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-c/Initialization.h"
40 #include "llvm/ADT/SmallPtrSet.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/ADT/StringSwitch.h"
43 #include "llvm/Analysis/ConstantFolding.h"
44 #include "llvm/Analysis/InstructionSimplify.h"
45 #include "llvm/Analysis/MemoryBuiltins.h"
46 #include "llvm/DataLayout.h"
47 #include "llvm/IntrinsicInst.h"
48 #include "llvm/Support/CFG.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/GetElementPtrTypeIterator.h"
52 #include "llvm/Support/PatternMatch.h"
53 #include "llvm/Support/ValueHandle.h"
54 #include "llvm/Target/TargetLibraryInfo.h"
55 #include "llvm/Transforms/Utils/Local.h"
59 using namespace llvm::PatternMatch;
61 STATISTIC(NumCombined , "Number of insts combined");
62 STATISTIC(NumConstProp, "Number of constant folds");
63 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
64 STATISTIC(NumSunkInst , "Number of instructions sunk");
65 STATISTIC(NumExpand, "Number of expansions");
66 STATISTIC(NumFactor , "Number of factorizations");
67 STATISTIC(NumReassoc , "Number of reassociations");
69 static cl::opt<bool> UnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
71 cl::desc("Enable unsafe double to float "
72 "shrinking for math lib calls"));
74 // Initialization Routines
75 void llvm::initializeInstCombine(PassRegistry &Registry) {
76 initializeInstCombinerPass(Registry);
79 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
80 initializeInstCombine(*unwrap(R));
83 char InstCombiner::ID = 0;
84 INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine",
85 "Combine redundant instructions", false, false)
86 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
87 INITIALIZE_PASS_END(InstCombiner, "instcombine",
88 "Combine redundant instructions", false, false)
90 void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
92 AU.addRequired<TargetLibraryInfo>();
96 Value *InstCombiner::EmitGEPOffset(User *GEP) {
97 return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP);
100 /// ShouldChangeType - Return true if it is desirable to convert a computation
101 /// from 'From' to 'To'. We don't want to convert from a legal to an illegal
102 /// type for example, or from a smaller to a larger illegal type.
103 bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
104 assert(From->isIntegerTy() && To->isIntegerTy());
106 // If we don't have TD, we don't know if the source/dest are legal.
107 if (!TD) return false;
109 unsigned FromWidth = From->getPrimitiveSizeInBits();
110 unsigned ToWidth = To->getPrimitiveSizeInBits();
111 bool FromLegal = TD->isLegalInteger(FromWidth);
112 bool ToLegal = TD->isLegalInteger(ToWidth);
114 // If this is a legal integer from type, and the result would be an illegal
115 // type, don't do the transformation.
116 if (FromLegal && !ToLegal)
119 // Otherwise, if both are illegal, do not increase the size of the result. We
120 // do allow things like i160 -> i64, but not i64 -> i160.
121 if (!FromLegal && !ToLegal && ToWidth > FromWidth)
127 // Return true, if No Signed Wrap should be maintained for I.
128 // The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
129 // where both B and C should be ConstantInts, results in a constant that does
130 // not overflow. This function only handles the Add and Sub opcodes. For
131 // all other opcodes, the function conservatively returns false.
132 static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
133 OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
134 if (!OBO || !OBO->hasNoSignedWrap()) {
138 // We reason about Add and Sub Only.
139 Instruction::BinaryOps Opcode = I.getOpcode();
140 if (Opcode != Instruction::Add &&
141 Opcode != Instruction::Sub) {
145 ConstantInt *CB = dyn_cast<ConstantInt>(B);
146 ConstantInt *CC = dyn_cast<ConstantInt>(C);
152 const APInt &BVal = CB->getValue();
153 const APInt &CVal = CC->getValue();
154 bool Overflow = false;
156 if (Opcode == Instruction::Add) {
157 BVal.sadd_ov(CVal, Overflow);
159 BVal.ssub_ov(CVal, Overflow);
165 /// SimplifyAssociativeOrCommutative - This performs a few simplifications for
166 /// operators which are associative or commutative:
168 // Commutative operators:
170 // 1. Order operands such that they are listed from right (least complex) to
171 // left (most complex). This puts constants before unary operators before
174 // Associative operators:
176 // 2. Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
177 // 3. Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
179 // Associative and commutative operators:
181 // 4. Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
182 // 5. Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
183 // 6. Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
184 // if C1 and C2 are constants.
186 bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) {
187 Instruction::BinaryOps Opcode = I.getOpcode();
188 bool Changed = false;
191 // Order operands such that they are listed from right (least complex) to
192 // left (most complex). This puts constants before unary operators before
194 if (I.isCommutative() && getComplexity(I.getOperand(0)) <
195 getComplexity(I.getOperand(1)))
196 Changed = !I.swapOperands();
198 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
199 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
201 if (I.isAssociative()) {
202 // Transform: "(A op B) op C" ==> "A op (B op C)" if "B op C" simplifies.
203 if (Op0 && Op0->getOpcode() == Opcode) {
204 Value *A = Op0->getOperand(0);
205 Value *B = Op0->getOperand(1);
206 Value *C = I.getOperand(1);
208 // Does "B op C" simplify?
209 if (Value *V = SimplifyBinOp(Opcode, B, C, TD)) {
210 // It simplifies to V. Form "A op V".
213 // Conservatively clear the optional flags, since they may not be
214 // preserved by the reassociation.
215 if (MaintainNoSignedWrap(I, B, C) &&
216 (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
217 // Note: this is only valid because SimplifyBinOp doesn't look at
218 // the operands to Op0.
219 I.clearSubclassOptionalData();
220 I.setHasNoSignedWrap(true);
222 I.clearSubclassOptionalData();
231 // Transform: "A op (B op C)" ==> "(A op B) op C" if "A op B" simplifies.
232 if (Op1 && Op1->getOpcode() == Opcode) {
233 Value *A = I.getOperand(0);
234 Value *B = Op1->getOperand(0);
235 Value *C = Op1->getOperand(1);
237 // Does "A op B" simplify?
238 if (Value *V = SimplifyBinOp(Opcode, A, B, TD)) {
239 // It simplifies to V. Form "V op C".
242 // Conservatively clear the optional flags, since they may not be
243 // preserved by the reassociation.
244 I.clearSubclassOptionalData();
252 if (I.isAssociative() && I.isCommutative()) {
253 // Transform: "(A op B) op C" ==> "(C op A) op B" if "C op A" simplifies.
254 if (Op0 && Op0->getOpcode() == Opcode) {
255 Value *A = Op0->getOperand(0);
256 Value *B = Op0->getOperand(1);
257 Value *C = I.getOperand(1);
259 // Does "C op A" simplify?
260 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
261 // It simplifies to V. Form "V op B".
264 // Conservatively clear the optional flags, since they may not be
265 // preserved by the reassociation.
266 I.clearSubclassOptionalData();
273 // Transform: "A op (B op C)" ==> "B op (C op A)" if "C op A" simplifies.
274 if (Op1 && Op1->getOpcode() == Opcode) {
275 Value *A = I.getOperand(0);
276 Value *B = Op1->getOperand(0);
277 Value *C = Op1->getOperand(1);
279 // Does "C op A" simplify?
280 if (Value *V = SimplifyBinOp(Opcode, C, A, TD)) {
281 // It simplifies to V. Form "B op V".
284 // Conservatively clear the optional flags, since they may not be
285 // preserved by the reassociation.
286 I.clearSubclassOptionalData();
293 // Transform: "(A op C1) op (B op C2)" ==> "(A op B) op (C1 op C2)"
294 // if C1 and C2 are constants.
296 Op0->getOpcode() == Opcode && Op1->getOpcode() == Opcode &&
297 isa<Constant>(Op0->getOperand(1)) &&
298 isa<Constant>(Op1->getOperand(1)) &&
299 Op0->hasOneUse() && Op1->hasOneUse()) {
300 Value *A = Op0->getOperand(0);
301 Constant *C1 = cast<Constant>(Op0->getOperand(1));
302 Value *B = Op1->getOperand(0);
303 Constant *C2 = cast<Constant>(Op1->getOperand(1));
305 Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
306 BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
307 InsertNewInstWith(New, I);
309 I.setOperand(0, New);
310 I.setOperand(1, Folded);
311 // Conservatively clear the optional flags, since they may not be
312 // preserved by the reassociation.
313 I.clearSubclassOptionalData();
320 // No further simplifications.
325 /// LeftDistributesOverRight - Whether "X LOp (Y ROp Z)" is always equal to
326 /// "(X LOp Y) ROp (X LOp Z)".
327 static bool LeftDistributesOverRight(Instruction::BinaryOps LOp,
328 Instruction::BinaryOps ROp) {
333 case Instruction::And:
334 // And distributes over Or and Xor.
338 case Instruction::Or:
339 case Instruction::Xor:
343 case Instruction::Mul:
344 // Multiplication distributes over addition and subtraction.
348 case Instruction::Add:
349 case Instruction::Sub:
353 case Instruction::Or:
354 // Or distributes over And.
358 case Instruction::And:
364 /// RightDistributesOverLeft - Whether "(X LOp Y) ROp Z" is always equal to
365 /// "(X ROp Z) LOp (Y ROp Z)".
366 static bool RightDistributesOverLeft(Instruction::BinaryOps LOp,
367 Instruction::BinaryOps ROp) {
368 if (Instruction::isCommutative(ROp))
369 return LeftDistributesOverRight(ROp, LOp);
370 // TODO: It would be nice to handle division, aka "(X + Y)/Z = X/Z + Y/Z",
371 // but this requires knowing that the addition does not overflow and other
376 /// SimplifyUsingDistributiveLaws - This tries to simplify binary operations
377 /// which some other binary operation distributes over either by factorizing
378 /// out common terms (eg "(A*B)+(A*C)" -> "A*(B+C)") or expanding out if this
379 /// results in simplifications (eg: "A & (B | C) -> (A&B) | (A&C)" if this is
380 /// a win). Returns the simplified value, or null if it didn't simplify.
381 Value *InstCombiner::SimplifyUsingDistributiveLaws(BinaryOperator &I) {
382 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
383 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
384 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
385 Instruction::BinaryOps TopLevelOpcode = I.getOpcode(); // op
388 if (Op0 && Op1 && Op0->getOpcode() == Op1->getOpcode()) {
389 // The instruction has the form "(A op' B) op (C op' D)". Try to factorize
391 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1);
392 Value *C = Op1->getOperand(0), *D = Op1->getOperand(1);
393 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
395 // Does "X op' Y" always equal "Y op' X"?
396 bool InnerCommutative = Instruction::isCommutative(InnerOpcode);
398 // Does "X op' (Y op Z)" always equal "(X op' Y) op (X op' Z)"?
399 if (LeftDistributesOverRight(InnerOpcode, TopLevelOpcode))
400 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
401 // commutative case, "(A op' B) op (C op' A)"?
402 if (A == C || (InnerCommutative && A == D)) {
405 // Consider forming "A op' (B op D)".
406 // If "B op D" simplifies then it can be formed with no cost.
407 Value *V = SimplifyBinOp(TopLevelOpcode, B, D, TD);
408 // If "B op D" doesn't simplify then only go on if both of the existing
409 // operations "A op' B" and "C op' D" will be zapped as no longer used.
410 if (!V && Op0->hasOneUse() && Op1->hasOneUse())
411 V = Builder->CreateBinOp(TopLevelOpcode, B, D, Op1->getName());
414 V = Builder->CreateBinOp(InnerOpcode, A, V);
420 // Does "(X op Y) op' Z" always equal "(X op' Z) op (Y op' Z)"?
421 if (RightDistributesOverLeft(TopLevelOpcode, InnerOpcode))
422 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
423 // commutative case, "(A op' B) op (B op' D)"?
424 if (B == D || (InnerCommutative && B == C)) {
427 // Consider forming "(A op C) op' B".
428 // If "A op C" simplifies then it can be formed with no cost.
429 Value *V = SimplifyBinOp(TopLevelOpcode, A, C, TD);
430 // If "A op C" doesn't simplify then only go on if both of the existing
431 // operations "A op' B" and "C op' D" will be zapped as no longer used.
432 if (!V && Op0->hasOneUse() && Op1->hasOneUse())
433 V = Builder->CreateBinOp(TopLevelOpcode, A, C, Op0->getName());
436 V = Builder->CreateBinOp(InnerOpcode, V, B);
444 if (Op0 && RightDistributesOverLeft(Op0->getOpcode(), TopLevelOpcode)) {
445 // The instruction has the form "(A op' B) op C". See if expanding it out
446 // to "(A op C) op' (B op C)" results in simplifications.
447 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
448 Instruction::BinaryOps InnerOpcode = Op0->getOpcode(); // op'
450 // Do "A op C" and "B op C" both simplify?
451 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, C, TD))
452 if (Value *R = SimplifyBinOp(TopLevelOpcode, B, C, TD)) {
453 // They do! Return "L op' R".
455 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
456 if ((L == A && R == B) ||
457 (Instruction::isCommutative(InnerOpcode) && L == B && R == A))
459 // Otherwise return "L op' R" if it simplifies.
460 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
462 // Otherwise, create a new instruction.
463 C = Builder->CreateBinOp(InnerOpcode, L, R);
469 if (Op1 && LeftDistributesOverRight(TopLevelOpcode, Op1->getOpcode())) {
470 // The instruction has the form "A op (B op' C)". See if expanding it out
471 // to "(A op B) op' (A op C)" results in simplifications.
472 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
473 Instruction::BinaryOps InnerOpcode = Op1->getOpcode(); // op'
475 // Do "A op B" and "A op C" both simplify?
476 if (Value *L = SimplifyBinOp(TopLevelOpcode, A, B, TD))
477 if (Value *R = SimplifyBinOp(TopLevelOpcode, A, C, TD)) {
478 // They do! Return "L op' R".
480 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
481 if ((L == B && R == C) ||
482 (Instruction::isCommutative(InnerOpcode) && L == C && R == B))
484 // Otherwise return "L op' R" if it simplifies.
485 if (Value *V = SimplifyBinOp(InnerOpcode, L, R, TD))
487 // Otherwise, create a new instruction.
488 A = Builder->CreateBinOp(InnerOpcode, L, R);
497 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
498 // if the LHS is a constant zero (which is the 'negate' form).
500 Value *InstCombiner::dyn_castNegVal(Value *V) const {
501 if (BinaryOperator::isNeg(V))
502 return BinaryOperator::getNegArgument(V);
504 // Constants can be considered to be negated values if they can be folded.
505 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
506 return ConstantExpr::getNeg(C);
508 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
509 if (C->getType()->getElementType()->isIntegerTy())
510 return ConstantExpr::getNeg(C);
515 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
516 // instruction if the LHS is a constant negative zero (which is the 'negate'
519 Value *InstCombiner::dyn_castFNegVal(Value *V) const {
520 if (BinaryOperator::isFNeg(V))
521 return BinaryOperator::getFNegArgument(V);
523 // Constants can be considered to be negated values if they can be folded.
524 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
525 return ConstantExpr::getFNeg(C);
527 if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
528 if (C->getType()->getElementType()->isFloatingPointTy())
529 return ConstantExpr::getFNeg(C);
534 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
536 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
537 return IC->Builder->CreateCast(CI->getOpcode(), SO, I.getType());
540 // Figure out if the constant is the left or the right argument.
541 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
542 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
544 if (Constant *SOC = dyn_cast<Constant>(SO)) {
546 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
547 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
550 Value *Op0 = SO, *Op1 = ConstOperand;
554 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
555 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
556 SO->getName()+".op");
557 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
558 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
559 SO->getName()+".cmp");
560 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
561 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
562 SO->getName()+".cmp");
563 llvm_unreachable("Unknown binary instruction type!");
566 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
567 // constant as the other operand, try to fold the binary operator into the
568 // select arguments. This also works for Cast instructions, which obviously do
569 // not have a second operand.
570 Instruction *InstCombiner::FoldOpIntoSelect(Instruction &Op, SelectInst *SI) {
571 // Don't modify shared select instructions
572 if (!SI->hasOneUse()) return 0;
573 Value *TV = SI->getOperand(1);
574 Value *FV = SI->getOperand(2);
576 if (isa<Constant>(TV) || isa<Constant>(FV)) {
577 // Bool selects with constant operands can be folded to logical ops.
578 if (SI->getType()->isIntegerTy(1)) return 0;
580 // If it's a bitcast involving vectors, make sure it has the same number of
581 // elements on both sides.
582 if (BitCastInst *BC = dyn_cast<BitCastInst>(&Op)) {
583 VectorType *DestTy = dyn_cast<VectorType>(BC->getDestTy());
584 VectorType *SrcTy = dyn_cast<VectorType>(BC->getSrcTy());
586 // Verify that either both or neither are vectors.
587 if ((SrcTy == NULL) != (DestTy == NULL)) return 0;
588 // If vectors, verify that they have the same number of elements.
589 if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
593 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
594 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
596 return SelectInst::Create(SI->getCondition(),
597 SelectTrueVal, SelectFalseVal);
603 /// FoldOpIntoPhi - Given a binary operator, cast instruction, or select which
604 /// has a PHI node as operand #0, see if we can fold the instruction into the
605 /// PHI (which is only possible if all operands to the PHI are constants).
607 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
608 PHINode *PN = cast<PHINode>(I.getOperand(0));
609 unsigned NumPHIValues = PN->getNumIncomingValues();
610 if (NumPHIValues == 0)
613 // We normally only transform phis with a single use. However, if a PHI has
614 // multiple uses and they are all the same operation, we can fold *all* of the
615 // uses into the PHI.
616 if (!PN->hasOneUse()) {
617 // Walk the use list for the instruction, comparing them to I.
618 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
620 Instruction *User = cast<Instruction>(*UI);
621 if (User != &I && !I.isIdenticalTo(User))
624 // Otherwise, we can replace *all* users with the new PHI we form.
627 // Check to see if all of the operands of the PHI are simple constants
628 // (constantint/constantfp/undef). If there is one non-constant value,
629 // remember the BB it is in. If there is more than one or if *it* is a PHI,
630 // bail out. We don't do arbitrary constant expressions here because moving
631 // their computation can be expensive without a cost model.
632 BasicBlock *NonConstBB = 0;
633 for (unsigned i = 0; i != NumPHIValues; ++i) {
634 Value *InVal = PN->getIncomingValue(i);
635 if (isa<Constant>(InVal) && !isa<ConstantExpr>(InVal))
638 if (isa<PHINode>(InVal)) return 0; // Itself a phi.
639 if (NonConstBB) return 0; // More than one non-const value.
641 NonConstBB = PN->getIncomingBlock(i);
643 // If the InVal is an invoke at the end of the pred block, then we can't
644 // insert a computation after it without breaking the edge.
645 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
646 if (II->getParent() == NonConstBB)
649 // If the incoming non-constant value is in I's block, we will remove one
650 // instruction, but insert another equivalent one, leading to infinite
652 if (NonConstBB == I.getParent())
656 // If there is exactly one non-constant value, we can insert a copy of the
657 // operation in that block. However, if this is a critical edge, we would be
658 // inserting the computation one some other paths (e.g. inside a loop). Only
659 // do this if the pred block is unconditionally branching into the phi block.
660 if (NonConstBB != 0) {
661 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
662 if (!BI || !BI->isUnconditional()) return 0;
665 // Okay, we can do the transformation: create the new PHI node.
666 PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
667 InsertNewInstBefore(NewPN, *PN);
670 // If we are going to have to insert a new computation, do so right before the
671 // predecessors terminator.
673 Builder->SetInsertPoint(NonConstBB->getTerminator());
675 // Next, add all of the operands to the PHI.
676 if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
677 // We only currently try to fold the condition of a select when it is a phi,
678 // not the true/false values.
679 Value *TrueV = SI->getTrueValue();
680 Value *FalseV = SI->getFalseValue();
681 BasicBlock *PhiTransBB = PN->getParent();
682 for (unsigned i = 0; i != NumPHIValues; ++i) {
683 BasicBlock *ThisBB = PN->getIncomingBlock(i);
684 Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
685 Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
687 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
688 InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
690 InV = Builder->CreateSelect(PN->getIncomingValue(i),
691 TrueVInPred, FalseVInPred, "phitmp");
692 NewPN->addIncoming(InV, ThisBB);
694 } else if (CmpInst *CI = dyn_cast<CmpInst>(&I)) {
695 Constant *C = cast<Constant>(I.getOperand(1));
696 for (unsigned i = 0; i != NumPHIValues; ++i) {
698 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
699 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
700 else if (isa<ICmpInst>(CI))
701 InV = Builder->CreateICmp(CI->getPredicate(), PN->getIncomingValue(i),
704 InV = Builder->CreateFCmp(CI->getPredicate(), PN->getIncomingValue(i),
706 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
708 } else if (I.getNumOperands() == 2) {
709 Constant *C = cast<Constant>(I.getOperand(1));
710 for (unsigned i = 0; i != NumPHIValues; ++i) {
712 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
713 InV = ConstantExpr::get(I.getOpcode(), InC, C);
715 InV = Builder->CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
716 PN->getIncomingValue(i), C, "phitmp");
717 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
720 CastInst *CI = cast<CastInst>(&I);
721 Type *RetTy = CI->getType();
722 for (unsigned i = 0; i != NumPHIValues; ++i) {
724 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
725 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
727 InV = Builder->CreateCast(CI->getOpcode(),
728 PN->getIncomingValue(i), I.getType(), "phitmp");
729 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
733 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
735 Instruction *User = cast<Instruction>(*UI++);
736 if (User == &I) continue;
737 ReplaceInstUsesWith(*User, NewPN);
738 EraseInstFromFunction(*User);
740 return ReplaceInstUsesWith(I, NewPN);
743 /// FindElementAtOffset - Given a type and a constant offset, determine whether
744 /// or not there is a sequence of GEP indices into the type that will land us at
745 /// the specified offset. If so, fill them into NewIndices and return the
746 /// resultant element type, otherwise return null.
747 Type *InstCombiner::FindElementAtOffset(Type *Ty, int64_t Offset,
748 SmallVectorImpl<Value*> &NewIndices) {
750 if (!Ty->isSized()) return 0;
752 // Start with the index over the outer type. Note that the type size
753 // might be zero (even if the offset isn't zero) if the indexed type
754 // is something like [0 x {int, int}]
755 Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
756 int64_t FirstIdx = 0;
757 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
758 FirstIdx = Offset/TySize;
759 Offset -= FirstIdx*TySize;
761 // Handle hosts where % returns negative instead of values [0..TySize).
767 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
770 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
772 // Index into the types. If we fail, set OrigBase to null.
774 // Indexing into tail padding between struct/array elements.
775 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
778 if (StructType *STy = dyn_cast<StructType>(Ty)) {
779 const StructLayout *SL = TD->getStructLayout(STy);
780 assert(Offset < (int64_t)SL->getSizeInBytes() &&
781 "Offset must stay within the indexed type");
783 unsigned Elt = SL->getElementContainingOffset(Offset);
784 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
787 Offset -= SL->getElementOffset(Elt);
788 Ty = STy->getElementType(Elt);
789 } else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
790 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
791 assert(EltSize && "Cannot index into a zero-sized array");
792 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
794 Ty = AT->getElementType();
796 // Otherwise, we can't index into the middle of this atomic type, bail.
804 static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
805 // If this GEP has only 0 indices, it is the same pointer as
806 // Src. If Src is not a trivial GEP too, don't combine
808 if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
814 /// Descale - Return a value X such that Val = X * Scale, or null if none. If
815 /// the multiplication is known not to overflow then NoSignedWrap is set.
816 Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
817 assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
818 assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
819 Scale.getBitWidth() && "Scale not compatible with value!");
821 // If Val is zero or Scale is one then Val = Val * Scale.
822 if (match(Val, m_Zero()) || Scale == 1) {
827 // If Scale is zero then it does not divide Val.
828 if (Scale.isMinValue())
831 // Look through chains of multiplications, searching for a constant that is
832 // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
833 // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
834 // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
837 // Val = M1 * X || Analysis starts here and works down
838 // M1 = M2 * Y || Doesn't descend into terms with more
839 // M2 = Z * 4 \/ than one use
841 // Then to modify a term at the bottom:
844 // M1 = Z * Y || Replaced M2 with Z
846 // Then to work back up correcting nsw flags.
848 // Op - the term we are currently analyzing. Starts at Val then drills down.
849 // Replaced with its descaled value before exiting from the drill down loop.
852 // Parent - initially null, but after drilling down notes where Op came from.
853 // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
854 // 0'th operand of Val.
855 std::pair<Instruction*, unsigned> Parent;
857 // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
858 // levels that doesn't overflow.
859 bool RequireNoSignedWrap = false;
861 // logScale - log base 2 of the scale. Negative if not a power of 2.
862 int32_t logScale = Scale.exactLogBase2();
864 for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
866 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
867 // If Op is a constant divisible by Scale then descale to the quotient.
868 APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
869 APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
870 if (!Remainder.isMinValue())
871 // Not divisible by Scale.
873 // Replace with the quotient in the parent.
874 Op = ConstantInt::get(CI->getType(), Quotient);
879 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
881 if (BO->getOpcode() == Instruction::Mul) {
883 NoSignedWrap = BO->hasNoSignedWrap();
884 if (RequireNoSignedWrap && !NoSignedWrap)
887 // There are three cases for multiplication: multiplication by exactly
888 // the scale, multiplication by a constant different to the scale, and
889 // multiplication by something else.
890 Value *LHS = BO->getOperand(0);
891 Value *RHS = BO->getOperand(1);
893 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
894 // Multiplication by a constant.
895 if (CI->getValue() == Scale) {
896 // Multiplication by exactly the scale, replace the multiplication
897 // by its left-hand side in the parent.
902 // Otherwise drill down into the constant.
903 if (!Op->hasOneUse())
906 Parent = std::make_pair(BO, 1);
910 // Multiplication by something else. Drill down into the left-hand side
911 // since that's where the reassociate pass puts the good stuff.
912 if (!Op->hasOneUse())
915 Parent = std::make_pair(BO, 0);
919 if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
920 isa<ConstantInt>(BO->getOperand(1))) {
921 // Multiplication by a power of 2.
922 NoSignedWrap = BO->hasNoSignedWrap();
923 if (RequireNoSignedWrap && !NoSignedWrap)
926 Value *LHS = BO->getOperand(0);
927 int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
928 getLimitedValue(Scale.getBitWidth());
931 if (Amt == logScale) {
932 // Multiplication by exactly the scale, replace the multiplication
933 // by its left-hand side in the parent.
937 if (Amt < logScale || !Op->hasOneUse())
940 // Multiplication by more than the scale. Reduce the multiplying amount
941 // by the scale in the parent.
942 Parent = std::make_pair(BO, 1);
943 Op = ConstantInt::get(BO->getType(), Amt - logScale);
948 if (!Op->hasOneUse())
951 if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
952 if (Cast->getOpcode() == Instruction::SExt) {
953 // Op is sign-extended from a smaller type, descale in the smaller type.
954 unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
955 APInt SmallScale = Scale.trunc(SmallSize);
956 // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
957 // descale Op as (sext Y) * Scale. In order to have
958 // sext (Y * SmallScale) = (sext Y) * Scale
959 // some conditions need to hold however: SmallScale must sign-extend to
960 // Scale and the multiplication Y * SmallScale should not overflow.
961 if (SmallScale.sext(Scale.getBitWidth()) != Scale)
962 // SmallScale does not sign-extend to Scale.
964 assert(SmallScale.exactLogBase2() == logScale);
965 // Require that Y * SmallScale must not overflow.
966 RequireNoSignedWrap = true;
968 // Drill down through the cast.
969 Parent = std::make_pair(Cast, 0);
974 if (Cast->getOpcode() == Instruction::Trunc) {
975 // Op is truncated from a larger type, descale in the larger type.
976 // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
977 // trunc (Y * sext Scale) = (trunc Y) * Scale
978 // always holds. However (trunc Y) * Scale may overflow even if
979 // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
980 // from this point up in the expression (see later).
981 if (RequireNoSignedWrap)
984 // Drill down through the cast.
985 unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
986 Parent = std::make_pair(Cast, 0);
987 Scale = Scale.sext(LargeSize);
988 if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
990 assert(Scale.exactLogBase2() == logScale);
995 // Unsupported expression, bail out.
999 // We know that we can successfully descale, so from here on we can safely
1000 // modify the IR. Op holds the descaled version of the deepest term in the
1001 // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
1005 // The expression only had one term.
1008 // Rewrite the parent using the descaled version of its operand.
1009 assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
1010 assert(Op != Parent.first->getOperand(Parent.second) &&
1011 "Descaling was a no-op?");
1012 Parent.first->setOperand(Parent.second, Op);
1013 Worklist.Add(Parent.first);
1015 // Now work back up the expression correcting nsw flags. The logic is based
1016 // on the following observation: if X * Y is known not to overflow as a signed
1017 // multiplication, and Y is replaced by a value Z with smaller absolute value,
1018 // then X * Z will not overflow as a signed multiplication either. As we work
1019 // our way up, having NoSignedWrap 'true' means that the descaled value at the
1020 // current level has strictly smaller absolute value than the original.
1021 Instruction *Ancestor = Parent.first;
1023 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
1024 // If the multiplication wasn't nsw then we can't say anything about the
1025 // value of the descaled multiplication, and we have to clear nsw flags
1026 // from this point on up.
1027 bool OpNoSignedWrap = BO->hasNoSignedWrap();
1028 NoSignedWrap &= OpNoSignedWrap;
1029 if (NoSignedWrap != OpNoSignedWrap) {
1030 BO->setHasNoSignedWrap(NoSignedWrap);
1031 Worklist.Add(Ancestor);
1033 } else if (Ancestor->getOpcode() == Instruction::Trunc) {
1034 // The fact that the descaled input to the trunc has smaller absolute
1035 // value than the original input doesn't tell us anything useful about
1036 // the absolute values of the truncations.
1037 NoSignedWrap = false;
1039 assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
1040 "Failed to keep proper track of nsw flags while drilling down?");
1042 if (Ancestor == Val)
1043 // Got to the top, all done!
1046 // Move up one level in the expression.
1047 assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
1048 Ancestor = Ancestor->use_back();
1052 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
1053 SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
1055 if (Value *V = SimplifyGEPInst(Ops, TD))
1056 return ReplaceInstUsesWith(GEP, V);
1058 Value *PtrOp = GEP.getOperand(0);
1060 // Eliminate unneeded casts for indices, and replace indices which displace
1061 // by multiples of a zero size type with zero.
1063 bool MadeChange = false;
1064 Type *IntPtrTy = TD->getIntPtrType(GEP.getPointerOperandType());
1066 gep_type_iterator GTI = gep_type_begin(GEP);
1067 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
1068 I != E; ++I, ++GTI) {
1069 // Skip indices into struct types.
1070 SequentialType *SeqTy = dyn_cast<SequentialType>(*GTI);
1071 if (!SeqTy) continue;
1073 // If the element type has zero size then any index over it is equivalent
1074 // to an index of zero, so replace it with zero if it is not zero already.
1075 if (SeqTy->getElementType()->isSized() &&
1076 TD->getTypeAllocSize(SeqTy->getElementType()) == 0)
1077 if (!isa<Constant>(*I) || !cast<Constant>(*I)->isNullValue()) {
1078 *I = Constant::getNullValue(IntPtrTy);
1082 Type *IndexTy = (*I)->getType();
1083 if (IndexTy != IntPtrTy) {
1084 // If we are using a wider index than needed for this platform, shrink
1085 // it to what we need. If narrower, sign-extend it to what we need.
1086 // This explicit cast can make subsequent optimizations more obvious.
1087 *I = Builder->CreateIntCast(*I, IntPtrTy, true);
1091 if (MadeChange) return &GEP;
1094 // Combine Indices - If the source pointer to this getelementptr instruction
1095 // is a getelementptr instruction, combine the indices of the two
1096 // getelementptr instructions into a single instruction.
1098 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
1099 if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
1102 // Note that if our source is a gep chain itself then we wait for that
1103 // chain to be resolved before we perform this transformation. This
1104 // avoids us creating a TON of code in some cases.
1105 if (GEPOperator *SrcGEP =
1106 dyn_cast<GEPOperator>(Src->getOperand(0)))
1107 if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
1108 return 0; // Wait until our source is folded to completion.
1110 SmallVector<Value*, 8> Indices;
1112 // Find out whether the last index in the source GEP is a sequential idx.
1113 bool EndsWithSequential = false;
1114 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
1116 EndsWithSequential = !(*I)->isStructTy();
1118 // Can we combine the two pointer arithmetics offsets?
1119 if (EndsWithSequential) {
1120 // Replace: gep (gep %P, long B), long A, ...
1121 // With: T = long A+B; gep %P, T, ...
1124 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
1125 Value *GO1 = GEP.getOperand(1);
1126 if (SO1 == Constant::getNullValue(SO1->getType())) {
1128 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
1131 // If they aren't the same type, then the input hasn't been processed
1132 // by the loop above yet (which canonicalizes sequential index types to
1133 // intptr_t). Just avoid transforming this until the input has been
1135 if (SO1->getType() != GO1->getType())
1137 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
1140 // Update the GEP in place if possible.
1141 if (Src->getNumOperands() == 2) {
1142 GEP.setOperand(0, Src->getOperand(0));
1143 GEP.setOperand(1, Sum);
1146 Indices.append(Src->op_begin()+1, Src->op_end()-1);
1147 Indices.push_back(Sum);
1148 Indices.append(GEP.op_begin()+2, GEP.op_end());
1149 } else if (isa<Constant>(*GEP.idx_begin()) &&
1150 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
1151 Src->getNumOperands() != 1) {
1152 // Otherwise we can do the fold if the first index of the GEP is a zero
1153 Indices.append(Src->op_begin()+1, Src->op_end());
1154 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
1157 if (!Indices.empty())
1158 return (GEP.isInBounds() && Src->isInBounds()) ?
1159 GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
1161 GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
1164 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
1165 Value *StrippedPtr = PtrOp->stripPointerCasts();
1166 PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
1168 // We do not handle pointer-vector geps here.
1172 if (StrippedPtr != PtrOp &&
1173 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
1175 bool HasZeroPointerIndex = false;
1176 if (ConstantInt *C = dyn_cast<ConstantInt>(GEP.getOperand(1)))
1177 HasZeroPointerIndex = C->isZero();
1179 // Transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
1180 // into : GEP [10 x i8]* X, i32 0, ...
1182 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
1183 // into : GEP i8* X, ...
1185 // This occurs when the program declares an array extern like "int X[];"
1186 if (HasZeroPointerIndex) {
1187 PointerType *CPTy = cast<PointerType>(PtrOp->getType());
1188 if (ArrayType *CATy =
1189 dyn_cast<ArrayType>(CPTy->getElementType())) {
1190 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
1191 if (CATy->getElementType() == StrippedPtrTy->getElementType()) {
1192 // -> GEP i8* X, ...
1193 SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
1194 GetElementPtrInst *Res =
1195 GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
1196 Res->setIsInBounds(GEP.isInBounds());
1200 if (ArrayType *XATy =
1201 dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
1202 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
1203 if (CATy->getElementType() == XATy->getElementType()) {
1204 // -> GEP [10 x i8]* X, i32 0, ...
1205 // At this point, we know that the cast source type is a pointer
1206 // to an array of the same type as the destination pointer
1207 // array. Because the array type is never stepped over (there
1208 // is a leading zero) we can fold the cast into this GEP.
1209 GEP.setOperand(0, StrippedPtr);
1214 } else if (GEP.getNumOperands() == 2) {
1215 // Transform things like:
1216 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
1217 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
1218 Type *SrcElTy = StrippedPtrTy->getElementType();
1219 Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
1220 if (TD && SrcElTy->isArrayTy() &&
1221 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
1222 TD->getTypeAllocSize(ResElTy)) {
1224 Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
1225 Idx[1] = GEP.getOperand(1);
1226 Value *NewGEP = GEP.isInBounds() ?
1227 Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
1228 Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
1229 // V and GEP are both pointer types --> BitCast
1230 return new BitCastInst(NewGEP, GEP.getType());
1233 // Transform things like:
1234 // %V = mul i64 %N, 4
1235 // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
1236 // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
1237 if (TD && ResElTy->isSized() && SrcElTy->isSized()) {
1238 // Check that changing the type amounts to dividing the index by a scale
1240 uint64_t ResSize = TD->getTypeAllocSize(ResElTy);
1241 uint64_t SrcSize = TD->getTypeAllocSize(SrcElTy);
1242 if (ResSize && SrcSize % ResSize == 0) {
1243 Value *Idx = GEP.getOperand(1);
1244 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1245 uint64_t Scale = SrcSize / ResSize;
1247 // Earlier transforms ensure that the index has type IntPtrType, which
1248 // considerably simplifies the logic by eliminating implicit casts.
1249 assert(Idx->getType() == TD->getIntPtrType(GEP.getContext()) &&
1250 "Index not cast to pointer width?");
1253 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1254 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1255 // If the multiplication NewIdx * Scale may overflow then the new
1256 // GEP may not be "inbounds".
1257 Value *NewGEP = GEP.isInBounds() && NSW ?
1258 Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
1259 Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
1260 // The NewGEP must be pointer typed, so must the old one -> BitCast
1261 return new BitCastInst(NewGEP, GEP.getType());
1266 // Similarly, transform things like:
1267 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
1268 // (where tmp = 8*tmp2) into:
1269 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
1270 if (TD && ResElTy->isSized() && SrcElTy->isSized() &&
1271 SrcElTy->isArrayTy()) {
1272 // Check that changing to the array element type amounts to dividing the
1273 // index by a scale factor.
1274 uint64_t ResSize = TD->getTypeAllocSize(ResElTy);
1275 uint64_t ArrayEltSize =
1276 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
1277 if (ResSize && ArrayEltSize % ResSize == 0) {
1278 Value *Idx = GEP.getOperand(1);
1279 unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
1280 uint64_t Scale = ArrayEltSize / ResSize;
1282 // Earlier transforms ensure that the index has type IntPtrType, which
1283 // considerably simplifies the logic by eliminating implicit casts.
1284 assert(Idx->getType() == TD->getIntPtrType(GEP.getContext()) &&
1285 "Index not cast to pointer width?");
1288 if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
1289 // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
1290 // If the multiplication NewIdx * Scale may overflow then the new
1291 // GEP may not be "inbounds".
1293 Off[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
1295 Value *NewGEP = GEP.isInBounds() && NSW ?
1296 Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
1297 Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
1298 // The NewGEP must be pointer typed, so must the old one -> BitCast
1299 return new BitCastInst(NewGEP, GEP.getType());
1306 /// See if we can simplify:
1307 /// X = bitcast A* to B*
1308 /// Y = gep X, <...constant indices...>
1309 /// into a gep of the original struct. This is important for SROA and alias
1310 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
1311 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
1312 APInt Offset(TD ? TD->getPointerSizeInBits() : 1, 0);
1314 !isa<BitCastInst>(BCI->getOperand(0)) &&
1315 GEP.accumulateConstantOffset(*TD, Offset) &&
1316 StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
1318 // If this GEP instruction doesn't move the pointer, just replace the GEP
1319 // with a bitcast of the real input to the dest type.
1321 // If the bitcast is of an allocation, and the allocation will be
1322 // converted to match the type of the cast, don't touch this.
1323 if (isa<AllocaInst>(BCI->getOperand(0)) ||
1324 isAllocationFn(BCI->getOperand(0), TLI)) {
1325 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
1326 if (Instruction *I = visitBitCast(*BCI)) {
1329 BCI->getParent()->getInstList().insert(BCI, I);
1330 ReplaceInstUsesWith(*BCI, I);
1335 return new BitCastInst(BCI->getOperand(0), GEP.getType());
1338 // Otherwise, if the offset is non-zero, we need to find out if there is a
1339 // field at Offset in 'A's type. If so, we can pull the cast through the
1341 SmallVector<Value*, 8> NewIndices;
1343 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
1344 if (FindElementAtOffset(InTy, Offset.getSExtValue(), NewIndices)) {
1345 Value *NGEP = GEP.isInBounds() ?
1346 Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices) :
1347 Builder->CreateGEP(BCI->getOperand(0), NewIndices);
1349 if (NGEP->getType() == GEP.getType())
1350 return ReplaceInstUsesWith(GEP, NGEP);
1351 NGEP->takeName(&GEP);
1352 return new BitCastInst(NGEP, GEP.getType());
1363 isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
1364 const TargetLibraryInfo *TLI) {
1365 SmallVector<Instruction*, 4> Worklist;
1366 Worklist.push_back(AI);
1369 Instruction *PI = Worklist.pop_back_val();
1370 for (Value::use_iterator UI = PI->use_begin(), UE = PI->use_end(); UI != UE;
1372 Instruction *I = cast<Instruction>(*UI);
1373 switch (I->getOpcode()) {
1375 // Give up the moment we see something we can't handle.
1378 case Instruction::BitCast:
1379 case Instruction::GetElementPtr:
1381 Worklist.push_back(I);
1384 case Instruction::ICmp: {
1385 ICmpInst *ICI = cast<ICmpInst>(I);
1386 // We can fold eq/ne comparisons with null to false/true, respectively.
1387 if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
1393 case Instruction::Call:
1394 // Ignore no-op and store intrinsics.
1395 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1396 switch (II->getIntrinsicID()) {
1400 case Intrinsic::memmove:
1401 case Intrinsic::memcpy:
1402 case Intrinsic::memset: {
1403 MemIntrinsic *MI = cast<MemIntrinsic>(II);
1404 if (MI->isVolatile() || MI->getRawDest() != PI)
1408 case Intrinsic::dbg_declare:
1409 case Intrinsic::dbg_value:
1410 case Intrinsic::invariant_start:
1411 case Intrinsic::invariant_end:
1412 case Intrinsic::lifetime_start:
1413 case Intrinsic::lifetime_end:
1414 case Intrinsic::objectsize:
1420 if (isFreeCall(I, TLI)) {
1426 case Instruction::Store: {
1427 StoreInst *SI = cast<StoreInst>(I);
1428 if (SI->isVolatile() || SI->getPointerOperand() != PI)
1434 llvm_unreachable("missing a return?");
1436 } while (!Worklist.empty());
1440 Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
1441 // If we have a malloc call which is only used in any amount of comparisons
1442 // to null and free calls, delete the calls and replace the comparisons with
1443 // true or false as appropriate.
1444 SmallVector<WeakVH, 64> Users;
1445 if (isAllocSiteRemovable(&MI, Users, TLI)) {
1446 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
1447 Instruction *I = cast_or_null<Instruction>(&*Users[i]);
1450 if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
1451 ReplaceInstUsesWith(*C,
1452 ConstantInt::get(Type::getInt1Ty(C->getContext()),
1453 C->isFalseWhenEqual()));
1454 } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
1455 ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
1456 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1457 if (II->getIntrinsicID() == Intrinsic::objectsize) {
1458 ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
1459 uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
1460 ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
1463 EraseInstFromFunction(*I);
1466 if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
1467 // Replace invoke with a NOP intrinsic to maintain the original CFG
1468 Module *M = II->getParent()->getParent()->getParent();
1469 Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
1470 InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
1471 ArrayRef<Value *>(), "", II->getParent());
1473 return EraseInstFromFunction(MI);
1480 Instruction *InstCombiner::visitFree(CallInst &FI) {
1481 Value *Op = FI.getArgOperand(0);
1483 // free undef -> unreachable.
1484 if (isa<UndefValue>(Op)) {
1485 // Insert a new store to null because we cannot modify the CFG here.
1486 Builder->CreateStore(ConstantInt::getTrue(FI.getContext()),
1487 UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
1488 return EraseInstFromFunction(FI);
1491 // If we have 'free null' delete the instruction. This can happen in stl code
1492 // when lots of inlining happens.
1493 if (isa<ConstantPointerNull>(Op))
1494 return EraseInstFromFunction(FI);
1501 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
1502 // Change br (not X), label True, label False to: br X, label False, True
1504 BasicBlock *TrueDest;
1505 BasicBlock *FalseDest;
1506 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
1507 !isa<Constant>(X)) {
1508 // Swap Destinations and condition...
1510 BI.swapSuccessors();
1514 // Cannonicalize fcmp_one -> fcmp_oeq
1515 FCmpInst::Predicate FPred; Value *Y;
1516 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
1517 TrueDest, FalseDest)) &&
1518 BI.getCondition()->hasOneUse())
1519 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
1520 FPred == FCmpInst::FCMP_OGE) {
1521 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
1522 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
1524 // Swap Destinations and condition.
1525 BI.swapSuccessors();
1530 // Cannonicalize icmp_ne -> icmp_eq
1531 ICmpInst::Predicate IPred;
1532 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
1533 TrueDest, FalseDest)) &&
1534 BI.getCondition()->hasOneUse())
1535 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
1536 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
1537 IPred == ICmpInst::ICMP_SGE) {
1538 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
1539 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
1540 // Swap Destinations and condition.
1541 BI.swapSuccessors();
1549 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
1550 Value *Cond = SI.getCondition();
1551 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
1552 if (I->getOpcode() == Instruction::Add)
1553 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1554 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
1555 // Skip the first item since that's the default case.
1556 for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
1558 ConstantInt* CaseVal = i.getCaseValue();
1559 Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal),
1561 assert(isa<ConstantInt>(NewCaseVal) &&
1562 "Result of expression should be constant");
1563 i.setValue(cast<ConstantInt>(NewCaseVal));
1565 SI.setCondition(I->getOperand(0));
1573 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
1574 Value *Agg = EV.getAggregateOperand();
1576 if (!EV.hasIndices())
1577 return ReplaceInstUsesWith(EV, Agg);
1579 if (Constant *C = dyn_cast<Constant>(Agg)) {
1580 if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
1581 if (EV.getNumIndices() == 0)
1582 return ReplaceInstUsesWith(EV, C2);
1583 // Extract the remaining indices out of the constant indexed by the
1585 return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
1587 return 0; // Can't handle other constants
1590 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
1591 // We're extracting from an insertvalue instruction, compare the indices
1592 const unsigned *exti, *exte, *insi, *inse;
1593 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
1594 exte = EV.idx_end(), inse = IV->idx_end();
1595 exti != exte && insi != inse;
1598 // The insert and extract both reference distinctly different elements.
1599 // This means the extract is not influenced by the insert, and we can
1600 // replace the aggregate operand of the extract with the aggregate
1601 // operand of the insert. i.e., replace
1602 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1603 // %E = extractvalue { i32, { i32 } } %I, 0
1605 // %E = extractvalue { i32, { i32 } } %A, 0
1606 return ExtractValueInst::Create(IV->getAggregateOperand(),
1609 if (exti == exte && insi == inse)
1610 // Both iterators are at the end: Index lists are identical. Replace
1611 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1612 // %C = extractvalue { i32, { i32 } } %B, 1, 0
1614 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
1616 // The extract list is a prefix of the insert list. i.e. replace
1617 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
1618 // %E = extractvalue { i32, { i32 } } %I, 1
1620 // %X = extractvalue { i32, { i32 } } %A, 1
1621 // %E = insertvalue { i32 } %X, i32 42, 0
1622 // by switching the order of the insert and extract (though the
1623 // insertvalue should be left in, since it may have other uses).
1624 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
1626 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
1627 makeArrayRef(insi, inse));
1630 // The insert list is a prefix of the extract list
1631 // We can simply remove the common indices from the extract and make it
1632 // operate on the inserted value instead of the insertvalue result.
1634 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
1635 // %E = extractvalue { i32, { i32 } } %I, 1, 0
1637 // %E extractvalue { i32 } { i32 42 }, 0
1638 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
1639 makeArrayRef(exti, exte));
1641 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
1642 // We're extracting from an intrinsic, see if we're the only user, which
1643 // allows us to simplify multiple result intrinsics to simpler things that
1644 // just get one value.
1645 if (II->hasOneUse()) {
1646 // Check if we're grabbing the overflow bit or the result of a 'with
1647 // overflow' intrinsic. If it's the latter we can remove the intrinsic
1648 // and replace it with a traditional binary instruction.
1649 switch (II->getIntrinsicID()) {
1650 case Intrinsic::uadd_with_overflow:
1651 case Intrinsic::sadd_with_overflow:
1652 if (*EV.idx_begin() == 0) { // Normal result.
1653 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1654 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
1655 EraseInstFromFunction(*II);
1656 return BinaryOperator::CreateAdd(LHS, RHS);
1659 // If the normal result of the add is dead, and the RHS is a constant,
1660 // we can transform this into a range comparison.
1661 // overflow = uadd a, -4 --> overflow = icmp ugt a, 3
1662 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
1663 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
1664 return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
1665 ConstantExpr::getNot(CI));
1667 case Intrinsic::usub_with_overflow:
1668 case Intrinsic::ssub_with_overflow:
1669 if (*EV.idx_begin() == 0) { // Normal result.
1670 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1671 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
1672 EraseInstFromFunction(*II);
1673 return BinaryOperator::CreateSub(LHS, RHS);
1676 case Intrinsic::umul_with_overflow:
1677 case Intrinsic::smul_with_overflow:
1678 if (*EV.idx_begin() == 0) { // Normal result.
1679 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
1680 ReplaceInstUsesWith(*II, UndefValue::get(II->getType()));
1681 EraseInstFromFunction(*II);
1682 return BinaryOperator::CreateMul(LHS, RHS);
1690 if (LoadInst *L = dyn_cast<LoadInst>(Agg))
1691 // If the (non-volatile) load only has one use, we can rewrite this to a
1692 // load from a GEP. This reduces the size of the load.
1693 // FIXME: If a load is used only by extractvalue instructions then this
1694 // could be done regardless of having multiple uses.
1695 if (L->isSimple() && L->hasOneUse()) {
1696 // extractvalue has integer indices, getelementptr has Value*s. Convert.
1697 SmallVector<Value*, 4> Indices;
1698 // Prefix an i32 0 since we need the first element.
1699 Indices.push_back(Builder->getInt32(0));
1700 for (ExtractValueInst::idx_iterator I = EV.idx_begin(), E = EV.idx_end();
1702 Indices.push_back(Builder->getInt32(*I));
1704 // We need to insert these at the location of the old load, not at that of
1705 // the extractvalue.
1706 Builder->SetInsertPoint(L->getParent(), L);
1707 Value *GEP = Builder->CreateInBoundsGEP(L->getPointerOperand(), Indices);
1708 // Returning the load directly will cause the main loop to insert it in
1709 // the wrong spot, so use ReplaceInstUsesWith().
1710 return ReplaceInstUsesWith(EV, Builder->CreateLoad(GEP));
1712 // We could simplify extracts from other values. Note that nested extracts may
1713 // already be simplified implicitly by the above: extract (extract (insert) )
1714 // will be translated into extract ( insert ( extract ) ) first and then just
1715 // the value inserted, if appropriate. Similarly for extracts from single-use
1716 // loads: extract (extract (load)) will be translated to extract (load (gep))
1717 // and if again single-use then via load (gep (gep)) to load (gep).
1718 // However, double extracts from e.g. function arguments or return values
1719 // aren't handled yet.
1723 enum Personality_Type {
1724 Unknown_Personality,
1725 GNU_Ada_Personality,
1726 GNU_CXX_Personality,
1727 GNU_ObjC_Personality
1730 /// RecognizePersonality - See if the given exception handling personality
1731 /// function is one that we understand. If so, return a description of it;
1732 /// otherwise return Unknown_Personality.
1733 static Personality_Type RecognizePersonality(Value *Pers) {
1734 Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
1736 return Unknown_Personality;
1737 return StringSwitch<Personality_Type>(F->getName())
1738 .Case("__gnat_eh_personality", GNU_Ada_Personality)
1739 .Case("__gxx_personality_v0", GNU_CXX_Personality)
1740 .Case("__objc_personality_v0", GNU_ObjC_Personality)
1741 .Default(Unknown_Personality);
1744 /// isCatchAll - Return 'true' if the given typeinfo will match anything.
1745 static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
1746 switch (Personality) {
1747 case Unknown_Personality:
1749 case GNU_Ada_Personality:
1750 // While __gnat_all_others_value will match any Ada exception, it doesn't
1751 // match foreign exceptions (or didn't, before gcc-4.7).
1753 case GNU_CXX_Personality:
1754 case GNU_ObjC_Personality:
1755 return TypeInfo->isNullValue();
1757 llvm_unreachable("Unknown personality!");
1760 static bool shorter_filter(const Value *LHS, const Value *RHS) {
1762 cast<ArrayType>(LHS->getType())->getNumElements()
1764 cast<ArrayType>(RHS->getType())->getNumElements();
1767 Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
1768 // The logic here should be correct for any real-world personality function.
1769 // However if that turns out not to be true, the offending logic can always
1770 // be conditioned on the personality function, like the catch-all logic is.
1771 Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
1773 // Simplify the list of clauses, eg by removing repeated catch clauses
1774 // (these are often created by inlining).
1775 bool MakeNewInstruction = false; // If true, recreate using the following:
1776 SmallVector<Value *, 16> NewClauses; // - Clauses for the new instruction;
1777 bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
1779 SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
1780 for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
1781 bool isLastClause = i + 1 == e;
1782 if (LI.isCatch(i)) {
1784 Value *CatchClause = LI.getClause(i);
1785 Constant *TypeInfo = cast<Constant>(CatchClause->stripPointerCasts());
1787 // If we already saw this clause, there is no point in having a second
1789 if (AlreadyCaught.insert(TypeInfo)) {
1790 // This catch clause was not already seen.
1791 NewClauses.push_back(CatchClause);
1793 // Repeated catch clause - drop the redundant copy.
1794 MakeNewInstruction = true;
1797 // If this is a catch-all then there is no point in keeping any following
1798 // clauses or marking the landingpad as having a cleanup.
1799 if (isCatchAll(Personality, TypeInfo)) {
1801 MakeNewInstruction = true;
1802 CleanupFlag = false;
1806 // A filter clause. If any of the filter elements were already caught
1807 // then they can be dropped from the filter. It is tempting to try to
1808 // exploit the filter further by saying that any typeinfo that does not
1809 // occur in the filter can't be caught later (and thus can be dropped).
1810 // However this would be wrong, since typeinfos can match without being
1811 // equal (for example if one represents a C++ class, and the other some
1812 // class derived from it).
1813 assert(LI.isFilter(i) && "Unsupported landingpad clause!");
1814 Value *FilterClause = LI.getClause(i);
1815 ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
1816 unsigned NumTypeInfos = FilterType->getNumElements();
1818 // An empty filter catches everything, so there is no point in keeping any
1819 // following clauses or marking the landingpad as having a cleanup. By
1820 // dealing with this case here the following code is made a bit simpler.
1821 if (!NumTypeInfos) {
1822 NewClauses.push_back(FilterClause);
1824 MakeNewInstruction = true;
1825 CleanupFlag = false;
1829 bool MakeNewFilter = false; // If true, make a new filter.
1830 SmallVector<Constant *, 16> NewFilterElts; // New elements.
1831 if (isa<ConstantAggregateZero>(FilterClause)) {
1832 // Not an empty filter - it contains at least one null typeinfo.
1833 assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
1834 Constant *TypeInfo =
1835 Constant::getNullValue(FilterType->getElementType());
1836 // If this typeinfo is a catch-all then the filter can never match.
1837 if (isCatchAll(Personality, TypeInfo)) {
1838 // Throw the filter away.
1839 MakeNewInstruction = true;
1843 // There is no point in having multiple copies of this typeinfo, so
1844 // discard all but the first copy if there is more than one.
1845 NewFilterElts.push_back(TypeInfo);
1846 if (NumTypeInfos > 1)
1847 MakeNewFilter = true;
1849 ConstantArray *Filter = cast<ConstantArray>(FilterClause);
1850 SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
1851 NewFilterElts.reserve(NumTypeInfos);
1853 // Remove any filter elements that were already caught or that already
1854 // occurred in the filter. While there, see if any of the elements are
1855 // catch-alls. If so, the filter can be discarded.
1856 bool SawCatchAll = false;
1857 for (unsigned j = 0; j != NumTypeInfos; ++j) {
1858 Value *Elt = Filter->getOperand(j);
1859 Constant *TypeInfo = cast<Constant>(Elt->stripPointerCasts());
1860 if (isCatchAll(Personality, TypeInfo)) {
1861 // This element is a catch-all. Bail out, noting this fact.
1865 if (AlreadyCaught.count(TypeInfo))
1866 // Already caught by an earlier clause, so having it in the filter
1869 // There is no point in having multiple copies of the same typeinfo in
1870 // a filter, so only add it if we didn't already.
1871 if (SeenInFilter.insert(TypeInfo))
1872 NewFilterElts.push_back(cast<Constant>(Elt));
1874 // A filter containing a catch-all cannot match anything by definition.
1876 // Throw the filter away.
1877 MakeNewInstruction = true;
1881 // If we dropped something from the filter, make a new one.
1882 if (NewFilterElts.size() < NumTypeInfos)
1883 MakeNewFilter = true;
1885 if (MakeNewFilter) {
1886 FilterType = ArrayType::get(FilterType->getElementType(),
1887 NewFilterElts.size());
1888 FilterClause = ConstantArray::get(FilterType, NewFilterElts);
1889 MakeNewInstruction = true;
1892 NewClauses.push_back(FilterClause);
1894 // If the new filter is empty then it will catch everything so there is
1895 // no point in keeping any following clauses or marking the landingpad
1896 // as having a cleanup. The case of the original filter being empty was
1897 // already handled above.
1898 if (MakeNewFilter && !NewFilterElts.size()) {
1899 assert(MakeNewInstruction && "New filter but not a new instruction!");
1900 CleanupFlag = false;
1906 // If several filters occur in a row then reorder them so that the shortest
1907 // filters come first (those with the smallest number of elements). This is
1908 // advantageous because shorter filters are more likely to match, speeding up
1909 // unwinding, but mostly because it increases the effectiveness of the other
1910 // filter optimizations below.
1911 for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
1913 // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
1914 for (j = i; j != e; ++j)
1915 if (!isa<ArrayType>(NewClauses[j]->getType()))
1918 // Check whether the filters are already sorted by length. We need to know
1919 // if sorting them is actually going to do anything so that we only make a
1920 // new landingpad instruction if it does.
1921 for (unsigned k = i; k + 1 < j; ++k)
1922 if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
1923 // Not sorted, so sort the filters now. Doing an unstable sort would be
1924 // correct too but reordering filters pointlessly might confuse users.
1925 std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
1927 MakeNewInstruction = true;
1931 // Look for the next batch of filters.
1935 // If typeinfos matched if and only if equal, then the elements of a filter L
1936 // that occurs later than a filter F could be replaced by the intersection of
1937 // the elements of F and L. In reality two typeinfos can match without being
1938 // equal (for example if one represents a C++ class, and the other some class
1939 // derived from it) so it would be wrong to perform this transform in general.
1940 // However the transform is correct and useful if F is a subset of L. In that
1941 // case L can be replaced by F, and thus removed altogether since repeating a
1942 // filter is pointless. So here we look at all pairs of filters F and L where
1943 // L follows F in the list of clauses, and remove L if every element of F is
1944 // an element of L. This can occur when inlining C++ functions with exception
1946 for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
1947 // Examine each filter in turn.
1948 Value *Filter = NewClauses[i];
1949 ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
1951 // Not a filter - skip it.
1953 unsigned FElts = FTy->getNumElements();
1954 // Examine each filter following this one. Doing this backwards means that
1955 // we don't have to worry about filters disappearing under us when removed.
1956 for (unsigned j = NewClauses.size() - 1; j != i; --j) {
1957 Value *LFilter = NewClauses[j];
1958 ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
1960 // Not a filter - skip it.
1962 // If Filter is a subset of LFilter, i.e. every element of Filter is also
1963 // an element of LFilter, then discard LFilter.
1964 SmallVector<Value *, 16>::iterator J = NewClauses.begin() + j;
1965 // If Filter is empty then it is a subset of LFilter.
1968 NewClauses.erase(J);
1969 MakeNewInstruction = true;
1970 // Move on to the next filter.
1973 unsigned LElts = LTy->getNumElements();
1974 // If Filter is longer than LFilter then it cannot be a subset of it.
1976 // Move on to the next filter.
1978 // At this point we know that LFilter has at least one element.
1979 if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
1980 // Filter is a subset of LFilter iff Filter contains only zeros (as we
1981 // already know that Filter is not longer than LFilter).
1982 if (isa<ConstantAggregateZero>(Filter)) {
1983 assert(FElts <= LElts && "Should have handled this case earlier!");
1985 NewClauses.erase(J);
1986 MakeNewInstruction = true;
1988 // Move on to the next filter.
1991 ConstantArray *LArray = cast<ConstantArray>(LFilter);
1992 if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
1993 // Since Filter is non-empty and contains only zeros, it is a subset of
1994 // LFilter iff LFilter contains a zero.
1995 assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
1996 for (unsigned l = 0; l != LElts; ++l)
1997 if (LArray->getOperand(l)->isNullValue()) {
1998 // LFilter contains a zero - discard it.
1999 NewClauses.erase(J);
2000 MakeNewInstruction = true;
2003 // Move on to the next filter.
2006 // At this point we know that both filters are ConstantArrays. Loop over
2007 // operands to see whether every element of Filter is also an element of
2008 // LFilter. Since filters tend to be short this is probably faster than
2009 // using a method that scales nicely.
2010 ConstantArray *FArray = cast<ConstantArray>(Filter);
2011 bool AllFound = true;
2012 for (unsigned f = 0; f != FElts; ++f) {
2013 Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
2015 for (unsigned l = 0; l != LElts; ++l) {
2016 Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
2017 if (LTypeInfo == FTypeInfo) {
2027 NewClauses.erase(J);
2028 MakeNewInstruction = true;
2030 // Move on to the next filter.
2034 // If we changed any of the clauses, replace the old landingpad instruction
2036 if (MakeNewInstruction) {
2037 LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
2038 LI.getPersonalityFn(),
2040 for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
2041 NLI->addClause(NewClauses[i]);
2042 // A landing pad with no clauses must have the cleanup flag set. It is
2043 // theoretically possible, though highly unlikely, that we eliminated all
2044 // clauses. If so, force the cleanup flag to true.
2045 if (NewClauses.empty())
2047 NLI->setCleanup(CleanupFlag);
2051 // Even if none of the clauses changed, we may nonetheless have understood
2052 // that the cleanup flag is pointless. Clear it if so.
2053 if (LI.isCleanup() != CleanupFlag) {
2054 assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
2055 LI.setCleanup(CleanupFlag);
2065 /// TryToSinkInstruction - Try to move the specified instruction from its
2066 /// current block into the beginning of DestBlock, which can only happen if it's
2067 /// safe to move the instruction past all of the instructions between it and the
2068 /// end of its block.
2069 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
2070 assert(I->hasOneUse() && "Invariants didn't hold!");
2072 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
2073 if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
2074 isa<TerminatorInst>(I))
2077 // Do not sink alloca instructions out of the entry block.
2078 if (isa<AllocaInst>(I) && I->getParent() ==
2079 &DestBlock->getParent()->getEntryBlock())
2082 // We can only sink load instructions if there is nothing between the load and
2083 // the end of block that could change the value.
2084 if (I->mayReadFromMemory()) {
2085 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
2087 if (Scan->mayWriteToMemory())
2091 BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
2092 I->moveBefore(InsertPos);
2098 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
2099 /// all reachable code to the worklist.
2101 /// This has a couple of tricks to make the code faster and more powerful. In
2102 /// particular, we constant fold and DCE instructions as we go, to avoid adding
2103 /// them to the worklist (this significantly speeds up instcombine on code where
2104 /// many instructions are dead or constant). Additionally, if we find a branch
2105 /// whose condition is a known constant, we only visit the reachable successors.
2107 static bool AddReachableCodeToWorklist(BasicBlock *BB,
2108 SmallPtrSet<BasicBlock*, 64> &Visited,
2110 const DataLayout *TD,
2111 const TargetLibraryInfo *TLI) {
2112 bool MadeIRChange = false;
2113 SmallVector<BasicBlock*, 256> Worklist;
2114 Worklist.push_back(BB);
2116 SmallVector<Instruction*, 128> InstrsForInstCombineWorklist;
2117 DenseMap<ConstantExpr*, Constant*> FoldedConstants;
2120 BB = Worklist.pop_back_val();
2122 // We have now visited this block! If we've already been here, ignore it.
2123 if (!Visited.insert(BB)) continue;
2125 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
2126 Instruction *Inst = BBI++;
2128 // DCE instruction if trivially dead.
2129 if (isInstructionTriviallyDead(Inst, TLI)) {
2131 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
2132 Inst->eraseFromParent();
2136 // ConstantProp instruction if trivially constant.
2137 if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
2138 if (Constant *C = ConstantFoldInstruction(Inst, TD, TLI)) {
2139 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
2141 Inst->replaceAllUsesWith(C);
2143 Inst->eraseFromParent();
2148 // See if we can constant fold its operands.
2149 for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
2151 ConstantExpr *CE = dyn_cast<ConstantExpr>(i);
2152 if (CE == 0) continue;
2154 Constant*& FoldRes = FoldedConstants[CE];
2156 FoldRes = ConstantFoldConstantExpression(CE, TD, TLI);
2160 if (FoldRes != CE) {
2162 MadeIRChange = true;
2167 InstrsForInstCombineWorklist.push_back(Inst);
2170 // Recursively visit successors. If this is a branch or switch on a
2171 // constant, only visit the reachable successor.
2172 TerminatorInst *TI = BB->getTerminator();
2173 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2174 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
2175 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
2176 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
2177 Worklist.push_back(ReachableBB);
2180 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2181 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
2182 // See if this is an explicit destination.
2183 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2185 if (i.getCaseValue() == Cond) {
2186 BasicBlock *ReachableBB = i.getCaseSuccessor();
2187 Worklist.push_back(ReachableBB);
2191 // Otherwise it is the default destination.
2192 Worklist.push_back(SI->getDefaultDest());
2197 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
2198 Worklist.push_back(TI->getSuccessor(i));
2199 } while (!Worklist.empty());
2201 // Once we've found all of the instructions to add to instcombine's worklist,
2202 // add them in reverse order. This way instcombine will visit from the top
2203 // of the function down. This jives well with the way that it adds all uses
2204 // of instructions to the worklist after doing a transformation, thus avoiding
2205 // some N^2 behavior in pathological cases.
2206 IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
2207 InstrsForInstCombineWorklist.size());
2209 return MadeIRChange;
2212 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
2213 MadeIRChange = false;
2215 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
2216 << F.getName() << "\n");
2219 // Do a depth-first traversal of the function, populate the worklist with
2220 // the reachable instructions. Ignore blocks that are not reachable. Keep
2221 // track of which blocks we visit.
2222 SmallPtrSet<BasicBlock*, 64> Visited;
2223 MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD,
2226 // Do a quick scan over the function. If we find any blocks that are
2227 // unreachable, remove any instructions inside of them. This prevents
2228 // the instcombine code from having to deal with some bad special cases.
2229 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
2230 if (Visited.count(BB)) continue;
2232 // Delete the instructions backwards, as it has a reduced likelihood of
2233 // having to update as many def-use and use-def chains.
2234 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
2235 while (EndInst != BB->begin()) {
2236 // Delete the next to last instruction.
2237 BasicBlock::iterator I = EndInst;
2238 Instruction *Inst = --I;
2239 if (!Inst->use_empty())
2240 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
2241 if (isa<LandingPadInst>(Inst)) {
2245 if (!isa<DbgInfoIntrinsic>(Inst)) {
2247 MadeIRChange = true;
2249 Inst->eraseFromParent();
2254 while (!Worklist.isEmpty()) {
2255 Instruction *I = Worklist.RemoveOne();
2256 if (I == 0) continue; // skip null values.
2258 // Check to see if we can DCE the instruction.
2259 if (isInstructionTriviallyDead(I, TLI)) {
2260 DEBUG(errs() << "IC: DCE: " << *I << '\n');
2261 EraseInstFromFunction(*I);
2263 MadeIRChange = true;
2267 // Instruction isn't dead, see if we can constant propagate it.
2268 if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
2269 if (Constant *C = ConstantFoldInstruction(I, TD, TLI)) {
2270 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
2272 // Add operands to the worklist.
2273 ReplaceInstUsesWith(*I, C);
2275 EraseInstFromFunction(*I);
2276 MadeIRChange = true;
2280 // See if we can trivially sink this instruction to a successor basic block.
2281 if (I->hasOneUse()) {
2282 BasicBlock *BB = I->getParent();
2283 Instruction *UserInst = cast<Instruction>(I->use_back());
2284 BasicBlock *UserParent;
2286 // Get the block the use occurs in.
2287 if (PHINode *PN = dyn_cast<PHINode>(UserInst))
2288 UserParent = PN->getIncomingBlock(I->use_begin().getUse());
2290 UserParent = UserInst->getParent();
2292 if (UserParent != BB) {
2293 bool UserIsSuccessor = false;
2294 // See if the user is one of our successors.
2295 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
2296 if (*SI == UserParent) {
2297 UserIsSuccessor = true;
2301 // If the user is one of our immediate successors, and if that successor
2302 // only has us as a predecessors (we'd have to split the critical edge
2303 // otherwise), we can keep going.
2304 if (UserIsSuccessor && UserParent->getSinglePredecessor())
2305 // Okay, the CFG is simple enough, try to sink this instruction.
2306 MadeIRChange |= TryToSinkInstruction(I, UserParent);
2310 // Now that we have an instruction, try combining it to simplify it.
2311 Builder->SetInsertPoint(I->getParent(), I);
2312 Builder->SetCurrentDebugLocation(I->getDebugLoc());
2317 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
2318 DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
2320 if (Instruction *Result = visit(*I)) {
2322 // Should we replace the old instruction with a new one?
2324 DEBUG(errs() << "IC: Old = " << *I << '\n'
2325 << " New = " << *Result << '\n');
2327 if (!I->getDebugLoc().isUnknown())
2328 Result->setDebugLoc(I->getDebugLoc());
2329 // Everything uses the new instruction now.
2330 I->replaceAllUsesWith(Result);
2332 // Move the name to the new instruction first.
2333 Result->takeName(I);
2335 // Push the new instruction and any users onto the worklist.
2336 Worklist.Add(Result);
2337 Worklist.AddUsersToWorkList(*Result);
2339 // Insert the new instruction into the basic block...
2340 BasicBlock *InstParent = I->getParent();
2341 BasicBlock::iterator InsertPos = I;
2343 // If we replace a PHI with something that isn't a PHI, fix up the
2345 if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
2346 InsertPos = InstParent->getFirstInsertionPt();
2348 InstParent->getInstList().insert(InsertPos, Result);
2350 EraseInstFromFunction(*I);
2353 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
2354 << " New = " << *I << '\n');
2357 // If the instruction was modified, it's possible that it is now dead.
2358 // if so, remove it.
2359 if (isInstructionTriviallyDead(I, TLI)) {
2360 EraseInstFromFunction(*I);
2363 Worklist.AddUsersToWorkList(*I);
2366 MadeIRChange = true;
2371 return MadeIRChange;
2375 class InstCombinerLibCallSimplifier : public LibCallSimplifier {
2378 InstCombinerLibCallSimplifier(const DataLayout *TD,
2379 const TargetLibraryInfo *TLI,
2381 : LibCallSimplifier(TD, TLI, UnsafeFPShrink) {
2385 /// replaceAllUsesWith - override so that instruction replacement
2386 /// can be defined in terms of the instruction combiner framework.
2387 virtual void replaceAllUsesWith(Instruction *I, Value *With) const {
2388 IC->ReplaceInstUsesWith(*I, With);
2393 bool InstCombiner::runOnFunction(Function &F) {
2394 TD = getAnalysisIfAvailable<DataLayout>();
2395 TLI = &getAnalysis<TargetLibraryInfo>();
2397 /// Builder - This is an IRBuilder that automatically inserts new
2398 /// instructions into the worklist when they are created.
2399 IRBuilder<true, TargetFolder, InstCombineIRInserter>
2400 TheBuilder(F.getContext(), TargetFolder(TD),
2401 InstCombineIRInserter(Worklist));
2402 Builder = &TheBuilder;
2404 InstCombinerLibCallSimplifier TheSimplifier(TD, TLI, this);
2405 Simplifier = &TheSimplifier;
2407 bool EverMadeChange = false;
2409 // Lower dbg.declare intrinsics otherwise their value may be clobbered
2411 EverMadeChange = LowerDbgDeclare(F);
2413 // Iterate while there is work to do.
2414 unsigned Iteration = 0;
2415 while (DoOneIteration(F, Iteration++))
2416 EverMadeChange = true;
2419 return EverMadeChange;
2422 FunctionPass *llvm::createInstructionCombiningPass() {
2423 return new InstCombiner();