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".
166 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
167 if (Op1 && Op1->getOpcode() == Opcode) {
168 Value *A = I.getOperand(0);
169 Value *B = Op1->getOperand(0);
170 Value *C = Op1->getOperand(1);
172 // Does "A op B" simplify?
173 if (Value *V = SimplifyBinOp(Opcode, A, B, TD)) {
174 // It simplifies to V. Form "V op C".
184 if (I.isAssociative() && I.isCommutative()) {
185 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
186 if (Op0 && Op0->getOpcode() == Opcode) {
187 Value *A = Op0->getOperand(0);
188 Value *B = Op0->getOperand(1);
189 Value *C = I.getOperand(1);
191 // Does "C op A" simplify?
192 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
193 // It simplifies to V. Form "V op B".
202 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
203 if (Op1 && Op1->getOpcode() == Opcode) {
204 Value *A = I.getOperand(0);
205 Value *B = Op1->getOperand(0);
206 Value *C = Op1->getOperand(1);
208 // Does "C op A" simplify?
209 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
210 // It simplifies to V. Form "B op V".
219 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
220 // if C1 and C2 are constants.
222 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
223 isa<Constant>(Op0->getOperand(1)) &&
224 isa<Constant>(Op1->getOperand(1)) &&
225 Op0->hasOneUse() && Op1->hasOneUse()) {
226 Value *A = Op0->getOperand(0);
227 Constant *C1 = cast<Constant>(Op0->getOperand(1));
228 Value *B = Op1->getOperand(0);
229 Constant *C2 = cast<Constant>(Op1->getOperand(1));
231 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
232 Instruction *New = BinaryOperator::Create(Opcode, A, B, Op1->getName(),
235 I.setOperand(0, New);
236 I.setOperand(1, Folded);
242 // No further simplifications.
247 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
248 /// "(X LOp Y) ROp (X LOp Z)".
249 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
250 Instruction::BinaryOps ROp) {
255 case Instruction::And:
256 // And distributes over Or and Xor.
260 case Instruction::Or:
261 case Instruction::Xor:
265 case Instruction::Mul:
266 // Multiplication distributes over addition and subtraction.
270 case Instruction::Add:
271 case Instruction::Sub:
275 case Instruction::Or:
276 // Or distributes over And.
280 case Instruction::And:
286 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
287 /// "(X ROp Z) LOp (Y ROp Z)".
288 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
289 Instruction::BinaryOps ROp) {
290 if (Instruction::isCommutative(ROp))
291 return LeftDistributesOverRight(ROp, LOp);
292 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
293 // but this requires knowing that the addition does not overflow and other
298 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
299 /// which some other binary operation distributes over either by factorizing
300 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
301 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
302 /// a win). Returns the simplified value, or null if it didn't simplify.
303 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
304 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
305 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
306 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
307 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); // op
310 if (Op0 && Op1 && Op0->getOpcode() == Op1->getOpcode()) {
311 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
313 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1);
314 Value *C = Op1->getOperand(0), *D = Op1->getOperand(1);
315 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
317 // Does "X op' Y" always equal "Y op' X"?
318 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
320 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
321 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
322 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
323 // commutative case, "(A op' B) op (C op' A)"?
324 if (A == C || (InnerCommutative && A == D)) {
327 // Consider forming "A op' (B op D)".
328 // If "B op D" simplifies then it can be formed with no cost.
329 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, TD);
330 // If "B op D" doesn't simplify then only go on if both of the existing
331 // operations "A op' B" and "C op' D" will be zapped as no longer used.
332 if (!V && Op0->hasOneUse() && Op1->hasOneUse())
333 V = Builder->CreateBinOp(TopLevelOpcode, B, D, Op1->getName());
336 V = Builder->CreateBinOp(InnerOpcode, A, V);
342 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
343 if (RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
344 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
345 // commutative case, "(A op' B) op (B op' D)"?
346 if (B == D || (InnerCommutative && B == C)) {
349 // Consider forming "(A op C) op' B".
350 // If "A op C" simplifies then it can be formed with no cost.
351 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, TD);
352 // If "A op C" doesn't simplify then only go on if both of the existing
353 // operations "A op' B" and "C op' D" will be zapped as no longer used.
354 if (!V && Op0->hasOneUse() && Op1->hasOneUse())
355 V = Builder->CreateBinOp(TopLevelOpcode, A, C, Op0->getName());
358 V = Builder->CreateBinOp(InnerOpcode, V, B);
366 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
367 // The instruction has the form "(A op' B) op C". See if expanding it out
368 // to "(A op C) op' (B op C)" results in simplifications.
369 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
370 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
372 // Do "A op C" and "B op C" both simplify?
373 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, TD))
374 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, TD)) {
375 // They do! Return "L op' R".
377 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
378 if ((L == A && R == B) ||
379 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
381 // Otherwise return "L op' R" if it simplifies.
382 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
384 // Otherwise, create a new instruction.
385 C = Builder->CreateBinOp(InnerOpcode, L, R);
391 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
392 // The instruction has the form "A op (B op' C)". See if expanding it out
393 // to "(A op B) op' (A op C)" results in simplifications.
394 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
395 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
397 // Do "A op B" and "A op C" both simplify?
398 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, TD))
399 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, TD)) {
400 // They do! Return "L op' R".
402 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
403 if ((L == B && R == C) ||
404 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
406 // Otherwise return "L op' R" if it simplifies.
407 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
409 // Otherwise, create a new instruction.
410 A = Builder->CreateBinOp(InnerOpcode, L, R);
419 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
420 // if the LHS is a constant zero (which is the 'negate' form).
422 Value *InstCombiner::dyn_castNegVal(Value *V) const {
423 if (BinaryOperator::isNeg(V))
424 return BinaryOperator::getNegArgument(V);
426 // Constants can be considered to be negated values if they can be folded.
427 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
428 return ConstantExpr::getNeg(C);
430 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
431 if (C->getType()->getElementType()->isIntegerTy())
432 return ConstantExpr::getNeg(C);
437 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
438 // instruction if the LHS is a constant negative zero (which is the 'negate'
441 Value *InstCombiner::dyn_castFNegVal(Value *V) const {
442 if (BinaryOperator::isFNeg(V))
443 return BinaryOperator::getFNegArgument(V);
445 // Constants can be considered to be negated values if they can be folded.
446 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
447 return ConstantExpr::getFNeg(C);
449 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
450 if (C->getType()->getElementType()->isFloatingPointTy())
451 return ConstantExpr::getFNeg(C);
456 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
458 if (CastInst *CI = dyn_cast<CastInst>(&I))
459 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
461 // Figure out if the constant is the left or the right argument.
462 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
463 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
465 if (Constant *SOC = dyn_cast<Constant>(SO)) {
467 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
468 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
471 Value *Op0 = SO, *Op1 = ConstOperand;
475 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
476 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
477 SO->getName()+".op");
478 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
479 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
480 SO->getName()+".cmp");
481 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
482 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
483 SO->getName()+".cmp");
484 llvm_unreachable("Unknown binary instruction type!");
487 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
488 // constant as the other operand, try to fold the binary operator into the
489 // select arguments. This also works for Cast instructions, which obviously do
490 // not have a second operand.
491 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
492 // Don't modify shared select instructions
493 if (!SI->hasOneUse()) return 0;
494 Value *TV = SI->getOperand(1);
495 Value *FV = SI->getOperand(2);
497 if (isa<Constant>(TV) || isa<Constant>(FV)) {
498 // Bool selects with constant operands can be folded to logical ops.
499 if (SI->getType()->isIntegerTy(1)) return 0;
501 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
502 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
504 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
511 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
512 /// has a PHI node as operand #0, see if we can fold the instruction into the
513 /// PHI (which is only possible if all operands to the PHI are constants).
515 /// If AllowAggressive is true, FoldOpIntoPhi will allow certain transforms
516 /// that would normally be unprofitable because they strongly encourage jump
518 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I,
519 bool AllowAggressive) {
520 AllowAggressive = false;
521 PHINode *PN = cast<PHINode>(I.getOperand(0));
522 unsigned NumPHIValues = PN->getNumIncomingValues();
523 if (NumPHIValues == 0 ||
524 // We normally only transform phis with a single use, unless we're trying
525 // hard to make jump threading happen.
526 (!PN->hasOneUse() && !AllowAggressive))
530 // Check to see if all of the operands of the PHI are simple constants
531 // (constantint/constantfp/undef). If there is one non-constant value,
532 // remember the BB it is in. If there is more than one or if *it* is a PHI,
533 // bail out. We don't do arbitrary constant expressions here because moving
534 // their computation can be expensive without a cost model.
535 BasicBlock *NonConstBB = 0;
536 for (unsigned i = 0; i != NumPHIValues; ++i)
537 if (!isa<Constant>(PN->getIncomingValue(i)) ||
538 isa<ConstantExpr>(PN->getIncomingValue(i))) {
539 if (NonConstBB) return 0; // More than one non-const value.
540 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
541 NonConstBB = PN->getIncomingBlock(i);
543 // If the incoming non-constant value is in I's block, we have an infinite
545 if (NonConstBB == I.getParent())
549 // If there is exactly one non-constant value, we can insert a copy of the
550 // operation in that block. However, if this is a critical edge, we would be
551 // inserting the computation one some other paths (e.g. inside a loop). Only
552 // do this if the pred block is unconditionally branching into the phi block.
553 if (NonConstBB != 0 && !AllowAggressive) {
554 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
555 if (!BI || !BI->isUnconditional()) return 0;
558 // Okay, we can do the transformation: create the new PHI node.
559 PHINode *NewPN = PHINode::Create(I.getType(), "");
560 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
561 InsertNewInstBefore(NewPN, *PN);
564 // Next, add all of the operands to the PHI.
565 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
566 // We only currently try to fold the condition of a select when it is a phi,
567 // not the true/false values.
568 Value *TrueV = SI->getTrueValue();
569 Value *FalseV = SI->getFalseValue();
570 BasicBlock *PhiTransBB = PN->getParent();
571 for (unsigned i = 0; i != NumPHIValues; ++i) {
572 BasicBlock *ThisBB = PN->getIncomingBlock(i);
573 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
574 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
576 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
577 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
579 assert(PN->getIncomingBlock(i) == NonConstBB);
580 InV = SelectInst::Create(PN->getIncomingValue(i), TrueVInPred,
582 "phitmp", NonConstBB->getTerminator());
583 Worklist.Add(cast<Instruction>(InV));
585 NewPN->addIncoming(InV, ThisBB);
587 } else if (I.getNumOperands() == 2) {
588 Constant *C = cast<Constant>(I.getOperand(1));
589 for (unsigned i = 0; i != NumPHIValues; ++i) {
591 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
592 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
593 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
595 InV = ConstantExpr::get(I.getOpcode(), InC, C);
597 assert(PN->getIncomingBlock(i) == NonConstBB);
598 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
599 InV = BinaryOperator::Create(BO->getOpcode(),
600 PN->getIncomingValue(i), C, "phitmp",
601 NonConstBB->getTerminator());
602 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
603 InV = CmpInst::Create(CI->getOpcode(),
605 PN->getIncomingValue(i), C, "phitmp",
606 NonConstBB->getTerminator());
608 llvm_unreachable("Unknown binop!");
610 Worklist.Add(cast<Instruction>(InV));
612 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
615 CastInst *CI = cast<CastInst>(&I);
616 const Type *RetTy = CI->getType();
617 for (unsigned i = 0; i != NumPHIValues; ++i) {
619 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
620 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
622 assert(PN->getIncomingBlock(i) == NonConstBB);
623 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
624 I.getType(), "phitmp",
625 NonConstBB->getTerminator());
626 Worklist.Add(cast<Instruction>(InV));
628 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
631 return ReplaceInstUsesWith(I, NewPN);
634 /// FindElementAtOffset - Given a type and a constant offset, determine whether
635 /// or not there is a sequence of GEP indices into the type that will land us at
636 /// the specified offset. If so, fill them into NewIndices and return the
637 /// resultant element type, otherwise return null.
638 const Type *InstCombiner::FindElementAtOffset(const Type *Ty, int64_t Offset,
639 SmallVectorImpl<Value*> &NewIndices) {
641 if (!Ty->isSized()) return 0;
643 // Start with the index over the outer type. Note that the type size
644 // might be zero (even if the offset isn't zero) if the indexed type
645 // is something like [0 x {int, int}]
646 const Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
647 int64_t FirstIdx = 0;
648 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
649 FirstIdx = Offset/TySize;
650 Offset -= FirstIdx*TySize;
652 // Handle hosts where % returns negative instead of values [0..TySize).
658 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
661 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
663 // Index into the types. If we fail, set OrigBase to null.
665 // Indexing into tail padding between struct/array elements.
666 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
669 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
670 const StructLayout *SL = TD->getStructLayout(STy);
671 assert(Offset < (int64_t)SL->getSizeInBytes() &&
672 "Offset must stay within the indexed type");
674 unsigned Elt = SL->getElementContainingOffset(Offset);
675 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
678 Offset -= SL->getElementOffset(Elt);
679 Ty = STy->getElementType(Elt);
680 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
681 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
682 assert(EltSize && "Cannot index into a zero-sized array");
683 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
685 Ty = AT->getElementType();
687 // Otherwise, we can't index into the middle of this atomic type, bail.
697 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
698 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
700 if (Value *V = SimplifyGEPInst(&Ops[0], Ops.size(), TD))
701 return ReplaceInstUsesWith(GEP, V);
703 Value *PtrOp = GEP.getOperand(0);
705 // Eliminate unneeded casts for indices, and replace indices which displace
706 // by multiples of a zero size type with zero.
708 bool MadeChange = false;
709 const Type *IntPtrTy = TD->getIntPtrType(GEP.getContext());
711 gep_type_iterator GTI = gep_type_begin(GEP);
712 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
713 I != E; ++I, ++GTI) {
714 // Skip indices into struct types.
715 const SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
716 if (!SeqTy) continue;
718 // If the element type has zero size then any index over it is equivalent
719 // to an index of zero, so replace it with zero if it is not zero already.
720 if (SeqTy->getElementType()->isSized() &&
721 TD->getTypeAllocSize(SeqTy->getElementType()) == 0)
722 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
723 *I = Constant::getNullValue(IntPtrTy);
727 if ((*I)->getType() != IntPtrTy) {
728 // If we are using a wider index than needed for this platform, shrink
729 // it to what we need. If narrower, sign-extend it to what we need.
730 // This explicit cast can make subsequent optimizations more obvious.
731 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
735 if (MadeChange) return &GEP;
738 // Combine Indices - If the source pointer to this getelementptr instruction
739 // is a getelementptr instruction, combine the indices of the two
740 // getelementptr instructions into a single instruction.
742 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
743 // Note that if our source is a gep chain itself that we wait for that
744 // chain to be resolved before we perform this transformation. This
745 // avoids us creating a TON of code in some cases.
747 if (GetElementPtrInst *SrcGEP =
748 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
749 if (SrcGEP->getNumOperands() == 2)
750 return 0; // Wait until our source is folded to completion.
752 SmallVector<Value*, 8> Indices;
754 // Find out whether the last index in the source GEP is a sequential idx.
755 bool EndsWithSequential = false;
756 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
758 EndsWithSequential = !(*I)->isStructTy();
760 // Can we combine the two pointer arithmetics offsets?
761 if (EndsWithSequential) {
762 // Replace: gep (gep %P, long B), long A, ...
763 // With: T = long A+B; gep %P, T, ...
766 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
767 Value *GO1 = GEP.getOperand(1);
768 if (SO1 == Constant::getNullValue(SO1->getType())) {
770 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
773 // If they aren't the same type, then the input hasn't been processed
774 // by the loop above yet (which canonicalizes sequential index types to
775 // intptr_t). Just avoid transforming this until the input has been
777 if (SO1->getType() != GO1->getType())
779 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
782 // Update the GEP in place if possible.
783 if (Src->getNumOperands() == 2) {
784 GEP.setOperand(0, Src->getOperand(0));
785 GEP.setOperand(1, Sum);
788 Indices.append(Src->op_begin()+1, Src->op_end()-1);
789 Indices.push_back(Sum);
790 Indices.append(GEP.op_begin()+2, GEP.op_end());
791 } else if (isa<Constant>(*GEP.idx_begin()) &&
792 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
793 Src->getNumOperands() != 1) {
794 // Otherwise we can do the fold if the first index of the GEP is a zero
795 Indices.append(Src->op_begin()+1, Src->op_end());
796 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
799 if (!Indices.empty())
800 return (GEP.isInBounds() && Src->isInBounds()) ?
801 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
802 Indices.end(), GEP.getName()) :
803 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
804 Indices.end(), GEP.getName());
807 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
808 Value *StrippedPtr = PtrOp->stripPointerCasts();
809 if (StrippedPtr != PtrOp) {
810 const PointerType *StrippedPtrTy =cast<PointerType>(StrippedPtr->getType());
812 bool HasZeroPointerIndex = false;
813 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
814 HasZeroPointerIndex = C->isZero();
816 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
817 // into : GEP [10 x i8]* X, i32 0, ...
819 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
820 // into : GEP i8* X, ...
822 // This occurs when the program declares an array extern like "int X[];"
823 if (HasZeroPointerIndex) {
824 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
825 if (const ArrayType *CATy =
826 dyn_cast<ArrayType>(CPTy->getElementType())) {
827 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
828 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
830 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
831 GetElementPtrInst *Res =
832 GetElementPtrInst::Create(StrippedPtr, Idx.begin(),
833 Idx.end(), GEP.getName());
834 Res->setIsInBounds(GEP.isInBounds());
838 if (const ArrayType *XATy =
839 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
840 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
841 if (CATy->getElementType() == XATy->getElementType()) {
842 // -> GEP [10 x i8]* X, i32 0, ...
843 // At this point, we know that the cast source type is a pointer
844 // to an array of the same type as the destination pointer
845 // array. Because the array type is never stepped over (there
846 // is a leading zero) we can fold the cast into this GEP.
847 GEP.setOperand(0, StrippedPtr);
852 } else if (GEP.getNumOperands() == 2) {
853 // Transform things like:
854 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
855 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
856 const Type *SrcElTy = StrippedPtrTy->getElementType();
857 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
858 if (TD && SrcElTy->isArrayTy() &&
859 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
860 TD->getTypeAllocSize(ResElTy)) {
862 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
863 Idx[1] = GEP.getOperand(1);
864 Value *NewGEP = GEP.isInBounds() ?
865 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2, GEP.getName()) :
866 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName());
867 // V and GEP are both pointer types --> BitCast
868 return new BitCastInst(NewGEP, GEP.getType());
871 // Transform things like:
872 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
873 // (where tmp = 8*tmp2) into:
874 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
876 if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) {
877 uint64_t ArrayEltSize =
878 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
880 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
881 // allow either a mul, shift, or constant here.
883 ConstantInt *Scale = 0;
884 if (ArrayEltSize == 1) {
885 NewIdx = GEP.getOperand(1);
886 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
887 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
888 NewIdx = ConstantInt::get(CI->getType(), 1);
890 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
891 if (Inst->getOpcode() == Instruction::Shl &&
892 isa<ConstantInt>(Inst->getOperand(1))) {
893 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
894 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
895 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
897 NewIdx = Inst->getOperand(0);
898 } else if (Inst->getOpcode() == Instruction::Mul &&
899 isa<ConstantInt>(Inst->getOperand(1))) {
900 Scale = cast<ConstantInt>(Inst->getOperand(1));
901 NewIdx = Inst->getOperand(0);
905 // If the index will be to exactly the right offset with the scale taken
906 // out, perform the transformation. Note, we don't know whether Scale is
907 // signed or not. We'll use unsigned version of division/modulo
908 // operation after making sure Scale doesn't have the sign bit set.
909 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
910 Scale->getZExtValue() % ArrayEltSize == 0) {
911 Scale = ConstantInt::get(Scale->getType(),
912 Scale->getZExtValue() / ArrayEltSize);
913 if (Scale->getZExtValue() != 1) {
914 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
916 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
919 // Insert the new GEP instruction.
921 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
923 Value *NewGEP = GEP.isInBounds() ?
924 Builder->CreateInBoundsGEP(StrippedPtr, Idx, Idx + 2,GEP.getName()):
925 Builder->CreateGEP(StrippedPtr, Idx, Idx + 2, GEP.getName());
926 // The NewGEP must be pointer typed, so must the old one -> BitCast
927 return new BitCastInst(NewGEP, GEP.getType());
933 /// See if we can simplify:
934 /// X = bitcast A* to B*
935 /// Y = gep X, <...constant indices...>
936 /// into a gep of the original struct. This is important for SROA and alias
937 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
938 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
940 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
941 // Determine how much the GEP moves the pointer. We are guaranteed to get
942 // a constant back from EmitGEPOffset.
943 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP));
944 int64_t Offset = OffsetV->getSExtValue();
946 // If this GEP instruction doesn't move the pointer, just replace the GEP
947 // with a bitcast of the real input to the dest type.
949 // If the bitcast is of an allocation, and the allocation will be
950 // converted to match the type of the cast, don't touch this.
951 if (isa<AllocaInst>(BCI->getOperand(0)) ||
952 isMalloc(BCI->getOperand(0))) {
953 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
954 if (Instruction *I = visitBitCast(*BCI)) {
957 BCI->getParent()->getInstList().insert(BCI, I);
958 ReplaceInstUsesWith(*BCI, I);
963 return new BitCastInst(BCI->getOperand(0), GEP.getType());
966 // Otherwise, if the offset is non-zero, we need to find out if there is a
967 // field at Offset in 'A's type. If so, we can pull the cast through the
969 SmallVector<Value*, 8> NewIndices;
971 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
972 if (FindElementAtOffset(InTy, Offset, NewIndices)) {
973 Value *NGEP = GEP.isInBounds() ?
974 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices.begin(),
976 Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
979 if (NGEP->getType() == GEP.getType())
980 return ReplaceInstUsesWith(GEP, NGEP);
981 NGEP->takeName(&GEP);
982 return new BitCastInst(NGEP, GEP.getType());
992 static bool IsOnlyNullComparedAndFreed(const Value &V) {
993 for (Value::const_use_iterator UI = V.use_begin(), UE = V.use_end();
998 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(U))
999 if (ICI->isEquality() && isa<ConstantPointerNull>(ICI->getOperand(1)))
1006 Instruction *InstCombiner::visitMalloc(Instruction &MI) {
1007 // If we have a malloc call which is only used in any amount of comparisons
1008 // to null and free calls, delete the calls and replace the comparisons with
1009 // true or false as appropriate.
1010 if (IsOnlyNullComparedAndFreed(MI)) {
1011 for (Value::use_iterator UI = MI.use_begin(), UE = MI.use_end();
1013 // We can assume that every remaining use is a free call or an icmp eq/ne
1014 // to null, so the cast is safe.
1015 Instruction *I = cast<Instruction>(*UI);
1017 // Early increment here, as we're about to get rid of the user.
1020 if (isFreeCall(I)) {
1021 EraseInstFromFunction(*cast<CallInst>(I));
1024 // Again, the cast is safe.
1025 ICmpInst *C = cast<ICmpInst>(I);
1026 ReplaceInstUsesWith(*C, ConstantInt::get(Type::getInt1Ty(C->getContext()),
1027 C->isFalseWhenEqual()));
1028 EraseInstFromFunction(*C);
1030 return EraseInstFromFunction(MI);
1037 Instruction *InstCombiner::visitFree(CallInst &FI) {
1038 Value *Op = FI.getArgOperand(0);
1040 // free undef -> unreachable.
1041 if (isa<UndefValue>(Op)) {
1042 // Insert a new store to null because we cannot modify the CFG here.
1043 new StoreInst(ConstantInt::getTrue(FI.getContext()),
1044 UndefValue::get(Type::getInt1PtrTy(FI.getContext())), &FI);
1045 return EraseInstFromFunction(FI);
1048 // If we have 'free null' delete the instruction. This can happen in stl code
1049 // when lots of inlining happens.
1050 if (isa<ConstantPointerNull>(Op))
1051 return EraseInstFromFunction(FI);
1058 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
1059 // Change br (not X), label True, label False to: br X, label False, True
1061 BasicBlock *TrueDest;
1062 BasicBlock *FalseDest;
1063 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
1064 !isa<Constant>(X)) {
1065 // Swap Destinations and condition...
1067 BI.setSuccessor(0, FalseDest);
1068 BI.setSuccessor(1, TrueDest);
1072 // Cannonicalize fcmp_one -> fcmp_oeq
1073 FCmpInst::Predicate FPred; Value *Y;
1074 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
1075 TrueDest, FalseDest)) &&
1076 BI.getCondition()->hasOneUse())
1077 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
1078 FPred == FCmpInst::FCMP_OGE) {
1079 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
1080 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
1082 // Swap Destinations and condition.
1083 BI.setSuccessor(0, FalseDest);
1084 BI.setSuccessor(1, TrueDest);
1089 // Cannonicalize icmp_ne -> icmp_eq
1090 ICmpInst::Predicate IPred;
1091 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
1092 TrueDest, FalseDest)) &&
1093 BI.getCondition()->hasOneUse())
1094 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
1095 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
1096 IPred == ICmpInst::ICMP_SGE) {
1097 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
1098 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
1099 // Swap Destinations and condition.
1100 BI.setSuccessor(0, FalseDest);
1101 BI.setSuccessor(1, TrueDest);
1109 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
1110 Value *Cond = SI.getCondition();
1111 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
1112 if (I->getOpcode() == Instruction::Add)
1113 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1114 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
1115 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
1117 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
1119 SI.setOperand(0, I->getOperand(0));
1127 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
1128 Value *Agg = EV.getAggregateOperand();
1130 if (!EV.hasIndices())
1131 return ReplaceInstUsesWith(EV, Agg);
1133 if (Constant *C = dyn_cast<Constant>(Agg)) {
1134 if (isa<UndefValue>(C))
1135 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
1137 if (isa<ConstantAggregateZero>(C))
1138 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
1140 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
1141 // Extract the element indexed by the first index out of the constant
1142 Value *V = C->getOperand(*EV.idx_begin());
1143 if (EV.getNumIndices() > 1)
1144 // Extract the remaining indices out of the constant indexed by the
1146 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
1148 return ReplaceInstUsesWith(EV, V);
1150 return 0; // Can't handle other constants
1152 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
1153 // We're extracting from an insertvalue instruction, compare the indices
1154 const unsigned *exti, *exte, *insi, *inse;
1155 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
1156 exte = EV.idx_end(), inse = IV->idx_end();
1157 exti != exte && insi != inse;
1160 // The insert and extract both reference distinctly different elements.
1161 // This means the extract is not influenced by the insert, and we can
1162 // replace the aggregate operand of the extract with the aggregate
1163 // operand of the insert. i.e., replace
1164 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1165 // %E = extractvalue { i32, { i32 } } %I, 0
1167 // %E = extractvalue { i32, { i32 } } %A, 0
1168 return ExtractValueInst::Create(IV->getAggregateOperand(),
1169 EV.idx_begin(), EV.idx_end());
1171 if (exti == exte && insi == inse)
1172 // Both iterators are at the end: Index lists are identical. Replace
1173 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1174 // %C = extractvalue { i32, { i32 } } %B, 1, 0
1176 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
1178 // The extract list is a prefix of the insert list. i.e. replace
1179 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1180 // %E = extractvalue { i32, { i32 } } %I, 1
1182 // %X = extractvalue { i32, { i32 } } %A, 1
1183 // %E = insertvalue { i32 } %X, i32 42, 0
1184 // by switching the order of the insert and extract (though the
1185 // insertvalue should be left in, since it may have other uses).
1186 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
1187 EV.idx_begin(), EV.idx_end());
1188 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
1192 // The insert list is a prefix of the extract list
1193 // We can simply remove the common indices from the extract and make it
1194 // operate on the inserted value instead of the insertvalue result.
1196 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1197 // %E = extractvalue { i32, { i32 } } %I, 1, 0
1199 // %E extractvalue { i32 } { i32 42 }, 0
1200 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
1203 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
1204 // We're extracting from an intrinsic, see if we're the only user, which
1205 // allows us to simplify multiple result intrinsics to simpler things that
1206 // just get one value.
1207 if (II->hasOneUse()) {
1208 // Check if we're grabbing the overflow bit or the result of a 'with
1209 // overflow' intrinsic. If it's the latter we can remove the intrinsic
1210 // and replace it with a traditional binary instruction.
1211 switch (II->getIntrinsicID()) {
1212 case Intrinsic::uadd_with_overflow:
1213 case Intrinsic::sadd_with_overflow:
1214 if (*EV.idx_begin() == 0) { // Normal result.
1215 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1216 II->replaceAllUsesWith(UndefValue::get(II->getType()));
1217 EraseInstFromFunction(*II);
1218 return BinaryOperator::CreateAdd(LHS, RHS);
1221 // If the normal result of the add is dead, and the RHS is a constant,
1222 // we can transform this into a range comparison.
1223 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
1224 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
1225 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
1226 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
1227 ConstantExpr::getNot(CI));
1229 case Intrinsic::usub_with_overflow:
1230 case Intrinsic::ssub_with_overflow:
1231 if (*EV.idx_begin() == 0) { // Normal result.
1232 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1233 II->replaceAllUsesWith(UndefValue::get(II->getType()));
1234 EraseInstFromFunction(*II);
1235 return BinaryOperator::CreateSub(LHS, RHS);
1238 case Intrinsic::umul_with_overflow:
1239 case Intrinsic::smul_with_overflow:
1240 if (*EV.idx_begin() == 0) { // Normal result.
1241 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1242 II->replaceAllUsesWith(UndefValue::get(II->getType()));
1243 EraseInstFromFunction(*II);
1244 return BinaryOperator::CreateMul(LHS, RHS);
1252 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
1253 // If the (non-volatile) load only has one use, we can rewrite this to a
1254 // load from a GEP. This reduces the size of the load.
1255 // FIXME: If a load is used only by extractvalue instructions then this
1256 // could be done regardless of having multiple uses.
1257 if (!L->isVolatile() && L->hasOneUse()) {
1258 // extractvalue has integer indices, getelementptr has Value*s. Convert.
1259 SmallVector<Value*, 4> Indices;
1260 // Prefix an i32 0 since we need the first element.
1261 Indices.push_back(Builder->getInt32(0));
1262 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
1264 Indices.push_back(Builder->getInt32(*I));
1266 // We need to insert these at the location of the old load, not at that of
1267 // the extractvalue.
1268 Builder->SetInsertPoint(L->getParent(), L);
1269 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(),
1270 Indices.begin(), Indices.end());
1271 // Returning the load directly will cause the main loop to insert it in
1272 // the wrong spot, so use ReplaceInstUsesWith().
1273 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
1275 // We could simplify extracts from other values. Note that nested extracts may
1276 // already be simplified implicitly by the above: extract (extract (insert) )
1277 // will be translated into extract ( insert ( extract ) ) first and then just
1278 // the value inserted, if appropriate. Similarly for extracts from single-use
1279 // loads: extract (extract (load)) will be translated to extract (load (gep))
1280 // and if again single-use then via load (gep (gep)) to load (gep).
1281 // However, double extracts from e.g. function arguments or return values
1282 // aren't handled yet.
1289 /// TryToSinkInstruction - Try to move the specified instruction from its
1290 /// current block into the beginning of DestBlock, which can only happen if it's
1291 /// safe to move the instruction past all of the instructions between it and the
1292 /// end of its block.
1293 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
1294 assert(I->hasOneUse() && "Invariants didn't hold!");
1296 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1297 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
1300 // Do not sink alloca instructions out of the entry block.
1301 if (isa<AllocaInst>(I) && I->getParent() ==
1302 &DestBlock->getParent()->getEntryBlock())
1305 // We can only sink load instructions if there is nothing between the load and
1306 // the end of block that could change the value.
1307 if (I->mayReadFromMemory()) {
1308 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
1310 if (Scan->mayWriteToMemory())
1314 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
1316 I->moveBefore(InsertPos);
1322 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
1323 /// all reachable code to the worklist.
1325 /// This has a couple of tricks to make the code faster and more powerful. In
1326 /// particular, we constant fold and DCE instructions as we go, to avoid adding
1327 /// them to the worklist (this significantly speeds up instcombine on code where
1328 /// many instructions are dead or constant). Additionally, if we find a branch
1329 /// whose condition is a known constant, we only visit the reachable successors.
1331 static bool AddReachableCodeToWorklist(BasicBlock *BB,
1332 SmallPtrSet<BasicBlock*, 64> &Visited,
1334 const TargetData *TD) {
1335 bool MadeIRChange = false;
1336 SmallVector<BasicBlock*, 256> Worklist;
1337 Worklist.push_back(BB);
1339 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
1340 SmallPtrSet<ConstantExpr*, 64> FoldedConstants;
1343 BB = Worklist.pop_back_val();
1345 // We have now visited this block! If we've already been here, ignore it.
1346 if (!Visited.insert(BB)) continue;
1348 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
1349 Instruction *Inst = BBI++;
1351 // DCE instruction if trivially dead.
1352 if (isInstructionTriviallyDead(Inst)) {
1354 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
1355 Inst->eraseFromParent();
1359 // ConstantProp instruction if trivially constant.
1360 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
1361 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
1362 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
1364 Inst->replaceAllUsesWith(C);
1366 Inst->eraseFromParent();
1371 // See if we can constant fold its operands.
1372 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
1374 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
1375 if (CE == 0) continue;
1377 // If we already folded this constant, don't try again.
1378 if (!FoldedConstants.insert(CE))
1381 Constant *NewC = ConstantFoldConstantExpression(CE, TD);
1382 if (NewC && NewC != CE) {
1384 MadeIRChange = true;
1389 InstrsForInstCombineWorklist.push_back(Inst);
1392 // Recursively visit successors. If this is a branch or switch on a
1393 // constant, only visit the reachable successor.
1394 TerminatorInst *TI = BB->getTerminator();
1395 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1396 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
1397 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
1398 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
1399 Worklist.push_back(ReachableBB);
1402 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1403 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
1404 // See if this is an explicit destination.
1405 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
1406 if (SI->getCaseValue(i) == Cond) {
1407 BasicBlock *ReachableBB = SI->getSuccessor(i);
1408 Worklist.push_back(ReachableBB);
1412 // Otherwise it is the default destination.
1413 Worklist.push_back(SI->getSuccessor(0));
1418 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
1419 Worklist.push_back(TI->getSuccessor(i));
1420 } while (!Worklist.empty());
1422 // Once we've found all of the instructions to add to instcombine's worklist,
1423 // add them in reverse order. This way instcombine will visit from the top
1424 // of the function down. This jives well with the way that it adds all uses
1425 // of instructions to the worklist after doing a transformation, thus avoiding
1426 // some N^2 behavior in pathological cases.
1427 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
1428 InstrsForInstCombineWorklist.size());
1430 return MadeIRChange;
1433 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
1434 MadeIRChange = false;
1436 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
1437 << F.getNameStr() << "\n");
1440 // Do a depth-first traversal of the function, populate the worklist with
1441 // the reachable instructions. Ignore blocks that are not reachable. Keep
1442 // track of which blocks we visit.
1443 SmallPtrSet<BasicBlock*, 64> Visited;
1444 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
1446 // Do a quick scan over the function. If we find any blocks that are
1447 // unreachable, remove any instructions inside of them. This prevents
1448 // the instcombine code from having to deal with some bad special cases.
1449 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1450 if (!Visited.count(BB)) {
1451 Instruction *Term = BB->getTerminator();
1452 while (Term != BB->begin()) { // Remove instrs bottom-up
1453 BasicBlock::iterator I = Term; --I;
1455 DEBUG(errs() << "IC: DCE: " << *I << '\n');
1456 // A debug intrinsic shouldn't force another iteration if we weren't
1457 // going to do one without it.
1458 if (!isa<DbgInfoIntrinsic>(I)) {
1460 MadeIRChange = true;
1463 // If I is not void type then replaceAllUsesWith undef.
1464 // This allows ValueHandlers and custom metadata to adjust itself.
1465 if (!I->getType()->isVoidTy())
1466 I->replaceAllUsesWith(UndefValue::get(I->getType()));
1467 I->eraseFromParent();
1472 while (!Worklist.isEmpty()) {
1473 Instruction *I = Worklist.RemoveOne();
1474 if (I == 0) continue; // skip null values.
1476 // Check to see if we can DCE the instruction.
1477 if (isInstructionTriviallyDead(I)) {
1478 DEBUG(errs() << "IC: DCE: " << *I << '\n');
1479 EraseInstFromFunction(*I);
1481 MadeIRChange = true;
1485 // Instruction isn't dead, see if we can constant propagate it.
1486 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
1487 if (Constant *C = ConstantFoldInstruction(I, TD)) {
1488 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
1490 // Add operands to the worklist.
1491 ReplaceInstUsesWith(*I, C);
1493 EraseInstFromFunction(*I);
1494 MadeIRChange = true;
1498 // See if we can trivially sink this instruction to a successor basic block.
1499 if (I->hasOneUse()) {
1500 BasicBlock *BB = I->getParent();
1501 Instruction *UserInst = cast<Instruction>(I->use_back());
1502 BasicBlock *UserParent;
1504 // Get the block the use occurs in.
1505 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
1506 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
1508 UserParent = UserInst->getParent();
1510 if (UserParent != BB) {
1511 bool UserIsSuccessor = false;
1512 // See if the user is one of our successors.
1513 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
1514 if (*SI == UserParent) {
1515 UserIsSuccessor = true;
1519 // If the user is one of our immediate successors, and if that successor
1520 // only has us as a predecessors (we'd have to split the critical edge
1521 // otherwise), we can keep going.
1522 if (UserIsSuccessor && UserParent->getSinglePredecessor())
1523 // Okay, the CFG is simple enough, try to sink this instruction.
1524 MadeIRChange |= TryToSinkInstruction(I, UserParent);
1528 // Now that we have an instruction, try combining it to simplify it.
1529 Builder->SetInsertPoint(I->getParent(), I);
1534 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
1535 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
1537 if (Instruction *Result = visit(*I)) {
1539 // Should we replace the old instruction with a new one?
1541 DEBUG(errs() << "IC: Old = " << *I << '\n'
1542 << " New = " << *Result << '\n');
1544 // Everything uses the new instruction now.
1545 I->replaceAllUsesWith(Result);
1547 // Push the new instruction and any users onto the worklist.
1548 Worklist.Add(Result);
1549 Worklist.AddUsersToWorkList(*Result);
1551 // Move the name to the new instruction first.
1552 Result->takeName(I);
1554 // Insert the new instruction into the basic block...
1555 BasicBlock *InstParent = I->getParent();
1556 BasicBlock::iterator InsertPos = I;
1558 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
1559 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
1562 InstParent->getInstList().insert(InsertPos, Result);
1564 EraseInstFromFunction(*I);
1567 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
1568 << " New = " << *I << '\n');
1571 // If the instruction was modified, it's possible that it is now dead.
1572 // if so, remove it.
1573 if (isInstructionTriviallyDead(I)) {
1574 EraseInstFromFunction(*I);
1577 Worklist.AddUsersToWorkList(*I);
1580 MadeIRChange = true;
1585 return MadeIRChange;
1589 bool InstCombiner::runOnFunction(Function &F) {
1590 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1591 TD = getAnalysisIfAvailable<TargetData>();
1594 /// Builder - This is an IRBuilder that automatically inserts new
1595 /// instructions into the worklist when they are created.
1596 IRBuilder<true, TargetFolder, InstCombineIRInserter>
1597 TheBuilder(F.getContext(), TargetFolder(TD),
1598 InstCombineIRInserter(Worklist));
1599 Builder = &TheBuilder;
1601 bool EverMadeChange = false;
1603 // Iterate while there is work to do.
1604 unsigned Iteration = 0;
1605 while (DoOneIteration(F, Iteration++))
1606 EverMadeChange = true;
1609 return EverMadeChange;
1612 FunctionPass *llvm::createInstructionCombiningPass() {
1613 return new InstCombiner();