1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic 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 "InstCombine.h"
39 #include "llvm/IntrinsicInst.h"
40 #include "llvm/Analysis/ConstantFolding.h"
41 #include "llvm/Analysis/InstructionSimplify.h"
42 #include "llvm/Analysis/MemoryBuiltins.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Support/CFG.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/GetElementPtrTypeIterator.h"
48 #include "llvm/Support/PatternMatch.h"
49 #include "llvm/ADT/SmallPtrSet.h"
50 #include "llvm/ADT/Statistic.h"
51 #include "llvm-c/Initialization.h"
55 using namespace llvm::PatternMatch;
57 STATISTIC(NumCombined , "Number of insts combined");
58 STATISTIC(NumConstProp, "Number of constant folds");
59 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
60 STATISTIC(NumSunkInst , "Number of instructions sunk");
62 // Initialization Routines
63 void llvm::initializeInstCombine(PassRegistry &Registry) {
64 initializeInstCombinerPass(Registry);
67 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
68 initializeInstCombine(*unwrap(R));
71 char InstCombiner::ID = 0;
72 INITIALIZE_PASS(InstCombiner, "instcombine",
73 "Combine redundant instructions", false, false)
75 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
76 AU.addPreservedID(LCSSAID);
81 /// ShouldChangeType - Return true if it is desirable to convert a computation
82 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
83 /// type for example, or from a smaller to a larger illegal type.
84 bool InstCombiner::ShouldChangeType(const Type *From, const Type *To) const {
85 assert(From->isIntegerTy() && To->isIntegerTy());
87 // If we don't have TD, we don't know if the source/dest are legal.
88 if (!TD) return false;
90 unsigned FromWidth = From->getPrimitiveSizeInBits();
91 unsigned ToWidth = To->getPrimitiveSizeInBits();
92 bool FromLegal = TD->isLegalInteger(FromWidth);
93 bool ToLegal = TD->isLegalInteger(ToWidth);
95 // If this is a legal integer from type, and the result would be an illegal
96 // type, don't do the transformation.
97 if (FromLegal && !ToLegal)
100 // Otherwise, if both are illegal, do not increase the size of the result. We
101 // do allow things like i160 -> i64, but not i64 -> i160.
102 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
109 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
110 /// operators which are associative or commutative:
112 // Commutative operators:
114 // 1. Order operands such that they are listed from right (least complex) to
115 // left (most complex). This puts constants before unary operators before
118 // Associative operators:
120 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
121 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
123 // Associative and commutative operators:
125 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
126 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
127 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
128 // if C1 and C2 are constants.
130 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
131 Instruction::BinaryOps Opcode = I.getOpcode();
132 bool Changed = false;
135 // Order operands such that they are listed from right (least complex) to
136 // left (most complex). This puts constants before unary operators before
138 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
139 getComplexity(I.getOperand(1)))
140 Changed = !I.swapOperands();
142 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
143 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
145 if (I.isAssociative()) {
146 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
147 if (Op0 && Op0->getOpcode() == Opcode) {
148 Value *A = Op0->getOperand(0);
149 Value *B = Op0->getOperand(1);
150 Value *C = I.getOperand(1);
152 // Does "B op C" simplify?
153 if (Value *V = SimplifyBinOp(Opcode, B, C, TD)) {
154 // It simplifies to V. Form "A op V".
162 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
163 if (Op1 && Op1->getOpcode() == Opcode) {
164 Value *A = I.getOperand(0);
165 Value *B = Op1->getOperand(0);
166 Value *C = Op1->getOperand(1);
168 // Does "A op B" simplify?
169 if (Value *V = SimplifyBinOp(Opcode, A, B, TD)) {
170 // It simplifies to V. Form "V op C".
179 if (I.isAssociative() && I.isCommutative()) {
180 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
181 if (Op0 && Op0->getOpcode() == Opcode) {
182 Value *A = Op0->getOperand(0);
183 Value *B = Op0->getOperand(1);
184 Value *C = I.getOperand(1);
186 // Does "C op A" simplify?
187 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
188 // It simplifies to V. Form "V op B".
196 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
197 if (Op1 && Op1->getOpcode() == Opcode) {
198 Value *A = I.getOperand(0);
199 Value *B = Op1->getOperand(0);
200 Value *C = Op1->getOperand(1);
202 // Does "C op A" simplify?
203 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
204 // It simplifies to V. Form "B op V".
212 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
213 // if C1 and C2 are constants.
215 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
216 isa<Constant>(Op0->getOperand(1)) &&
217 isa<Constant>(Op1->getOperand(1)) &&
218 Op0->hasOneUse() && Op1->hasOneUse()) {
219 Value *A = Op0->getOperand(0);
220 Constant *C1 = cast<Constant>(Op0->getOperand(1));
221 Value *B = Op1->getOperand(0);
222 Constant *C2 = cast<Constant>(Op1->getOperand(1));
224 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
225 Instruction *New = BinaryOperator::Create(Opcode, A, B, Op1->getName(),
228 I.setOperand(0, New);
229 I.setOperand(1, Folded);
235 // No further simplifications.
240 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
241 /// "(X LOp Y) ROp (X LOp Z)".
242 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
243 Instruction::BinaryOps ROp) {
248 case Instruction::And:
249 // And distributes over Or and Xor.
253 case Instruction::Or:
254 case Instruction::Xor:
258 case Instruction::Mul:
259 // Multiplication distributes over addition and subtraction.
263 case Instruction::Add:
264 case Instruction::Sub:
268 case Instruction::Or:
269 // Or distributes over And.
273 case Instruction::And:
279 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
280 /// "(X ROp Z) LOp (Y ROp Z)".
281 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
282 Instruction::BinaryOps ROp) {
283 if (Instruction::isCommutative(ROp))
284 return LeftDistributesOverRight(ROp, LOp);
285 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
286 // but this requires knowing that the addition does not overflow and other
291 /// SimplifyDistributed - This tries to simplify binary operations which some
292 /// other binary operation distributes over (eg "A*B+A*C" -> "A*(B+C)" since
293 /// addition is distributed over by multiplication). Returns the result of
294 /// the simplification, or null if no simplification was performed.
295 Instruction *InstCombiner::SimplifyDistributed(BinaryOperator &I) {
296 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
297 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
298 if (!Op0 || !Op1 || Op0->getOpcode() != Op1->getOpcode())
301 // The instruction has the form "(A op' B) op (C op' D)".
302 Value *A = Op0->getOperand(0); Value *B = Op0->getOperand(1);
303 Value *C = Op1->getOperand(0); Value *D = Op1->getOperand(1);
304 Instruction::BinaryOps OuterOpcode = I.getOpcode(); // op
305 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
307 // Does "X op' Y" always equal "Y op' X"?
308 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
310 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
311 if (LeftDistributesOverRight(InnerOpcode, OuterOpcode))
312 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
313 // commutative case, "(A op' B) op (C op' A)"?
314 if (A == C || (InnerCommutative && A == D)) {
317 // Consider forming "A op' (B op D)".
318 // If "B op D" simplifies then it can be formed with no cost.
319 Value *RHS = SimplifyBinOp(OuterOpcode, B, D, TD);
320 // If "B op D" doesn't simplify then only proceed if both of the existing
321 // operations "A op' B" and "C op' D" will be zapped since no longer used.
322 if (!RHS && Op0->hasOneUse() && Op1->hasOneUse())
323 RHS = Builder->CreateBinOp(OuterOpcode, B, D, Op1->getName());
325 return BinaryOperator::Create(InnerOpcode, A, RHS);
328 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
329 if (RightDistributesOverLeft(OuterOpcode, InnerOpcode))
330 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
331 // commutative case, "(A op' B) op (B op' D)"?
332 if (B == D || (InnerCommutative && B == C)) {
335 // Consider forming "(A op C) op' B".
336 // If "A op C" simplifies then it can be formed with no cost.
337 Value *LHS = SimplifyBinOp(OuterOpcode, A, C, TD);
338 // If "A op C" doesn't simplify then only proceed if both of the existing
339 // operations "A op' B" and "C op' D" will be zapped since no longer used.
340 if (!LHS && Op0->hasOneUse() && Op1->hasOneUse())
341 LHS = Builder->CreateBinOp(OuterOpcode, A, C, Op0->getName());
343 return BinaryOperator::Create(InnerOpcode, LHS, B);
349 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
350 // if the LHS is a constant zero (which is the 'negate' form).
352 Value *InstCombiner::dyn_castNegVal(Value *V) const {
353 if (BinaryOperator::isNeg(V))
354 return BinaryOperator::getNegArgument(V);
356 // Constants can be considered to be negated values if they can be folded.
357 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
358 return ConstantExpr::getNeg(C);
360 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
361 if (C->getType()->getElementType()->isIntegerTy())
362 return ConstantExpr::getNeg(C);
367 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
368 // instruction if the LHS is a constant negative zero (which is the 'negate'
371 Value *InstCombiner::dyn_castFNegVal(Value *V) const {
372 if (BinaryOperator::isFNeg(V))
373 return BinaryOperator::getFNegArgument(V);
375 // Constants can be considered to be negated values if they can be folded.
376 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
377 return ConstantExpr::getFNeg(C);
379 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
380 if (C->getType()->getElementType()->isFloatingPointTy())
381 return ConstantExpr::getFNeg(C);
386 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
388 if (CastInst *CI = dyn_cast<CastInst>(&I))
389 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
391 // Figure out if the constant is the left or the right argument.
392 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
393 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
395 if (Constant *SOC = dyn_cast<Constant>(SO)) {
397 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
398 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
401 Value *Op0 = SO, *Op1 = ConstOperand;
405 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
406 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
407 SO->getName()+".op");
408 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
409 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
410 SO->getName()+".cmp");
411 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
412 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
413 SO->getName()+".cmp");
414 llvm_unreachable("Unknown binary instruction type!");
417 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
418 // constant as the other operand, try to fold the binary operator into the
419 // select arguments. This also works for Cast instructions, which obviously do
420 // not have a second operand.
421 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
422 // Don't modify shared select instructions
423 if (!SI->hasOneUse()) return 0;
424 Value *TV = SI->getOperand(1);
425 Value *FV = SI->getOperand(2);
427 if (isa<Constant>(TV) || isa<Constant>(FV)) {
428 // Bool selects with constant operands can be folded to logical ops.
429 if (SI->getType()->isIntegerTy(1)) return 0;
431 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
432 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
434 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
441 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
442 /// has a PHI node as operand #0, see if we can fold the instruction into the
443 /// PHI (which is only possible if all operands to the PHI are constants).
445 /// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
446 /// that would normally be unprofitable because they strongly encourage jump
448 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
449 bool AllowAggressive) {
450 AllowAggressive = false;
451 PHINode *PN = cast<PHINode>(I.getOperand(0));
452 unsigned NumPHIValues = PN->getNumIncomingValues();
453 if (NumPHIValues == 0 ||
454 // We normally only transform phis with a single use, unless we're trying
455 // hard to make jump threading happen.
456 (!PN->hasOneUse() && !AllowAggressive))
460 // Check to see if all of the operands of the PHI are simple constants
461 // (constantint/constantfp/undef). If there is one non-constant value,
462 // remember the BB it is in. If there is more than one or if *it* is a PHI,
463 // bail out. We don't do arbitrary constant expressions here because moving
464 // their computation can be expensive without a cost model.
465 BasicBlock *NonConstBB = 0;
466 for (unsigned i = 0; i != NumPHIValues; ++i)
467 if (!isa<Constant>(PN->getIncomingValue(i)) ||
468 isa<ConstantExpr>(PN->getIncomingValue(i))) {
469 if (NonConstBB) return 0; // More than one non-const value.
470 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
471 NonConstBB = PN->getIncomingBlock(i);
473 // If the incoming non-constant value is in I's block, we have an infinite
475 if (NonConstBB == I.getParent())
479 // If there is exactly one non-constant value, we can insert a copy of the
480 // operation in that block. However, if this is a critical edge, we would be
481 // inserting the computation one some other paths (e.g. inside a loop). Only
482 // do this if the pred block is unconditionally branching into the phi block.
483 if (NonConstBB != 0 && !AllowAggressive) {
484 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
485 if (!BI || !BI->isUnconditional()) return 0;
488 // Okay, we can do the transformation: create the new PHI node.
489 PHINode *NewPN = PHINode::Create(I.getType(), "");
490 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
491 InsertNewInstBefore(NewPN, *PN);
494 // Next, add all of the operands to the PHI.
495 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
496 // We only currently try to fold the condition of a select when it is a phi,
497 // not the true/false values.
498 Value *TrueV = SI->getTrueValue();
499 Value *FalseV = SI->getFalseValue();
500 BasicBlock *PhiTransBB = PN->getParent();
501 for (unsigned i = 0; i != NumPHIValues; ++i) {
502 BasicBlock *ThisBB = PN->getIncomingBlock(i);
503 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
504 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
506 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
507 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
509 assert(PN->getIncomingBlock(i) == NonConstBB);
510 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
512 "phitmp", NonConstBB->getTerminator());
513 Worklist.Add(cast<Instruction>(InV));
515 NewPN->addIncoming(InV, ThisBB);
517 } else if (I.getNumOperands() == 2) {
518 Constant *C = cast<Constant>(I.getOperand(1));
519 for (unsigned i = 0; i != NumPHIValues; ++i) {
521 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
522 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
523 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
525 InV = ConstantExpr::get(I.getOpcode(), InC, C);
527 assert(PN->getIncomingBlock(i) == NonConstBB);
528 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
529 InV = BinaryOperator::Create(BO->getOpcode(),
530 PN->getIncomingValue(i), C, "phitmp",
531 NonConstBB->getTerminator());
532 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
533 InV = CmpInst::Create(CI->getOpcode(),
535 PN->getIncomingValue(i), C, "phitmp",
536 NonConstBB->getTerminator());
538 llvm_unreachable("Unknown binop!");
540 Worklist.Add(cast<Instruction>(InV));
542 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
545 CastInst *CI = cast<CastInst>(&I);
546 const Type *RetTy = CI->getType();
547 for (unsigned i = 0; i != NumPHIValues; ++i) {
549 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
550 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
552 assert(PN->getIncomingBlock(i) == NonConstBB);
553 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
554 I.getType(), "phitmp",
555 NonConstBB->getTerminator());
556 Worklist.Add(cast<Instruction>(InV));
558 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
561 return ReplaceInstUsesWith(I, NewPN);
564 /// FindElementAtOffset - Given a type and a constant offset, determine whether
565 /// or not there is a sequence of GEP indices into the type that will land us at
566 /// the specified offset. If so, fill them into NewIndices and return the
567 /// resultant element type, otherwise return null.
568 const Type *InstCombiner::FindElementAtOffset(const Type *Ty, int64_t Offset,
569 SmallVectorImpl<Value*> &NewIndices) {
571 if (!Ty->isSized()) return 0;
573 // Start with the index over the outer type. Note that the type size
574 // might be zero (even if the offset isn't zero) if the indexed type
575 // is something like [0 x {int, int}]
576 const Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
577 int64_t FirstIdx = 0;
578 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
579 FirstIdx = Offset/TySize;
580 Offset -= FirstIdx*TySize;
582 // Handle hosts where % returns negative instead of values [0..TySize).
588 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
591 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
593 // Index into the types. If we fail, set OrigBase to null.
595 // Indexing into tail padding between struct/array elements.
596 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
599 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
600 const StructLayout *SL = TD->getStructLayout(STy);
601 assert(Offset < (int64_t)SL->getSizeInBytes() &&
602 "Offset must stay within the indexed type");
604 unsigned Elt = SL->getElementContainingOffset(Offset);
605 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
608 Offset -= SL->getElementOffset(Elt);
609 Ty = STy->getElementType(Elt);
610 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
611 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
612 assert(EltSize && "Cannot index into a zero-sized array");
613 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
615 Ty = AT->getElementType();
617 // Otherwise, we can't index into the middle of this atomic type, bail.
627 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
628 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
630 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD))
631 return ReplaceInstUsesWith(GEP, V);
633 Value *PtrOp = GEP.getOperand(0);
635 // Eliminate unneeded casts for indices, and replace indices which displace
636 // by multiples of a zero size type with zero.
638 bool MadeChange = false;
639 const Type *IntPtrTy = TD->getIntPtrType(GEP.getContext());
641 gep_type_iterator GTI = gep_type_begin(GEP);
642 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
643 I != E; ++I, ++GTI) {
644 // Skip indices into struct types.
645 const SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
646 if (!SeqTy) continue;
648 // If the element type has zero size then any index over it is equivalent
649 // to an index of zero, so replace it with zero if it is not zero already.
650 if (SeqTy->getElementType()->isSized() &&
651 TD->getTypeAllocSize(SeqTy->getElementType()) == 0)
652 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
653 *I = Constant::getNullValue(IntPtrTy);
657 if ((*I)->getType() != IntPtrTy) {
658 // If we are using a wider index than needed for this platform, shrink
659 // it to what we need. If narrower, sign-extend it to what we need.
660 // This explicit cast can make subsequent optimizations more obvious.
661 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
665 if (MadeChange) return &GEP;
668 // Combine Indices - If the source pointer to this getelementptr instruction
669 // is a getelementptr instruction, combine the indices of the two
670 // getelementptr instructions into a single instruction.
672 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
673 // Note that if our source is a gep chain itself that we wait for that
674 // chain to be resolved before we perform this transformation. This
675 // avoids us creating a TON of code in some cases.
677 if (GetElementPtrInst *SrcGEP =
678 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
679 if (SrcGEP->getNumOperands() == 2)
680 return 0; // Wait until our source is folded to completion.
682 SmallVector<Value*, 8> Indices;
684 // Find out whether the last index in the source GEP is a sequential idx.
685 bool EndsWithSequential = false;
686 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
688 EndsWithSequential = !(*I)->isStructTy();
690 // Can we combine the two pointer arithmetics offsets?
691 if (EndsWithSequential) {
692 // Replace: gep (gep %P, long B), long A, ...
693 // With: T = long A+B; gep %P, T, ...
696 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
697 Value *GO1 = GEP.getOperand(1);
698 if (SO1 == Constant::getNullValue(SO1->getType())) {
700 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
703 // If they aren't the same type, then the input hasn't been processed
704 // by the loop above yet (which canonicalizes sequential index types to
705 // intptr_t). Just avoid transforming this until the input has been
707 if (SO1->getType() != GO1->getType())
709 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
712 // Update the GEP in place if possible.
713 if (Src->getNumOperands() == 2) {
714 GEP.setOperand(0, Src->getOperand(0));
715 GEP.setOperand(1, Sum);
718 Indices.append(Src->op_begin()+1, Src->op_end()-1);
719 Indices.push_back(Sum);
720 Indices.append(GEP.op_begin()+2, GEP.op_end());
721 } else if (isa<Constant>(*GEP.idx_begin()) &&
722 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
723 Src->getNumOperands() != 1) {
724 // Otherwise we can do the fold if the first index of the GEP is a zero
725 Indices.append(Src->op_begin()+1, Src->op_end());
726 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
729 if (!Indices.empty())
730 return (GEP.isInBounds() && Src->isInBounds()) ?
731 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
732 Indices.end(), GEP.getName()) :
733 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
734 Indices.end(), GEP.getName());
737 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
738 Value *StrippedPtr = PtrOp->stripPointerCasts();
739 if (StrippedPtr != PtrOp) {
740 const PointerType *StrippedPtrTy =cast<PointerType>(StrippedPtr->getType());
742 bool HasZeroPointerIndex = false;
743 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
744 HasZeroPointerIndex = C->isZero();
746 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
747 // into : GEP [10 x i8]* X, i32 0, ...
749 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
750 // into : GEP i8* X, ...
752 // This occurs when the program declares an array extern like "int X[];"
753 if (HasZeroPointerIndex) {
754 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
755 if (const ArrayType *CATy =
756 dyn_cast<ArrayType>(CPTy->getElementType())) {
757 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
758 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
760 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
761 GetElementPtrInst *Res =
762 GetElementPtrInst::Create(StrippedPtr, Idx.begin(),
763 Idx.end(), GEP.getName());
764 Res->setIsInBounds(GEP.isInBounds());
768 if (const ArrayType *XATy =
769 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
770 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
771 if (CATy->getElementType() == XATy->getElementType()) {
772 // -> GEP [10 x i8]* X, i32 0, ...
773 // At this point, we know that the cast source type is a pointer
774 // to an array of the same type as the destination pointer
775 // array. Because the array type is never stepped over (there
776 // is a leading zero) we can fold the cast into this GEP.
777 GEP.setOperand(0, StrippedPtr);
782 } else if (GEP.getNumOperands() == 2) {
783 // Transform things like:
784 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
785 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
786 const Type *SrcElTy = StrippedPtrTy->getElementType();
787 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
788 if (TD && SrcElTy->isArrayTy() &&
789 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
790 TD->getTypeAllocSize(ResElTy)) {
792 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
793 Idx[1] = GEP.getOperand(1);
794 Value *NewGEP = GEP.isInBounds() ?
795 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()) :
796 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName());
797 // V and GEP are both pointer types --> BitCast
798 return new BitCastInst(NewGEP, GEP.getType());
801 // Transform things like:
802 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
803 // (where tmp = 8*tmp2) into:
804 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
806 if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) {
807 uint64_t ArrayEltSize =
808 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
810 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
811 // allow either a mul, shift, or constant here.
813 ConstantInt *Scale = 0;
814 if (ArrayEltSize == 1) {
815 NewIdx = GEP.getOperand(1);
816 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
817 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
818 NewIdx = ConstantInt::get(CI->getType(), 1);
820 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
821 if (Inst->getOpcode() == Instruction::Shl &&
822 isa<ConstantInt>(Inst->getOperand(1))) {
823 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
824 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
825 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
827 NewIdx = Inst->getOperand(0);
828 } else if (Inst->getOpcode() == Instruction::Mul &&
829 isa<ConstantInt>(Inst->getOperand(1))) {
830 Scale = cast<ConstantInt>(Inst->getOperand(1));
831 NewIdx = Inst->getOperand(0);
835 // If the index will be to exactly the right offset with the scale taken
836 // out, perform the transformation. Note, we don't know whether Scale is
837 // signed or not. We'll use unsigned version of division/modulo
838 // operation after making sure Scale doesn't have the sign bit set.
839 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
840 Scale->getZExtValue() % ArrayEltSize == 0) {
841 Scale = ConstantInt::get(Scale->getType(),
842 Scale->getZExtValue() / ArrayEltSize);
843 if (Scale->getZExtValue() != 1) {
844 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
846 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
849 // Insert the new GEP instruction.
851 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
853 Value *NewGEP = GEP.isInBounds() ?
854 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2,GEP.getName()):
855 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName());
856 // The NewGEP must be pointer typed, so must the old one -> BitCast
857 return new BitCastInst(NewGEP, GEP.getType());
863 /// See if we can simplify:
864 /// X = bitcast A* to B*
865 /// Y = gep X, <...constant indices...>
866 /// into a gep of the original struct. This is important for SROA and alias
867 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
868 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
870 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
871 // Determine how much the GEP moves the pointer. We are guaranteed to get
872 // a constant back from EmitGEPOffset.
873 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP));
874 int64_t Offset = OffsetV->getSExtValue();
876 // If this GEP instruction doesn't move the pointer, just replace the GEP
877 // with a bitcast of the real input to the dest type.
879 // If the bitcast is of an allocation, and the allocation will be
880 // converted to match the type of the cast, don't touch this.
881 if (isa<AllocaInst>(BCI->getOperand(0)) ||
882 isMalloc(BCI->getOperand(0))) {
883 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
884 if (Instruction *I = visitBitCast(*BCI)) {
887 BCI->getParent()->getInstList().insert(BCI, I);
888 ReplaceInstUsesWith(*BCI, I);
893 return new BitCastInst(BCI->getOperand(0), GEP.getType());
896 // Otherwise, if the offset is non-zero, we need to find out if there is a
897 // field at Offset in 'A's type. If so, we can pull the cast through the
899 SmallVector<Value*, 8> NewIndices;
901 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
902 if (FindElementAtOffset(InTy, Offset, NewIndices)) {
903 Value *NGEP = GEP.isInBounds() ?
904 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
906 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
909 if (NGEP->getType() == GEP.getType())
910 return ReplaceInstUsesWith(GEP, NGEP);
911 NGEP->takeName(&GEP);
912 return new BitCastInst(NGEP, GEP.getType());
922 static bool IsOnlyNullComparedAndFreed(const Value &V) {
923 for (Value::const_use_iterator UI = V.use_begin(), UE = V.use_end();
928 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(U))
929 if (ICI->isEquality() && isa<ConstantPointerNull>(ICI->getOperand(1)))
936 Instruction *InstCombiner::visitMalloc(Instruction &MI) {
937 // If we have a malloc call which is only used in any amount of comparisons
938 // to null and free calls, delete the calls and replace the comparisons with
939 // true or false as appropriate.
940 if (IsOnlyNullComparedAndFreed(MI)) {
941 for (Value::use_iterator UI = MI.use_begin(), UE = MI.use_end();
943 // We can assume that every remaining use is a free call or an icmp eq/ne
944 // to null, so the cast is safe.
945 Instruction *I = cast<Instruction>(*UI);
947 // Early increment here, as we're about to get rid of the user.
951 EraseInstFromFunction(*cast<CallInst>(I));
954 // Again, the cast is safe.
955 ICmpInst *C = cast<ICmpInst>(I);
956 ReplaceInstUsesWith(*C, ConstantInt::get(Type::getInt1Ty(C->getContext()),
957 C->isFalseWhenEqual()));
958 EraseInstFromFunction(*C);
960 return EraseInstFromFunction(MI);
967 Instruction *InstCombiner::visitFree(CallInst &FI) {
968 Value *Op = FI.getArgOperand(0);
970 // free undef -> unreachable.
971 if (isa<UndefValue>(Op)) {
972 // Insert a new store to null because we cannot modify the CFG here.
973 new StoreInst(ConstantInt::getTrue(FI.getContext()),
974 UndefValue::get(Type::getInt1PtrTy(FI.getContext())), &FI);
975 return EraseInstFromFunction(FI);
978 // If we have 'free null' delete the instruction. This can happen in stl code
979 // when lots of inlining happens.
980 if (isa<ConstantPointerNull>(Op))
981 return EraseInstFromFunction(FI);
988 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
989 // Change br (not X), label True, label False to: br X, label False, True
991 BasicBlock *TrueDest;
992 BasicBlock *FalseDest;
993 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
995 // Swap Destinations and condition...
997 BI.setSuccessor(0, FalseDest);
998 BI.setSuccessor(1, TrueDest);
1002 // Cannonicalize fcmp_one -> fcmp_oeq
1003 FCmpInst::Predicate FPred; Value *Y;
1004 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
1005 TrueDest, FalseDest)) &&
1006 BI.getCondition()->hasOneUse())
1007 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
1008 FPred == FCmpInst::FCMP_OGE) {
1009 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
1010 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
1012 // Swap Destinations and condition.
1013 BI.setSuccessor(0, FalseDest);
1014 BI.setSuccessor(1, TrueDest);
1019 // Cannonicalize icmp_ne -> icmp_eq
1020 ICmpInst::Predicate IPred;
1021 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
1022 TrueDest, FalseDest)) &&
1023 BI.getCondition()->hasOneUse())
1024 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
1025 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
1026 IPred == ICmpInst::ICMP_SGE) {
1027 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
1028 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
1029 // Swap Destinations and condition.
1030 BI.setSuccessor(0, FalseDest);
1031 BI.setSuccessor(1, TrueDest);
1039 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
1040 Value *Cond = SI.getCondition();
1041 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
1042 if (I->getOpcode() == Instruction::Add)
1043 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1044 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
1045 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
1047 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
1049 SI.setOperand(0, I->getOperand(0));
1057 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
1058 Value *Agg = EV.getAggregateOperand();
1060 if (!EV.hasIndices())
1061 return ReplaceInstUsesWith(EV, Agg);
1063 if (Constant *C = dyn_cast<Constant>(Agg)) {
1064 if (isa<UndefValue>(C))
1065 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
1067 if (isa<ConstantAggregateZero>(C))
1068 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
1070 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
1071 // Extract the element indexed by the first index out of the constant
1072 Value *V = C->getOperand(*EV.idx_begin());
1073 if (EV.getNumIndices() > 1)
1074 // Extract the remaining indices out of the constant indexed by the
1076 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
1078 return ReplaceInstUsesWith(EV, V);
1080 return 0; // Can't handle other constants
1082 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
1083 // We're extracting from an insertvalue instruction, compare the indices
1084 const unsigned *exti, *exte, *insi, *inse;
1085 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
1086 exte = EV.idx_end(), inse = IV->idx_end();
1087 exti != exte && insi != inse;
1090 // The insert and extract both reference distinctly different elements.
1091 // This means the extract is not influenced by the insert, and we can
1092 // replace the aggregate operand of the extract with the aggregate
1093 // operand of the insert. i.e., replace
1094 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1095 // %E = extractvalue { i32, { i32 } } %I, 0
1097 // %E = extractvalue { i32, { i32 } } %A, 0
1098 return ExtractValueInst::Create(IV->getAggregateOperand(),
1099 EV.idx_begin(), EV.idx_end());
1101 if (exti == exte && insi == inse)
1102 // Both iterators are at the end: Index lists are identical. Replace
1103 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1104 // %C = extractvalue { i32, { i32 } } %B, 1, 0
1106 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
1108 // The extract list is a prefix of the insert list. i.e. replace
1109 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1110 // %E = extractvalue { i32, { i32 } } %I, 1
1112 // %X = extractvalue { i32, { i32 } } %A, 1
1113 // %E = insertvalue { i32 } %X, i32 42, 0
1114 // by switching the order of the insert and extract (though the
1115 // insertvalue should be left in, since it may have other uses).
1116 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
1117 EV.idx_begin(), EV.idx_end());
1118 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
1122 // The insert list is a prefix of the extract list
1123 // We can simply remove the common indices from the extract and make it
1124 // operate on the inserted value instead of the insertvalue result.
1126 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1127 // %E = extractvalue { i32, { i32 } } %I, 1, 0
1129 // %E extractvalue { i32 } { i32 42 }, 0
1130 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
1133 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
1134 // We're extracting from an intrinsic, see if we're the only user, which
1135 // allows us to simplify multiple result intrinsics to simpler things that
1136 // just get one value.
1137 if (II->hasOneUse()) {
1138 // Check if we're grabbing the overflow bit or the result of a 'with
1139 // overflow' intrinsic. If it's the latter we can remove the intrinsic
1140 // and replace it with a traditional binary instruction.
1141 switch (II->getIntrinsicID()) {
1142 case Intrinsic::uadd_with_overflow:
1143 case Intrinsic::sadd_with_overflow:
1144 if (*EV.idx_begin() == 0) { // Normal result.
1145 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1146 II->replaceAllUsesWith(UndefValue::get(II->getType()));
1147 EraseInstFromFunction(*II);
1148 return BinaryOperator::CreateAdd(LHS, RHS);
1151 case Intrinsic::usub_with_overflow:
1152 case Intrinsic::ssub_with_overflow:
1153 if (*EV.idx_begin() == 0) { // Normal result.
1154 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1155 II->replaceAllUsesWith(UndefValue::get(II->getType()));
1156 EraseInstFromFunction(*II);
1157 return BinaryOperator::CreateSub(LHS, RHS);
1160 case Intrinsic::umul_with_overflow:
1161 case Intrinsic::smul_with_overflow:
1162 if (*EV.idx_begin() == 0) { // Normal result.
1163 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1164 II->replaceAllUsesWith(UndefValue::get(II->getType()));
1165 EraseInstFromFunction(*II);
1166 return BinaryOperator::CreateMul(LHS, RHS);
1174 // Can't simplify extracts from other values. Note that nested extracts are
1175 // already simplified implicitely by the above (extract ( extract (insert) )
1176 // will be translated into extract ( insert ( extract ) ) first and then just
1177 // the value inserted, if appropriate).
1184 /// TryToSinkInstruction - Try to move the specified instruction from its
1185 /// current block into the beginning of DestBlock, which can only happen if it's
1186 /// safe to move the instruction past all of the instructions between it and the
1187 /// end of its block.
1188 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
1189 assert(I->hasOneUse() && "Invariants didn't hold!");
1191 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1192 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
1195 // Do not sink alloca instructions out of the entry block.
1196 if (isa<AllocaInst>(I) && I->getParent() ==
1197 &DestBlock->getParent()->getEntryBlock())
1200 // We can only sink load instructions if there is nothing between the load and
1201 // the end of block that could change the value.
1202 if (I->mayReadFromMemory()) {
1203 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
1205 if (Scan->mayWriteToMemory())
1209 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
1211 I->moveBefore(InsertPos);
1217 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
1218 /// all reachable code to the worklist.
1220 /// This has a couple of tricks to make the code faster and more powerful. In
1221 /// particular, we constant fold and DCE instructions as we go, to avoid adding
1222 /// them to the worklist (this significantly speeds up instcombine on code where
1223 /// many instructions are dead or constant). Additionally, if we find a branch
1224 /// whose condition is a known constant, we only visit the reachable successors.
1226 static bool AddReachableCodeToWorklist(BasicBlock *BB,
1227 SmallPtrSet<BasicBlock*, 64> &Visited,
1229 const TargetData *TD) {
1230 bool MadeIRChange = false;
1231 SmallVector<BasicBlock*, 256> Worklist;
1232 Worklist.push_back(BB);
1234 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
1235 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
1238 BB = Worklist.pop_back_val();
1240 // We have now visited this block! If we've already been here, ignore it.
1241 if (!Visited.insert(BB)) continue;
1243 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
1244 Instruction *Inst = BBI++;
1246 // DCE instruction if trivially dead.
1247 if (isInstructionTriviallyDead(Inst)) {
1249 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
1250 Inst->eraseFromParent();
1254 // ConstantProp instruction if trivially constant.
1255 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
1256 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
1257 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
1259 Inst->replaceAllUsesWith(C);
1261 Inst->eraseFromParent();
1266 // See if we can constant fold its operands.
1267 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
1269 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
1270 if (CE == 0) continue;
1272 // If we already folded this constant, don't try again.
1273 if (!FoldedConstants.insert(CE))
1276 Constant *NewC = ConstantFoldConstantExpression(CE, TD);
1277 if (NewC && NewC != CE) {
1279 MadeIRChange = true;
1284 InstrsForInstCombineWorklist.push_back(Inst);
1287 // Recursively visit successors. If this is a branch or switch on a
1288 // constant, only visit the reachable successor.
1289 TerminatorInst *TI = BB->getTerminator();
1290 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1291 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
1292 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
1293 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
1294 Worklist.push_back(ReachableBB);
1297 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1298 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
1299 // See if this is an explicit destination.
1300 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
1301 if (SI->getCaseValue(i) == Cond) {
1302 BasicBlock *ReachableBB = SI->getSuccessor(i);
1303 Worklist.push_back(ReachableBB);
1307 // Otherwise it is the default destination.
1308 Worklist.push_back(SI->getSuccessor(0));
1313 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
1314 Worklist.push_back(TI->getSuccessor(i));
1315 } while (!Worklist.empty());
1317 // Once we've found all of the instructions to add to instcombine's worklist,
1318 // add them in reverse order. This way instcombine will visit from the top
1319 // of the function down. This jives well with the way that it adds all uses
1320 // of instructions to the worklist after doing a transformation, thus avoiding
1321 // some N^2 behavior in pathological cases.
1322 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
1323 InstrsForInstCombineWorklist.size());
1325 return MadeIRChange;
1328 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
1329 MadeIRChange = false;
1331 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
1332 << F.getNameStr() << "\n");
1335 // Do a depth-first traversal of the function, populate the worklist with
1336 // the reachable instructions. Ignore blocks that are not reachable. Keep
1337 // track of which blocks we visit.
1338 SmallPtrSet<BasicBlock*, 64> Visited;
1339 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
1341 // Do a quick scan over the function. If we find any blocks that are
1342 // unreachable, remove any instructions inside of them. This prevents
1343 // the instcombine code from having to deal with some bad special cases.
1344 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1345 if (!Visited.count(BB)) {
1346 Instruction *Term = BB->getTerminator();
1347 while (Term != BB->begin()) { // Remove instrs bottom-up
1348 BasicBlock::iterator I = Term; --I;
1350 DEBUG(errs() << "IC: DCE: " << *I << '\n');
1351 // A debug intrinsic shouldn't force another iteration if we weren't
1352 // going to do one without it.
1353 if (!isa<DbgInfoIntrinsic>(I)) {
1355 MadeIRChange = true;
1358 // If I is not void type then replaceAllUsesWith undef.
1359 // This allows ValueHandlers and custom metadata to adjust itself.
1360 if (!I->getType()->isVoidTy())
1361 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1362 I->eraseFromParent();
1367 while (!Worklist.isEmpty()) {
1368 Instruction *I = Worklist.RemoveOne();
1369 if (I == 0) continue; // skip null values.
1371 // Check to see if we can DCE the instruction.
1372 if (isInstructionTriviallyDead(I)) {
1373 DEBUG(errs() << "IC: DCE: " << *I << '\n');
1374 EraseInstFromFunction(*I);
1376 MadeIRChange = true;
1380 // Instruction isn't dead, see if we can constant propagate it.
1381 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
1382 if (Constant *C = ConstantFoldInstruction(I, TD)) {
1383 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
1385 // Add operands to the worklist.
1386 ReplaceInstUsesWith(*I, C);
1388 EraseInstFromFunction(*I);
1389 MadeIRChange = true;
1393 // See if we can trivially sink this instruction to a successor basic block.
1394 if (I->hasOneUse()) {
1395 BasicBlock *BB = I->getParent();
1396 Instruction *UserInst = cast<Instruction>(I->use_back());
1397 BasicBlock *UserParent;
1399 // Get the block the use occurs in.
1400 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
1401 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
1403 UserParent = UserInst->getParent();
1405 if (UserParent != BB) {
1406 bool UserIsSuccessor = false;
1407 // See if the user is one of our successors.
1408 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
1409 if (*SI == UserParent) {
1410 UserIsSuccessor = true;
1414 // If the user is one of our immediate successors, and if that successor
1415 // only has us as a predecessors (we'd have to split the critical edge
1416 // otherwise), we can keep going.
1417 if (UserIsSuccessor && UserParent->getSinglePredecessor())
1418 // Okay, the CFG is simple enough, try to sink this instruction.
1419 MadeIRChange |= TryToSinkInstruction(I, UserParent);
1423 // Now that we have an instruction, try combining it to simplify it.
1424 Builder->SetInsertPoint(I->getParent(), I);
1429 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
1430 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
1432 if (Instruction *Result = visit(*I)) {
1434 // Should we replace the old instruction with a new one?
1436 DEBUG(errs() << "IC: Old = " << *I << '\n'
1437 << " New = " << *Result << '\n');
1439 // Everything uses the new instruction now.
1440 I->replaceAllUsesWith(Result);
1442 // Push the new instruction and any users onto the worklist.
1443 Worklist.Add(Result);
1444 Worklist.AddUsersToWorkList(*Result);
1446 // Move the name to the new instruction first.
1447 Result->takeName(I);
1449 // Insert the new instruction into the basic block...
1450 BasicBlock *InstParent = I->getParent();
1451 BasicBlock::iterator InsertPos = I;
1453 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
1454 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
1457 InstParent->getInstList().insert(InsertPos, Result);
1459 EraseInstFromFunction(*I);
1462 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
1463 << " New = " << *I << '\n');
1466 // If the instruction was modified, it's possible that it is now dead.
1467 // if so, remove it.
1468 if (isInstructionTriviallyDead(I)) {
1469 EraseInstFromFunction(*I);
1472 Worklist.AddUsersToWorkList(*I);
1475 MadeIRChange = true;
1480 return MadeIRChange;
1484 bool InstCombiner::runOnFunction(Function &F) {
1485 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1486 TD = getAnalysisIfAvailable<TargetData>();
1489 /// Builder - This is an IRBuilder that automatically inserts new
1490 /// instructions into the worklist when they are created.
1491 IRBuilder<true, TargetFolder, InstCombineIRInserter>
1492 TheBuilder(F.getContext(), TargetFolder(TD),
1493 InstCombineIRInserter(Worklist));
1494 Builder = &TheBuilder;
1496 bool EverMadeChange = false;
1498 // Iterate while there is work to do.
1499 unsigned Iteration = 0;
1500 while (DoOneIteration(F, Iteration++))
1501 EverMadeChange = true;
1504 return EverMadeChange;
1507 FunctionPass *llvm::createInstructionCombiningPass() {
1508 return new InstCombiner();