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");
61 STATISTIC(NumExpand, "Number of expansions");
62 STATISTIC(NumFactor , "Number of factorizations");
63 STATISTIC(NumReassoc , "Number of reassociations");
65 // Initialization Routines
66 void llvm::initializeInstCombine(PassRegistry &Registry) {
67 initializeInstCombinerPass(Registry);
70 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
71 initializeInstCombine(*unwrap(R));
74 char InstCombiner::ID = 0;
75 INITIALIZE_PASS(InstCombiner, "instcombine",
76 "Combine redundant instructions", false, false)
78 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
79 AU.addPreservedID(LCSSAID);
84 /// ShouldChangeType - Return true if it is desirable to convert a computation
85 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
86 /// type for example, or from a smaller to a larger illegal type.
87 bool InstCombiner::ShouldChangeType(const Type *From, const Type *To) const {
88 assert(From->isIntegerTy() && To->isIntegerTy());
90 // If we don't have TD, we don't know if the source/dest are legal.
91 if (!TD) return false;
93 unsigned FromWidth = From->getPrimitiveSizeInBits();
94 unsigned ToWidth = To->getPrimitiveSizeInBits();
95 bool FromLegal = TD->isLegalInteger(FromWidth);
96 bool ToLegal = TD->isLegalInteger(ToWidth);
98 // If this is a legal integer from type, and the result would be an illegal
99 // type, don't do the transformation.
100 if (FromLegal && !ToLegal)
103 // Otherwise, if both are illegal, do not increase the size of the result. We
104 // do allow things like i160 -> i64, but not i64 -> i160.
105 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
112 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
113 /// operators which are associative or commutative:
115 // Commutative operators:
117 // 1. Order operands such that they are listed from right (least complex) to
118 // left (most complex). This puts constants before unary operators before
121 // Associative operators:
123 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
124 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
126 // Associative and commutative operators:
128 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
129 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
130 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
131 // if C1 and C2 are constants.
133 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
134 Instruction::BinaryOps Opcode = I.getOpcode();
135 bool Changed = false;
138 // Order operands such that they are listed from right (least complex) to
139 // left (most complex). This puts constants before unary operators before
141 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
142 getComplexity(I.getOperand(1)))
143 Changed = !I.swapOperands();
145 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
146 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
148 if (I.isAssociative()) {
149 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
150 if (Op0 && Op0->getOpcode() == Opcode) {
151 Value *A = Op0->getOperand(0);
152 Value *B = Op0->getOperand(1);
153 Value *C = I.getOperand(1);
155 // Does "B op C" simplify?
156 if (Value *V = SimplifyBinOp(Opcode, B, C, TD)) {
157 // It simplifies to V. Form "A op V".
160 // Conservatively clear the optional flags, since they may not be
161 // preserved by the reassociation.
162 I.clearSubclassOptionalData();
169 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
170 if (Op1 && Op1->getOpcode() == Opcode) {
171 Value *A = I.getOperand(0);
172 Value *B = Op1->getOperand(0);
173 Value *C = Op1->getOperand(1);
175 // Does "A op B" simplify?
176 if (Value *V = SimplifyBinOp(Opcode, A, B, TD)) {
177 // It simplifies to V. Form "V op C".
180 // Conservatively clear the optional flags, since they may not be
181 // preserved by the reassociation.
182 I.clearSubclassOptionalData();
190 if (I.isAssociative() && I.isCommutative()) {
191 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
192 if (Op0 && Op0->getOpcode() == Opcode) {
193 Value *A = Op0->getOperand(0);
194 Value *B = Op0->getOperand(1);
195 Value *C = I.getOperand(1);
197 // Does "C op A" simplify?
198 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
199 // It simplifies to V. Form "V op B".
202 // Conservatively clear the optional flags, since they may not be
203 // preserved by the reassociation.
204 I.clearSubclassOptionalData();
211 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
212 if (Op1 && Op1->getOpcode() == Opcode) {
213 Value *A = I.getOperand(0);
214 Value *B = Op1->getOperand(0);
215 Value *C = Op1->getOperand(1);
217 // Does "C op A" simplify?
218 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
219 // It simplifies to V. Form "B op V".
222 // Conservatively clear the optional flags, since they may not be
223 // preserved by the reassociation.
224 I.clearSubclassOptionalData();
231 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
232 // if C1 and C2 are constants.
234 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
235 isa<Constant>(Op0->getOperand(1)) &&
236 isa<Constant>(Op1->getOperand(1)) &&
237 Op0->hasOneUse() && Op1->hasOneUse()) {
238 Value *A = Op0->getOperand(0);
239 Constant *C1 = cast<Constant>(Op0->getOperand(1));
240 Value *B = Op1->getOperand(0);
241 Constant *C2 = cast<Constant>(Op1->getOperand(1));
243 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
244 Instruction *New = BinaryOperator::Create(Opcode, A, B, Op1->getName(),
247 I.setOperand(0, New);
248 I.setOperand(1, Folded);
249 // Conservatively clear the optional flags, since they may not be
250 // preserved by the reassociation.
251 I.clearSubclassOptionalData();
257 // No further simplifications.
262 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
263 /// "(X LOp Y) ROp (X LOp Z)".
264 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
265 Instruction::BinaryOps ROp) {
270 case Instruction::And:
271 // And distributes over Or and Xor.
275 case Instruction::Or:
276 case Instruction::Xor:
280 case Instruction::Mul:
281 // Multiplication distributes over addition and subtraction.
285 case Instruction::Add:
286 case Instruction::Sub:
290 case Instruction::Or:
291 // Or distributes over And.
295 case Instruction::And:
301 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
302 /// "(X ROp Z) LOp (Y ROp Z)".
303 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
304 Instruction::BinaryOps ROp) {
305 if (Instruction::isCommutative(ROp))
306 return LeftDistributesOverRight(ROp, LOp);
307 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
308 // but this requires knowing that the addition does not overflow and other
313 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
314 /// which some other binary operation distributes over either by factorizing
315 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
316 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
317 /// a win). Returns the simplified value, or null if it didn't simplify.
318 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
319 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
320 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
321 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
322 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); // op
325 if (Op0 && Op1 && Op0->getOpcode() == Op1->getOpcode()) {
326 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
328 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1);
329 Value *C = Op1->getOperand(0), *D = Op1->getOperand(1);
330 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
332 // Does "X op' Y" always equal "Y op' X"?
333 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
335 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
336 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
337 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
338 // commutative case, "(A op' B) op (C op' A)"?
339 if (A == C || (InnerCommutative && A == D)) {
342 // Consider forming "A op' (B op D)".
343 // If "B op D" simplifies then it can be formed with no cost.
344 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, TD);
345 // If "B op D" doesn't simplify then only go on if both of the existing
346 // operations "A op' B" and "C op' D" will be zapped as no longer used.
347 if (!V && Op0->hasOneUse() && Op1->hasOneUse())
348 V = Builder->CreateBinOp(TopLevelOpcode, B, D, Op1->getName());
351 V = Builder->CreateBinOp(InnerOpcode, A, V);
357 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
358 if (RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
359 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
360 // commutative case, "(A op' B) op (B op' D)"?
361 if (B == D || (InnerCommutative && B == C)) {
364 // Consider forming "(A op C) op' B".
365 // If "A op C" simplifies then it can be formed with no cost.
366 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, TD);
367 // If "A op C" doesn't simplify then only go on if both of the existing
368 // operations "A op' B" and "C op' D" will be zapped as no longer used.
369 if (!V && Op0->hasOneUse() && Op1->hasOneUse())
370 V = Builder->CreateBinOp(TopLevelOpcode, A, C, Op0->getName());
373 V = Builder->CreateBinOp(InnerOpcode, V, B);
381 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
382 // The instruction has the form "(A op' B) op C". See if expanding it out
383 // to "(A op C) op' (B op C)" results in simplifications.
384 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
385 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
387 // Do "A op C" and "B op C" both simplify?
388 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, TD))
389 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, TD)) {
390 // They do! Return "L op' R".
392 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
393 if ((L == A && R == B) ||
394 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
396 // Otherwise return "L op' R" if it simplifies.
397 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
399 // Otherwise, create a new instruction.
400 C = Builder->CreateBinOp(InnerOpcode, L, R);
406 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
407 // The instruction has the form "A op (B op' C)". See if expanding it out
408 // to "(A op B) op' (A op C)" results in simplifications.
409 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
410 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
412 // Do "A op B" and "A op C" both simplify?
413 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, TD))
414 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, TD)) {
415 // They do! Return "L op' R".
417 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
418 if ((L == B && R == C) ||
419 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
421 // Otherwise return "L op' R" if it simplifies.
422 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
424 // Otherwise, create a new instruction.
425 A = Builder->CreateBinOp(InnerOpcode, L, R);
434 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
435 // if the LHS is a constant zero (which is the 'negate' form).
437 Value *InstCombiner::dyn_castNegVal(Value *V) const {
438 if (BinaryOperator::isNeg(V))
439 return BinaryOperator::getNegArgument(V);
441 // Constants can be considered to be negated values if they can be folded.
442 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
443 return ConstantExpr::getNeg(C);
445 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
446 if (C->getType()->getElementType()->isIntegerTy())
447 return ConstantExpr::getNeg(C);
452 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
453 // instruction if the LHS is a constant negative zero (which is the 'negate'
456 Value *InstCombiner::dyn_castFNegVal(Value *V) const {
457 if (BinaryOperator::isFNeg(V))
458 return BinaryOperator::getFNegArgument(V);
460 // Constants can be considered to be negated values if they can be folded.
461 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
462 return ConstantExpr::getFNeg(C);
464 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
465 if (C->getType()->getElementType()->isFloatingPointTy())
466 return ConstantExpr::getFNeg(C);
471 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
473 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
474 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
477 // Figure out if the constant is the left or the right argument.
478 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
479 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
481 if (Constant *SOC = dyn_cast<Constant>(SO)) {
483 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
484 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
487 Value *Op0 = SO, *Op1 = ConstOperand;
491 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
492 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
493 SO->getName()+".op");
494 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
495 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
496 SO->getName()+".cmp");
497 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
498 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
499 SO->getName()+".cmp");
500 llvm_unreachable("Unknown binary instruction type!");
503 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
504 // constant as the other operand, try to fold the binary operator into the
505 // select arguments. This also works for Cast instructions, which obviously do
506 // not have a second operand.
507 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
508 // Don't modify shared select instructions
509 if (!SI->hasOneUse()) return 0;
510 Value *TV = SI->getOperand(1);
511 Value *FV = SI->getOperand(2);
513 if (isa<Constant>(TV) || isa<Constant>(FV)) {
514 // Bool selects with constant operands can be folded to logical ops.
515 if (SI->getType()->isIntegerTy(1)) return 0;
517 // If it's a bitcast involving vectors, make sure it has the same number of
518 // elements on both sides.
519 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
520 const VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
521 const VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
523 // Verify that either both or neither are vectors.
524 if ((SrcTy == NULL) != (DestTy == NULL)) return 0;
525 // If vectors, verify that they have the same number of elements.
526 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
530 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
531 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
533 return SelectInst::Create(SI->getCondition(),
534 SelectTrueVal, SelectFalseVal);
540 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
541 /// has a PHI node as operand #0, see if we can fold the instruction into the
542 /// PHI (which is only possible if all operands to the PHI are constants).
544 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
545 PHINode *PN = cast<PHINode>(I.getOperand(0));
546 unsigned NumPHIValues = PN->getNumIncomingValues();
547 if (NumPHIValues == 0)
550 // We normally only transform phis with a single use. However, if a PHI has
551 // multiple uses and they are all the same operation, we can fold *all* of the
552 // uses into the PHI.
553 if (!PN->hasOneUse()) {
554 // Walk the use list for the instruction, comparing them to I.
555 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
557 Instruction *User = cast<Instruction>(*UI);
558 if (User != &I && !I.isIdenticalTo(User))
561 // Otherwise, we can replace *all* users with the new PHI we form.
564 // Check to see if all of the operands of the PHI are simple constants
565 // (constantint/constantfp/undef). If there is one non-constant value,
566 // remember the BB it is in. If there is more than one or if *it* is a PHI,
567 // bail out. We don't do arbitrary constant expressions here because moving
568 // their computation can be expensive without a cost model.
569 BasicBlock *NonConstBB = 0;
570 for (unsigned i = 0; i != NumPHIValues; ++i) {
571 Value *InVal = PN->getIncomingValue(i);
572 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
575 if (isa<PHINode>(InVal)) return 0; // Itself a phi.
576 if (NonConstBB) return 0; // More than one non-const value.
578 NonConstBB = PN->getIncomingBlock(i);
580 // If the InVal is an invoke at the end of the pred block, then we can't
581 // insert a computation after it without breaking the edge.
582 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
583 if (II->getParent() == NonConstBB)
586 // If the incoming non-constant value is in I's block, we will remove one
587 // instruction, but insert another equivalent one, leading to infinite
589 if (NonConstBB == I.getParent())
593 // If there is exactly one non-constant value, we can insert a copy of the
594 // operation in that block. However, if this is a critical edge, we would be
595 // inserting the computation one some other paths (e.g. inside a loop). Only
596 // do this if the pred block is unconditionally branching into the phi block.
597 if (NonConstBB != 0) {
598 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
599 if (!BI || !BI->isUnconditional()) return 0;
602 // Okay, we can do the transformation: create the new PHI node.
603 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues(), "");
604 InsertNewInstBefore(NewPN, *PN);
607 // If we are going to have to insert a new computation, do so right before the
608 // predecessors terminator.
610 Builder->SetInsertPoint(NonConstBB->getTerminator());
612 // Next, add all of the operands to the PHI.
613 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
614 // We only currently try to fold the condition of a select when it is a phi,
615 // not the true/false values.
616 Value *TrueV = SI->getTrueValue();
617 Value *FalseV = SI->getFalseValue();
618 BasicBlock *PhiTransBB = PN->getParent();
619 for (unsigned i = 0; i != NumPHIValues; ++i) {
620 BasicBlock *ThisBB = PN->getIncomingBlock(i);
621 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
622 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
624 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
625 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
627 InV = Builder->CreateSelect(PN->getIncomingValue(i),
628 TrueVInPred, FalseVInPred, "phitmp");
629 NewPN->addIncoming(InV, ThisBB);
631 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
632 Constant *C = cast<Constant>(I.getOperand(1));
633 for (unsigned i = 0; i != NumPHIValues; ++i) {
635 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
636 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
637 else if (isa<ICmpInst>(CI))
638 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
641 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
643 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
645 } else if (I.getNumOperands() == 2) {
646 Constant *C = cast<Constant>(I.getOperand(1));
647 for (unsigned i = 0; i != NumPHIValues; ++i) {
649 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
650 InV = ConstantExpr::get(I.getOpcode(), InC, C);
652 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
653 PN->getIncomingValue(i), C, "phitmp");
654 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
657 CastInst *CI = cast<CastInst>(&I);
658 const Type *RetTy = CI->getType();
659 for (unsigned i = 0; i != NumPHIValues; ++i) {
661 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
662 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
664 InV = Builder->CreateCast(CI->getOpcode(),
665 PN->getIncomingValue(i), I.getType(), "phitmp");
666 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
670 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
672 Instruction *User = cast<Instruction>(*UI++);
673 if (User == &I) continue;
674 ReplaceInstUsesWith(*User, NewPN);
675 EraseInstFromFunction(*User);
677 return ReplaceInstUsesWith(I, NewPN);
680 /// FindElementAtOffset - Given a type and a constant offset, determine whether
681 /// or not there is a sequence of GEP indices into the type that will land us at
682 /// the specified offset. If so, fill them into NewIndices and return the
683 /// resultant element type, otherwise return null.
684 const Type *InstCombiner::FindElementAtOffset(const Type *Ty, int64_t Offset,
685 SmallVectorImpl<Value*> &NewIndices) {
687 if (!Ty->isSized()) return 0;
689 // Start with the index over the outer type. Note that the type size
690 // might be zero (even if the offset isn't zero) if the indexed type
691 // is something like [0 x {int, int}]
692 const Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
693 int64_t FirstIdx = 0;
694 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
695 FirstIdx = Offset/TySize;
696 Offset -= FirstIdx*TySize;
698 // Handle hosts where % returns negative instead of values [0..TySize).
704 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
707 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
709 // Index into the types. If we fail, set OrigBase to null.
711 // Indexing into tail padding between struct/array elements.
712 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
715 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
716 const StructLayout *SL = TD->getStructLayout(STy);
717 assert(Offset < (int64_t)SL->getSizeInBytes() &&
718 "Offset must stay within the indexed type");
720 unsigned Elt = SL->getElementContainingOffset(Offset);
721 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
724 Offset -= SL->getElementOffset(Elt);
725 Ty = STy->getElementType(Elt);
726 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
727 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
728 assert(EltSize && "Cannot index into a zero-sized array");
729 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
731 Ty = AT->getElementType();
733 // Otherwise, we can't index into the middle of this atomic type, bail.
743 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
744 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
746 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD))
747 return ReplaceInstUsesWith(GEP, V);
749 Value *PtrOp = GEP.getOperand(0);
751 // Eliminate unneeded casts for indices, and replace indices which displace
752 // by multiples of a zero size type with zero.
754 bool MadeChange = false;
755 const Type *IntPtrTy = TD->getIntPtrType(GEP.getContext());
757 gep_type_iterator GTI = gep_type_begin(GEP);
758 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
759 I != E; ++I, ++GTI) {
760 // Skip indices into struct types.
761 const SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
762 if (!SeqTy) continue;
764 // If the element type has zero size then any index over it is equivalent
765 // to an index of zero, so replace it with zero if it is not zero already.
766 if (SeqTy->getElementType()->isSized() &&
767 TD->getTypeAllocSize(SeqTy->getElementType()) == 0)
768 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
769 *I = Constant::getNullValue(IntPtrTy);
773 if ((*I)->getType() != IntPtrTy) {
774 // If we are using a wider index than needed for this platform, shrink
775 // it to what we need. If narrower, sign-extend it to what we need.
776 // This explicit cast can make subsequent optimizations more obvious.
777 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
781 if (MadeChange) return &GEP;
784 // Combine Indices - If the source pointer to this getelementptr instruction
785 // is a getelementptr instruction, combine the indices of the two
786 // getelementptr instructions into a single instruction.
788 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
789 // Note that if our source is a gep chain itself that we wait for that
790 // chain to be resolved before we perform this transformation. This
791 // avoids us creating a TON of code in some cases.
793 if (GetElementPtrInst *SrcGEP =
794 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
795 if (SrcGEP->getNumOperands() == 2)
796 return 0; // Wait until our source is folded to completion.
798 SmallVector<Value*, 8> Indices;
800 // Find out whether the last index in the source GEP is a sequential idx.
801 bool EndsWithSequential = false;
802 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
804 EndsWithSequential = !(*I)->isStructTy();
806 // Can we combine the two pointer arithmetics offsets?
807 if (EndsWithSequential) {
808 // Replace: gep (gep %P, long B), long A, ...
809 // With: T = long A+B; gep %P, T, ...
812 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
813 Value *GO1 = GEP.getOperand(1);
814 if (SO1 == Constant::getNullValue(SO1->getType())) {
816 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
819 // If they aren't the same type, then the input hasn't been processed
820 // by the loop above yet (which canonicalizes sequential index types to
821 // intptr_t). Just avoid transforming this until the input has been
823 if (SO1->getType() != GO1->getType())
825 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
828 // Update the GEP in place if possible.
829 if (Src->getNumOperands() == 2) {
830 GEP.setOperand(0, Src->getOperand(0));
831 GEP.setOperand(1, Sum);
834 Indices.append(Src->op_begin()+1, Src->op_end()-1);
835 Indices.push_back(Sum);
836 Indices.append(GEP.op_begin()+2, GEP.op_end());
837 } else if (isa<Constant>(*GEP.idx_begin()) &&
838 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
839 Src->getNumOperands() != 1) {
840 // Otherwise we can do the fold if the first index of the GEP is a zero
841 Indices.append(Src->op_begin()+1, Src->op_end());
842 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
845 if (!Indices.empty())
846 return (GEP.isInBounds() && Src->isInBounds()) ?
847 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
848 Indices.end(), GEP.getName()) :
849 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
850 Indices.end(), GEP.getName());
853 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
854 Value *StrippedPtr = PtrOp->stripPointerCasts();
855 const PointerType *StrippedPtrTy =cast<PointerType>(StrippedPtr->getType());
856 if (StrippedPtr != PtrOp &&
857 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
859 bool HasZeroPointerIndex = false;
860 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
861 HasZeroPointerIndex = C->isZero();
863 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
864 // into : GEP [10 x i8]* X, i32 0, ...
866 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
867 // into : GEP i8* X, ...
869 // This occurs when the program declares an array extern like "int X[];"
870 if (HasZeroPointerIndex) {
871 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
872 if (const ArrayType *CATy =
873 dyn_cast<ArrayType>(CPTy->getElementType())) {
874 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
875 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
877 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
878 GetElementPtrInst *Res =
879 GetElementPtrInst::Create(StrippedPtr, Idx.begin(),
880 Idx.end(), GEP.getName());
881 Res->setIsInBounds(GEP.isInBounds());
885 if (const ArrayType *XATy =
886 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
887 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
888 if (CATy->getElementType() == XATy->getElementType()) {
889 // -> GEP [10 x i8]* X, i32 0, ...
890 // At this point, we know that the cast source type is a pointer
891 // to an array of the same type as the destination pointer
892 // array. Because the array type is never stepped over (there
893 // is a leading zero) we can fold the cast into this GEP.
894 GEP.setOperand(0, StrippedPtr);
899 } else if (GEP.getNumOperands() == 2) {
900 // Transform things like:
901 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
902 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
903 const Type *SrcElTy = StrippedPtrTy->getElementType();
904 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
905 if (TD && SrcElTy->isArrayTy() &&
906 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
907 TD->getTypeAllocSize(ResElTy)) {
909 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
910 Idx[1] = GEP.getOperand(1);
911 Value *NewGEP = GEP.isInBounds() ?
912 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()) :
913 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName());
914 // V and GEP are both pointer types --> BitCast
915 return new BitCastInst(NewGEP, GEP.getType());
918 // Transform things like:
919 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
920 // (where tmp = 8*tmp2) into:
921 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
923 if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) {
924 uint64_t ArrayEltSize =
925 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
927 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
928 // allow either a mul, shift, or constant here.
930 ConstantInt *Scale = 0;
931 if (ArrayEltSize == 1) {
932 NewIdx = GEP.getOperand(1);
933 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
934 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
935 NewIdx = ConstantInt::get(CI->getType(), 1);
937 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
938 if (Inst->getOpcode() == Instruction::Shl &&
939 isa<ConstantInt>(Inst->getOperand(1))) {
940 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
941 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
942 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
944 NewIdx = Inst->getOperand(0);
945 } else if (Inst->getOpcode() == Instruction::Mul &&
946 isa<ConstantInt>(Inst->getOperand(1))) {
947 Scale = cast<ConstantInt>(Inst->getOperand(1));
948 NewIdx = Inst->getOperand(0);
952 // If the index will be to exactly the right offset with the scale taken
953 // out, perform the transformation. Note, we don't know whether Scale is
954 // signed or not. We'll use unsigned version of division/modulo
955 // operation after making sure Scale doesn't have the sign bit set.
956 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
957 Scale->getZExtValue() % ArrayEltSize == 0) {
958 Scale = ConstantInt::get(Scale->getType(),
959 Scale->getZExtValue() / ArrayEltSize);
960 if (Scale->getZExtValue() != 1) {
961 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
963 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
966 // Insert the new GEP instruction.
968 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
970 Value *NewGEP = GEP.isInBounds() ?
971 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2,GEP.getName()):
972 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName());
973 // The NewGEP must be pointer typed, so must the old one -> BitCast
974 return new BitCastInst(NewGEP, GEP.getType());
980 /// See if we can simplify:
981 /// X = bitcast A* to B*
982 /// Y = gep X, <...constant indices...>
983 /// into a gep of the original struct. This is important for SROA and alias
984 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
985 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
987 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices() &&
988 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
990 // Determine how much the GEP moves the pointer. We are guaranteed to get
991 // a constant back from EmitGEPOffset.
992 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP));
993 int64_t Offset = OffsetV->getSExtValue();
995 // If this GEP instruction doesn't move the pointer, just replace the GEP
996 // with a bitcast of the real input to the dest type.
998 // If the bitcast is of an allocation, and the allocation will be
999 // converted to match the type of the cast, don't touch this.
1000 if (isa<AllocaInst>(BCI->getOperand(0)) ||
1001 isMalloc(BCI->getOperand(0))) {
1002 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1003 if (Instruction *I = visitBitCast(*BCI)) {
1006 BCI->getParent()->getInstList().insert(BCI, I);
1007 ReplaceInstUsesWith(*BCI, I);
1012 return new BitCastInst(BCI->getOperand(0), GEP.getType());
1015 // Otherwise, if the offset is non-zero, we need to find out if there is a
1016 // field at Offset in 'A's type. If so, we can pull the cast through the
1018 SmallVector<Value*, 8> NewIndices;
1020 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
1021 if (FindElementAtOffset(InTy, Offset, NewIndices)) {
1022 Value *NGEP = GEP.isInBounds() ?
1023 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
1025 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
1028 if (NGEP->getType() == GEP.getType())
1029 return ReplaceInstUsesWith(GEP, NGEP);
1030 NGEP->takeName(&GEP);
1031 return new BitCastInst(NGEP, GEP.getType());
1041 static bool IsOnlyNullComparedAndFreed(const Value &V) {
1042 for (Value::const_use_iterator UI = V.use_begin(), UE = V.use_end();
1044 const User *U = *UI;
1047 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(U))
1048 if (ICI->isEquality() && isa<ConstantPointerNull>(ICI->getOperand(1)))
1055 Instruction *InstCombiner::visitMalloc(Instruction &MI) {
1056 // If we have a malloc call which is only used in any amount of comparisons
1057 // to null and free calls, delete the calls and replace the comparisons with
1058 // true or false as appropriate.
1059 if (IsOnlyNullComparedAndFreed(MI)) {
1060 for (Value::use_iterator UI = MI.use_begin(), UE = MI.use_end();
1062 // We can assume that every remaining use is a free call or an icmp eq/ne
1063 // to null, so the cast is safe.
1064 Instruction *I = cast<Instruction>(*UI);
1066 // Early increment here, as we're about to get rid of the user.
1069 if (isFreeCall(I)) {
1070 EraseInstFromFunction(*cast<CallInst>(I));
1073 // Again, the cast is safe.
1074 ICmpInst *C = cast<ICmpInst>(I);
1075 ReplaceInstUsesWith(*C, ConstantInt::get(Type::getInt1Ty(C->getContext()),
1076 C->isFalseWhenEqual()));
1077 EraseInstFromFunction(*C);
1079 return EraseInstFromFunction(MI);
1086 Instruction *InstCombiner::visitFree(CallInst &FI) {
1087 Value *Op = FI.getArgOperand(0);
1089 // free undef -> unreachable.
1090 if (isa<UndefValue>(Op)) {
1091 // Insert a new store to null because we cannot modify the CFG here.
1092 new StoreInst(ConstantInt::getTrue(FI.getContext()),
1093 UndefValue::get(Type::getInt1PtrTy(FI.getContext())), &FI);
1094 return EraseInstFromFunction(FI);
1097 // If we have 'free null' delete the instruction. This can happen in stl code
1098 // when lots of inlining happens.
1099 if (isa<ConstantPointerNull>(Op))
1100 return EraseInstFromFunction(FI);
1107 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
1108 // Change br (not X), label True, label False to: br X, label False, True
1110 BasicBlock *TrueDest;
1111 BasicBlock *FalseDest;
1112 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
1113 !isa<Constant>(X)) {
1114 // Swap Destinations and condition...
1116 BI.setSuccessor(0, FalseDest);
1117 BI.setSuccessor(1, TrueDest);
1121 // Cannonicalize fcmp_one -> fcmp_oeq
1122 FCmpInst::Predicate FPred; Value *Y;
1123 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
1124 TrueDest, FalseDest)) &&
1125 BI.getCondition()->hasOneUse())
1126 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
1127 FPred == FCmpInst::FCMP_OGE) {
1128 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
1129 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
1131 // Swap Destinations and condition.
1132 BI.setSuccessor(0, FalseDest);
1133 BI.setSuccessor(1, TrueDest);
1138 // Cannonicalize icmp_ne -> icmp_eq
1139 ICmpInst::Predicate IPred;
1140 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
1141 TrueDest, FalseDest)) &&
1142 BI.getCondition()->hasOneUse())
1143 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
1144 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
1145 IPred == ICmpInst::ICMP_SGE) {
1146 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
1147 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
1148 // Swap Destinations and condition.
1149 BI.setSuccessor(0, FalseDest);
1150 BI.setSuccessor(1, TrueDest);
1158 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
1159 Value *Cond = SI.getCondition();
1160 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
1161 if (I->getOpcode() == Instruction::Add)
1162 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1163 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
1164 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
1166 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
1168 SI.setOperand(0, I->getOperand(0));
1176 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
1177 Value *Agg = EV.getAggregateOperand();
1179 if (!EV.hasIndices())
1180 return ReplaceInstUsesWith(EV, Agg);
1182 if (Constant *C = dyn_cast<Constant>(Agg)) {
1183 if (isa<UndefValue>(C))
1184 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
1186 if (isa<ConstantAggregateZero>(C))
1187 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
1189 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
1190 // Extract the element indexed by the first index out of the constant
1191 Value *V = C->getOperand(*EV.idx_begin());
1192 if (EV.getNumIndices() > 1)
1193 // Extract the remaining indices out of the constant indexed by the
1195 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
1197 return ReplaceInstUsesWith(EV, V);
1199 return 0; // Can't handle other constants
1201 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
1202 // We're extracting from an insertvalue instruction, compare the indices
1203 const unsigned *exti, *exte, *insi, *inse;
1204 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
1205 exte = EV.idx_end(), inse = IV->idx_end();
1206 exti != exte && insi != inse;
1209 // The insert and extract both reference distinctly different elements.
1210 // This means the extract is not influenced by the insert, and we can
1211 // replace the aggregate operand of the extract with the aggregate
1212 // operand of the insert. i.e., replace
1213 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1214 // %E = extractvalue { i32, { i32 } } %I, 0
1216 // %E = extractvalue { i32, { i32 } } %A, 0
1217 return ExtractValueInst::Create(IV->getAggregateOperand(),
1218 EV.idx_begin(), EV.idx_end());
1220 if (exti == exte && insi == inse)
1221 // Both iterators are at the end: Index lists are identical. Replace
1222 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1223 // %C = extractvalue { i32, { i32 } } %B, 1, 0
1225 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
1227 // The extract list is a prefix of the insert list. i.e. replace
1228 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1229 // %E = extractvalue { i32, { i32 } } %I, 1
1231 // %X = extractvalue { i32, { i32 } } %A, 1
1232 // %E = insertvalue { i32 } %X, i32 42, 0
1233 // by switching the order of the insert and extract (though the
1234 // insertvalue should be left in, since it may have other uses).
1235 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
1236 EV.idx_begin(), EV.idx_end());
1237 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
1241 // The insert list is a prefix of the extract list
1242 // We can simply remove the common indices from the extract and make it
1243 // operate on the inserted value instead of the insertvalue result.
1245 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1246 // %E = extractvalue { i32, { i32 } } %I, 1, 0
1248 // %E extractvalue { i32 } { i32 42 }, 0
1249 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
1252 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
1253 // We're extracting from an intrinsic, see if we're the only user, which
1254 // allows us to simplify multiple result intrinsics to simpler things that
1255 // just get one value.
1256 if (II->hasOneUse()) {
1257 // Check if we're grabbing the overflow bit or the result of a 'with
1258 // overflow' intrinsic. If it's the latter we can remove the intrinsic
1259 // and replace it with a traditional binary instruction.
1260 switch (II->getIntrinsicID()) {
1261 case Intrinsic::uadd_with_overflow:
1262 case Intrinsic::sadd_with_overflow:
1263 if (*EV.idx_begin() == 0) { // Normal result.
1264 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1265 II->replaceAllUsesWith(UndefValue::get(II->getType()));
1266 EraseInstFromFunction(*II);
1267 return BinaryOperator::CreateAdd(LHS, RHS);
1270 // If the normal result of the add is dead, and the RHS is a constant,
1271 // we can transform this into a range comparison.
1272 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
1273 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
1274 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
1275 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
1276 ConstantExpr::getNot(CI));
1278 case Intrinsic::usub_with_overflow:
1279 case Intrinsic::ssub_with_overflow:
1280 if (*EV.idx_begin() == 0) { // Normal result.
1281 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1282 II->replaceAllUsesWith(UndefValue::get(II->getType()));
1283 EraseInstFromFunction(*II);
1284 return BinaryOperator::CreateSub(LHS, RHS);
1287 case Intrinsic::umul_with_overflow:
1288 case Intrinsic::smul_with_overflow:
1289 if (*EV.idx_begin() == 0) { // Normal result.
1290 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1291 II->replaceAllUsesWith(UndefValue::get(II->getType()));
1292 EraseInstFromFunction(*II);
1293 return BinaryOperator::CreateMul(LHS, RHS);
1301 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
1302 // If the (non-volatile) load only has one use, we can rewrite this to a
1303 // load from a GEP. This reduces the size of the load.
1304 // FIXME: If a load is used only by extractvalue instructions then this
1305 // could be done regardless of having multiple uses.
1306 if (!L->isVolatile() && L->hasOneUse()) {
1307 // extractvalue has integer indices, getelementptr has Value*s. Convert.
1308 SmallVector<Value*, 4> Indices;
1309 // Prefix an i32 0 since we need the first element.
1310 Indices.push_back(Builder->getInt32(0));
1311 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
1313 Indices.push_back(Builder->getInt32(*I));
1315 // We need to insert these at the location of the old load, not at that of
1316 // the extractvalue.
1317 Builder->SetInsertPoint(L->getParent(), L);
1318 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(),
1319 Indices.begin(), Indices.end());
1320 // Returning the load directly will cause the main loop to insert it in
1321 // the wrong spot, so use ReplaceInstUsesWith().
1322 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
1324 // We could simplify extracts from other values. Note that nested extracts may
1325 // already be simplified implicitly by the above: extract (extract (insert) )
1326 // will be translated into extract ( insert ( extract ) ) first and then just
1327 // the value inserted, if appropriate. Similarly for extracts from single-use
1328 // loads: extract (extract (load)) will be translated to extract (load (gep))
1329 // and if again single-use then via load (gep (gep)) to load (gep).
1330 // However, double extracts from e.g. function arguments or return values
1331 // aren't handled yet.
1338 /// TryToSinkInstruction - Try to move the specified instruction from its
1339 /// current block into the beginning of DestBlock, which can only happen if it's
1340 /// safe to move the instruction past all of the instructions between it and the
1341 /// end of its block.
1342 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
1343 assert(I->hasOneUse() && "Invariants didn't hold!");
1345 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1346 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
1349 // Do not sink alloca instructions out of the entry block.
1350 if (isa<AllocaInst>(I) && I->getParent() ==
1351 &DestBlock->getParent()->getEntryBlock())
1354 // We can only sink load instructions if there is nothing between the load and
1355 // the end of block that could change the value.
1356 if (I->mayReadFromMemory()) {
1357 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
1359 if (Scan->mayWriteToMemory())
1363 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
1365 I->moveBefore(InsertPos);
1371 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
1372 /// all reachable code to the worklist.
1374 /// This has a couple of tricks to make the code faster and more powerful. In
1375 /// particular, we constant fold and DCE instructions as we go, to avoid adding
1376 /// them to the worklist (this significantly speeds up instcombine on code where
1377 /// many instructions are dead or constant). Additionally, if we find a branch
1378 /// whose condition is a known constant, we only visit the reachable successors.
1380 static bool AddReachableCodeToWorklist(BasicBlock *BB,
1381 SmallPtrSet<BasicBlock*, 64> &Visited,
1383 const TargetData *TD) {
1384 bool MadeIRChange = false;
1385 SmallVector<BasicBlock*, 256> Worklist;
1386 Worklist.push_back(BB);
1388 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
1389 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
1392 BB = Worklist.pop_back_val();
1394 // We have now visited this block! If we've already been here, ignore it.
1395 if (!Visited.insert(BB)) continue;
1397 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
1398 Instruction *Inst = BBI++;
1400 // DCE instruction if trivially dead.
1401 if (isInstructionTriviallyDead(Inst)) {
1403 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
1404 Inst->eraseFromParent();
1408 // ConstantProp instruction if trivially constant.
1409 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
1410 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
1411 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
1413 Inst->replaceAllUsesWith(C);
1415 Inst->eraseFromParent();
1420 // See if we can constant fold its operands.
1421 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
1423 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
1424 if (CE == 0) continue;
1426 // If we already folded this constant, don't try again.
1427 if (!FoldedConstants.insert(CE))
1430 Constant *NewC = ConstantFoldConstantExpression(CE, TD);
1431 if (NewC && NewC != CE) {
1433 MadeIRChange = true;
1438 InstrsForInstCombineWorklist.push_back(Inst);
1441 // Recursively visit successors. If this is a branch or switch on a
1442 // constant, only visit the reachable successor.
1443 TerminatorInst *TI = BB->getTerminator();
1444 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1445 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
1446 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
1447 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
1448 Worklist.push_back(ReachableBB);
1451 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1452 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
1453 // See if this is an explicit destination.
1454 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
1455 if (SI->getCaseValue(i) == Cond) {
1456 BasicBlock *ReachableBB = SI->getSuccessor(i);
1457 Worklist.push_back(ReachableBB);
1461 // Otherwise it is the default destination.
1462 Worklist.push_back(SI->getSuccessor(0));
1467 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
1468 Worklist.push_back(TI->getSuccessor(i));
1469 } while (!Worklist.empty());
1471 // Once we've found all of the instructions to add to instcombine's worklist,
1472 // add them in reverse order. This way instcombine will visit from the top
1473 // of the function down. This jives well with the way that it adds all uses
1474 // of instructions to the worklist after doing a transformation, thus avoiding
1475 // some N^2 behavior in pathological cases.
1476 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
1477 InstrsForInstCombineWorklist.size());
1479 return MadeIRChange;
1482 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
1483 MadeIRChange = false;
1485 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
1486 << F.getNameStr() << "\n");
1489 // Do a depth-first traversal of the function, populate the worklist with
1490 // the reachable instructions. Ignore blocks that are not reachable. Keep
1491 // track of which blocks we visit.
1492 SmallPtrSet<BasicBlock*, 64> Visited;
1493 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
1495 // Do a quick scan over the function. If we find any blocks that are
1496 // unreachable, remove any instructions inside of them. This prevents
1497 // the instcombine code from having to deal with some bad special cases.
1498 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1499 if (!Visited.count(BB)) {
1500 Instruction *Term = BB->getTerminator();
1501 while (Term != BB->begin()) { // Remove instrs bottom-up
1502 BasicBlock::iterator I = Term; --I;
1504 DEBUG(errs() << "IC: DCE: " << *I << '\n');
1505 // A debug intrinsic shouldn't force another iteration if we weren't
1506 // going to do one without it.
1507 if (!isa<DbgInfoIntrinsic>(I)) {
1509 MadeIRChange = true;
1512 // If I is not void type then replaceAllUsesWith undef.
1513 // This allows ValueHandlers and custom metadata to adjust itself.
1514 if (!I->getType()->isVoidTy())
1515 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1516 I->eraseFromParent();
1521 while (!Worklist.isEmpty()) {
1522 Instruction *I = Worklist.RemoveOne();
1523 if (I == 0) continue; // skip null values.
1525 // Check to see if we can DCE the instruction.
1526 if (isInstructionTriviallyDead(I)) {
1527 DEBUG(errs() << "IC: DCE: " << *I << '\n');
1528 EraseInstFromFunction(*I);
1530 MadeIRChange = true;
1534 // Instruction isn't dead, see if we can constant propagate it.
1535 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
1536 if (Constant *C = ConstantFoldInstruction(I, TD)) {
1537 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
1539 // Add operands to the worklist.
1540 ReplaceInstUsesWith(*I, C);
1542 EraseInstFromFunction(*I);
1543 MadeIRChange = true;
1547 // See if we can trivially sink this instruction to a successor basic block.
1548 if (I->hasOneUse()) {
1549 BasicBlock *BB = I->getParent();
1550 Instruction *UserInst = cast<Instruction>(I->use_back());
1551 BasicBlock *UserParent;
1553 // Get the block the use occurs in.
1554 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
1555 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
1557 UserParent = UserInst->getParent();
1559 if (UserParent != BB) {
1560 bool UserIsSuccessor = false;
1561 // See if the user is one of our successors.
1562 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
1563 if (*SI == UserParent) {
1564 UserIsSuccessor = true;
1568 // If the user is one of our immediate successors, and if that successor
1569 // only has us as a predecessors (we'd have to split the critical edge
1570 // otherwise), we can keep going.
1571 if (UserIsSuccessor && UserParent->getSinglePredecessor())
1572 // Okay, the CFG is simple enough, try to sink this instruction.
1573 MadeIRChange |= TryToSinkInstruction(I, UserParent);
1577 // Now that we have an instruction, try combining it to simplify it.
1578 Builder->SetInsertPoint(I->getParent(), I);
1583 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
1584 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
1586 if (Instruction *Result = visit(*I)) {
1588 // Should we replace the old instruction with a new one?
1590 DEBUG(errs() << "IC: Old = " << *I << '\n'
1591 << " New = " << *Result << '\n');
1593 Result->setDebugLoc(I->getDebugLoc());
1594 // Everything uses the new instruction now.
1595 I->replaceAllUsesWith(Result);
1597 // Push the new instruction and any users onto the worklist.
1598 Worklist.Add(Result);
1599 Worklist.AddUsersToWorkList(*Result);
1601 // Move the name to the new instruction first.
1602 Result->takeName(I);
1604 // Insert the new instruction into the basic block...
1605 BasicBlock *InstParent = I->getParent();
1606 BasicBlock::iterator InsertPos = I;
1608 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
1609 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
1612 InstParent->getInstList().insert(InsertPos, Result);
1614 EraseInstFromFunction(*I);
1617 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
1618 << " New = " << *I << '\n');
1621 // If the instruction was modified, it's possible that it is now dead.
1622 // if so, remove it.
1623 if (isInstructionTriviallyDead(I)) {
1624 EraseInstFromFunction(*I);
1627 Worklist.AddUsersToWorkList(*I);
1630 MadeIRChange = true;
1635 return MadeIRChange;
1639 bool InstCombiner::runOnFunction(Function &F) {
1640 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1641 TD = getAnalysisIfAvailable<TargetData>();
1644 /// Builder - This is an IRBuilder that automatically inserts new
1645 /// instructions into the worklist when they are created.
1646 IRBuilder<true, TargetFolder, InstCombineIRInserter>
1647 TheBuilder(F.getContext(), TargetFolder(TD),
1648 InstCombineIRInserter(Worklist));
1649 Builder = &TheBuilder;
1651 bool EverMadeChange = false;
1653 // Lower dbg.declare intrinsics otherwise their value may be clobbered
1655 EverMadeChange = LowerDbgDeclare(F);
1657 // Iterate while there is work to do.
1658 unsigned Iteration = 0;
1659 while (DoOneIteration(F, Iteration++))
1660 EverMadeChange = true;
1663 return EverMadeChange;
1666 FunctionPass *llvm::createInstructionCombiningPass() {
1667 return new InstCombiner();