-//===- ADCE.cpp - Code to perform agressive dead code elimination ---------===//
+//===- ADCE.cpp - Code to perform aggressive dead code elimination --------===//
+//
+// The LLVM Compiler Infrastructure
//
-// This file implements "agressive" dead code elimination. ADCE is DCe where
+// This file was developed by the LLVM research group and is distributed under
+// the University of Illinois Open Source License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements "aggressive" dead code elimination. ADCE is DCe where
// values are assumed to be dead until proven otherwise. This is similar to
// SCCP, except applied to the liveness of values.
//
//===----------------------------------------------------------------------===//
-#include "llvm/Optimizations/DCE.h"
-#include "llvm/Instruction.h"
+#include "llvm/Transforms/Scalar.h"
+#include "llvm/Constant.h"
+#include "llvm/Instructions.h"
#include "llvm/Type.h"
-#include "llvm/Analysis/Dominators.h"
-#include <set>
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/PostDominators.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Transforms/Utils/UnifyFunctionExitNodes.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/ADT/DepthFirstIterator.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/STLExtras.h"
+#include <algorithm>
+using namespace llvm;
-#include "llvm/Analysis/Writer.h"
+namespace {
+ Statistic<> NumBlockRemoved("adce", "Number of basic blocks removed");
+ Statistic<> NumInstRemoved ("adce", "Number of instructions removed");
+ Statistic<> NumCallRemoved ("adce", "Number of calls and invokes removed");
//===----------------------------------------------------------------------===//
// ADCE Class
//
-// This class does all of the work of Agressive Dead Code Elimination.
+// This class does all of the work of Aggressive Dead Code Elimination.
// It's public interface consists of a constructor and a doADCE() method.
//
-class ADCE {
- Method *M; // The method that we are working on...
- vector<Instruction*> WorkList; // Instructions that just became live
- set<Instruction*> LiveSet; // The set of live instructions
+class ADCE : public FunctionPass {
+ Function *Func; // The function that we are working on
+ std::vector<Instruction*> WorkList; // Instructions that just became live
+ std::set<Instruction*> LiveSet; // The set of live instructions
//===--------------------------------------------------------------------===//
// The public interface for this class
//
public:
- // ADCE Ctor - Save the method to operate on...
- inline ADCE(Method *m) : M(m) {}
+ // Execute the Aggressive Dead Code Elimination Algorithm
+ //
+ virtual bool runOnFunction(Function &F) {
+ Func = &F;
+ bool Changed = doADCE();
+ assert(WorkList.empty());
+ LiveSet.clear();
+ return Changed;
+ }
+ // getAnalysisUsage - We require post dominance frontiers (aka Control
+ // Dependence Graph)
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ // We require that all function nodes are unified, because otherwise code
+ // can be marked live that wouldn't necessarily be otherwise.
+ AU.addRequired<UnifyFunctionExitNodes>();
+ AU.addRequired<AliasAnalysis>();
+ AU.addRequired<PostDominatorTree>();
+ AU.addRequired<PostDominanceFrontier>();
+ }
- // doADCE() - Run the Agressive Dead Code Elimination algorithm, returning
- // true if the method was modified.
- bool doADCE();
//===--------------------------------------------------------------------===//
// The implementation of this class
//
private:
+ // doADCE() - Run the Aggressive Dead Code Elimination algorithm, returning
+ // true if the function was modified.
+ //
+ bool doADCE();
+
+ void markBlockAlive(BasicBlock *BB);
+
+
+ // dropReferencesOfDeadInstructionsInLiveBlock - Loop over all of the
+ // instructions in the specified basic block, dropping references on
+ // instructions that are dead according to LiveSet.
+ bool dropReferencesOfDeadInstructionsInLiveBlock(BasicBlock *BB);
+
+ TerminatorInst *convertToUnconditionalBranch(TerminatorInst *TI);
+
inline void markInstructionLive(Instruction *I) {
if (LiveSet.count(I)) return;
- cerr << "Insn Live: " << I;
+ DEBUG(std::cerr << "Insn Live: " << *I);
LiveSet.insert(I);
WorkList.push_back(I);
}
-
+ inline void markTerminatorLive(const BasicBlock *BB) {
+ DEBUG(std::cerr << "Terminator Live: " << *BB->getTerminator());
+ markInstructionLive(const_cast<TerminatorInst*>(BB->getTerminator()));
+ }
};
+ RegisterOpt<ADCE> X("adce", "Aggressive Dead Code Elimination");
+} // End of anonymous namespace
+
+Pass *llvm::createAggressiveDCEPass() { return new ADCE(); }
+
+void ADCE::markBlockAlive(BasicBlock *BB) {
+ // Mark the basic block as being newly ALIVE... and mark all branches that
+ // this block is control dependent on as being alive also...
+ //
+ PostDominanceFrontier &CDG = getAnalysis<PostDominanceFrontier>();
+
+ PostDominanceFrontier::const_iterator It = CDG.find(BB);
+ if (It != CDG.end()) {
+ // Get the blocks that this node is control dependent on...
+ const PostDominanceFrontier::DomSetType &CDB = It->second;
+ for_each(CDB.begin(), CDB.end(), // Mark all their terminators as live
+ bind_obj(this, &ADCE::markTerminatorLive));
+ }
+
+ // If this basic block is live, and it ends in an unconditional branch, then
+ // the branch is alive as well...
+ if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
+ if (BI->isUnconditional())
+ markTerminatorLive(BB);
+}
+
+// dropReferencesOfDeadInstructionsInLiveBlock - Loop over all of the
+// instructions in the specified basic block, dropping references on
+// instructions that are dead according to LiveSet.
+bool ADCE::dropReferencesOfDeadInstructionsInLiveBlock(BasicBlock *BB) {
+ bool Changed = false;
+ for (BasicBlock::iterator I = BB->begin(), E = --BB->end(); I != E; )
+ if (!LiveSet.count(I)) { // Is this instruction alive?
+ I->dropAllReferences(); // Nope, drop references...
+ if (PHINode *PN = dyn_cast<PHINode>(I)) {
+ // We don't want to leave PHI nodes in the program that have
+ // #arguments != #predecessors, so we remove them now.
+ //
+ PN->replaceAllUsesWith(Constant::getNullValue(PN->getType()));
+
+ // Delete the instruction...
+ ++I;
+ BB->getInstList().erase(PN);
+ Changed = true;
+ ++NumInstRemoved;
+ } else {
+ ++I;
+ }
+ } else {
+ ++I;
+ }
+ return Changed;
+}
+
+
+/// convertToUnconditionalBranch - Transform this conditional terminator
+/// instruction into an unconditional branch because we don't care which of the
+/// successors it goes to. This eliminate a use of the condition as well.
+///
+TerminatorInst *ADCE::convertToUnconditionalBranch(TerminatorInst *TI) {
+ BranchInst *NB = new BranchInst(TI->getSuccessor(0), TI);
+ BasicBlock *BB = TI->getParent();
+ // Remove entries from PHI nodes to avoid confusing ourself later...
+ for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
+ TI->getSuccessor(i)->removePredecessor(BB);
+
+ // Delete the old branch itself...
+ BB->getInstList().erase(TI);
+ return NB;
+}
-// doADCE() - Run the Agressive Dead Code Elimination algorithm, returning
-// true if the method was modified.
+
+// doADCE() - Run the Aggressive Dead Code Elimination algorithm, returning
+// true if the function was modified.
//
bool ADCE::doADCE() {
- // Iterate over all of the instructions in the method, eliminating trivially
+ bool MadeChanges = false;
+
+ AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
+
+
+ // Iterate over all invokes in the function, turning invokes into calls if
+ // they cannot throw.
+ for (Function::iterator BB = Func->begin(), E = Func->end(); BB != E; ++BB)
+ if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator()))
+ if (Function *F = II->getCalledFunction())
+ if (AA.onlyReadsMemory(F)) {
+ // The function cannot unwind. Convert it to a call with a branch
+ // after it to the normal destination.
+ std::vector<Value*> Args(II->op_begin()+3, II->op_end());
+ std::string Name = II->getName(); II->setName("");
+ Instruction *NewCall = new CallInst(F, Args, Name, II);
+ II->replaceAllUsesWith(NewCall);
+ new BranchInst(II->getNormalDest(), II);
+
+ // Update PHI nodes in the unwind destination
+ II->getUnwindDest()->removePredecessor(BB);
+ BB->getInstList().erase(II);
+
+ if (NewCall->use_empty()) {
+ BB->getInstList().erase(NewCall);
+ ++NumCallRemoved;
+ }
+ }
+
+ // Iterate over all of the instructions in the function, eliminating trivially
// dead instructions, and marking instructions live that are known to be
- // needed.
+ // needed. Perform the walk in depth first order so that we avoid marking any
+ // instructions live in basic blocks that are unreachable. These blocks will
+ // be eliminated later, along with the instructions inside.
//
- for (Method::inst_iterator II = M->inst_begin(); II != M->inst_end(); ) {
- Instruction *I = *II;
- switch (I->getInstType()) {
- case Instruction::Call:
- case Instruction::Store:
- markInstructionLive(I);
- break;
- default:
- if (I->getType() == Type::VoidTy) {
- markInstructionLive(I); // Catches terminators and friends
- } else {
- if (I->use_size() == 0) { // Check to see if anything is trivially dead
- // Remove the instruction from it's basic block...
- BasicBlock *BB = I->getParent();
- delete BB->getInstList().remove(II.getInstructionIterator());
-
- // Make sure to sync up the iterator again...
- II.resyncInstructionIterator();
- continue; // Don't increment the iterator past the current slot
- }
+ std::set<BasicBlock*> ReachableBBs;
+ for (df_ext_iterator<BasicBlock*>
+ BBI = df_ext_begin(&Func->front(), ReachableBBs),
+ BBE = df_ext_end(&Func->front(), ReachableBBs); BBI != BBE; ++BBI) {
+ BasicBlock *BB = *BBI;
+ for (BasicBlock::iterator II = BB->begin(), EI = BB->end(); II != EI; ) {
+ Instruction *I = II++;
+ if (CallInst *CI = dyn_cast<CallInst>(I)) {
+ Function *F = CI->getCalledFunction();
+ if (F && AA.onlyReadsMemory(F)) {
+ if (CI->use_empty()) {
+ BB->getInstList().erase(CI);
+ ++NumCallRemoved;
+ }
+ } else {
+ markInstructionLive(I);
+ }
+ } else if (I->mayWriteToMemory() || isa<ReturnInst>(I) ||
+ isa<UnwindInst>(I)) {
+ markInstructionLive(I);
+ } else if (isInstructionTriviallyDead(I)) {
+ // Remove the instruction from it's basic block...
+ BB->getInstList().erase(I);
+ ++NumInstRemoved;
}
}
+ }
- ++II; // Increment the iterator
+ // Check to ensure we have an exit node for this CFG. If we don't, we won't
+ // have any post-dominance information, thus we cannot perform our
+ // transformations safely.
+ //
+ PostDominatorTree &DT = getAnalysis<PostDominatorTree>();
+ if (DT[&Func->getEntryBlock()] == 0) {
+ WorkList.clear();
+ return MadeChanges;
}
+ // Scan the function marking blocks without post-dominance information as
+ // live. Blocks without post-dominance information occur when there is an
+ // infinite loop in the program. Because the infinite loop could contain a
+ // function which unwinds, exits or has side-effects, we don't want to delete
+ // the infinite loop or those blocks leading up to it.
+ for (Function::iterator I = Func->begin(), E = Func->end(); I != E; ++I)
+ if (DT[I] == 0)
+ for (pred_iterator PI = pred_begin(I), E = pred_end(I); PI != E; ++PI)
+ markInstructionLive((*PI)->getTerminator());
+
- cerr << "Processing work list\n";
+
+ DEBUG(std::cerr << "Processing work list\n");
+
+ // AliveBlocks - Set of basic blocks that we know have instructions that are
+ // alive in them...
+ //
+ std::set<BasicBlock*> AliveBlocks;
// Process the work list of instructions that just became live... if they
- // became live, then that means that all of their operands are neccesary as
+ // became live, then that means that all of their operands are necessary as
// well... make them live as well.
//
while (!WorkList.empty()) {
- Instruction *I = WorkList.back();
+ Instruction *I = WorkList.back(); // Get an instruction that became live...
WorkList.pop_back();
- for (unsigned op = 0; Value *Op = I->getOperand(op); ++op) {
- Instruction *Operand = Op->castInstruction();
- if (Operand) markInstructionLive(Operand);
+ BasicBlock *BB = I->getParent();
+ if (!ReachableBBs.count(BB)) continue;
+ if (!AliveBlocks.count(BB)) { // Basic block not alive yet...
+ AliveBlocks.insert(BB); // Block is now ALIVE!
+ markBlockAlive(BB); // Make it so now!
}
+
+ // PHI nodes are a special case, because the incoming values are actually
+ // defined in the predecessor nodes of this block, meaning that the PHI
+ // makes the predecessors alive.
+ //
+ if (PHINode *PN = dyn_cast<PHINode>(I))
+ for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI)
+ if (!AliveBlocks.count(*PI)) {
+ AliveBlocks.insert(BB); // Block is now ALIVE!
+ markBlockAlive(*PI);
+ }
+
+ // Loop over all of the operands of the live instruction, making sure that
+ // they are known to be alive as well...
+ //
+ for (unsigned op = 0, End = I->getNumOperands(); op != End; ++op)
+ if (Instruction *Operand = dyn_cast<Instruction>(I->getOperand(op)))
+ markInstructionLive(Operand);
}
- // After the worklist is processed, loop through the instructions again,
- // removing any that are not live... by the definition of the LiveSet.
+ DEBUG(
+ std::cerr << "Current Function: X = Live\n";
+ for (Function::iterator I = Func->begin(), E = Func->end(); I != E; ++I){
+ std::cerr << I->getName() << ":\t"
+ << (AliveBlocks.count(I) ? "LIVE\n" : "DEAD\n");
+ for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE; ++BI){
+ if (LiveSet.count(BI)) std::cerr << "X ";
+ std::cerr << *BI;
+ }
+ });
+
+ // Find the first postdominator of the entry node that is alive. Make it the
+ // new entry node...
//
- for (Method::inst_iterator II = M->inst_begin(); II != M->inst_end(); ) {
- Instruction *I = *II;
- if (!LiveSet.count(I)) {
- cerr << "Instruction Dead: " << I;
+ if (AliveBlocks.size() == Func->size()) { // No dead blocks?
+ for (Function::iterator I = Func->begin(), E = Func->end(); I != E; ++I) {
+ // Loop over all of the instructions in the function, telling dead
+ // instructions to drop their references. This is so that the next sweep
+ // over the program can safely delete dead instructions without other dead
+ // instructions still referring to them.
+ //
+ dropReferencesOfDeadInstructionsInLiveBlock(I);
+
+ // Check to make sure the terminator instruction is live. If it isn't,
+ // this means that the condition that it branches on (we know it is not an
+ // unconditional branch), is not needed to make the decision of where to
+ // go to, because all outgoing edges go to the same place. We must remove
+ // the use of the condition (because it's probably dead), so we convert
+ // the terminator to a conditional branch.
+ //
+ TerminatorInst *TI = I->getTerminator();
+ if (!LiveSet.count(TI))
+ convertToUnconditionalBranch(TI);
+ }
+
+ } else { // If there are some blocks dead...
+ // If the entry node is dead, insert a new entry node to eliminate the entry
+ // node as a special case.
+ //
+ if (!AliveBlocks.count(&Func->front())) {
+ BasicBlock *NewEntry = new BasicBlock();
+ new BranchInst(&Func->front(), NewEntry);
+ Func->getBasicBlockList().push_front(NewEntry);
+ AliveBlocks.insert(NewEntry); // This block is always alive!
+ LiveSet.insert(NewEntry->getTerminator()); // The branch is live
}
+
+ // Loop over all of the alive blocks in the function. If any successor
+ // blocks are not alive, we adjust the outgoing branches to branch to the
+ // first live postdominator of the live block, adjusting any PHI nodes in
+ // the block to reflect this.
+ //
+ for (Function::iterator I = Func->begin(), E = Func->end(); I != E; ++I)
+ if (AliveBlocks.count(I)) {
+ BasicBlock *BB = I;
+ TerminatorInst *TI = BB->getTerminator();
+
+ // If the terminator instruction is alive, but the block it is contained
+ // in IS alive, this means that this terminator is a conditional branch
+ // on a condition that doesn't matter. Make it an unconditional branch
+ // to ONE of the successors. This has the side effect of dropping a use
+ // of the conditional value, which may also be dead.
+ if (!LiveSet.count(TI))
+ TI = convertToUnconditionalBranch(TI);
+
+ // Loop over all of the successors, looking for ones that are not alive.
+ // We cannot save the number of successors in the terminator instruction
+ // here because we may remove them if we don't have a postdominator...
+ //
+ for (unsigned i = 0; i != TI->getNumSuccessors(); ++i)
+ if (!AliveBlocks.count(TI->getSuccessor(i))) {
+ // Scan up the postdominator tree, looking for the first
+ // postdominator that is alive, and the last postdominator that is
+ // dead...
+ //
+ PostDominatorTree::Node *LastNode = DT[TI->getSuccessor(i)];
+
+ // There is a special case here... if there IS no post-dominator for
+ // the block we have no owhere to point our branch to. Instead,
+ // convert it to a return. This can only happen if the code
+ // branched into an infinite loop. Note that this may not be
+ // desirable, because we _are_ altering the behavior of the code.
+ // This is a well known drawback of ADCE, so in the future if we
+ // choose to revisit the decision, this is where it should be.
+ //
+ if (LastNode == 0) { // No postdominator!
+ // Call RemoveSuccessor to transmogrify the terminator instruction
+ // to not contain the outgoing branch, or to create a new
+ // terminator if the form fundamentally changes (i.e.,
+ // unconditional branch to return). Note that this will change a
+ // branch into an infinite loop into a return instruction!
+ //
+ RemoveSuccessor(TI, i);
+
+ // RemoveSuccessor may replace TI... make sure we have a fresh
+ // pointer... and e variable.
+ //
+ TI = BB->getTerminator();
+
+ // Rescan this successor...
+ --i;
+ } else {
+ PostDominatorTree::Node *NextNode = LastNode->getIDom();
- ++II; // Increment the iterator
+ while (!AliveBlocks.count(NextNode->getBlock())) {
+ LastNode = NextNode;
+ NextNode = NextNode->getIDom();
+ }
+
+ // Get the basic blocks that we need...
+ BasicBlock *LastDead = LastNode->getBlock();
+ BasicBlock *NextAlive = NextNode->getBlock();
+
+ // Make the conditional branch now go to the next alive block...
+ TI->getSuccessor(i)->removePredecessor(BB);
+ TI->setSuccessor(i, NextAlive);
+
+ // If there are PHI nodes in NextAlive, we need to add entries to
+ // the PHI nodes for the new incoming edge. The incoming values
+ // should be identical to the incoming values for LastDead.
+ //
+ for (BasicBlock::iterator II = NextAlive->begin();
+ PHINode *PN = dyn_cast<PHINode>(II); ++II)
+ if (LiveSet.count(PN)) { // Only modify live phi nodes
+ // Get the incoming value for LastDead...
+ int OldIdx = PN->getBasicBlockIndex(LastDead);
+ assert(OldIdx != -1 &&"LastDead is not a pred of NextAlive!");
+ Value *InVal = PN->getIncomingValue(OldIdx);
+
+ // Add an incoming value for BB now...
+ PN->addIncoming(InVal, BB);
+ }
+ }
+ }
+
+ // Now loop over all of the instructions in the basic block, telling
+ // dead instructions to drop their references. This is so that the next
+ // sweep over the program can safely delete dead instructions without
+ // other dead instructions still referring to them.
+ //
+ dropReferencesOfDeadInstructionsInLiveBlock(BB);
+ }
}
- return false;
-}
+ // We make changes if there are any dead blocks in the function...
+ if (unsigned NumDeadBlocks = Func->size() - AliveBlocks.size()) {
+ MadeChanges = true;
+ NumBlockRemoved += NumDeadBlocks;
+ }
+ // Loop over all of the basic blocks in the function, removing control flow
+ // edges to live blocks (also eliminating any entries in PHI functions in
+ // referenced blocks).
+ //
+ for (Function::iterator BB = Func->begin(), E = Func->end(); BB != E; ++BB)
+ if (!AliveBlocks.count(BB)) {
+ // Remove all outgoing edges from this basic block and convert the
+ // terminator into a return instruction.
+ std::vector<BasicBlock*> Succs(succ_begin(BB), succ_end(BB));
+
+ if (!Succs.empty()) {
+ // Loop over all of the successors, removing this block from PHI node
+ // entries that might be in the block...
+ while (!Succs.empty()) {
+ Succs.back()->removePredecessor(BB);
+ Succs.pop_back();
+ }
+
+ // Delete the old terminator instruction...
+ const Type *TermTy = BB->getTerminator()->getType();
+ if (TermTy != Type::VoidTy)
+ BB->getTerminator()->replaceAllUsesWith(
+ Constant::getNullValue(TermTy));
+ BB->getInstList().pop_back();
+ const Type *RetTy = Func->getReturnType();
+ new ReturnInst(RetTy != Type::VoidTy ?
+ Constant::getNullValue(RetTy) : 0, BB);
+ }
+ }
-// DoADCE - Execute the Agressive Dead Code Elimination Algorithm
-//
-bool opt::DoADCE(Method *M) {
- ADCE DCE(M);
- return DCE.doADCE();
+
+ // Loop over all of the basic blocks in the function, dropping references of
+ // the dead basic blocks. We must do this after the previous step to avoid
+ // dropping references to PHIs which still have entries...
+ //
+ for (Function::iterator BB = Func->begin(), E = Func->end(); BB != E; ++BB)
+ if (!AliveBlocks.count(BB))
+ BB->dropAllReferences();
+
+ // Now loop through all of the blocks and delete the dead ones. We can safely
+ // do this now because we know that there are no references to dead blocks
+ // (because they have dropped all of their references... we also remove dead
+ // instructions from alive blocks.
+ //
+ for (Function::iterator BI = Func->begin(); BI != Func->end(); )
+ if (!AliveBlocks.count(BI)) { // Delete dead blocks...
+ BI = Func->getBasicBlockList().erase(BI);
+ } else { // Scan alive blocks...
+ for (BasicBlock::iterator II = BI->begin(); II != --BI->end(); )
+ if (!LiveSet.count(II)) { // Is this instruction alive?
+ // Nope... remove the instruction from it's basic block...
+ if (isa<CallInst>(II))
+ ++NumCallRemoved;
+ else
+ ++NumInstRemoved;
+ II = BI->getInstList().erase(II);
+ MadeChanges = true;
+ } else {
+ ++II;
+ }
+
+ ++BI; // Increment iterator...
+ }
+
+ return MadeChanges;
}
projects / oota-llvm.git / blobdiff