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/DataLayout.h"
44 #include "llvm/Target/TargetLibraryInfo.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Support/CFG.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/GetElementPtrTypeIterator.h"
49 #include "llvm/Support/PatternMatch.h"
50 #include "llvm/Support/ValueHandle.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/Statistic.h"
53 #include "llvm/ADT/StringSwitch.h"
54 #include "llvm-c/Initialization.h"
58 using namespace llvm::PatternMatch;
60 STATISTIC(NumCombined , "Number of insts combined");
61 STATISTIC(NumConstProp, "Number of constant folds");
62 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
63 STATISTIC(NumSunkInst , "Number of instructions sunk");
64 STATISTIC(NumExpand, "Number of expansions");
65 STATISTIC(NumFactor , "Number of factorizations");
66 STATISTIC(NumReassoc , "Number of reassociations");
68 // Initialization Routines
69 void llvm::initializeInstCombine(PassRegistry &Registry) {
70 initializeInstCombinerPass(Registry);
73 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
74 initializeInstCombine(*unwrap(R));
77 char InstCombiner::ID = 0;
78 INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine",
79 "Combine redundant instructions", false, false)
80 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
81 INITIALIZE_PASS_END(InstCombiner, "instcombine",
82 "Combine redundant instructions", false, false)
84 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
86 AU.addRequired<TargetLibraryInfo>();
90 Value *InstCombiner::EmitGEPOffset(User *GEP) {
91 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP);
94 /// ShouldChangeType - Return true if it is desirable to convert a computation
95 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
96 /// type for example, or from a smaller to a larger illegal type.
97 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
98 assert(From->isIntegerTy() && To->isIntegerTy());
100 // If we don't have TD, we don't know if the source/dest are legal.
101 if (!TD) return false;
103 unsigned FromWidth = From->getPrimitiveSizeInBits();
104 unsigned ToWidth = To->getPrimitiveSizeInBits();
105 bool FromLegal = TD->isLegalInteger(FromWidth);
106 bool ToLegal = TD->isLegalInteger(ToWidth);
108 // If this is a legal integer from type, and the result would be an illegal
109 // type, don't do the transformation.
110 if (FromLegal && !ToLegal)
113 // Otherwise, if both are illegal, do not increase the size of the result. We
114 // do allow things like i160 -> i64, but not i64 -> i160.
115 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
121 // Return true, if No Signed Wrap should be maintained for I.
122 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
123 // where both B and C should be ConstantInts, results in a constant that does
124 // not overflow. This function only handles the Add and Sub opcodes. For
125 // all other opcodes, the function conservatively returns false.
126 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
127 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
128 if (!OBO || !OBO->hasNoSignedWrap()) {
132 // We reason about Add and Sub Only.
133 Instruction::BinaryOps Opcode = I.getOpcode();
134 if (Opcode != Instruction::Add &&
135 Opcode != Instruction::Sub) {
139 ConstantInt *CB = dyn_cast<ConstantInt>(B);
140 ConstantInt *CC = dyn_cast<ConstantInt>(C);
146 const APInt &BVal = CB->getValue();
147 const APInt &CVal = CC->getValue();
148 bool Overflow = false;
150 if (Opcode == Instruction::Add) {
151 BVal.sadd_ov(CVal, Overflow);
153 BVal.ssub_ov(CVal, Overflow);
159 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
160 /// operators which are associative or commutative:
162 // Commutative operators:
164 // 1. Order operands such that they are listed from right (least complex) to
165 // left (most complex). This puts constants before unary operators before
168 // Associative operators:
170 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
171 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
173 // Associative and commutative operators:
175 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
176 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
177 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
178 // if C1 and C2 are constants.
180 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
181 Instruction::BinaryOps Opcode = I.getOpcode();
182 bool Changed = false;
185 // Order operands such that they are listed from right (least complex) to
186 // left (most complex). This puts constants before unary operators before
188 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
189 getComplexity(I.getOperand(1)))
190 Changed = !I.swapOperands();
192 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
193 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
195 if (I.isAssociative()) {
196 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
197 if (Op0 && Op0->getOpcode() == Opcode) {
198 Value *A = Op0->getOperand(0);
199 Value *B = Op0->getOperand(1);
200 Value *C = I.getOperand(1);
202 // Does "B op C" simplify?
203 if (Value *V = SimplifyBinOp(Opcode, B, C, TD)) {
204 // It simplifies to V. Form "A op V".
207 // Conservatively clear the optional flags, since they may not be
208 // preserved by the reassociation.
209 if (MaintainNoSignedWrap(I, B, C) &&
210 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
211 // Note: this is only valid because SimplifyBinOp doesn't look at
212 // the operands to Op0.
213 I.clearSubclassOptionalData();
214 I.setHasNoSignedWrap(true);
216 I.clearSubclassOptionalData();
225 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
226 if (Op1 && Op1->getOpcode() == Opcode) {
227 Value *A = I.getOperand(0);
228 Value *B = Op1->getOperand(0);
229 Value *C = Op1->getOperand(1);
231 // Does "A op B" simplify?
232 if (Value *V = SimplifyBinOp(Opcode, A, B, TD)) {
233 // It simplifies to V. Form "V op C".
236 // Conservatively clear the optional flags, since they may not be
237 // preserved by the reassociation.
238 I.clearSubclassOptionalData();
246 if (I.isAssociative() && I.isCommutative()) {
247 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
248 if (Op0 && Op0->getOpcode() == Opcode) {
249 Value *A = Op0->getOperand(0);
250 Value *B = Op0->getOperand(1);
251 Value *C = I.getOperand(1);
253 // Does "C op A" simplify?
254 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
255 // It simplifies to V. Form "V op B".
258 // Conservatively clear the optional flags, since they may not be
259 // preserved by the reassociation.
260 I.clearSubclassOptionalData();
267 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
268 if (Op1 && Op1->getOpcode() == Opcode) {
269 Value *A = I.getOperand(0);
270 Value *B = Op1->getOperand(0);
271 Value *C = Op1->getOperand(1);
273 // Does "C op A" simplify?
274 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
275 // It simplifies to V. Form "B op V".
278 // Conservatively clear the optional flags, since they may not be
279 // preserved by the reassociation.
280 I.clearSubclassOptionalData();
287 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
288 // if C1 and C2 are constants.
290 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
291 isa<Constant>(Op0->getOperand(1)) &&
292 isa<Constant>(Op1->getOperand(1)) &&
293 Op0->hasOneUse() && Op1->hasOneUse()) {
294 Value *A = Op0->getOperand(0);
295 Constant *C1 = cast<Constant>(Op0->getOperand(1));
296 Value *B = Op1->getOperand(0);
297 Constant *C2 = cast<Constant>(Op1->getOperand(1));
299 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
300 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
301 InsertNewInstWith(New, I);
303 I.setOperand(0, New);
304 I.setOperand(1, Folded);
305 // Conservatively clear the optional flags, since they may not be
306 // preserved by the reassociation.
307 I.clearSubclassOptionalData();
314 // No further simplifications.
319 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
320 /// "(X LOp Y) ROp (X LOp Z)".
321 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
322 Instruction::BinaryOps ROp) {
327 case Instruction::And:
328 // And distributes over Or and Xor.
332 case Instruction::Or:
333 case Instruction::Xor:
337 case Instruction::Mul:
338 // Multiplication distributes over addition and subtraction.
342 case Instruction::Add:
343 case Instruction::Sub:
347 case Instruction::Or:
348 // Or distributes over And.
352 case Instruction::And:
358 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
359 /// "(X ROp Z) LOp (Y ROp Z)".
360 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
361 Instruction::BinaryOps ROp) {
362 if (Instruction::isCommutative(ROp))
363 return LeftDistributesOverRight(ROp, LOp);
364 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
365 // but this requires knowing that the addition does not overflow and other
370 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
371 /// which some other binary operation distributes over either by factorizing
372 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
373 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
374 /// a win). Returns the simplified value, or null if it didn't simplify.
375 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
376 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
377 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
378 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
379 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); // op
382 if (Op0 && Op1 && Op0->getOpcode() == Op1->getOpcode()) {
383 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
385 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1);
386 Value *C = Op1->getOperand(0), *D = Op1->getOperand(1);
387 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
389 // Does "X op' Y" always equal "Y op' X"?
390 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
392 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
393 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
394 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
395 // commutative case, "(A op' B) op (C op' A)"?
396 if (A == C || (InnerCommutative && A == D)) {
399 // Consider forming "A op' (B op D)".
400 // If "B op D" simplifies then it can be formed with no cost.
401 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, TD);
402 // If "B op D" doesn't simplify then only go on if both of the existing
403 // operations "A op' B" and "C op' D" will be zapped as no longer used.
404 if (!V && Op0->hasOneUse() && Op1->hasOneUse())
405 V = Builder->CreateBinOp(TopLevelOpcode, B, D, Op1->getName());
408 V = Builder->CreateBinOp(InnerOpcode, A, V);
414 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
415 if (RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
416 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
417 // commutative case, "(A op' B) op (B op' D)"?
418 if (B == D || (InnerCommutative && B == C)) {
421 // Consider forming "(A op C) op' B".
422 // If "A op C" simplifies then it can be formed with no cost.
423 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, TD);
424 // If "A op C" doesn't simplify then only go on if both of the existing
425 // operations "A op' B" and "C op' D" will be zapped as no longer used.
426 if (!V && Op0->hasOneUse() && Op1->hasOneUse())
427 V = Builder->CreateBinOp(TopLevelOpcode, A, C, Op0->getName());
430 V = Builder->CreateBinOp(InnerOpcode, V, B);
438 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
439 // The instruction has the form "(A op' B) op C". See if expanding it out
440 // to "(A op C) op' (B op C)" results in simplifications.
441 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
442 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
444 // Do "A op C" and "B op C" both simplify?
445 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, TD))
446 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, TD)) {
447 // They do! Return "L op' R".
449 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
450 if ((L == A && R == B) ||
451 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
453 // Otherwise return "L op' R" if it simplifies.
454 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
456 // Otherwise, create a new instruction.
457 C = Builder->CreateBinOp(InnerOpcode, L, R);
463 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
464 // The instruction has the form "A op (B op' C)". See if expanding it out
465 // to "(A op B) op' (A op C)" results in simplifications.
466 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
467 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
469 // Do "A op B" and "A op C" both simplify?
470 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, TD))
471 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, TD)) {
472 // They do! Return "L op' R".
474 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
475 if ((L == B && R == C) ||
476 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
478 // Otherwise return "L op' R" if it simplifies.
479 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
481 // Otherwise, create a new instruction.
482 A = Builder->CreateBinOp(InnerOpcode, L, R);
491 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
492 // if the LHS is a constant zero (which is the 'negate' form).
494 Value *InstCombiner::dyn_castNegVal(Value *V) const {
495 if (BinaryOperator::isNeg(V))
496 return BinaryOperator::getNegArgument(V);
498 // Constants can be considered to be negated values if they can be folded.
499 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
500 return ConstantExpr::getNeg(C);
502 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
503 if (C->getType()->getElementType()->isIntegerTy())
504 return ConstantExpr::getNeg(C);
509 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
510 // instruction if the LHS is a constant negative zero (which is the 'negate'
513 Value *InstCombiner::dyn_castFNegVal(Value *V) const {
514 if (BinaryOperator::isFNeg(V))
515 return BinaryOperator::getFNegArgument(V);
517 // Constants can be considered to be negated values if they can be folded.
518 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
519 return ConstantExpr::getFNeg(C);
521 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
522 if (C->getType()->getElementType()->isFloatingPointTy())
523 return ConstantExpr::getFNeg(C);
528 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
530 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
531 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
534 // Figure out if the constant is the left or the right argument.
535 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
536 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
538 if (Constant *SOC = dyn_cast<Constant>(SO)) {
540 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
541 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
544 Value *Op0 = SO, *Op1 = ConstOperand;
548 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
549 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
550 SO->getName()+".op");
551 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
552 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
553 SO->getName()+".cmp");
554 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
555 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
556 SO->getName()+".cmp");
557 llvm_unreachable("Unknown binary instruction type!");
560 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
561 // constant as the other operand, try to fold the binary operator into the
562 // select arguments. This also works for Cast instructions, which obviously do
563 // not have a second operand.
564 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
565 // Don't modify shared select instructions
566 if (!SI->hasOneUse()) return 0;
567 Value *TV = SI->getOperand(1);
568 Value *FV = SI->getOperand(2);
570 if (isa<Constant>(TV) || isa<Constant>(FV)) {
571 // Bool selects with constant operands can be folded to logical ops.
572 if (SI->getType()->isIntegerTy(1)) return 0;
574 // If it's a bitcast involving vectors, make sure it has the same number of
575 // elements on both sides.
576 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
577 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
578 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
580 // Verify that either both or neither are vectors.
581 if ((SrcTy == NULL) != (DestTy == NULL)) return 0;
582 // If vectors, verify that they have the same number of elements.
583 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
587 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
588 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
590 return SelectInst::Create(SI->getCondition(),
591 SelectTrueVal, SelectFalseVal);
597 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
598 /// has a PHI node as operand #0, see if we can fold the instruction into the
599 /// PHI (which is only possible if all operands to the PHI are constants).
601 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
602 PHINode *PN = cast<PHINode>(I.getOperand(0));
603 unsigned NumPHIValues = PN->getNumIncomingValues();
604 if (NumPHIValues == 0)
607 // We normally only transform phis with a single use. However, if a PHI has
608 // multiple uses and they are all the same operation, we can fold *all* of the
609 // uses into the PHI.
610 if (!PN->hasOneUse()) {
611 // Walk the use list for the instruction, comparing them to I.
612 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
614 Instruction *User = cast<Instruction>(*UI);
615 if (User != &I && !I.isIdenticalTo(User))
618 // Otherwise, we can replace *all* users with the new PHI we form.
621 // Check to see if all of the operands of the PHI are simple constants
622 // (constantint/constantfp/undef). If there is one non-constant value,
623 // remember the BB it is in. If there is more than one or if *it* is a PHI,
624 // bail out. We don't do arbitrary constant expressions here because moving
625 // their computation can be expensive without a cost model.
626 BasicBlock *NonConstBB = 0;
627 for (unsigned i = 0; i != NumPHIValues; ++i) {
628 Value *InVal = PN->getIncomingValue(i);
629 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
632 if (isa<PHINode>(InVal)) return 0; // Itself a phi.
633 if (NonConstBB) return 0; // More than one non-const value.
635 NonConstBB = PN->getIncomingBlock(i);
637 // If the InVal is an invoke at the end of the pred block, then we can't
638 // insert a computation after it without breaking the edge.
639 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
640 if (II->getParent() == NonConstBB)
643 // If the incoming non-constant value is in I's block, we will remove one
644 // instruction, but insert another equivalent one, leading to infinite
646 if (NonConstBB == I.getParent())
650 // If there is exactly one non-constant value, we can insert a copy of the
651 // operation in that block. However, if this is a critical edge, we would be
652 // inserting the computation one some other paths (e.g. inside a loop). Only
653 // do this if the pred block is unconditionally branching into the phi block.
654 if (NonConstBB != 0) {
655 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
656 if (!BI || !BI->isUnconditional()) return 0;
659 // Okay, we can do the transformation: create the new PHI node.
660 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
661 InsertNewInstBefore(NewPN, *PN);
664 // If we are going to have to insert a new computation, do so right before the
665 // predecessors terminator.
667 Builder->SetInsertPoint(NonConstBB->getTerminator());
669 // Next, add all of the operands to the PHI.
670 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
671 // We only currently try to fold the condition of a select when it is a phi,
672 // not the true/false values.
673 Value *TrueV = SI->getTrueValue();
674 Value *FalseV = SI->getFalseValue();
675 BasicBlock *PhiTransBB = PN->getParent();
676 for (unsigned i = 0; i != NumPHIValues; ++i) {
677 BasicBlock *ThisBB = PN->getIncomingBlock(i);
678 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
679 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
681 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
682 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
684 InV = Builder->CreateSelect(PN->getIncomingValue(i),
685 TrueVInPred, FalseVInPred, "phitmp");
686 NewPN->addIncoming(InV, ThisBB);
688 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
689 Constant *C = cast<Constant>(I.getOperand(1));
690 for (unsigned i = 0; i != NumPHIValues; ++i) {
692 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
693 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
694 else if (isa<ICmpInst>(CI))
695 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
698 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
700 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
702 } else if (I.getNumOperands() == 2) {
703 Constant *C = cast<Constant>(I.getOperand(1));
704 for (unsigned i = 0; i != NumPHIValues; ++i) {
706 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
707 InV = ConstantExpr::get(I.getOpcode(), InC, C);
709 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
710 PN->getIncomingValue(i), C, "phitmp");
711 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
714 CastInst *CI = cast<CastInst>(&I);
715 Type *RetTy = CI->getType();
716 for (unsigned i = 0; i != NumPHIValues; ++i) {
718 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
719 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
721 InV = Builder->CreateCast(CI->getOpcode(),
722 PN->getIncomingValue(i), I.getType(), "phitmp");
723 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
727 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
729 Instruction *User = cast<Instruction>(*UI++);
730 if (User == &I) continue;
731 ReplaceInstUsesWith(*User, NewPN);
732 EraseInstFromFunction(*User);
734 return ReplaceInstUsesWith(I, NewPN);
737 /// FindElementAtOffset - Given a type and a constant offset, determine whether
738 /// or not there is a sequence of GEP indices into the type that will land us at
739 /// the specified offset. If so, fill them into NewIndices and return the
740 /// resultant element type, otherwise return null.
741 Type *InstCombiner::FindElementAtOffset(Type *Ty, int64_t Offset,
742 SmallVectorImpl<Value*> &NewIndices) {
744 if (!Ty->isSized()) return 0;
746 // Start with the index over the outer type. Note that the type size
747 // might be zero (even if the offset isn't zero) if the indexed type
748 // is something like [0 x {int, int}]
749 Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
750 int64_t FirstIdx = 0;
751 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
752 FirstIdx = Offset/TySize;
753 Offset -= FirstIdx*TySize;
755 // Handle hosts where % returns negative instead of values [0..TySize).
761 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
764 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
766 // Index into the types. If we fail, set OrigBase to null.
768 // Indexing into tail padding between struct/array elements.
769 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
772 if (StructType *STy = dyn_cast<StructType>(Ty)) {
773 const StructLayout *SL = TD->getStructLayout(STy);
774 assert(Offset < (int64_t)SL->getSizeInBytes() &&
775 "Offset must stay within the indexed type");
777 unsigned Elt = SL->getElementContainingOffset(Offset);
778 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
781 Offset -= SL->getElementOffset(Elt);
782 Ty = STy->getElementType(Elt);
783 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
784 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
785 assert(EltSize && "Cannot index into a zero-sized array");
786 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
788 Ty = AT->getElementType();
790 // Otherwise, we can't index into the middle of this atomic type, bail.
798 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
799 // If this GEP has only 0 indices, it is the same pointer as
800 // Src. If Src is not a trivial GEP too, don't combine
802 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
808 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
809 /// the multiplication is known not to overflow then NoSignedWrap is set.
810 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
811 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
812 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
813 Scale.getBitWidth() && "Scale not compatible with value!");
815 // If Val is zero or Scale is one then Val = Val * Scale.
816 if (match(Val, m_Zero()) || Scale == 1) {
821 // If Scale is zero then it does not divide Val.
822 if (Scale.isMinValue())
825 // Look through chains of multiplications, searching for a constant that is
826 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
827 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
828 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
831 // Val = M1 * X || Analysis starts here and works down
832 // M1 = M2 * Y || Doesn't descend into terms with more
833 // M2 = Z * 4 \/ than one use
835 // Then to modify a term at the bottom:
838 // M1 = Z * Y || Replaced M2 with Z
840 // Then to work back up correcting nsw flags.
842 // Op - the term we are currently analyzing. Starts at Val then drills down.
843 // Replaced with its descaled value before exiting from the drill down loop.
846 // Parent - initially null, but after drilling down notes where Op came from.
847 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
848 // 0'th operand of Val.
849 std::pair<Instruction*, unsigned> Parent;
851 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
852 // levels that doesn't overflow.
853 bool RequireNoSignedWrap = false;
855 // logScale - log base 2 of the scale. Negative if not a power of 2.
856 int32_t logScale = Scale.exactLogBase2();
858 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
860 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
861 // If Op is a constant divisible by Scale then descale to the quotient.
862 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
863 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
864 if (!Remainder.isMinValue())
865 // Not divisible by Scale.
867 // Replace with the quotient in the parent.
868 Op = ConstantInt::get(CI->getType(), Quotient);
873 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
875 if (BO->getOpcode() == Instruction::Mul) {
877 NoSignedWrap = BO->hasNoSignedWrap();
878 if (RequireNoSignedWrap && !NoSignedWrap)
881 // There are three cases for multiplication: multiplication by exactly
882 // the scale, multiplication by a constant different to the scale, and
883 // multiplication by something else.
884 Value *LHS = BO->getOperand(0);
885 Value *RHS = BO->getOperand(1);
887 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
888 // Multiplication by a constant.
889 if (CI->getValue() == Scale) {
890 // Multiplication by exactly the scale, replace the multiplication
891 // by its left-hand side in the parent.
896 // Otherwise drill down into the constant.
897 if (!Op->hasOneUse())
900 Parent = std::make_pair(BO, 1);
904 // Multiplication by something else. Drill down into the left-hand side
905 // since that's where the reassociate pass puts the good stuff.
906 if (!Op->hasOneUse())
909 Parent = std::make_pair(BO, 0);
913 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
914 isa<ConstantInt>(BO->getOperand(1))) {
915 // Multiplication by a power of 2.
916 NoSignedWrap = BO->hasNoSignedWrap();
917 if (RequireNoSignedWrap && !NoSignedWrap)
920 Value *LHS = BO->getOperand(0);
921 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
922 getLimitedValue(Scale.getBitWidth());
925 if (Amt == logScale) {
926 // Multiplication by exactly the scale, replace the multiplication
927 // by its left-hand side in the parent.
931 if (Amt < logScale || !Op->hasOneUse())
934 // Multiplication by more than the scale. Reduce the multiplying amount
935 // by the scale in the parent.
936 Parent = std::make_pair(BO, 1);
937 Op = ConstantInt::get(BO->getType(), Amt - logScale);
942 if (!Op->hasOneUse())
945 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
946 if (Cast->getOpcode() == Instruction::SExt) {
947 // Op is sign-extended from a smaller type, descale in the smaller type.
948 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
949 APInt SmallScale = Scale.trunc(SmallSize);
950 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
951 // descale Op as (sext Y) * Scale. In order to have
952 // sext (Y * SmallScale) = (sext Y) * Scale
953 // some conditions need to hold however: SmallScale must sign-extend to
954 // Scale and the multiplication Y * SmallScale should not overflow.
955 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
956 // SmallScale does not sign-extend to Scale.
958 assert(SmallScale.exactLogBase2() == logScale);
959 // Require that Y * SmallScale must not overflow.
960 RequireNoSignedWrap = true;
962 // Drill down through the cast.
963 Parent = std::make_pair(Cast, 0);
968 if (Cast->getOperand(0)) {
969 // Op is truncated from a larger type, descale in the larger type.
970 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
971 // trunc (Y * sext Scale) = (trunc Y) * Scale
972 // always holds. However (trunc Y) * Scale may overflow even if
973 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
974 // from this point up in the expression (see later).
975 if (RequireNoSignedWrap)
978 // Drill down through the cast.
979 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
980 Parent = std::make_pair(Cast, 0);
981 Scale = Scale.sext(LargeSize);
982 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
984 assert(Scale.exactLogBase2() == logScale);
989 // Unsupported expression, bail out.
993 // We know that we can successfully descale, so from here on we can safely
994 // modify the IR. Op holds the descaled version of the deepest term in the
995 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
999 // The expression only had one term.
1002 // Rewrite the parent using the descaled version of its operand.
1003 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1004 assert(Op != Parent.first->getOperand(Parent.second) &&
1005 "Descaling was a no-op?");
1006 Parent.first->setOperand(Parent.second, Op);
1007 Worklist.Add(Parent.first);
1009 // Now work back up the expression correcting nsw flags. The logic is based
1010 // on the following observation: if X * Y is known not to overflow as a signed
1011 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1012 // then X * Z will not overflow as a signed multiplication either. As we work
1013 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1014 // current level has strictly smaller absolute value than the original.
1015 Instruction *Ancestor = Parent.first;
1017 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1018 // If the multiplication wasn't nsw then we can't say anything about the
1019 // value of the descaled multiplication, and we have to clear nsw flags
1020 // from this point on up.
1021 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1022 NoSignedWrap &= OpNoSignedWrap;
1023 if (NoSignedWrap != OpNoSignedWrap) {
1024 BO->setHasNoSignedWrap(NoSignedWrap);
1025 Worklist.Add(Ancestor);
1027 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1028 // The fact that the descaled input to the trunc has smaller absolute
1029 // value than the original input doesn't tell us anything useful about
1030 // the absolute values of the truncations.
1031 NoSignedWrap = false;
1033 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1034 "Failed to keep proper track of nsw flags while drilling down?");
1036 if (Ancestor == Val)
1037 // Got to the top, all done!
1040 // Move up one level in the expression.
1041 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1042 Ancestor = Ancestor->use_back();
1046 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1047 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1049 if (Value *V = SimplifyGEPInst(Ops, TD))
1050 return ReplaceInstUsesWith(GEP, V);
1052 Value *PtrOp = GEP.getOperand(0);
1054 // Eliminate unneeded casts for indices, and replace indices which displace
1055 // by multiples of a zero size type with zero.
1057 bool MadeChange = false;
1058 Type *IntPtrTy = TD->getIntPtrType(GEP.getContext());
1060 gep_type_iterator GTI = gep_type_begin(GEP);
1061 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
1062 I != E; ++I, ++GTI) {
1063 // Skip indices into struct types.
1064 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1065 if (!SeqTy) continue;
1067 // If the element type has zero size then any index over it is equivalent
1068 // to an index of zero, so replace it with zero if it is not zero already.
1069 if (SeqTy->getElementType()->isSized() &&
1070 TD->getTypeAllocSize(SeqTy->getElementType()) == 0)
1071 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1072 *I = Constant::getNullValue(IntPtrTy);
1076 Type *IndexTy = (*I)->getType();
1077 if (IndexTy != IntPtrTy && !IndexTy->isVectorTy()) {
1078 // If we are using a wider index than needed for this platform, shrink
1079 // it to what we need. If narrower, sign-extend it to what we need.
1080 // This explicit cast can make subsequent optimizations more obvious.
1081 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1085 if (MadeChange) return &GEP;
1088 // Combine Indices - If the source pointer to this getelementptr instruction
1089 // is a getelementptr instruction, combine the indices of the two
1090 // getelementptr instructions into a single instruction.
1092 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1093 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1096 // Note that if our source is a gep chain itself then we wait for that
1097 // chain to be resolved before we perform this transformation. This
1098 // avoids us creating a TON of code in some cases.
1099 if (GEPOperator *SrcGEP =
1100 dyn_cast<GEPOperator>(Src->getOperand(0)))
1101 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1102 return 0; // Wait until our source is folded to completion.
1104 SmallVector<Value*, 8> Indices;
1106 // Find out whether the last index in the source GEP is a sequential idx.
1107 bool EndsWithSequential = false;
1108 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1110 EndsWithSequential = !(*I)->isStructTy();
1112 // Can we combine the two pointer arithmetics offsets?
1113 if (EndsWithSequential) {
1114 // Replace: gep (gep %P, long B), long A, ...
1115 // With: T = long A+B; gep %P, T, ...
1118 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1119 Value *GO1 = GEP.getOperand(1);
1120 if (SO1 == Constant::getNullValue(SO1->getType())) {
1122 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1125 // If they aren't the same type, then the input hasn't been processed
1126 // by the loop above yet (which canonicalizes sequential index types to
1127 // intptr_t). Just avoid transforming this until the input has been
1129 if (SO1->getType() != GO1->getType())
1131 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1134 // Update the GEP in place if possible.
1135 if (Src->getNumOperands() == 2) {
1136 GEP.setOperand(0, Src->getOperand(0));
1137 GEP.setOperand(1, Sum);
1140 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1141 Indices.push_back(Sum);
1142 Indices.append(GEP.op_begin()+2, GEP.op_end());
1143 } else if (isa<Constant>(*GEP.idx_begin()) &&
1144 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1145 Src->getNumOperands() != 1) {
1146 // Otherwise we can do the fold if the first index of the GEP is a zero
1147 Indices.append(Src->op_begin()+1, Src->op_end());
1148 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1151 if (!Indices.empty())
1152 return (GEP.isInBounds() && Src->isInBounds()) ?
1153 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
1155 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
1158 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1159 Value *StrippedPtr = PtrOp->stripPointerCasts();
1160 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1162 // We do not handle pointer-vector geps here.
1166 if (StrippedPtr != PtrOp &&
1167 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
1169 bool HasZeroPointerIndex = false;
1170 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1171 HasZeroPointerIndex = C->isZero();
1173 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1174 // into : GEP [10 x i8]* X, i32 0, ...
1176 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1177 // into : GEP i8* X, ...
1179 // This occurs when the program declares an array extern like "int X[];"
1180 if (HasZeroPointerIndex) {
1181 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1182 if (ArrayType *CATy =
1183 dyn_cast<ArrayType>(CPTy->getElementType())) {
1184 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1185 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1186 // -> GEP i8* X, ...
1187 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1188 GetElementPtrInst *Res =
1189 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
1190 Res->setIsInBounds(GEP.isInBounds());
1194 if (ArrayType *XATy =
1195 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1196 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1197 if (CATy->getElementType() == XATy->getElementType()) {
1198 // -> GEP [10 x i8]* X, i32 0, ...
1199 // At this point, we know that the cast source type is a pointer
1200 // to an array of the same type as the destination pointer
1201 // array. Because the array type is never stepped over (there
1202 // is a leading zero) we can fold the cast into this GEP.
1203 GEP.setOperand(0, StrippedPtr);
1208 } else if (GEP.getNumOperands() == 2) {
1209 // Transform things like:
1210 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1211 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1212 Type *SrcElTy = StrippedPtrTy->getElementType();
1213 Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
1214 if (TD && SrcElTy->isArrayTy() &&
1215 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
1216 TD->getTypeAllocSize(ResElTy)) {
1218 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
1219 Idx[1] = GEP.getOperand(1);
1220 Value *NewGEP = GEP.isInBounds() ?
1221 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1222 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1223 // V and GEP are both pointer types --> BitCast
1224 return new BitCastInst(NewGEP, GEP.getType());
1227 // Transform things like:
1228 // %V = mul i64 %N, 4
1229 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1230 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1231 if (TD && ResElTy->isSized() && SrcElTy->isSized()) {
1232 // Check that changing the type amounts to dividing the index by a scale
1234 uint64_t ResSize = TD->getTypeAllocSize(ResElTy);
1235 uint64_t SrcSize = TD->getTypeAllocSize(SrcElTy);
1236 if (ResSize && SrcSize % ResSize == 0) {
1237 Value *Idx = GEP.getOperand(1);
1238 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1239 uint64_t Scale = SrcSize / ResSize;
1241 // Earlier transforms ensure that the index has type IntPtrType, which
1242 // considerably simplifies the logic by eliminating implicit casts.
1243 assert(Idx->getType() == TD->getIntPtrType(GEP.getContext()) &&
1244 "Index not cast to pointer width?");
1247 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1248 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1249 // If the multiplication NewIdx * Scale may overflow then the new
1250 // GEP may not be "inbounds".
1251 Value *NewGEP = GEP.isInBounds() && NSW ?
1252 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
1253 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
1254 // The NewGEP must be pointer typed, so must the old one -> BitCast
1255 return new BitCastInst(NewGEP, GEP.getType());
1260 // Similarly, transform things like:
1261 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1262 // (where tmp = 8*tmp2) into:
1263 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1264 if (TD && ResElTy->isSized() && SrcElTy->isSized() &&
1265 SrcElTy->isArrayTy()) {
1266 // Check that changing to the array element type amounts to dividing the
1267 // index by a scale factor.
1268 uint64_t ResSize = TD->getTypeAllocSize(ResElTy);
1269 uint64_t ArrayEltSize =
1270 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
1271 if (ResSize && ArrayEltSize % ResSize == 0) {
1272 Value *Idx = GEP.getOperand(1);
1273 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1274 uint64_t Scale = ArrayEltSize / ResSize;
1276 // Earlier transforms ensure that the index has type IntPtrType, which
1277 // considerably simplifies the logic by eliminating implicit casts.
1278 assert(Idx->getType() == TD->getIntPtrType(GEP.getContext()) &&
1279 "Index not cast to pointer width?");
1282 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1283 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1284 // If the multiplication NewIdx * Scale may overflow then the new
1285 // GEP may not be "inbounds".
1287 Off[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
1289 Value *NewGEP = GEP.isInBounds() && NSW ?
1290 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
1291 Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
1292 // The NewGEP must be pointer typed, so must the old one -> BitCast
1293 return new BitCastInst(NewGEP, GEP.getType());
1300 /// See if we can simplify:
1301 /// X = bitcast A* to B*
1302 /// Y = gep X, <...constant indices...>
1303 /// into a gep of the original struct. This is important for SROA and alias
1304 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1305 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1307 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices() &&
1308 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
1310 // Determine how much the GEP moves the pointer.
1311 SmallVector<Value*, 8> Ops(GEP.idx_begin(), GEP.idx_end());
1312 int64_t Offset = TD->getIndexedOffset(GEP.getPointerOperandType(), Ops);
1314 // If this GEP instruction doesn't move the pointer, just replace the GEP
1315 // with a bitcast of the real input to the dest type.
1317 // If the bitcast is of an allocation, and the allocation will be
1318 // converted to match the type of the cast, don't touch this.
1319 if (isa<AllocaInst>(BCI->getOperand(0)) ||
1320 isAllocationFn(BCI->getOperand(0), TLI)) {
1321 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1322 if (Instruction *I = visitBitCast(*BCI)) {
1325 BCI->getParent()->getInstList().insert(BCI, I);
1326 ReplaceInstUsesWith(*BCI, I);
1331 return new BitCastInst(BCI->getOperand(0), GEP.getType());
1334 // Otherwise, if the offset is non-zero, we need to find out if there is a
1335 // field at Offset in 'A's type. If so, we can pull the cast through the
1337 SmallVector<Value*, 8> NewIndices;
1339 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
1340 if (FindElementAtOffset(InTy, Offset, NewIndices)) {
1341 Value *NGEP = GEP.isInBounds() ?
1342 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices) :
1343 Builder->CreateGEP(BCI->getOperand(0), NewIndices);
1345 if (NGEP->getType() == GEP.getType())
1346 return ReplaceInstUsesWith(GEP, NGEP);
1347 NGEP->takeName(&GEP);
1348 return new BitCastInst(NGEP, GEP.getType());
1359 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1360 const TargetLibraryInfo *TLI) {
1361 SmallVector<Instruction*, 4> Worklist;
1362 Worklist.push_back(AI);
1365 Instruction *PI = Worklist.pop_back_val();
1366 for (Value::use_iterator UI = PI->use_begin(), UE = PI->use_end(); UI != UE;
1368 Instruction *I = cast<Instruction>(*UI);
1369 switch (I->getOpcode()) {
1371 // Give up the moment we see something we can't handle.
1374 case Instruction::BitCast:
1375 case Instruction::GetElementPtr:
1377 Worklist.push_back(I);
1380 case Instruction::ICmp: {
1381 ICmpInst *ICI = cast<ICmpInst>(I);
1382 // We can fold eq/ne comparisons with null to false/true, respectively.
1383 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1389 case Instruction::Call:
1390 // Ignore no-op and store intrinsics.
1391 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1392 switch (II->getIntrinsicID()) {
1396 case Intrinsic::memmove:
1397 case Intrinsic::memcpy:
1398 case Intrinsic::memset: {
1399 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1400 if (MI->isVolatile() || MI->getRawDest() != PI)
1404 case Intrinsic::dbg_declare:
1405 case Intrinsic::dbg_value:
1406 case Intrinsic::invariant_start:
1407 case Intrinsic::invariant_end:
1408 case Intrinsic::lifetime_start:
1409 case Intrinsic::lifetime_end:
1410 case Intrinsic::objectsize:
1416 if (isFreeCall(I, TLI)) {
1422 case Instruction::Store: {
1423 StoreInst *SI = cast<StoreInst>(I);
1424 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1430 llvm_unreachable("missing a return?");
1432 } while (!Worklist.empty());
1436 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1437 // If we have a malloc call which is only used in any amount of comparisons
1438 // to null and free calls, delete the calls and replace the comparisons with
1439 // true or false as appropriate.
1440 SmallVector<WeakVH, 64> Users;
1441 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1442 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1443 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1446 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1447 ReplaceInstUsesWith(*C,
1448 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1449 C->isFalseWhenEqual()));
1450 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1451 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1452 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1453 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1454 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1455 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1456 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1459 EraseInstFromFunction(*I);
1462 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1463 // Replace invoke with a NOP intrinsic to maintain the original CFG
1464 Module *M = II->getParent()->getParent()->getParent();
1465 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1466 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1467 ArrayRef<Value *>(), "", II->getParent());
1469 return EraseInstFromFunction(MI);
1476 Instruction *InstCombiner::visitFree(CallInst &FI) {
1477 Value *Op = FI.getArgOperand(0);
1479 // free undef -> unreachable.
1480 if (isa<UndefValue>(Op)) {
1481 // Insert a new store to null because we cannot modify the CFG here.
1482 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
1483 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
1484 return EraseInstFromFunction(FI);
1487 // If we have 'free null' delete the instruction. This can happen in stl code
1488 // when lots of inlining happens.
1489 if (isa<ConstantPointerNull>(Op))
1490 return EraseInstFromFunction(FI);
1497 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
1498 // Change br (not X), label True, label False to: br X, label False, True
1500 BasicBlock *TrueDest;
1501 BasicBlock *FalseDest;
1502 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
1503 !isa<Constant>(X)) {
1504 // Swap Destinations and condition...
1506 BI.swapSuccessors();
1510 // Cannonicalize fcmp_one -> fcmp_oeq
1511 FCmpInst::Predicate FPred; Value *Y;
1512 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
1513 TrueDest, FalseDest)) &&
1514 BI.getCondition()->hasOneUse())
1515 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
1516 FPred == FCmpInst::FCMP_OGE) {
1517 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
1518 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
1520 // Swap Destinations and condition.
1521 BI.swapSuccessors();
1526 // Cannonicalize icmp_ne -> icmp_eq
1527 ICmpInst::Predicate IPred;
1528 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
1529 TrueDest, FalseDest)) &&
1530 BI.getCondition()->hasOneUse())
1531 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
1532 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
1533 IPred == ICmpInst::ICMP_SGE) {
1534 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
1535 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
1536 // Swap Destinations and condition.
1537 BI.swapSuccessors();
1545 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
1546 Value *Cond = SI.getCondition();
1547 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
1548 if (I->getOpcode() == Instruction::Add)
1549 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1550 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
1551 // Skip the first item since that's the default case.
1552 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
1554 ConstantInt* CaseVal = i.getCaseValue();
1555 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal),
1557 assert(isa<ConstantInt>(NewCaseVal) &&
1558 "Result of expression should be constant");
1559 i.setValue(cast<ConstantInt>(NewCaseVal));
1561 SI.setCondition(I->getOperand(0));
1569 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
1570 Value *Agg = EV.getAggregateOperand();
1572 if (!EV.hasIndices())
1573 return ReplaceInstUsesWith(EV, Agg);
1575 if (Constant *C = dyn_cast<Constant>(Agg)) {
1576 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
1577 if (EV.getNumIndices() == 0)
1578 return ReplaceInstUsesWith(EV, C2);
1579 // Extract the remaining indices out of the constant indexed by the
1581 return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
1583 return 0; // Can't handle other constants
1586 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
1587 // We're extracting from an insertvalue instruction, compare the indices
1588 const unsigned *exti, *exte, *insi, *inse;
1589 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
1590 exte = EV.idx_end(), inse = IV->idx_end();
1591 exti != exte && insi != inse;
1594 // The insert and extract both reference distinctly different elements.
1595 // This means the extract is not influenced by the insert, and we can
1596 // replace the aggregate operand of the extract with the aggregate
1597 // operand of the insert. i.e., replace
1598 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1599 // %E = extractvalue { i32, { i32 } } %I, 0
1601 // %E = extractvalue { i32, { i32 } } %A, 0
1602 return ExtractValueInst::Create(IV->getAggregateOperand(),
1605 if (exti == exte && insi == inse)
1606 // Both iterators are at the end: Index lists are identical. Replace
1607 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1608 // %C = extractvalue { i32, { i32 } } %B, 1, 0
1610 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
1612 // The extract list is a prefix of the insert list. i.e. replace
1613 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1614 // %E = extractvalue { i32, { i32 } } %I, 1
1616 // %X = extractvalue { i32, { i32 } } %A, 1
1617 // %E = insertvalue { i32 } %X, i32 42, 0
1618 // by switching the order of the insert and extract (though the
1619 // insertvalue should be left in, since it may have other uses).
1620 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
1622 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
1623 makeArrayRef(insi, inse));
1626 // The insert list is a prefix of the extract list
1627 // We can simply remove the common indices from the extract and make it
1628 // operate on the inserted value instead of the insertvalue result.
1630 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1631 // %E = extractvalue { i32, { i32 } } %I, 1, 0
1633 // %E extractvalue { i32 } { i32 42 }, 0
1634 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
1635 makeArrayRef(exti, exte));
1637 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
1638 // We're extracting from an intrinsic, see if we're the only user, which
1639 // allows us to simplify multiple result intrinsics to simpler things that
1640 // just get one value.
1641 if (II->hasOneUse()) {
1642 // Check if we're grabbing the overflow bit or the result of a 'with
1643 // overflow' intrinsic. If it's the latter we can remove the intrinsic
1644 // and replace it with a traditional binary instruction.
1645 switch (II->getIntrinsicID()) {
1646 case Intrinsic::uadd_with_overflow:
1647 case Intrinsic::sadd_with_overflow:
1648 if (*EV.idx_begin() == 0) { // Normal result.
1649 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1650 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
1651 EraseInstFromFunction(*II);
1652 return BinaryOperator::CreateAdd(LHS, RHS);
1655 // If the normal result of the add is dead, and the RHS is a constant,
1656 // we can transform this into a range comparison.
1657 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
1658 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
1659 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
1660 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
1661 ConstantExpr::getNot(CI));
1663 case Intrinsic::usub_with_overflow:
1664 case Intrinsic::ssub_with_overflow:
1665 if (*EV.idx_begin() == 0) { // Normal result.
1666 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1667 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
1668 EraseInstFromFunction(*II);
1669 return BinaryOperator::CreateSub(LHS, RHS);
1672 case Intrinsic::umul_with_overflow:
1673 case Intrinsic::smul_with_overflow:
1674 if (*EV.idx_begin() == 0) { // Normal result.
1675 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1676 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
1677 EraseInstFromFunction(*II);
1678 return BinaryOperator::CreateMul(LHS, RHS);
1686 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
1687 // If the (non-volatile) load only has one use, we can rewrite this to a
1688 // load from a GEP. This reduces the size of the load.
1689 // FIXME: If a load is used only by extractvalue instructions then this
1690 // could be done regardless of having multiple uses.
1691 if (L->isSimple() && L->hasOneUse()) {
1692 // extractvalue has integer indices, getelementptr has Value*s. Convert.
1693 SmallVector<Value*, 4> Indices;
1694 // Prefix an i32 0 since we need the first element.
1695 Indices.push_back(Builder->getInt32(0));
1696 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
1698 Indices.push_back(Builder->getInt32(*I));
1700 // We need to insert these at the location of the old load, not at that of
1701 // the extractvalue.
1702 Builder->SetInsertPoint(L->getParent(), L);
1703 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
1704 // Returning the load directly will cause the main loop to insert it in
1705 // the wrong spot, so use ReplaceInstUsesWith().
1706 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
1708 // We could simplify extracts from other values. Note that nested extracts may
1709 // already be simplified implicitly by the above: extract (extract (insert) )
1710 // will be translated into extract ( insert ( extract ) ) first and then just
1711 // the value inserted, if appropriate. Similarly for extracts from single-use
1712 // loads: extract (extract (load)) will be translated to extract (load (gep))
1713 // and if again single-use then via load (gep (gep)) to load (gep).
1714 // However, double extracts from e.g. function arguments or return values
1715 // aren't handled yet.
1719 enum Personality_Type {
1720 Unknown_Personality,
1721 GNU_Ada_Personality,
1722 GNU_CXX_Personality,
1723 GNU_ObjC_Personality
1726 /// RecognizePersonality - See if the given exception handling personality
1727 /// function is one that we understand. If so, return a description of it;
1728 /// otherwise return Unknown_Personality.
1729 static Personality_Type RecognizePersonality(Value *Pers) {
1730 Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
1732 return Unknown_Personality;
1733 return StringSwitch<Personality_Type>(F->getName())
1734 .Case("__gnat_eh_personality", GNU_Ada_Personality)
1735 .Case("__gxx_personality_v0", GNU_CXX_Personality)
1736 .Case("__objc_personality_v0", GNU_ObjC_Personality)
1737 .Default(Unknown_Personality);
1740 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
1741 static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
1742 switch (Personality) {
1743 case Unknown_Personality:
1745 case GNU_Ada_Personality:
1746 // While __gnat_all_others_value will match any Ada exception, it doesn't
1747 // match foreign exceptions (or didn't, before gcc-4.7).
1749 case GNU_CXX_Personality:
1750 case GNU_ObjC_Personality:
1751 return TypeInfo->isNullValue();
1753 llvm_unreachable("Unknown personality!");
1756 static bool shorter_filter(const Value *LHS, const Value *RHS) {
1758 cast<ArrayType>(LHS->getType())->getNumElements()
1760 cast<ArrayType>(RHS->getType())->getNumElements();
1763 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
1764 // The logic here should be correct for any real-world personality function.
1765 // However if that turns out not to be true, the offending logic can always
1766 // be conditioned on the personality function, like the catch-all logic is.
1767 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
1769 // Simplify the list of clauses, eg by removing repeated catch clauses
1770 // (these are often created by inlining).
1771 bool MakeNewInstruction = false; // If true, recreate using the following:
1772 SmallVector<Value *, 16> NewClauses; // - Clauses for the new instruction;
1773 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
1775 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
1776 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
1777 bool isLastClause = i + 1 == e;
1778 if (LI.isCatch(i)) {
1780 Value *CatchClause = LI.getClause(i);
1781 Constant *TypeInfo = cast<Constant>(CatchClause->stripPointerCasts());
1783 // If we already saw this clause, there is no point in having a second
1785 if (AlreadyCaught.insert(TypeInfo)) {
1786 // This catch clause was not already seen.
1787 NewClauses.push_back(CatchClause);
1789 // Repeated catch clause - drop the redundant copy.
1790 MakeNewInstruction = true;
1793 // If this is a catch-all then there is no point in keeping any following
1794 // clauses or marking the landingpad as having a cleanup.
1795 if (isCatchAll(Personality, TypeInfo)) {
1797 MakeNewInstruction = true;
1798 CleanupFlag = false;
1802 // A filter clause. If any of the filter elements were already caught
1803 // then they can be dropped from the filter. It is tempting to try to
1804 // exploit the filter further by saying that any typeinfo that does not
1805 // occur in the filter can't be caught later (and thus can be dropped).
1806 // However this would be wrong, since typeinfos can match without being
1807 // equal (for example if one represents a C++ class, and the other some
1808 // class derived from it).
1809 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
1810 Value *FilterClause = LI.getClause(i);
1811 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
1812 unsigned NumTypeInfos = FilterType->getNumElements();
1814 // An empty filter catches everything, so there is no point in keeping any
1815 // following clauses or marking the landingpad as having a cleanup. By
1816 // dealing with this case here the following code is made a bit simpler.
1817 if (!NumTypeInfos) {
1818 NewClauses.push_back(FilterClause);
1820 MakeNewInstruction = true;
1821 CleanupFlag = false;
1825 bool MakeNewFilter = false; // If true, make a new filter.
1826 SmallVector<Constant *, 16> NewFilterElts; // New elements.
1827 if (isa<ConstantAggregateZero>(FilterClause)) {
1828 // Not an empty filter - it contains at least one null typeinfo.
1829 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
1830 Constant *TypeInfo =
1831 Constant::getNullValue(FilterType->getElementType());
1832 // If this typeinfo is a catch-all then the filter can never match.
1833 if (isCatchAll(Personality, TypeInfo)) {
1834 // Throw the filter away.
1835 MakeNewInstruction = true;
1839 // There is no point in having multiple copies of this typeinfo, so
1840 // discard all but the first copy if there is more than one.
1841 NewFilterElts.push_back(TypeInfo);
1842 if (NumTypeInfos > 1)
1843 MakeNewFilter = true;
1845 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
1846 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
1847 NewFilterElts.reserve(NumTypeInfos);
1849 // Remove any filter elements that were already caught or that already
1850 // occurred in the filter. While there, see if any of the elements are
1851 // catch-alls. If so, the filter can be discarded.
1852 bool SawCatchAll = false;
1853 for (unsigned j = 0; j != NumTypeInfos; ++j) {
1854 Value *Elt = Filter->getOperand(j);
1855 Constant *TypeInfo = cast<Constant>(Elt->stripPointerCasts());
1856 if (isCatchAll(Personality, TypeInfo)) {
1857 // This element is a catch-all. Bail out, noting this fact.
1861 if (AlreadyCaught.count(TypeInfo))
1862 // Already caught by an earlier clause, so having it in the filter
1865 // There is no point in having multiple copies of the same typeinfo in
1866 // a filter, so only add it if we didn't already.
1867 if (SeenInFilter.insert(TypeInfo))
1868 NewFilterElts.push_back(cast<Constant>(Elt));
1870 // A filter containing a catch-all cannot match anything by definition.
1872 // Throw the filter away.
1873 MakeNewInstruction = true;
1877 // If we dropped something from the filter, make a new one.
1878 if (NewFilterElts.size() < NumTypeInfos)
1879 MakeNewFilter = true;
1881 if (MakeNewFilter) {
1882 FilterType = ArrayType::get(FilterType->getElementType(),
1883 NewFilterElts.size());
1884 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
1885 MakeNewInstruction = true;
1888 NewClauses.push_back(FilterClause);
1890 // If the new filter is empty then it will catch everything so there is
1891 // no point in keeping any following clauses or marking the landingpad
1892 // as having a cleanup. The case of the original filter being empty was
1893 // already handled above.
1894 if (MakeNewFilter && !NewFilterElts.size()) {
1895 assert(MakeNewInstruction && "New filter but not a new instruction!");
1896 CleanupFlag = false;
1902 // If several filters occur in a row then reorder them so that the shortest
1903 // filters come first (those with the smallest number of elements). This is
1904 // advantageous because shorter filters are more likely to match, speeding up
1905 // unwinding, but mostly because it increases the effectiveness of the other
1906 // filter optimizations below.
1907 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
1909 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
1910 for (j = i; j != e; ++j)
1911 if (!isa<ArrayType>(NewClauses[j]->getType()))
1914 // Check whether the filters are already sorted by length. We need to know
1915 // if sorting them is actually going to do anything so that we only make a
1916 // new landingpad instruction if it does.
1917 for (unsigned k = i; k + 1 < j; ++k)
1918 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
1919 // Not sorted, so sort the filters now. Doing an unstable sort would be
1920 // correct too but reordering filters pointlessly might confuse users.
1921 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
1923 MakeNewInstruction = true;
1927 // Look for the next batch of filters.
1931 // If typeinfos matched if and only if equal, then the elements of a filter L
1932 // that occurs later than a filter F could be replaced by the intersection of
1933 // the elements of F and L. In reality two typeinfos can match without being
1934 // equal (for example if one represents a C++ class, and the other some class
1935 // derived from it) so it would be wrong to perform this transform in general.
1936 // However the transform is correct and useful if F is a subset of L. In that
1937 // case L can be replaced by F, and thus removed altogether since repeating a
1938 // filter is pointless. So here we look at all pairs of filters F and L where
1939 // L follows F in the list of clauses, and remove L if every element of F is
1940 // an element of L. This can occur when inlining C++ functions with exception
1942 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
1943 // Examine each filter in turn.
1944 Value *Filter = NewClauses[i];
1945 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
1947 // Not a filter - skip it.
1949 unsigned FElts = FTy->getNumElements();
1950 // Examine each filter following this one. Doing this backwards means that
1951 // we don't have to worry about filters disappearing under us when removed.
1952 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
1953 Value *LFilter = NewClauses[j];
1954 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
1956 // Not a filter - skip it.
1958 // If Filter is a subset of LFilter, i.e. every element of Filter is also
1959 // an element of LFilter, then discard LFilter.
1960 SmallVector<Value *, 16>::iterator J = NewClauses.begin() + j;
1961 // If Filter is empty then it is a subset of LFilter.
1964 NewClauses.erase(J);
1965 MakeNewInstruction = true;
1966 // Move on to the next filter.
1969 unsigned LElts = LTy->getNumElements();
1970 // If Filter is longer than LFilter then it cannot be a subset of it.
1972 // Move on to the next filter.
1974 // At this point we know that LFilter has at least one element.
1975 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
1976 // Filter is a subset of LFilter iff Filter contains only zeros (as we
1977 // already know that Filter is not longer than LFilter).
1978 if (isa<ConstantAggregateZero>(Filter)) {
1979 assert(FElts <= LElts && "Should have handled this case earlier!");
1981 NewClauses.erase(J);
1982 MakeNewInstruction = true;
1984 // Move on to the next filter.
1987 ConstantArray *LArray = cast<ConstantArray>(LFilter);
1988 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
1989 // Since Filter is non-empty and contains only zeros, it is a subset of
1990 // LFilter iff LFilter contains a zero.
1991 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
1992 for (unsigned l = 0; l != LElts; ++l)
1993 if (LArray->getOperand(l)->isNullValue()) {
1994 // LFilter contains a zero - discard it.
1995 NewClauses.erase(J);
1996 MakeNewInstruction = true;
1999 // Move on to the next filter.
2002 // At this point we know that both filters are ConstantArrays. Loop over
2003 // operands to see whether every element of Filter is also an element of
2004 // LFilter. Since filters tend to be short this is probably faster than
2005 // using a method that scales nicely.
2006 ConstantArray *FArray = cast<ConstantArray>(Filter);
2007 bool AllFound = true;
2008 for (unsigned f = 0; f != FElts; ++f) {
2009 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2011 for (unsigned l = 0; l != LElts; ++l) {
2012 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2013 if (LTypeInfo == FTypeInfo) {
2023 NewClauses.erase(J);
2024 MakeNewInstruction = true;
2026 // Move on to the next filter.
2030 // If we changed any of the clauses, replace the old landingpad instruction
2032 if (MakeNewInstruction) {
2033 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2034 LI.getPersonalityFn(),
2036 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2037 NLI->addClause(NewClauses[i]);
2038 // A landing pad with no clauses must have the cleanup flag set. It is
2039 // theoretically possible, though highly unlikely, that we eliminated all
2040 // clauses. If so, force the cleanup flag to true.
2041 if (NewClauses.empty())
2043 NLI->setCleanup(CleanupFlag);
2047 // Even if none of the clauses changed, we may nonetheless have understood
2048 // that the cleanup flag is pointless. Clear it if so.
2049 if (LI.isCleanup() != CleanupFlag) {
2050 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2051 LI.setCleanup(CleanupFlag);
2061 /// TryToSinkInstruction - Try to move the specified instruction from its
2062 /// current block into the beginning of DestBlock, which can only happen if it's
2063 /// safe to move the instruction past all of the instructions between it and the
2064 /// end of its block.
2065 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2066 assert(I->hasOneUse() && "Invariants didn't hold!");
2068 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2069 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
2070 isa<TerminatorInst>(I))
2073 // Do not sink alloca instructions out of the entry block.
2074 if (isa<AllocaInst>(I) && I->getParent() ==
2075 &DestBlock->getParent()->getEntryBlock())
2078 // We can only sink load instructions if there is nothing between the load and
2079 // the end of block that could change the value.
2080 if (I->mayReadFromMemory()) {
2081 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2083 if (Scan->mayWriteToMemory())
2087 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2088 I->moveBefore(InsertPos);
2094 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2095 /// all reachable code to the worklist.
2097 /// This has a couple of tricks to make the code faster and more powerful. In
2098 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2099 /// them to the worklist (this significantly speeds up instcombine on code where
2100 /// many instructions are dead or constant). Additionally, if we find a branch
2101 /// whose condition is a known constant, we only visit the reachable successors.
2103 static bool AddReachableCodeToWorklist(BasicBlock *BB,
2104 SmallPtrSet<BasicBlock*, 64> &Visited,
2106 const DataLayout *TD,
2107 const TargetLibraryInfo *TLI) {
2108 bool MadeIRChange = false;
2109 SmallVector<BasicBlock*, 256> Worklist;
2110 Worklist.push_back(BB);
2112 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2113 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2116 BB = Worklist.pop_back_val();
2118 // We have now visited this block! If we've already been here, ignore it.
2119 if (!Visited.insert(BB)) continue;
2121 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2122 Instruction *Inst = BBI++;
2124 // DCE instruction if trivially dead.
2125 if (isInstructionTriviallyDead(Inst, TLI)) {
2127 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
2128 Inst->eraseFromParent();
2132 // ConstantProp instruction if trivially constant.
2133 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
2134 if (Constant *C = ConstantFoldInstruction(Inst, TD, TLI)) {
2135 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
2137 Inst->replaceAllUsesWith(C);
2139 Inst->eraseFromParent();
2144 // See if we can constant fold its operands.
2145 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
2147 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2148 if (CE == 0) continue;
2150 Constant*& FoldRes = FoldedConstants[CE];
2152 FoldRes = ConstantFoldConstantExpression(CE, TD, TLI);
2156 if (FoldRes != CE) {
2158 MadeIRChange = true;
2163 InstrsForInstCombineWorklist.push_back(Inst);
2166 // Recursively visit successors. If this is a branch or switch on a
2167 // constant, only visit the reachable successor.
2168 TerminatorInst *TI = BB->getTerminator();
2169 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2170 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2171 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2172 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2173 Worklist.push_back(ReachableBB);
2176 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2177 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2178 // See if this is an explicit destination.
2179 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2181 if (i.getCaseValue() == Cond) {
2182 BasicBlock *ReachableBB = i.getCaseSuccessor();
2183 Worklist.push_back(ReachableBB);
2187 // Otherwise it is the default destination.
2188 Worklist.push_back(SI->getDefaultDest());
2193 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
2194 Worklist.push_back(TI->getSuccessor(i));
2195 } while (!Worklist.empty());
2197 // Once we've found all of the instructions to add to instcombine's worklist,
2198 // add them in reverse order. This way instcombine will visit from the top
2199 // of the function down. This jives well with the way that it adds all uses
2200 // of instructions to the worklist after doing a transformation, thus avoiding
2201 // some N^2 behavior in pathological cases.
2202 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2203 InstrsForInstCombineWorklist.size());
2205 return MadeIRChange;
2208 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
2209 MadeIRChange = false;
2211 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
2212 << F.getName() << "\n");
2215 // Do a depth-first traversal of the function, populate the worklist with
2216 // the reachable instructions. Ignore blocks that are not reachable. Keep
2217 // track of which blocks we visit.
2218 SmallPtrSet<BasicBlock*, 64> Visited;
2219 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD,
2222 // Do a quick scan over the function. If we find any blocks that are
2223 // unreachable, remove any instructions inside of them. This prevents
2224 // the instcombine code from having to deal with some bad special cases.
2225 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2226 if (Visited.count(BB)) continue;
2228 // Delete the instructions backwards, as it has a reduced likelihood of
2229 // having to update as many def-use and use-def chains.
2230 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2231 while (EndInst != BB->begin()) {
2232 // Delete the next to last instruction.
2233 BasicBlock::iterator I = EndInst;
2234 Instruction *Inst = --I;
2235 if (!Inst->use_empty())
2236 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2237 if (isa<LandingPadInst>(Inst)) {
2241 if (!isa<DbgInfoIntrinsic>(Inst)) {
2243 MadeIRChange = true;
2245 Inst->eraseFromParent();
2250 while (!Worklist.isEmpty()) {
2251 Instruction *I = Worklist.RemoveOne();
2252 if (I == 0) continue; // skip null values.
2254 // Check to see if we can DCE the instruction.
2255 if (isInstructionTriviallyDead(I, TLI)) {
2256 DEBUG(errs() << "IC: DCE: " << *I << '\n');
2257 EraseInstFromFunction(*I);
2259 MadeIRChange = true;
2263 // Instruction isn't dead, see if we can constant propagate it.
2264 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
2265 if (Constant *C = ConstantFoldInstruction(I, TD, TLI)) {
2266 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2268 // Add operands to the worklist.
2269 ReplaceInstUsesWith(*I, C);
2271 EraseInstFromFunction(*I);
2272 MadeIRChange = true;
2276 // See if we can trivially sink this instruction to a successor basic block.
2277 if (I->hasOneUse()) {
2278 BasicBlock *BB = I->getParent();
2279 Instruction *UserInst = cast<Instruction>(I->use_back());
2280 BasicBlock *UserParent;
2282 // Get the block the use occurs in.
2283 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2284 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
2286 UserParent = UserInst->getParent();
2288 if (UserParent != BB) {
2289 bool UserIsSuccessor = false;
2290 // See if the user is one of our successors.
2291 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2292 if (*SI == UserParent) {
2293 UserIsSuccessor = true;
2297 // If the user is one of our immediate successors, and if that successor
2298 // only has us as a predecessors (we'd have to split the critical edge
2299 // otherwise), we can keep going.
2300 if (UserIsSuccessor && UserParent->getSinglePredecessor())
2301 // Okay, the CFG is simple enough, try to sink this instruction.
2302 MadeIRChange |= TryToSinkInstruction(I, UserParent);
2306 // Now that we have an instruction, try combining it to simplify it.
2307 Builder->SetInsertPoint(I->getParent(), I);
2308 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2313 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2314 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
2316 if (Instruction *Result = visit(*I)) {
2318 // Should we replace the old instruction with a new one?
2320 DEBUG(errs() << "IC: Old = " << *I << '\n'
2321 << " New = " << *Result << '\n');
2323 if (!I->getDebugLoc().isUnknown())
2324 Result->setDebugLoc(I->getDebugLoc());
2325 // Everything uses the new instruction now.
2326 I->replaceAllUsesWith(Result);
2328 // Move the name to the new instruction first.
2329 Result->takeName(I);
2331 // Push the new instruction and any users onto the worklist.
2332 Worklist.Add(Result);
2333 Worklist.AddUsersToWorkList(*Result);
2335 // Insert the new instruction into the basic block...
2336 BasicBlock *InstParent = I->getParent();
2337 BasicBlock::iterator InsertPos = I;
2339 // If we replace a PHI with something that isn't a PHI, fix up the
2341 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2342 InsertPos = InstParent->getFirstInsertionPt();
2344 InstParent->getInstList().insert(InsertPos, Result);
2346 EraseInstFromFunction(*I);
2349 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
2350 << " New = " << *I << '\n');
2353 // If the instruction was modified, it's possible that it is now dead.
2354 // if so, remove it.
2355 if (isInstructionTriviallyDead(I, TLI)) {
2356 EraseInstFromFunction(*I);
2359 Worklist.AddUsersToWorkList(*I);
2362 MadeIRChange = true;
2367 return MadeIRChange;
2371 bool InstCombiner::runOnFunction(Function &F) {
2372 TD = getAnalysisIfAvailable<DataLayout>();
2373 TLI = &getAnalysis<TargetLibraryInfo>();
2375 /// Builder - This is an IRBuilder that automatically inserts new
2376 /// instructions into the worklist when they are created.
2377 IRBuilder<true, TargetFolder, InstCombineIRInserter>
2378 TheBuilder(F.getContext(), TargetFolder(TD),
2379 InstCombineIRInserter(Worklist));
2380 Builder = &TheBuilder;
2382 LibCallSimplifier TheSimplifier(TD, TLI);
2383 Simplifier = &TheSimplifier;
2385 bool EverMadeChange = false;
2387 // Lower dbg.declare intrinsics otherwise their value may be clobbered
2389 EverMadeChange = LowerDbgDeclare(F);
2391 // Iterate while there is work to do.
2392 unsigned Iteration = 0;
2393 while (DoOneIteration(F, Iteration++))
2394 EverMadeChange = true;
2397 return EverMadeChange;
2400 FunctionPass *llvm::createInstructionCombiningPass() {
2401 return new InstCombiner();