
6007 if (SimplifyDemandedBits(&CI, DestTy->getIntegralTypeMask(),
6008 KnownZero, KnownOne))
6011 // If the source isn't an instruction or has more than one use then we
6012 // can't do anything more.
6013 Instruction *SrcI = dyn_cast<Instruction>(Src);
6014 if (!SrcI || !Src->hasOneUse())
6017 // Attempt to propagate the cast into the instruction.
6018 int NumCastsRemoved = 0;
6019 if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
6020 // If this cast is a truncate, evaluting in a different type always
6021 // eliminates the cast, so it is always a win. If this is a noop-cast
6022 // this just removes a noop cast which isn't pointful, but simplifies
6023 // the code. If this is a zero-extension, we need to do an AND to
6024 // maintain the clear top-part of the computation, so we require that
6025 // the input have eliminated at least one cast. If this is a sign
6026 // extension, we insert two new casts (to do the extension) so we
6027 // require that two casts have been eliminated.
6028 bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
6030 switch (CI.getOpcode()) {
6031 case Instruction::Trunc:
6034 case Instruction::ZExt:
6035 DoXForm = NumCastsRemoved >= 1;
6037 case Instruction::SExt:
6038 DoXForm = NumCastsRemoved >= 2;
6040 case Instruction::BitCast:
6044 // All the others use floating point so we shouldn't actually
6045 // get here because of the check above.
6046 assert(!"Unknown cast type .. unreachable");
6052 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6053 CI.getOpcode() == Instruction::SExt);
6054 assert(Res->getType() == DestTy);
6055 switch (CI.getOpcode()) {
6056 default: assert(0 && "Unknown cast type!");
6057 case Instruction::Trunc:
6058 case Instruction::BitCast:
6059 // Just replace this cast with the result.
6060 return ReplaceInstUsesWith(CI, Res);
6061 case Instruction::ZExt: {
6062 // We need to emit an AND to clear the high bits.
6063 assert(SrcBitSize < DestBitSize && "Not a zext?");
6065 ConstantInt::get(Type::Int64Ty, (1ULL << SrcBitSize)-1);
6066 if (DestBitSize < 64)
6067 C = ConstantExpr::getTrunc(C, DestTy);
6068 return BinaryOperator::createAnd(Res, C);
6070 case Instruction::SExt:
6071 // We need to emit a cast to truncate, then a cast to sext.
6072 return CastInst::create(Instruction::SExt,
6073 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6079 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6080 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6082 switch (SrcI->getOpcode()) {
6083 case Instruction::Add:
6084 case Instruction::Mul:
6085 case Instruction::And:
6086 case Instruction::Or:
6087 case Instruction::Xor:
6088 // If we are discarding information, or just changing the sign,
6090 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6091 // Don't insert two casts if they cannot be eliminated. We allow
6092 // two casts to be inserted if the sizes are the same. This could
6093 // only be converting signedness, which is a noop.
6094 if (DestBitSize == SrcBitSize ||
6095 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6096 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6097 Instruction::CastOps opcode = CI.getOpcode();
6098 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6099 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6100 return BinaryOperator::create(
6101 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6105 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6106 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6107 SrcI->getOpcode() == Instruction::Xor &&
6108 Op1 == ConstantInt::getTrue() &&
6109 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6110 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6111 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6114 case Instruction::SDiv:
6115 case Instruction::UDiv:
6116 case Instruction::SRem:
6117 case Instruction::URem:
6118 // If we are just changing the sign, rewrite.
6119 if (DestBitSize == SrcBitSize) {
6120 // Don't insert two casts if they cannot be eliminated. We allow
6121 // two casts to be inserted if the sizes are the same. This could
6122 // only be converting signedness, which is a noop.
6123 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6124 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6125 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6127 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6129 return BinaryOperator::create(
6130 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6135 case Instruction::Shl:
6136 // Allow changing the sign of the source operand. Do not allow
6137 // changing the size of the shift, UNLESS the shift amount is a
6138 // constant. We must not change variable sized shifts to a smaller
6139 // size, because it is undefined to shift more bits out than exist
6141 if (DestBitSize == SrcBitSize ||
6142 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6143 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6144 Instruction::BitCast : Instruction::Trunc);
6145 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6146 return new ShiftInst(Instruction::Shl, Op0c, Op1);
6149 case Instruction::AShr:
6150 // If this is a signed shr, and if all bits shifted in are about to be
6151 // truncated off, turn it into an unsigned shr to allow greater
6153 if (DestBitSize < SrcBitSize &&
6154 isa<ConstantInt>(Op1)) {
6155 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6156 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6157 // Insert the new logical shift right.
6158 return new ShiftInst(Instruction::LShr, Op0, Op1);
6163 case Instruction::ICmp:
6164 // If we are just checking for a icmp eq of a single bit and casting it
6165 // to an integer, then shift the bit to the appropriate place and then
6166 // cast to integer to avoid the comparison.
6167 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6168 uint64_t Op1CV = Op1C->getZExtValue();
6169 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6170 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6171 // cast (X == 1) to int --> X iff X has only the low bit set.
6172 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6173 // cast (X != 0) to int --> X iff X has only the low bit set.
6174 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6175 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6176 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6177 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
6178 // If Op1C some other power of two, convert:
6179 uint64_t KnownZero, KnownOne;
6180 uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
6181 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6183 // This only works for EQ and NE
6184 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6185 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6188 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
6189 bool isNE = pred == ICmpInst::ICMP_NE;
6190 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
6191 // (X&4) == 2 --> false
6192 // (X&4) != 2 --> true
6193 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6194 Res = ConstantExpr::getZExt(Res, CI.getType());
6195 return ReplaceInstUsesWith(CI, Res);
6198 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
6201 // Perform a logical shr by shiftamt.
6202 // Insert the shift to put the result in the low bit.
6203 In = InsertNewInstBefore(
6204 new ShiftInst(Instruction::LShr, In,
6205 ConstantInt::get(Type::Int8Ty, ShiftAmt),
6206 In->getName()+".lobit"), CI);
6209 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6210 Constant *One = ConstantInt::get(In->getType(), 1);
6211 In = BinaryOperator::createXor(In, One, "tmp");
6212 InsertNewInstBefore(cast<Instruction>(In), CI);
6215 if (CI.getType() == In->getType())
6216 return ReplaceInstUsesWith(CI, In);
6218 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6227 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6228 if (Instruction *Result = commonIntCastTransforms(CI))
6231 Value *Src = CI.getOperand(0);
6232 const Type *Ty = CI.getType();
6233 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6235 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6236 switch (SrcI->getOpcode()) {
6238 case Instruction::LShr:
6239 // We can shrink lshr to something smaller if we know the bits shifted in
6240 // are already zeros.
6241 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6242 unsigned ShAmt = ShAmtV->getZExtValue();
6244 // Get a mask for the bits shifting in.
6245 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
6246 Value* SrcIOp0 = SrcI->getOperand(0);
6247 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6248 if (ShAmt >= DestBitWidth) // All zeros.
6249 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6251 // Okay, we can shrink this. Truncate the input, then return a new
6253 Value *V = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6254 return new ShiftInst(Instruction::LShr, V, SrcI->getOperand(1));
6256 } else { // This is a variable shr.
6258 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6259 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6260 // loop-invariant and CSE'd.
6261 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6262 Value *One = ConstantInt::get(SrcI->getType(), 1);
6264 Value *V = InsertNewInstBefore(new ShiftInst(Instruction::Shl, One,
6265 SrcI->getOperand(1),
6267 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6268 SrcI->getOperand(0),
6270 Value *Zero = Constant::getNullValue(V->getType());
6271 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6281 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6282 // If one of the common conversion will work ..
6283 if (Instruction *Result = commonIntCastTransforms(CI))
6286 Value *Src = CI.getOperand(0);
6288 // If this is a cast of a cast
6289 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6290 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6291 // types and if the sizes are just right we can convert this into a logical
6292 // 'and' which will be much cheaper than the pair of casts.
6293 if (isa<TruncInst>(CSrc)) {
6294 // Get the sizes of the types involved
6295 Value *A = CSrc->getOperand(0);
6296 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6297 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6298 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6299 // If we're actually extending zero bits and the trunc is a no-op
6300 if (MidSize < DstSize && SrcSize == DstSize) {
6301 // Replace both of the casts with an And of the type mask.
6302 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
6303 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6305 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6306 // Unfortunately, if the type changed, we need to cast it back.
6307 if (And->getType() != CI.getType()) {
6308 And->setName(CSrc->getName()+".mask");
6309 InsertNewInstBefore(And, CI);
6310 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6320 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6321 return commonIntCastTransforms(CI);
6324 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6325 return commonCastTransforms(CI);
6328 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6329 return commonCastTransforms(CI);
6332 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6333 return commonCastTransforms(CI);
6336 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6337 return commonCastTransforms(CI);
6340 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6341 return commonCastTransforms(CI);
6344 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6345 return commonCastTransforms(CI);
6348 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6349 return commonCastTransforms(CI);
6352 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6353 return commonCastTransforms(CI);
6356 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6358 // If the operands are integer typed then apply the integer transforms,
6359 // otherwise just apply the common ones.
6360 Value *Src = CI.getOperand(0);
6361 const Type *SrcTy = Src->getType();
6362 const Type *DestTy = CI.getType();
6364 if (SrcTy->isInteger() && DestTy->isInteger()) {
6365 if (Instruction *Result = commonIntCastTransforms(CI))
6368 if (Instruction *Result = commonCastTransforms(CI))
6373 // Get rid of casts from one type to the same type. These are useless and can
6374 // be replaced by the operand.
6375 if (DestTy == Src->getType())
6376 return ReplaceInstUsesWith(CI, Src);
6378 // If the source and destination are pointers, and this cast is equivalent to
6379 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6380 // This can enhance SROA and other transforms that want type-safe pointers.
6381 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6382 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6383 const Type *DstElTy = DstPTy->getElementType();
6384 const Type *SrcElTy = SrcPTy->getElementType();
6386 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6387 unsigned NumZeros = 0;
6388 while (SrcElTy != DstElTy &&
6389 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6390 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6391 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6395 // If we found a path from the src to dest, create the getelementptr now.
6396 if (SrcElTy == DstElTy) {
6397 std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
6398 return new GetElementPtrInst(Src, Idxs);
6403 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6404 if (SVI->hasOneUse()) {
6405 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6406 // a bitconvert to a vector with the same # elts.
6407 if (isa<PackedType>(DestTy) &&
6408 cast<PackedType>(DestTy)->getNumElements() ==
6409 SVI->getType()->getNumElements()) {
6411 // If either of the operands is a cast from CI.getType(), then
6412 // evaluating the shuffle in the casted destination's type will allow
6413 // us to eliminate at least one cast.
6414 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6415 Tmp->getOperand(0)->getType() == DestTy) ||
6416 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6417 Tmp->getOperand(0)->getType() == DestTy)) {
6418 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6419 SVI->getOperand(0), DestTy, &CI);
6420 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6421 SVI->getOperand(1), DestTy, &CI);
6422 // Return a new shuffle vector. Use the same element ID's, as we
6423 // know the vector types match #elts.
6424 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6432 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6434 /// %D = select %cond, %C, %A
6436 /// %C = select %cond, %B, 0
6439 /// Assuming that the specified instruction is an operand to the select, return
6440 /// a bitmask indicating which operands of this instruction are foldable if they
6441 /// equal the other incoming value of the select.
6443 static unsigned GetSelectFoldableOperands(Instruction *I) {
6444 switch (I->getOpcode()) {
6445 case Instruction::Add:
6446 case Instruction::Mul:
6447 case Instruction::And:
6448 case Instruction::Or:
6449 case Instruction::Xor:
6450 return 3; // Can fold through either operand.
6451 case Instruction::Sub: // Can only fold on the amount subtracted.
6452 case Instruction::Shl: // Can only fold on the shift amount.
6453 case Instruction::LShr:
6454 case Instruction::AShr:
6457 return 0; // Cannot fold
6461 /// GetSelectFoldableConstant - For the same transformation as the previous
6462 /// function, return the identity constant that goes into the select.
6463 static Constant *GetSelectFoldableConstant(Instruction *I) {
6464 switch (I->getOpcode()) {
6465 default: assert(0 && "This cannot happen!"); abort();
6466 case Instruction::Add:
6467 case Instruction::Sub:
6468 case Instruction::Or:
6469 case Instruction::Xor:
6470 return Constant::getNullValue(I->getType());
6471 case Instruction::Shl:
6472 case Instruction::LShr:
6473 case Instruction::AShr:
6474 return Constant::getNullValue(Type::Int8Ty);
6475 case Instruction::And:
6476 return ConstantInt::getAllOnesValue(I->getType());
6477 case Instruction::Mul:
6478 return ConstantInt::get(I->getType(), 1);
6482 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6483 /// have the same opcode and only one use each. Try to simplify this.
6484 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6486 if (TI->getNumOperands() == 1) {
6487 // If this is a non-volatile load or a cast from the same type,
6490 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6493 return 0; // unknown unary op.
6496 // Fold this by inserting a select from the input values.
6497 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6498 FI->getOperand(0), SI.getName()+".v");
6499 InsertNewInstBefore(NewSI, SI);
6500 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6504 // Only handle binary, compare and shift operators here.
6505 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
6508 // Figure out if the operations have any operands in common.
6509 Value *MatchOp, *OtherOpT, *OtherOpF;
6511 if (TI->getOperand(0) == FI->getOperand(0)) {
6512 MatchOp = TI->getOperand(0);
6513 OtherOpT = TI->getOperand(1);
6514 OtherOpF = FI->getOperand(1);
6515 MatchIsOpZero = true;
6516 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6517 MatchOp = TI->getOperand(1);
6518 OtherOpT = TI->getOperand(0);
6519 OtherOpF = FI->getOperand(0);
6520 MatchIsOpZero = false;
6521 } else if (!TI->isCommutative()) {
6523 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6524 MatchOp = TI->getOperand(0);
6525 OtherOpT = TI->getOperand(1);
6526 OtherOpF = FI->getOperand(0);
6527 MatchIsOpZero = true;
6528 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6529 MatchOp = TI->getOperand(1);
6530 OtherOpT = TI->getOperand(0);
6531 OtherOpF = FI->getOperand(1);
6532 MatchIsOpZero = true;
6537 // If we reach here, they do have operations in common.
6538 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6539 OtherOpF, SI.getName()+".v");
6540 InsertNewInstBefore(NewSI, SI);
6542 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6544 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6546 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6549 assert(isa<ShiftInst>(TI) && "Should only have Shift here");
6551 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
6553 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
6556 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6557 Value *CondVal = SI.getCondition();
6558 Value *TrueVal = SI.getTrueValue();
6559 Value *FalseVal = SI.getFalseValue();
6561 // select true, X, Y -> X
6562 // select false, X, Y -> Y
6563 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
6564 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
6566 // select C, X, X -> X
6567 if (TrueVal == FalseVal)
6568 return ReplaceInstUsesWith(SI, TrueVal);
6570 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6571 return ReplaceInstUsesWith(SI, FalseVal);
6572 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6573 return ReplaceInstUsesWith(SI, TrueVal);
6574 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6575 if (isa<Constant>(TrueVal))
6576 return ReplaceInstUsesWith(SI, TrueVal);
6578 return ReplaceInstUsesWith(SI, FalseVal);
6581 if (SI.getType() == Type::Int1Ty) {
6582 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
6583 if (C->getZExtValue()) {
6584 // Change: A = select B, true, C --> A = or B, C
6585 return BinaryOperator::createOr(CondVal, FalseVal);
6587 // Change: A = select B, false, C --> A = and !B, C
6589 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6590 "not."+CondVal->getName()), SI);
6591 return BinaryOperator::createAnd(NotCond, FalseVal);
6593 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
6594 if (C->getZExtValue() == false) {
6595 // Change: A = select B, C, false --> A = and B, C
6596 return BinaryOperator::createAnd(CondVal, TrueVal);
6598 // Change: A = select B, C, true --> A = or !B, C
6600 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6601 "not."+CondVal->getName()), SI);
6602 return BinaryOperator::createOr(NotCond, TrueVal);
6607 // Selecting between two integer constants?
6608 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6609 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6610 // select C, 1, 0 -> cast C to int
6611 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6612 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6613 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6614 // select C, 0, 1 -> cast !C to int
6616 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6617 "not."+CondVal->getName()), SI);
6618 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6621 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6623 // (x <s 0) ? -1 : 0 -> ashr x, 31
6624 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6625 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6626 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6627 bool CanXForm = false;
6628 if (IC->isSignedPredicate())
6629 CanXForm = CmpCst->isNullValue() &&
6630 IC->getPredicate() == ICmpInst::ICMP_SLT;
6632 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6633 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6634 IC->getPredicate() == ICmpInst::ICMP_UGT;
6638 // The comparison constant and the result are not neccessarily the
6639 // same width. Make an all-ones value by inserting a AShr.
6640 Value *X = IC->getOperand(0);
6641 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6642 Constant *ShAmt = ConstantInt::get(Type::Int8Ty, Bits-1);
6643 Instruction *SRA = new ShiftInst(Instruction::AShr, X,
6645 InsertNewInstBefore(SRA, SI);
6647 // Finally, convert to the type of the select RHS. We figure out
6648 // if this requires a SExt, Trunc or BitCast based on the sizes.
6649 Instruction::CastOps opc = Instruction::BitCast;
6650 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6651 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6652 if (SRASize < SISize)
6653 opc = Instruction::SExt;
6654 else if (SRASize > SISize)
6655 opc = Instruction::Trunc;
6656 return CastInst::create(opc, SRA, SI.getType());
6661 // If one of the constants is zero (we know they can't both be) and we
6662 // have a fcmp instruction with zero, and we have an 'and' with the
6663 // non-constant value, eliminate this whole mess. This corresponds to
6664 // cases like this: ((X & 27) ? 27 : 0)
6665 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6666 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6667 cast<Constant>(IC->getOperand(1))->isNullValue())
6668 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6669 if (ICA->getOpcode() == Instruction::And &&
6670 isa<ConstantInt>(ICA->getOperand(1)) &&
6671 (ICA->getOperand(1) == TrueValC ||
6672 ICA->getOperand(1) == FalseValC) &&
6673 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6674 // Okay, now we know that everything is set up, we just don't
6675 // know whether we have a icmp_ne or icmp_eq and whether the
6676 // true or false val is the zero.
6677 bool ShouldNotVal = !TrueValC->isNullValue();
6678 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6681 V = InsertNewInstBefore(BinaryOperator::create(
6682 Instruction::Xor, V, ICA->getOperand(1)), SI);
6683 return ReplaceInstUsesWith(SI, V);
6688 // See if we are selecting two values based on a comparison of the two values.
6689 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6690 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6691 // Transform (X == Y) ? X : Y -> Y
6692 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6693 return ReplaceInstUsesWith(SI, FalseVal);
6694 // Transform (X != Y) ? X : Y -> X
6695 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6696 return ReplaceInstUsesWith(SI, TrueVal);
6697 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6699 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6700 // Transform (X == Y) ? Y : X -> X
6701 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6702 return ReplaceInstUsesWith(SI, FalseVal);
6703 // Transform (X != Y) ? Y : X -> Y
6704 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6705 return ReplaceInstUsesWith(SI, TrueVal);
6706 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6710 // See if we are selecting two values based on a comparison of the two values.
6711 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6712 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6713 // Transform (X == Y) ? X : Y -> Y
6714 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6715 return ReplaceInstUsesWith(SI, FalseVal);
6716 // Transform (X != Y) ? X : Y -> X
6717 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6718 return ReplaceInstUsesWith(SI, TrueVal);
6719 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6721 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6722 // Transform (X == Y) ? Y : X -> X
6723 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6724 return ReplaceInstUsesWith(SI, FalseVal);
6725 // Transform (X != Y) ? Y : X -> Y
6726 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6727 return ReplaceInstUsesWith(SI, TrueVal);
6728 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6732 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6733 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6734 if (TI->hasOneUse() && FI->hasOneUse()) {
6735 Instruction *AddOp = 0, *SubOp = 0;
6737 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6738 if (TI->getOpcode() == FI->getOpcode())
6739 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6742 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6743 // even legal for FP.
6744 if (TI->getOpcode() == Instruction::Sub &&
6745 FI->getOpcode() == Instruction::Add) {
6746 AddOp = FI; SubOp = TI;
6747 } else if (FI->getOpcode() == Instruction::Sub &&
6748 TI->getOpcode() == Instruction::Add) {
6749 AddOp = TI; SubOp = FI;
6753 Value *OtherAddOp = 0;
6754 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6755 OtherAddOp = AddOp->getOperand(1);
6756 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6757 OtherAddOp = AddOp->getOperand(0);
6761 // So at this point we know we have (Y -> OtherAddOp):
6762 // select C, (add X, Y), (sub X, Z)
6763 Value *NegVal; // Compute -Z
6764 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6765 NegVal = ConstantExpr::getNeg(C);
6767 NegVal = InsertNewInstBefore(
6768 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6771 Value *NewTrueOp = OtherAddOp;
6772 Value *NewFalseOp = NegVal;
6774 std::swap(NewTrueOp, NewFalseOp);
6775 Instruction *NewSel =
6776 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6778 NewSel = InsertNewInstBefore(NewSel, SI);
6779 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6784 // See if we can fold the select into one of our operands.
6785 if (SI.getType()->isInteger()) {
6786 // See the comment above GetSelectFoldableOperands for a description of the
6787 // transformation we are doing here.
6788 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6789 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6790 !isa<Constant>(FalseVal))
6791 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6792 unsigned OpToFold = 0;
6793 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6795 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6800 Constant *C = GetSelectFoldableConstant(TVI);
6801 std::string Name = TVI->getName(); TVI->setName("");
6802 Instruction *NewSel =
6803 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
6805 InsertNewInstBefore(NewSel, SI);
6806 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6807 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6808 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
6809 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
6811 assert(0 && "Unknown instruction!!");
6816 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6817 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6818 !isa<Constant>(TrueVal))
6819 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6820 unsigned OpToFold = 0;
6821 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6823 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6828 Constant *C = GetSelectFoldableConstant(FVI);
6829 std::string Name = FVI->getName(); FVI->setName("");
6830 Instruction *NewSel =
6831 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
6833 InsertNewInstBefore(NewSel, SI);
6834 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6835 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6836 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
6837 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
6839 assert(0 && "Unknown instruction!!");
6845 if (BinaryOperator::isNot(CondVal)) {
6846 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6847 SI.setOperand(1, FalseVal);
6848 SI.setOperand(2, TrueVal);
6855 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6856 /// determine, return it, otherwise return 0.
6857 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6858 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6859 unsigned Align = GV->getAlignment();
6860 if (Align == 0 && TD)
6861 Align = TD->getTypeAlignment(GV->getType()->getElementType());
6863 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6864 unsigned Align = AI->getAlignment();
6865 if (Align == 0 && TD) {
6866 if (isa<AllocaInst>(AI))
6867 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6868 else if (isa<MallocInst>(AI)) {
6869 // Malloc returns maximally aligned memory.
6870 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6871 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::DoubleTy));
6872 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::Int64Ty));
6876 } else if (isa<BitCastInst>(V) ||
6877 (isa<ConstantExpr>(V) &&
6878 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
6879 User *CI = cast<User>(V);
6880 if (isa<PointerType>(CI->getOperand(0)->getType()))
6881 return GetKnownAlignment(CI->getOperand(0), TD);
6883 } else if (isa<GetElementPtrInst>(V) ||
6884 (isa<ConstantExpr>(V) &&
6885 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
6886 User *GEPI = cast<User>(V);
6887 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
6888 if (BaseAlignment == 0) return 0;
6890 // If all indexes are zero, it is just the alignment of the base pointer.
6891 bool AllZeroOperands = true;
6892 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
6893 if (!isa<Constant>(GEPI->getOperand(i)) ||
6894 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
6895 AllZeroOperands = false;
6898 if (AllZeroOperands)
6899 return BaseAlignment;
6901 // Otherwise, if the base alignment is >= the alignment we expect for the
6902 // base pointer type, then we know that the resultant pointer is aligned at
6903 // least as much as its type requires.
6906 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
6907 if (TD->getTypeAlignment(cast<PointerType>(BasePtrTy)->getElementType())
6909 const Type *GEPTy = GEPI->getType();
6910 return TD->getTypeAlignment(cast<PointerType>(GEPTy)->getElementType());
6918 /// visitCallInst - CallInst simplification. This mostly only handles folding
6919 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
6920 /// the heavy lifting.
6922 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
6923 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
6924 if (!II) return visitCallSite(&CI);
6926 // Intrinsics cannot occur in an invoke, so handle them here instead of in
6928 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
6929 bool Changed = false;
6931 // memmove/cpy/set of zero bytes is a noop.
6932 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
6933 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
6935 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
6936 if (CI->getZExtValue() == 1) {
6937 // Replace the instruction with just byte operations. We would
6938 // transform other cases to loads/stores, but we don't know if
6939 // alignment is sufficient.
6943 // If we have a memmove and the source operation is a constant global,
6944 // then the source and dest pointers can't alias, so we can change this
6945 // into a call to memcpy.
6946 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
6947 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
6948 if (GVSrc->isConstant()) {
6949 Module *M = CI.getParent()->getParent()->getParent();
6951 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
6953 Name = "llvm.memcpy.i32";
6955 Name = "llvm.memcpy.i64";
6956 Constant *MemCpy = M->getOrInsertFunction(Name,
6957 CI.getCalledFunction()->getFunctionType());
6958 CI.setOperand(0, MemCpy);
6963 // If we can determine a pointer alignment that is bigger than currently
6964 // set, update the alignment.
6965 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
6966 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
6967 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
6968 unsigned Align = std::min(Alignment1, Alignment2);
6969 if (MI->getAlignment()->getZExtValue() < Align) {
6970 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
6973 } else if (isa<MemSetInst>(MI)) {
6974 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
6975 if (MI->getAlignment()->getZExtValue() < Alignment) {
6976 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
6981 if (Changed) return II;
6983 switch (II->getIntrinsicID()) {
6985 case Intrinsic::ppc_altivec_lvx:
6986 case Intrinsic::ppc_altivec_lvxl:
6987 case Intrinsic::x86_sse_loadu_ps:
6988 case Intrinsic::x86_sse2_loadu_pd:
6989 case Intrinsic::x86_sse2_loadu_dq:
6990 // Turn PPC lvx -> load if the pointer is known aligned.
6991 // Turn X86 loadups -> load if the pointer is known aligned.
6992 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6993 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
6994 PointerType::get(II->getType()), CI);
6995 return new LoadInst(Ptr);
6998 case Intrinsic::ppc_altivec_stvx:
6999 case Intrinsic::ppc_altivec_stvxl:
7000 // Turn stvx -> store if the pointer is known aligned.
7001 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7002 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7003 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7005 return new StoreInst(II->getOperand(1), Ptr);
7008 case Intrinsic::x86_sse_storeu_ps:
7009 case Intrinsic::x86_sse2_storeu_pd:
7010 case Intrinsic::x86_sse2_storeu_dq:
7011 case Intrinsic::x86_sse2_storel_dq:
7012 // Turn X86 storeu -> store if the pointer is known aligned.
7013 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7014 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7015 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7017 return new StoreInst(II->getOperand(2), Ptr);
7021 case Intrinsic::x86_sse_cvttss2si: {
7022 // These intrinsics only demands the 0th element of its input vector. If
7023 // we can simplify the input based on that, do so now.
7025 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7027 II->setOperand(1, V);
7033 case Intrinsic::ppc_altivec_vperm:
7034 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7035 if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
7036 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7038 // Check that all of the elements are integer constants or undefs.
7039 bool AllEltsOk = true;
7040 for (unsigned i = 0; i != 16; ++i) {
7041 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7042 !isa<UndefValue>(Mask->getOperand(i))) {
7049 // Cast the input vectors to byte vectors.
7050 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7051 II->getOperand(1), Mask->getType(), CI);
7052 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7053 II->getOperand(2), Mask->getType(), CI);
7054 Value *Result = UndefValue::get(Op0->getType());
7056 // Only extract each element once.
7057 Value *ExtractedElts[32];
7058 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7060 for (unsigned i = 0; i != 16; ++i) {
7061 if (isa<UndefValue>(Mask->getOperand(i)))
7063 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7064 Idx &= 31; // Match the hardware behavior.
7066 if (ExtractedElts[Idx] == 0) {
7068 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7069 InsertNewInstBefore(Elt, CI);
7070 ExtractedElts[Idx] = Elt;
7073 // Insert this value into the result vector.
7074 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7075 InsertNewInstBefore(cast<Instruction>(Result), CI);
7077 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7082 case Intrinsic::stackrestore: {
7083 // If the save is right next to the restore, remove the restore. This can
7084 // happen when variable allocas are DCE'd.
7085 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7086 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7087 BasicBlock::iterator BI = SS;
7089 return EraseInstFromFunction(CI);
7093 // If the stack restore is in a return/unwind block and if there are no
7094 // allocas or calls between the restore and the return, nuke the restore.
7095 TerminatorInst *TI = II->getParent()->getTerminator();
7096 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7097 BasicBlock::iterator BI = II;
7098 bool CannotRemove = false;
7099 for (++BI; &*BI != TI; ++BI) {
7100 if (isa<AllocaInst>(BI) ||
7101 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7102 CannotRemove = true;
7107 return EraseInstFromFunction(CI);
7114 return visitCallSite(II);
7117 // InvokeInst simplification
7119 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7120 return visitCallSite(&II);
7123 // visitCallSite - Improvements for call and invoke instructions.
7125 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7126 bool Changed = false;
7128 // If the callee is a constexpr cast of a function, attempt to move the cast
7129 // to the arguments of the call/invoke.
7130 if (transformConstExprCastCall(CS)) return 0;
7132 Value *Callee = CS.getCalledValue();
7134 if (Function *CalleeF = dyn_cast<Function>(Callee))
7135 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7136 Instruction *OldCall = CS.getInstruction();
7137 // If the call and callee calling conventions don't match, this call must
7138 // be unreachable, as the call is undefined.
7139 new StoreInst(ConstantInt::getTrue(),
7140 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7141 if (!OldCall->use_empty())
7142 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7143 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7144 return EraseInstFromFunction(*OldCall);
7148 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7149 // This instruction is not reachable, just remove it. We insert a store to
7150 // undef so that we know that this code is not reachable, despite the fact
7151 // that we can't modify the CFG here.
7152 new StoreInst(ConstantInt::getTrue(),
7153 UndefValue::get(PointerType::get(Type::Int1Ty)),
7154 CS.getInstruction());
7156 if (!CS.getInstruction()->use_empty())
7157 CS.getInstruction()->
7158 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7160 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7161 // Don't break the CFG, insert a dummy cond branch.
7162 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7163 ConstantInt::getTrue(), II);
7165 return EraseInstFromFunction(*CS.getInstruction());
7168 const PointerType *PTy = cast<PointerType>(Callee->getType());
7169 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7170 if (FTy->isVarArg()) {
7171 // See if we can optimize any arguments passed through the varargs area of
7173 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7174 E = CS.arg_end(); I != E; ++I)
7175 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7176 // If this cast does not effect the value passed through the varargs
7177 // area, we can eliminate the use of the cast.
7178 Value *Op = CI->getOperand(0);
7179 if (CI->isLosslessCast()) {
7186 return Changed ? CS.getInstruction() : 0;
7189 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7190 // attempt to move the cast to the arguments of the call/invoke.
7192 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7193 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7194 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7195 if (CE->getOpcode() != Instruction::BitCast ||
7196 !isa<Function>(CE->getOperand(0)))
7198 Function *Callee = cast<Function>(CE->getOperand(0));
7199 Instruction *Caller = CS.getInstruction();
7201 // Okay, this is a cast from a function to a different type. Unless doing so
7202 // would cause a type conversion of one of our arguments, change this call to
7203 // be a direct call with arguments casted to the appropriate types.
7205 const FunctionType *FT = Callee->getFunctionType();
7206 const Type *OldRetTy = Caller->getType();
7208 // Check to see if we are changing the return type...
7209 if (OldRetTy != FT->getReturnType()) {
7210 if (Callee->isExternal() && !Caller->use_empty() &&
7211 OldRetTy != FT->getReturnType() &&
7212 // Conversion is ok if changing from pointer to int of same size.
7213 !(isa<PointerType>(FT->getReturnType()) &&
7214 TD->getIntPtrType() == OldRetTy))
7215 return false; // Cannot transform this return value.
7217 // If the callsite is an invoke instruction, and the return value is used by
7218 // a PHI node in a successor, we cannot change the return type of the call
7219 // because there is no place to put the cast instruction (without breaking
7220 // the critical edge). Bail out in this case.
7221 if (!Caller->use_empty())
7222 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7223 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7225 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7226 if (PN->getParent() == II->getNormalDest() ||
7227 PN->getParent() == II->getUnwindDest())
7231 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7232 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7234 CallSite::arg_iterator AI = CS.arg_begin();
7235 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7236 const Type *ParamTy = FT->getParamType(i);
7237 const Type *ActTy = (*AI)->getType();
7238 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7239 //Either we can cast directly, or we can upconvert the argument
7240 bool isConvertible = ActTy == ParamTy ||
7241 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7242 (ParamTy->isIntegral() && ActTy->isIntegral() &&
7243 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7244 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7245 && c->getSExtValue() > 0);
7246 if (Callee->isExternal() && !isConvertible) return false;
7249 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7250 Callee->isExternal())
7251 return false; // Do not delete arguments unless we have a function body...
7253 // Okay, we decided that this is a safe thing to do: go ahead and start
7254 // inserting cast instructions as necessary...
7255 std::vector<Value*> Args;
7256 Args.reserve(NumActualArgs);
7258 AI = CS.arg_begin();
7259 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7260 const Type *ParamTy = FT->getParamType(i);
7261 if ((*AI)->getType() == ParamTy) {
7262 Args.push_back(*AI);
7264 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7265 false, ParamTy, false);
7266 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7267 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7271 // If the function takes more arguments than the call was taking, add them
7273 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7274 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7276 // If we are removing arguments to the function, emit an obnoxious warning...
7277 if (FT->getNumParams() < NumActualArgs)
7278 if (!FT->isVarArg()) {
7279 cerr << "WARNING: While resolving call to function '"
7280 << Callee->getName() << "' arguments were dropped!\n";
7282 // Add all of the arguments in their promoted form to the arg list...
7283 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7284 const Type *PTy = getPromotedType((*AI)->getType());
7285 if (PTy != (*AI)->getType()) {
7286 // Must promote to pass through va_arg area!
7287 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7289 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7290 InsertNewInstBefore(Cast, *Caller);
7291 Args.push_back(Cast);
7293 Args.push_back(*AI);
7298 if (FT->getReturnType() == Type::VoidTy)
7299 Caller->setName(""); // Void type should not have a name...
7302 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7303 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7304 Args, Caller->getName(), Caller);
7305 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7307 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
7308 if (cast<CallInst>(Caller)->isTailCall())
7309 cast<CallInst>(NC)->setTailCall();
7310 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7313 // Insert a cast of the return type as necessary...
7315 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7316 if (NV->getType() != Type::VoidTy) {
7317 const Type *CallerTy = Caller->getType();
7318 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7320 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7322 // If this is an invoke instruction, we should insert it after the first
7323 // non-phi, instruction in the normal successor block.
7324 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7325 BasicBlock::iterator I = II->getNormalDest()->begin();
7326 while (isa<PHINode>(I)) ++I;
7327 InsertNewInstBefore(NC, *I);
7329 // Otherwise, it's a call, just insert cast right after the call instr
7330 InsertNewInstBefore(NC, *Caller);
7332 AddUsersToWorkList(*Caller);
7334 NV = UndefValue::get(Caller->getType());
7338 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7339 Caller->replaceAllUsesWith(NV);
7340 Caller->getParent()->getInstList().erase(Caller);
7341 removeFromWorkList(Caller);
7345 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7346 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7347 /// and a single binop.
7348 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7349 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7350 assert(isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7351 isa<GetElementPtrInst>(FirstInst) || isa<CmpInst>(FirstInst));
7352 unsigned Opc = FirstInst->getOpcode();
7353 Value *LHSVal = FirstInst->getOperand(0);
7354 Value *RHSVal = FirstInst->getOperand(1);
7356 const Type *LHSType = LHSVal->getType();
7357 const Type *RHSType = RHSVal->getType();
7359 // Scan to see if all operands are the same opcode, all have one use, and all
7360 // kill their operands (i.e. the operands have one use).
7361 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7362 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7363 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7364 // Verify type of the LHS matches so we don't fold cmp's of different
7365 // types or GEP's with different index types.
7366 I->getOperand(0)->getType() != LHSType ||
7367 I->getOperand(1)->getType() != RHSType)
7370 // If they are CmpInst instructions, check their predicates
7371 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7372 if (cast<CmpInst>(I)->getPredicate() !=
7373 cast<CmpInst>(FirstInst)->getPredicate())
7376 // Keep track of which operand needs a phi node.
7377 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7378 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7381 // Otherwise, this is safe to transform, determine if it is profitable.
7383 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7384 // Indexes are often folded into load/store instructions, so we don't want to
7385 // hide them behind a phi.
7386 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7389 Value *InLHS = FirstInst->getOperand(0);
7390 Value *InRHS = FirstInst->getOperand(1);
7391 PHINode *NewLHS = 0, *NewRHS = 0;
7393 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7394 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7395 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7396 InsertNewInstBefore(NewLHS, PN);
7401 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7402 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7403 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7404 InsertNewInstBefore(NewRHS, PN);
7408 // Add all operands to the new PHIs.
7409 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7411 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7412 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7415 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7416 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7420 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7421 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7422 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7423 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7425 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FirstInst))
7426 return new ShiftInst(SI->getOpcode(), LHSVal, RHSVal);
7428 assert(isa<GetElementPtrInst>(FirstInst));
7429 return new GetElementPtrInst(LHSVal, RHSVal);
7433 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7434 /// of the block that defines it. This means that it must be obvious the value
7435 /// of the load is not changed from the point of the load to the end of the
7437 static bool isSafeToSinkLoad(LoadInst *L) {
7438 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7440 for (++BBI; BBI != E; ++BBI)
7441 if (BBI->mayWriteToMemory())
7447 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7448 // operator and they all are only used by the PHI, PHI together their
7449 // inputs, and do the operation once, to the result of the PHI.
7450 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7451 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7453 // Scan the instruction, looking for input operations that can be folded away.
7454 // If all input operands to the phi are the same instruction (e.g. a cast from
7455 // the same type or "+42") we can pull the operation through the PHI, reducing
7456 // code size and simplifying code.
7457 Constant *ConstantOp = 0;
7458 const Type *CastSrcTy = 0;
7459 bool isVolatile = false;
7460 if (isa<CastInst>(FirstInst)) {
7461 CastSrcTy = FirstInst->getOperand(0)->getType();
7462 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7463 isa<CmpInst>(FirstInst)) {
7464 // Can fold binop, compare or shift here if the RHS is a constant,
7465 // otherwise call FoldPHIArgBinOpIntoPHI.
7466 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7467 if (ConstantOp == 0)
7468 return FoldPHIArgBinOpIntoPHI(PN);
7469 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7470 isVolatile = LI->isVolatile();
7471 // We can't sink the load if the loaded value could be modified between the
7472 // load and the PHI.
7473 if (LI->getParent() != PN.getIncomingBlock(0) ||
7474 !isSafeToSinkLoad(LI))
7476 } else if (isa<GetElementPtrInst>(FirstInst)) {
7477 if (FirstInst->getNumOperands() == 2)
7478 return FoldPHIArgBinOpIntoPHI(PN);
7479 // Can't handle general GEPs yet.
7482 return 0; // Cannot fold this operation.
7485 // Check to see if all arguments are the same operation.
7486 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7487 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7488 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7489 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7492 if (I->getOperand(0)->getType() != CastSrcTy)
7493 return 0; // Cast operation must match.
7494 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7495 // We can't sink the load if the loaded value could be modified between
7496 // the load and the PHI.
7497 if (LI->isVolatile() != isVolatile ||
7498 LI->getParent() != PN.getIncomingBlock(i) ||
7499 !isSafeToSinkLoad(LI))
7501 } else if (I->getOperand(1) != ConstantOp) {
7506 // Okay, they are all the same operation. Create a new PHI node of the
7507 // correct type, and PHI together all of the LHS's of the instructions.
7508 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7509 PN.getName()+".in");
7510 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7512 Value *InVal = FirstInst->getOperand(0);
7513 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7515 // Add all operands to the new PHI.
7516 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7517 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7518 if (NewInVal != InVal)
7520 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7525 // The new PHI unions all of the same values together. This is really
7526 // common, so we handle it intelligently here for compile-time speed.
7530 InsertNewInstBefore(NewPN, PN);
7534 // Insert and return the new operation.
7535 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7536 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7537 else if (isa<LoadInst>(FirstInst))
7538 return new LoadInst(PhiVal, "", isVolatile);
7539 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7540 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7541 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7542 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7543 PhiVal, ConstantOp);
7545 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
7546 PhiVal, ConstantOp);
7549 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7551 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7552 if (PN->use_empty()) return true;
7553 if (!PN->hasOneUse()) return false;
7555 // Remember this node, and if we find the cycle, return.
7556 if (!PotentiallyDeadPHIs.insert(PN).second)
7559 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7560 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7565 // PHINode simplification
7567 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7568 // If LCSSA is around, don't mess with Phi nodes
7569 if (mustPreserveAnalysisID(LCSSAID)) return 0;
7571 if (Value *V = PN.hasConstantValue())
7572 return ReplaceInstUsesWith(PN, V);
7574 // If all PHI operands are the same operation, pull them through the PHI,
7575 // reducing code size.
7576 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7577 PN.getIncomingValue(0)->hasOneUse())
7578 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7581 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7582 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7583 // PHI)... break the cycle.
7585 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
7586 std::set<PHINode*> PotentiallyDeadPHIs;
7587 PotentiallyDeadPHIs.insert(&PN);
7588 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7589 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7595 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7596 Instruction *InsertPoint,
7598 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
7599 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
7600 // We must cast correctly to the pointer type. Ensure that we
7601 // sign extend the integer value if it is smaller as this is
7602 // used for address computation.
7603 Instruction::CastOps opcode =
7604 (VTySize < PtrSize ? Instruction::SExt :
7605 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7606 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7610 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7611 Value *PtrOp = GEP.getOperand(0);
7612 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7613 // If so, eliminate the noop.
7614 if (GEP.getNumOperands() == 1)
7615 return ReplaceInstUsesWith(GEP, PtrOp);
7617 if (isa<UndefValue>(GEP.getOperand(0)))
7618 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7620 bool HasZeroPointerIndex = false;
7621 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7622 HasZeroPointerIndex = C->isNullValue();
7624 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7625 return ReplaceInstUsesWith(GEP, PtrOp);
7627 // Eliminate unneeded casts for indices.
7628 bool MadeChange = false;
7629 gep_type_iterator GTI = gep_type_begin(GEP);
7630 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7631 if (isa<SequentialType>(*GTI)) {
7632 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7633 Value *Src = CI->getOperand(0);
7634 const Type *SrcTy = Src->getType();
7635 const Type *DestTy = CI->getType();
7636 if (Src->getType()->isInteger()) {
7637 if (SrcTy->getPrimitiveSizeInBits() ==
7638 DestTy->getPrimitiveSizeInBits()) {
7639 // We can always eliminate a cast from ulong or long to the other.
7640 // We can always eliminate a cast from uint to int or the other on
7641 // 32-bit pointer platforms.
7642 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
7644 GEP.setOperand(i, Src);
7646 } else if (SrcTy->getPrimitiveSizeInBits() <
7647 DestTy->getPrimitiveSizeInBits() &&
7648 SrcTy->getPrimitiveSizeInBits() == 32) {
7649 // We can eliminate a cast from [u]int to [u]long iff the target
7650 // is a 32-bit pointer target.
7651 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7653 GEP.setOperand(i, Src);
7658 // If we are using a wider index than needed for this platform, shrink it
7659 // to what we need. If the incoming value needs a cast instruction,
7660 // insert it. This explicit cast can make subsequent optimizations more
7662 Value *Op = GEP.getOperand(i);
7663 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
7664 if (Constant *C = dyn_cast<Constant>(Op)) {
7665 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7668 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7670 GEP.setOperand(i, Op);
7674 if (MadeChange) return &GEP;
7676 // Combine Indices - If the source pointer to this getelementptr instruction
7677 // is a getelementptr instruction, combine the indices of the two
7678 // getelementptr instructions into a single instruction.
7680 std::vector<Value*> SrcGEPOperands;
7681 if (User *Src = dyn_castGetElementPtr(PtrOp))
7682 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
7684 if (!SrcGEPOperands.empty()) {
7685 // Note that if our source is a gep chain itself that we wait for that
7686 // chain to be resolved before we perform this transformation. This
7687 // avoids us creating a TON of code in some cases.
7689 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7690 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7691 return 0; // Wait until our source is folded to completion.
7693 std::vector<Value *> Indices;
7695 // Find out whether the last index in the source GEP is a sequential idx.
7696 bool EndsWithSequential = false;
7697 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7698 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7699 EndsWithSequential = !isa<StructType>(*I);
7701 // Can we combine the two pointer arithmetics offsets?
7702 if (EndsWithSequential) {
7703 // Replace: gep (gep %P, long B), long A, ...
7704 // With: T = long A+B; gep %P, T, ...
7706 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7707 if (SO1 == Constant::getNullValue(SO1->getType())) {
7709 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7712 // If they aren't the same type, convert both to an integer of the
7713 // target's pointer size.
7714 if (SO1->getType() != GO1->getType()) {
7715 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7716 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7717 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7718 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7720 unsigned PS = TD->getPointerSize();
7721 if (TD->getTypeSize(SO1->getType()) == PS) {
7722 // Convert GO1 to SO1's type.
7723 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
7725 } else if (TD->getTypeSize(GO1->getType()) == PS) {
7726 // Convert SO1 to GO1's type.
7727 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
7729 const Type *PT = TD->getIntPtrType();
7730 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
7731 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
7735 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7736 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7738 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7739 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7743 // Recycle the GEP we already have if possible.
7744 if (SrcGEPOperands.size() == 2) {
7745 GEP.setOperand(0, SrcGEPOperands[0]);
7746 GEP.setOperand(1, Sum);
7749 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7750 SrcGEPOperands.end()-1);
7751 Indices.push_back(Sum);
7752 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7754 } else if (isa<Constant>(*GEP.idx_begin()) &&
7755 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7756 SrcGEPOperands.size() != 1) {
7757 // Otherwise we can do the fold if the first index of the GEP is a zero
7758 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7759 SrcGEPOperands.end());
7760 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7763 if (!Indices.empty())
7764 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
7766 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7767 // GEP of global variable. If all of the indices for this GEP are
7768 // constants, we can promote this to a constexpr instead of an instruction.
7770 // Scan for nonconstants...
7771 std::vector<Constant*> Indices;
7772 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7773 for (; I != E && isa<Constant>(*I); ++I)
7774 Indices.push_back(cast<Constant>(*I));
7776 if (I == E) { // If they are all constants...
7777 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
7779 // Replace all uses of the GEP with the new constexpr...
7780 return ReplaceInstUsesWith(GEP, CE);
7782 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7783 if (!isa<PointerType>(X->getType())) {
7784 // Not interesting. Source pointer must be a cast from pointer.
7785 } else if (HasZeroPointerIndex) {
7786 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7787 // into : GEP [10 x ubyte]* X, long 0, ...
7789 // This occurs when the program declares an array extern like "int X[];"
7791 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7792 const PointerType *XTy = cast<PointerType>(X->getType());
7793 if (const ArrayType *XATy =
7794 dyn_cast<ArrayType>(XTy->getElementType()))
7795 if (const ArrayType *CATy =
7796 dyn_cast<ArrayType>(CPTy->getElementType()))
7797 if (CATy->getElementType() == XATy->getElementType()) {
7798 // At this point, we know that the cast source type is a pointer
7799 // to an array of the same type as the destination pointer
7800 // array. Because the array type is never stepped over (there
7801 // is a leading zero) we can fold the cast into this GEP.
7802 GEP.setOperand(0, X);
7805 } else if (GEP.getNumOperands() == 2) {
7806 // Transform things like:
7807 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7808 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7809 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7810 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7811 if (isa<ArrayType>(SrcElTy) &&
7812 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7813 TD->getTypeSize(ResElTy)) {
7814 Value *V = InsertNewInstBefore(
7815 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7816 GEP.getOperand(1), GEP.getName()), GEP);
7817 // V and GEP are both pointer types --> BitCast
7818 return new BitCastInst(V, GEP.getType());
7821 // Transform things like:
7822 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7823 // (where tmp = 8*tmp2) into:
7824 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7826 if (isa<ArrayType>(SrcElTy) &&
7827 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
7828 uint64_t ArrayEltSize =
7829 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7831 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7832 // allow either a mul, shift, or constant here.
7834 ConstantInt *Scale = 0;
7835 if (ArrayEltSize == 1) {
7836 NewIdx = GEP.getOperand(1);
7837 Scale = ConstantInt::get(NewIdx->getType(), 1);
7838 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7839 NewIdx = ConstantInt::get(CI->getType(), 1);
7841 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7842 if (Inst->getOpcode() == Instruction::Shl &&
7843 isa<ConstantInt>(Inst->getOperand(1))) {
7845 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7846 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7847 NewIdx = Inst->getOperand(0);
7848 } else if (Inst->getOpcode() == Instruction::Mul &&
7849 isa<ConstantInt>(Inst->getOperand(1))) {
7850 Scale = cast<ConstantInt>(Inst->getOperand(1));
7851 NewIdx = Inst->getOperand(0);
7855 // If the index will be to exactly the right offset with the scale taken
7856 // out, perform the transformation.
7857 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7858 if (isa<ConstantInt>(Scale))
7859 Scale = ConstantInt::get(Scale->getType(),
7860 Scale->getZExtValue() / ArrayEltSize);
7861 if (Scale->getZExtValue() != 1) {
7862 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
7864 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7865 NewIdx = InsertNewInstBefore(Sc, GEP);
7868 // Insert the new GEP instruction.
7869 Instruction *NewGEP =
7870 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
7871 NewIdx, GEP.getName());
7872 NewGEP = InsertNewInstBefore(NewGEP, GEP);
7873 // The NewGEP must be pointer typed, so must the old one -> BitCast
7874 return new BitCastInst(NewGEP, GEP.getType());
7883 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7884 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7885 if (AI.isArrayAllocation()) // Check C != 1
7886 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7888 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7889 AllocationInst *New = 0;
7891 // Create and insert the replacement instruction...
7892 if (isa<MallocInst>(AI))
7893 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7895 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7896 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7899 InsertNewInstBefore(New, AI);
7901 // Scan to the end of the allocation instructions, to skip over a block of
7902 // allocas if possible...
7904 BasicBlock::iterator It = New;
7905 while (isa<AllocationInst>(*It)) ++It;
7907 // Now that I is pointing to the first non-allocation-inst in the block,
7908 // insert our getelementptr instruction...
7910 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
7911 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
7912 New->getName()+".sub", It);
7914 // Now make everything use the getelementptr instead of the original
7916 return ReplaceInstUsesWith(AI, V);
7917 } else if (isa<UndefValue>(AI.getArraySize())) {
7918 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7921 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7922 // Note that we only do this for alloca's, because malloc should allocate and
7923 // return a unique pointer, even for a zero byte allocation.
7924 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
7925 TD->getTypeSize(AI.getAllocatedType()) == 0)
7926 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7931 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
7932 Value *Op = FI.getOperand(0);
7934 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
7935 if (CastInst *CI = dyn_cast<CastInst>(Op))
7936 if (isa<PointerType>(CI->getOperand(0)->getType())) {
7937 FI.setOperand(0, CI->getOperand(0));
7941 // free undef -> unreachable.
7942 if (isa<UndefValue>(Op)) {
7943 // Insert a new store to null because we cannot modify the CFG here.
7944 new StoreInst(ConstantInt::getTrue(),
7945 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
7946 return EraseInstFromFunction(FI);
7949 // If we have 'free null' delete the instruction. This can happen in stl code
7950 // when lots of inlining happens.
7951 if (isa<ConstantPointerNull>(Op))
7952 return EraseInstFromFunction(FI);
7958 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
7959 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
7960 User *CI = cast<User>(LI.getOperand(0));
7961 Value *CastOp = CI->getOperand(0);
7963 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7964 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7965 const Type *SrcPTy = SrcTy->getElementType();
7967 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
7968 isa<PackedType>(DestPTy)) {
7969 // If the source is an array, the code below will not succeed. Check to
7970 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7972 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7973 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7974 if (ASrcTy->getNumElements() != 0) {
7975 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::Int32Ty));
7976 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7977 SrcTy = cast<PointerType>(CastOp->getType());
7978 SrcPTy = SrcTy->getElementType();
7981 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
7982 isa<PackedType>(SrcPTy)) &&
7983 // Do not allow turning this into a load of an integer, which is then
7984 // casted to a pointer, this pessimizes pointer analysis a lot.
7985 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
7986 IC.getTargetData().getTypeSize(SrcPTy) ==
7987 IC.getTargetData().getTypeSize(DestPTy)) {
7989 // Okay, we are casting from one integer or pointer type to another of
7990 // the same size. Instead of casting the pointer before the load, cast
7991 // the result of the loaded value.
7992 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
7994 LI.isVolatile()),LI);
7995 // Now cast the result of the load.
7996 return new BitCastInst(NewLoad, LI.getType());
8003 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8004 /// from this value cannot trap. If it is not obviously safe to load from the
8005 /// specified pointer, we do a quick local scan of the basic block containing
8006 /// ScanFrom, to determine if the address is already accessed.
8007 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8008 // If it is an alloca or global variable, it is always safe to load from.
8009 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8011 // Otherwise, be a little bit agressive by scanning the local block where we
8012 // want to check to see if the pointer is already being loaded or stored
8013 // from/to. If so, the previous load or store would have already trapped,
8014 // so there is no harm doing an extra load (also, CSE will later eliminate
8015 // the load entirely).
8016 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8021 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8022 if (LI->getOperand(0) == V) return true;
8023 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8024 if (SI->getOperand(1) == V) return true;
8030 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8031 Value *Op = LI.getOperand(0);
8033 // load (cast X) --> cast (load X) iff safe
8034 if (isa<CastInst>(Op))
8035 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8038 // None of the following transforms are legal for volatile loads.
8039 if (LI.isVolatile()) return 0;
8041 if (&LI.getParent()->front() != &LI) {
8042 BasicBlock::iterator BBI = &LI; --BBI;
8043 // If the instruction immediately before this is a store to the same
8044 // address, do a simple form of store->load forwarding.
8045 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8046 if (SI->getOperand(1) == LI.getOperand(0))
8047 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8048 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8049 if (LIB->getOperand(0) == LI.getOperand(0))
8050 return ReplaceInstUsesWith(LI, LIB);
8053 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8054 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8055 isa<UndefValue>(GEPI->getOperand(0))) {
8056 // Insert a new store to null instruction before the load to indicate
8057 // that this code is not reachable. We do this instead of inserting
8058 // an unreachable instruction directly because we cannot modify the
8060 new StoreInst(UndefValue::get(LI.getType()),
8061 Constant::getNullValue(Op->getType()), &LI);
8062 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8065 if (Constant *C = dyn_cast<Constant>(Op)) {
8066 // load null/undef -> undef
8067 if ((C->isNullValue() || isa<UndefValue>(C))) {
8068 // Insert a new store to null instruction before the load to indicate that
8069 // this code is not reachable. We do this instead of inserting an
8070 // unreachable instruction directly because we cannot modify the CFG.
8071 new StoreInst(UndefValue::get(LI.getType()),
8072 Constant::getNullValue(Op->getType()), &LI);
8073 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8076 // Instcombine load (constant global) into the value loaded.
8077 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8078 if (GV->isConstant() && !GV->isExternal())
8079 return ReplaceInstUsesWith(LI, GV->getInitializer());
8081 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8082 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8083 if (CE->getOpcode() == Instruction::GetElementPtr) {
8084 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8085 if (GV->isConstant() && !GV->isExternal())
8087 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8088 return ReplaceInstUsesWith(LI, V);
8089 if (CE->getOperand(0)->isNullValue()) {
8090 // Insert a new store to null instruction before the load to indicate
8091 // that this code is not reachable. We do this instead of inserting
8092 // an unreachable instruction directly because we cannot modify the
8094 new StoreInst(UndefValue::get(LI.getType()),
8095 Constant::getNullValue(Op->getType()), &LI);
8096 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8099 } else if (CE->isCast()) {
8100 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8105 if (Op->hasOneUse()) {
8106 // Change select and PHI nodes to select values instead of addresses: this
8107 // helps alias analysis out a lot, allows many others simplifications, and
8108 // exposes redundancy in the code.
8110 // Note that we cannot do the transformation unless we know that the
8111 // introduced loads cannot trap! Something like this is valid as long as
8112 // the condition is always false: load (select bool %C, int* null, int* %G),
8113 // but it would not be valid if we transformed it to load from null
8116 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8117 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8118 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8119 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8120 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8121 SI->getOperand(1)->getName()+".val"), LI);
8122 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8123 SI->getOperand(2)->getName()+".val"), LI);
8124 return new SelectInst(SI->getCondition(), V1, V2);
8127 // load (select (cond, null, P)) -> load P
8128 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8129 if (C->isNullValue()) {
8130 LI.setOperand(0, SI->getOperand(2));
8134 // load (select (cond, P, null)) -> load P
8135 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8136 if (C->isNullValue()) {
8137 LI.setOperand(0, SI->getOperand(1));
8145 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
8147 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8148 User *CI = cast<User>(SI.getOperand(1));
8149 Value *CastOp = CI->getOperand(0);
8151 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8152 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8153 const Type *SrcPTy = SrcTy->getElementType();
8155 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8156 // If the source is an array, the code below will not succeed. Check to
8157 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8159 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8160 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8161 if (ASrcTy->getNumElements() != 0) {
8162 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::Int32Ty));
8163 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
8164 SrcTy = cast<PointerType>(CastOp->getType());
8165 SrcPTy = SrcTy->getElementType();
8168 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8169 IC.getTargetData().getTypeSize(SrcPTy) ==
8170 IC.getTargetData().getTypeSize(DestPTy)) {
8172 // Okay, we are casting from one integer or pointer type to another of
8173 // the same size. Instead of casting the pointer before the store, cast
8174 // the value to be stored.
8176 Instruction::CastOps opcode = Instruction::BitCast;
8177 Value *SIOp0 = SI.getOperand(0);
8178 if (isa<PointerType>(SrcPTy)) {
8179 if (SIOp0->getType()->isIntegral())
8180 opcode = Instruction::IntToPtr;
8181 } else if (SrcPTy->isIntegral()) {
8182 if (isa<PointerType>(SIOp0->getType()))
8183 opcode = Instruction::PtrToInt;
8185 if (Constant *C = dyn_cast<Constant>(SIOp0))
8186 NewCast = ConstantExpr::getCast(opcode, C, SrcPTy);
8188 NewCast = IC.InsertNewInstBefore(
8189 CastInst::create(opcode, SIOp0, SrcPTy, SIOp0->getName()+".c"), SI);
8190 return new StoreInst(NewCast, CastOp);
8197 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8198 Value *Val = SI.getOperand(0);
8199 Value *Ptr = SI.getOperand(1);
8201 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8202 EraseInstFromFunction(SI);
8207 // Do really simple DSE, to catch cases where there are several consequtive
8208 // stores to the same location, separated by a few arithmetic operations. This
8209 // situation often occurs with bitfield accesses.
8210 BasicBlock::iterator BBI = &SI;
8211 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8215 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8216 // Prev store isn't volatile, and stores to the same location?
8217 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8220 EraseInstFromFunction(*PrevSI);
8226 // If this is a load, we have to stop. However, if the loaded value is from
8227 // the pointer we're loading and is producing the pointer we're storing,
8228 // then *this* store is dead (X = load P; store X -> P).
8229 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8230 if (LI == Val && LI->getOperand(0) == Ptr) {
8231 EraseInstFromFunction(SI);
8235 // Otherwise, this is a load from some other location. Stores before it
8240 // Don't skip over loads or things that can modify memory.
8241 if (BBI->mayWriteToMemory())
8246 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8248 // store X, null -> turns into 'unreachable' in SimplifyCFG
8249 if (isa<ConstantPointerNull>(Ptr)) {
8250 if (!isa<UndefValue>(Val)) {
8251 SI.setOperand(0, UndefValue::get(Val->getType()));
8252 if (Instruction *U = dyn_cast<Instruction>(Val))
8253 WorkList.push_back(U); // Dropped a use.
8256 return 0; // Do not modify these!
8259 // store undef, Ptr -> noop
8260 if (isa<UndefValue>(Val)) {
8261 EraseInstFromFunction(SI);
8266 // If the pointer destination is a cast, see if we can fold the cast into the
8268 if (isa<CastInst>(Ptr))
8269 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8271 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8273 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8277 // If this store is the last instruction in the basic block, and if the block
8278 // ends with an unconditional branch, try to move it to the successor block.
8280 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8281 if (BI->isUnconditional()) {
8282 // Check to see if the successor block has exactly two incoming edges. If
8283 // so, see if the other predecessor contains a store to the same location.
8284 // if so, insert a PHI node (if needed) and move the stores down.
8285 BasicBlock *Dest = BI->getSuccessor(0);
8287 pred_iterator PI = pred_begin(Dest);
8288 BasicBlock *Other = 0;
8289 if (*PI != BI->getParent())
8292 if (PI != pred_end(Dest)) {
8293 if (*PI != BI->getParent())
8298 if (++PI != pred_end(Dest))
8301 if (Other) { // If only one other pred...
8302 BBI = Other->getTerminator();
8303 // Make sure this other block ends in an unconditional branch and that
8304 // there is an instruction before the branch.
8305 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8306 BBI != Other->begin()) {
8308 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8310 // If this instruction is a store to the same location.
8311 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8312 // Okay, we know we can perform this transformation. Insert a PHI
8313 // node now if we need it.
8314 Value *MergedVal = OtherStore->getOperand(0);
8315 if (MergedVal != SI.getOperand(0)) {
8316 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8317 PN->reserveOperandSpace(2);
8318 PN->addIncoming(SI.getOperand(0), SI.getParent());
8319 PN->addIncoming(OtherStore->getOperand(0), Other);
8320 MergedVal = InsertNewInstBefore(PN, Dest->front());
8323 // Advance to a place where it is safe to insert the new store and
8325 BBI = Dest->begin();
8326 while (isa<PHINode>(BBI)) ++BBI;
8327 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8328 OtherStore->isVolatile()), *BBI);
8330 // Nuke the old stores.
8331 EraseInstFromFunction(SI);
8332 EraseInstFromFunction(*OtherStore);
8344 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8345 // Change br (not X), label True, label False to: br X, label False, True
8347 BasicBlock *TrueDest;
8348 BasicBlock *FalseDest;
8349 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8350 !isa<Constant>(X)) {
8351 // Swap Destinations and condition...
8353 BI.setSuccessor(0, FalseDest);
8354 BI.setSuccessor(1, TrueDest);
8358 // Cannonicalize fcmp_one -> fcmp_oeq
8359 FCmpInst::Predicate FPred; Value *Y;
8360 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8361 TrueDest, FalseDest)))
8362 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8363 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8364 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8365 std::string Name = I->getName(); I->setName("");
8366 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8367 Value *NewSCC = new FCmpInst(NewPred, X, Y, Name, I);
8368 // Swap Destinations and condition...
8369 BI.setCondition(NewSCC);
8370 BI.setSuccessor(0, FalseDest);
8371 BI.setSuccessor(1, TrueDest);
8372 removeFromWorkList(I);
8373 I->getParent()->getInstList().erase(I);
8374 WorkList.push_back(cast<Instruction>(NewSCC));
8378 // Cannonicalize icmp_ne -> icmp_eq
8379 ICmpInst::Predicate IPred;
8380 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8381 TrueDest, FalseDest)))
8382 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8383 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8384 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8385 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8386 std::string Name = I->getName(); I->setName("");
8387 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8388 Value *NewSCC = new ICmpInst(NewPred, X, Y, Name, I);
8389 // Swap Destinations and condition...
8390 BI.setCondition(NewSCC);
8391 BI.setSuccessor(0, FalseDest);
8392 BI.setSuccessor(1, TrueDest);
8393 removeFromWorkList(I);
8394 I->getParent()->getInstList().erase(I);
8395 WorkList.push_back(cast<Instruction>(NewSCC));
8402 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8403 Value *Cond = SI.getCondition();
8404 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8405 if (I->getOpcode() == Instruction::Add)
8406 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8407 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8408 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8409 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8411 SI.setOperand(0, I->getOperand(0));
8412 WorkList.push_back(I);
8419 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8420 /// is to leave as a vector operation.
8421 static bool CheapToScalarize(Value *V, bool isConstant) {
8422 if (isa<ConstantAggregateZero>(V))
8424 if (ConstantPacked *C = dyn_cast<ConstantPacked>(V)) {
8425 if (isConstant) return true;
8426 // If all elts are the same, we can extract.
8427 Constant *Op0 = C->getOperand(0);
8428 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8429 if (C->getOperand(i) != Op0)
8433 Instruction *I = dyn_cast<Instruction>(V);
8434 if (!I) return false;
8436 // Insert element gets simplified to the inserted element or is deleted if
8437 // this is constant idx extract element and its a constant idx insertelt.
8438 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8439 isa<ConstantInt>(I->getOperand(2)))
8441 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8443 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8444 if (BO->hasOneUse() &&
8445 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8446 CheapToScalarize(BO->getOperand(1), isConstant)))
8448 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8449 if (CI->hasOneUse() &&
8450 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8451 CheapToScalarize(CI->getOperand(1), isConstant)))
8457 /// getShuffleMask - Read and decode a shufflevector mask. It turns undef
8458 /// elements into values that are larger than the #elts in the input.
8459 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8460 unsigned NElts = SVI->getType()->getNumElements();
8461 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8462 return std::vector<unsigned>(NElts, 0);
8463 if (isa<UndefValue>(SVI->getOperand(2)))
8464 return std::vector<unsigned>(NElts, 2*NElts);
8466 std::vector<unsigned> Result;
8467 const ConstantPacked *CP = cast<ConstantPacked>(SVI->getOperand(2));
8468 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8469 if (isa<UndefValue>(CP->getOperand(i)))
8470 Result.push_back(NElts*2); // undef -> 8
8472 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8476 /// FindScalarElement - Given a vector and an element number, see if the scalar
8477 /// value is already around as a register, for example if it were inserted then
8478 /// extracted from the vector.
8479 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8480 assert(isa<PackedType>(V->getType()) && "Not looking at a vector?");
8481 const PackedType *PTy = cast<PackedType>(V->getType());
8482 unsigned Width = PTy->getNumElements();
8483 if (EltNo >= Width) // Out of range access.
8484 return UndefValue::get(PTy->getElementType());
8486 if (isa<UndefValue>(V))
8487 return UndefValue::get(PTy->getElementType());
8488 else if (isa<ConstantAggregateZero>(V))
8489 return Constant::getNullValue(PTy->getElementType());
8490 else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V))
8491 return CP->getOperand(EltNo);
8492 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8493 // If this is an insert to a variable element, we don't know what it is.
8494 if (!isa<ConstantInt>(III->getOperand(2)))
8496 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8498 // If this is an insert to the element we are looking for, return the
8501 return III->getOperand(1);
8503 // Otherwise, the insertelement doesn't modify the value, recurse on its
8505 return FindScalarElement(III->getOperand(0), EltNo);
8506 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8507 unsigned InEl = getShuffleMask(SVI)[EltNo];
8509 return FindScalarElement(SVI->getOperand(0), InEl);
8510 else if (InEl < Width*2)
8511 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8513 return UndefValue::get(PTy->getElementType());
8516 // Otherwise, we don't know.
8520 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8522 // If packed val is undef, replace extract with scalar undef.
8523 if (isa<UndefValue>(EI.getOperand(0)))
8524 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8526 // If packed val is constant 0, replace extract with scalar 0.
8527 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8528 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8530 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
8531 // If packed val is constant with uniform operands, replace EI
8532 // with that operand
8533 Constant *op0 = C->getOperand(0);
8534 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8535 if (C->getOperand(i) != op0) {
8540 return ReplaceInstUsesWith(EI, op0);
8543 // If extracting a specified index from the vector, see if we can recursively
8544 // find a previously computed scalar that was inserted into the vector.
8545 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8546 // This instruction only demands the single element from the input vector.
8547 // If the input vector has a single use, simplify it based on this use
8549 uint64_t IndexVal = IdxC->getZExtValue();
8550 if (EI.getOperand(0)->hasOneUse()) {
8552 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8555 EI.setOperand(0, V);
8560 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8561 return ReplaceInstUsesWith(EI, Elt);
8564 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8565 if (I->hasOneUse()) {
8566 // Push extractelement into predecessor operation if legal and
8567 // profitable to do so
8568 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8569 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8570 if (CheapToScalarize(BO, isConstantElt)) {
8571 ExtractElementInst *newEI0 =
8572 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8573 EI.getName()+".lhs");
8574 ExtractElementInst *newEI1 =
8575 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8576 EI.getName()+".rhs");
8577 InsertNewInstBefore(newEI0, EI);
8578 InsertNewInstBefore(newEI1, EI);
8579 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8581 } else if (isa<LoadInst>(I)) {
8582 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8583 PointerType::get(EI.getType()), EI);
8584 GetElementPtrInst *GEP =
8585 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8586 InsertNewInstBefore(GEP, EI);
8587 return new LoadInst(GEP);
8590 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8591 // Extracting the inserted element?
8592 if (IE->getOperand(2) == EI.getOperand(1))
8593 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8594 // If the inserted and extracted elements are constants, they must not
8595 // be the same value, extract from the pre-inserted value instead.
8596 if (isa<Constant>(IE->getOperand(2)) &&
8597 isa<Constant>(EI.getOperand(1))) {
8598 AddUsesToWorkList(EI);
8599 EI.setOperand(0, IE->getOperand(0));
8602 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8603 // If this is extracting an element from a shufflevector, figure out where
8604 // it came from and extract from the appropriate input element instead.
8605 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8606 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8608 if (SrcIdx < SVI->getType()->getNumElements())
8609 Src = SVI->getOperand(0);
8610 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8611 SrcIdx -= SVI->getType()->getNumElements();
8612 Src = SVI->getOperand(1);
8614 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8616 return new ExtractElementInst(Src, SrcIdx);
8623 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8624 /// elements from either LHS or RHS, return the shuffle mask and true.
8625 /// Otherwise, return false.
8626 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8627 std::vector<Constant*> &Mask) {
8628 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8629 "Invalid CollectSingleShuffleElements");
8630 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8632 if (isa<UndefValue>(V)) {
8633 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8635 } else if (V == LHS) {
8636 for (unsigned i = 0; i != NumElts; ++i)
8637 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8639 } else if (V == RHS) {
8640 for (unsigned i = 0; i != NumElts; ++i)
8641 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
8643 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8644 // If this is an insert of an extract from some other vector, include it.
8645 Value *VecOp = IEI->getOperand(0);
8646 Value *ScalarOp = IEI->getOperand(1);
8647 Value *IdxOp = IEI->getOperand(2);
8649 if (!isa<ConstantInt>(IdxOp))
8651 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8653 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8654 // Okay, we can handle this if the vector we are insertinting into is
8656 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8657 // If so, update the mask to reflect the inserted undef.
8658 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
8661 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8662 if (isa<ConstantInt>(EI->getOperand(1)) &&
8663 EI->getOperand(0)->getType() == V->getType()) {
8664 unsigned ExtractedIdx =
8665 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8667 // This must be extracting from either LHS or RHS.
8668 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8669 // Okay, we can handle this if the vector we are insertinting into is
8671 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8672 // If so, update the mask to reflect the inserted value.
8673 if (EI->getOperand(0) == LHS) {
8674 Mask[InsertedIdx & (NumElts-1)] =
8675 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8677 assert(EI->getOperand(0) == RHS);
8678 Mask[InsertedIdx & (NumElts-1)] =
8679 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
8688 // TODO: Handle shufflevector here!
8693 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8694 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8695 /// that computes V and the LHS value of the shuffle.
8696 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8698 assert(isa<PackedType>(V->getType()) &&
8699 (RHS == 0 || V->getType() == RHS->getType()) &&
8700 "Invalid shuffle!");
8701 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8703 if (isa<UndefValue>(V)) {
8704 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
8706 } else if (isa<ConstantAggregateZero>(V)) {
8707 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
8709 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8710 // If this is an insert of an extract from some other vector, include it.
8711 Value *VecOp = IEI->getOperand(0);
8712 Value *ScalarOp = IEI->getOperand(1);
8713 Value *IdxOp = IEI->getOperand(2);
8715 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8716 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8717 EI->getOperand(0)->getType() == V->getType()) {
8718 unsigned ExtractedIdx =
8719 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8720 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8722 // Either the extracted from or inserted into vector must be RHSVec,
8723 // otherwise we'd end up with a shuffle of three inputs.
8724 if (EI->getOperand(0) == RHS || RHS == 0) {
8725 RHS = EI->getOperand(0);
8726 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8727 Mask[InsertedIdx & (NumElts-1)] =
8728 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
8733 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8734 // Everything but the extracted element is replaced with the RHS.
8735 for (unsigned i = 0; i != NumElts; ++i) {
8736 if (i != InsertedIdx)
8737 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
8742 // If this insertelement is a chain that comes from exactly these two
8743 // vectors, return the vector and the effective shuffle.
8744 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8745 return EI->getOperand(0);
8750 // TODO: Handle shufflevector here!
8752 // Otherwise, can't do anything fancy. Return an identity vector.
8753 for (unsigned i = 0; i != NumElts; ++i)
8754 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
8758 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8759 Value *VecOp = IE.getOperand(0);
8760 Value *ScalarOp = IE.getOperand(1);
8761 Value *IdxOp = IE.getOperand(2);
8763 // If the inserted element was extracted from some other vector, and if the
8764 // indexes are constant, try to turn this into a shufflevector operation.
8765 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8766 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8767 EI->getOperand(0)->getType() == IE.getType()) {
8768 unsigned NumVectorElts = IE.getType()->getNumElements();
8769 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8770 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8772 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8773 return ReplaceInstUsesWith(IE, VecOp);
8775 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8776 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8778 // If we are extracting a value from a vector, then inserting it right
8779 // back into the same place, just use the input vector.
8780 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8781 return ReplaceInstUsesWith(IE, VecOp);
8783 // We could theoretically do this for ANY input. However, doing so could
8784 // turn chains of insertelement instructions into a chain of shufflevector
8785 // instructions, and right now we do not merge shufflevectors. As such,
8786 // only do this in a situation where it is clear that there is benefit.
8787 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8788 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8789 // the values of VecOp, except then one read from EIOp0.
8790 // Build a new shuffle mask.
8791 std::vector<Constant*> Mask;
8792 if (isa<UndefValue>(VecOp))
8793 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
8795 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8796 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
8799 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
8800 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8801 ConstantPacked::get(Mask));
8804 // If this insertelement isn't used by some other insertelement, turn it
8805 // (and any insertelements it points to), into one big shuffle.
8806 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8807 std::vector<Constant*> Mask;
8809 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8810 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8811 // We now have a shuffle of LHS, RHS, Mask.
8812 return new ShuffleVectorInst(LHS, RHS, ConstantPacked::get(Mask));
8821 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8822 Value *LHS = SVI.getOperand(0);
8823 Value *RHS = SVI.getOperand(1);
8824 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8826 bool MadeChange = false;
8828 // Undefined shuffle mask -> undefined value.
8829 if (isa<UndefValue>(SVI.getOperand(2)))
8830 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8832 // If we have shuffle(x, undef, mask) and any elements of mask refer to
8833 // the undef, change them to undefs.
8834 if (isa<UndefValue>(SVI.getOperand(1))) {
8835 // Scan to see if there are any references to the RHS. If so, replace them
8836 // with undef element refs and set MadeChange to true.
8837 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8838 if (Mask[i] >= e && Mask[i] != 2*e) {
8845 // Remap any references to RHS to use LHS.
8846 std::vector<Constant*> Elts;
8847 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8849 Elts.push_back(UndefValue::get(Type::Int32Ty));
8851 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8853 SVI.setOperand(2, ConstantPacked::get(Elts));
8857 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8858 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8859 if (LHS == RHS || isa<UndefValue>(LHS)) {
8860 if (isa<UndefValue>(LHS) && LHS == RHS) {
8861 // shuffle(undef,undef,mask) -> undef.
8862 return ReplaceInstUsesWith(SVI, LHS);
8865 // Remap any references to RHS to use LHS.
8866 std::vector<Constant*> Elts;
8867 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8869 Elts.push_back(UndefValue::get(Type::Int32Ty));
8871 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8872 (Mask[i] < e && isa<UndefValue>(LHS)))
8873 Mask[i] = 2*e; // Turn into undef.
8875 Mask[i] &= (e-1); // Force to LHS.
8876 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
8879 SVI.setOperand(0, SVI.getOperand(1));
8880 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8881 SVI.setOperand(2, ConstantPacked::get(Elts));
8882 LHS = SVI.getOperand(0);
8883 RHS = SVI.getOperand(1);
8887 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8888 bool isLHSID = true, isRHSID = true;
8890 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8891 if (Mask[i] >= e*2) continue; // Ignore undef values.
8892 // Is this an identity shuffle of the LHS value?
8893 isLHSID &= (Mask[i] == i);
8895 // Is this an identity shuffle of the RHS value?
8896 isRHSID &= (Mask[i]-e == i);
8899 // Eliminate identity shuffles.
8900 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
8901 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
8903 // If the LHS is a shufflevector itself, see if we can combine it with this
8904 // one without producing an unusual shuffle. Here we are really conservative:
8905 // we are absolutely afraid of producing a shuffle mask not in the input
8906 // program, because the code gen may not be smart enough to turn a merged
8907 // shuffle into two specific shuffles: it may produce worse code. As such,
8908 // we only merge two shuffles if the result is one of the two input shuffle
8909 // masks. In this case, merging the shuffles just removes one instruction,
8910 // which we know is safe. This is good for things like turning:
8911 // (splat(splat)) -> splat.
8912 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
8913 if (isa<UndefValue>(RHS)) {
8914 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
8916 std::vector<unsigned> NewMask;
8917 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
8919 NewMask.push_back(2*e);
8921 NewMask.push_back(LHSMask[Mask[i]]);
8923 // If the result mask is equal to the src shuffle or this shuffle mask, do
8925 if (NewMask == LHSMask || NewMask == Mask) {
8926 std::vector<Constant*> Elts;
8927 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
8928 if (NewMask[i] >= e*2) {
8929 Elts.push_back(UndefValue::get(Type::Int32Ty));
8931 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
8934 return new ShuffleVectorInst(LHSSVI->getOperand(0),
8935 LHSSVI->getOperand(1),
8936 ConstantPacked::get(Elts));
8941 return MadeChange ? &SVI : 0;
8946 void InstCombiner::removeFromWorkList(Instruction *I) {
8947 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
8952 /// TryToSinkInstruction - Try to move the specified instruction from its
8953 /// current block into the beginning of DestBlock, which can only happen if it's
8954 /// safe to move the instruction past all of the instructions between it and the
8955 /// end of its block.
8956 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
8957 assert(I->hasOneUse() && "Invariants didn't hold!");
8959 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
8960 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
8962 // Do not sink alloca instructions out of the entry block.
8963 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
8966 // We can only sink load instructions if there is nothing between the load and
8967 // the end of block that could change the value.
8968 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8969 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
8971 if (Scan->mayWriteToMemory())
8975 BasicBlock::iterator InsertPos = DestBlock->begin();
8976 while (isa<PHINode>(InsertPos)) ++InsertPos;
8978 I->moveBefore(InsertPos);
8983 /// OptimizeConstantExpr - Given a constant expression and target data layout
8984 /// information, symbolically evaluate the constant expr to something simpler
8986 static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
8989 Constant *Ptr = CE->getOperand(0);
8990 if (CE->getOpcode() == Instruction::GetElementPtr && Ptr->isNullValue() &&
8991 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
8992 // If this is a constant expr gep that is effectively computing an
8993 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
8994 bool isFoldableGEP = true;
8995 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
8996 if (!isa<ConstantInt>(CE->getOperand(i)))
8997 isFoldableGEP = false;
8998 if (isFoldableGEP) {
8999 std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
9000 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
9001 Constant *C = ConstantInt::get(TD->getIntPtrType(), Offset);
9002 return ConstantExpr::getIntToPtr(C, CE->getType());
9010 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9011 /// all reachable code to the worklist.
9013 /// This has a couple of tricks to make the code faster and more powerful. In
9014 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9015 /// them to the worklist (this significantly speeds up instcombine on code where
9016 /// many instructions are dead or constant). Additionally, if we find a branch
9017 /// whose condition is a known constant, we only visit the reachable successors.
9019 static void AddReachableCodeToWorklist(BasicBlock *BB,
9020 std::set<BasicBlock*> &Visited,
9021 std::vector<Instruction*> &WorkList,
9022 const TargetData *TD) {
9023 // We have now visited this block! If we've already been here, bail out.
9024 if (!Visited.insert(BB).second) return;
9026 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9027 Instruction *Inst = BBI++;
9029 // DCE instruction if trivially dead.
9030 if (isInstructionTriviallyDead(Inst)) {
9032 DOUT << "IC: DCE: " << *Inst;
9033 Inst->eraseFromParent();
9037 // ConstantProp instruction if trivially constant.
9038 if (Constant *C = ConstantFoldInstruction(Inst)) {
9039 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
9040 C = OptimizeConstantExpr(CE, TD);
9041 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9042 Inst->replaceAllUsesWith(C);
9044 Inst->eraseFromParent();
9048 WorkList.push_back(Inst);
9051 // Recursively visit successors. If this is a branch or switch on a constant,
9052 // only visit the reachable successor.
9053 TerminatorInst *TI = BB->getTerminator();
9054 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9055 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9056 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9057 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
9061 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9062 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9063 // See if this is an explicit destination.
9064 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9065 if (SI->getCaseValue(i) == Cond) {
9066 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
9070 // Otherwise it is the default destination.
9071 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
9076 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9077 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
9080 bool InstCombiner::runOnFunction(Function &F) {
9081 bool Changed = false;
9082 TD = &getAnalysis<TargetData>();
9085 // Do a depth-first traversal of the function, populate the worklist with
9086 // the reachable instructions. Ignore blocks that are not reachable. Keep
9087 // track of which blocks we visit.
9088 std::set<BasicBlock*> Visited;
9089 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
9091 // Do a quick scan over the function. If we find any blocks that are
9092 // unreachable, remove any instructions inside of them. This prevents
9093 // the instcombine code from having to deal with some bad special cases.
9094 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9095 if (!Visited.count(BB)) {
9096 Instruction *Term = BB->getTerminator();
9097 while (Term != BB->begin()) { // Remove instrs bottom-up
9098 BasicBlock::iterator I = Term; --I;
9100 DOUT << "IC: DCE: " << *I;
9103 if (!I->use_empty())
9104 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9105 I->eraseFromParent();
9110 while (!WorkList.empty()) {
9111 Instruction *I = WorkList.back(); // Get an instruction from the worklist
9112 WorkList.pop_back();
9114 // Check to see if we can DCE the instruction.
9115 if (isInstructionTriviallyDead(I)) {
9116 // Add operands to the worklist.
9117 if (I->getNumOperands() < 4)
9118 AddUsesToWorkList(*I);
9121 DOUT << "IC: DCE: " << *I;
9123 I->eraseFromParent();
9124 removeFromWorkList(I);
9128 // Instruction isn't dead, see if we can constant propagate it.
9129 if (Constant *C = ConstantFoldInstruction(I)) {
9130 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
9131 C = OptimizeConstantExpr(CE, TD);
9132 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9134 // Add operands to the worklist.
9135 AddUsesToWorkList(*I);
9136 ReplaceInstUsesWith(*I, C);
9139 I->eraseFromParent();
9140 removeFromWorkList(I);
9144 // See if we can trivially sink this instruction to a successor basic block.
9145 if (I->hasOneUse()) {
9146 BasicBlock *BB = I->getParent();
9147 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9148 if (UserParent != BB) {
9149 bool UserIsSuccessor = false;
9150 // See if the user is one of our successors.
9151 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9152 if (*SI == UserParent) {
9153 UserIsSuccessor = true;
9157 // If the user is one of our immediate successors, and if that successor
9158 // only has us as a predecessors (we'd have to split the critical edge
9159 // otherwise), we can keep going.
9160 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9161 next(pred_begin(UserParent)) == pred_end(UserParent))
9162 // Okay, the CFG is simple enough, try to sink this instruction.
9163 Changed |= TryToSinkInstruction(I, UserParent);
9167 // Now that we have an instruction, try combining it to simplify it...
9168 if (Instruction *Result = visit(*I)) {
9170 // Should we replace the old instruction with a new one?
9172 DOUT << "IC: Old = " << *I
9173 << " New = " << *Result;
9175 // Everything uses the new instruction now.
9176 I->replaceAllUsesWith(Result);
9178 // Push the new instruction and any users onto the worklist.
9179 WorkList.push_back(Result);
9180 AddUsersToWorkList(*Result);
9182 // Move the name to the new instruction first...
9183 std::string OldName = I->getName(); I->setName("");
9184 Result->setName(OldName);
9186 // Insert the new instruction into the basic block...
9187 BasicBlock *InstParent = I->getParent();
9188 BasicBlock::iterator InsertPos = I;
9190 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9191 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9194 InstParent->getInstList().insert(InsertPos, Result);
9196 // Make sure that we reprocess all operands now that we reduced their
9198 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9199 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9200 WorkList.push_back(OpI);
9202 // Instructions can end up on the worklist more than once. Make sure
9203 // we do not process an instruction that has been deleted.
9204 removeFromWorkList(I);
9206 // Erase the old instruction.
9207 InstParent->getInstList().erase(I);
9209 DOUT << "IC: MOD = " << *I;
9211 // If the instruction was modified, it's possible that it is now dead.
9212 // if so, remove it.
9213 if (isInstructionTriviallyDead(I)) {
9214 // Make sure we process all operands now that we are reducing their
9216 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9217 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9218 WorkList.push_back(OpI);
9220 // Instructions may end up in the worklist more than once. Erase all
9221 // occurrences of this instruction.
9222 removeFromWorkList(I);
9223 I->eraseFromParent();
9225 WorkList.push_back(Result);
9226 AddUsersToWorkList(*Result);
9236 FunctionPass *llvm::createInstructionCombiningPass() {
9237 return new InstCombiner();