Replace the old ADCE implementation with a new one that more simply solves

the one case that ADCE catches that normal DCE doesn't: non-induction variable
loop computations.

This implementation handles this problem without using postdominators.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51668 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Owen Anderson 2008-05-29 08:45:13 +00:00
parent f26ffe987c
commit 038a8746c9
3 changed files with 62 additions and 501 deletions

View File

@ -1,4 +1,4 @@
//===- ADCE.cpp - Code to perform aggressive dead code elimination --------===//
//===- DCE.cpp - Code to perform dead code elimination --------------------===//
//
// The LLVM Compiler Infrastructure
//
@ -7,481 +7,86 @@
//
//===----------------------------------------------------------------------===//
//
// 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.
// This file implements the Aggressive Dead Code Elimination pass. This pass
// optimistically assumes that all instructions are dead until proven otherwise,
// allowing it to eliminate dead computations that other DCE passes do not
// catch, particularly involving loop computations.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "adce"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Constants.h"
#include "llvm/Instructions.h"
#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/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Pass.h"
#include "llvm/Support/Compiler.h"
#include <algorithm>
#include "llvm/Support/InstIterator.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/SmallPtrSet.h"
using namespace llvm;
STATISTIC(NumBlockRemoved, "Number of basic blocks removed");
STATISTIC(NumInstRemoved , "Number of instructions removed");
STATISTIC(NumCallRemoved , "Number of calls removed");
STATISTIC(NumRemoved, "Number of instructions removed");
namespace {
//===----------------------------------------------------------------------===//
// ADCE Class
//
// 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 VISIBILITY_HIDDEN 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:
static char ID; // Pass identification, replacement for typeid
ADCE() : FunctionPass((intptr_t)&ID) {}
// 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>();
}
//===--------------------------------------------------------------------===//
// 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);
// deleteDeadInstructionsInLiveBlock - Loop over all of the instructions in
// the specified basic block, deleting ones that are dead according to
// LiveSet.
bool deleteDeadInstructionsInLiveBlock(BasicBlock *BB);
TerminatorInst *convertToUnconditionalBranch(TerminatorInst *TI);
inline void markInstructionLive(Instruction *I) {
if (!LiveSet.insert(I).second) return;
DOUT << "Insn Live: " << *I;
WorkList.push_back(I);
}
inline void markTerminatorLive(const BasicBlock *BB) {
DOUT << "Terminator Live: " << *BB->getTerminator();
markInstructionLive(const_cast<TerminatorInst*>(BB->getTerminator()));
}
};
} // End of anonymous namespace
struct VISIBILITY_HIDDEN ADCE : public FunctionPass {
static char ID; // Pass identification, replacement for typeid
ADCE() : FunctionPass((intptr_t)&ID) {}
virtual bool runOnFunction(Function& F);
virtual void getAnalysisUsage(AnalysisUsage& AU) const {
AU.setPreservesCFG();
}
};
}
char ADCE::ID = 0;
static RegisterPass<ADCE> X("adce", "Aggressive Dead Code Elimination");
FunctionPass *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 (PostDominanceFrontier::DomSetType::const_iterator I =
CDB.begin(), E = CDB.end(); I != E; ++I)
markTerminatorLive(*I); // Mark all their terminators as live
}
// 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);
}
// deleteDeadInstructionsInLiveBlock - Loop over all of the instructions in the
// specified basic block, deleting ones that are dead according to LiveSet.
bool ADCE::deleteDeadInstructionsInLiveBlock(BasicBlock *BB) {
bool Changed = false;
for (BasicBlock::iterator II = BB->begin(), E = --BB->end(); II != E; ) {
Instruction *I = II++;
if (!LiveSet.count(I)) { // Is this instruction alive?
if (!I->use_empty())
I->replaceAllUsesWith(UndefValue::get(I->getType()));
// Nope... remove the instruction from it's basic block...
if (isa<CallInst>(I))
++NumCallRemoved;
else
++NumInstRemoved;
BB->getInstList().erase(I);
Changed = true;
}
}
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 = BranchInst::Create(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 Aggressive Dead Code Elimination algorithm, returning
// true if the function was modified.
//
bool ADCE::doADCE() {
bool MadeChanges = false;
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
// Iterate over all of the instructions in the function, eliminating trivially
// dead instructions, and marking instructions live that are known to be
// 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.
//
std::set<BasicBlock*> ReachableBBs;
std::vector<BasicBlock*> Stack;
Stack.push_back(&Func->getEntryBlock());
bool ADCE::runOnFunction(Function& F) {
SmallPtrSet<Instruction*, 32> alive;
std::vector<Instruction*> worklist;
while (!Stack.empty()) {
BasicBlock* BB = Stack.back();
if (ReachableBBs.count(BB)) {
Stack.pop_back();
continue;
} else {
ReachableBBs.insert(BB);
// Collect the set of "root" instructions that are known live.
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
if (isa<TerminatorInst>(I.getInstructionIterator()) ||
I->mayWriteToMemory()) {
alive.insert(I.getInstructionIterator());
worklist.push_back(I.getInstructionIterator());
}
// Propagate liveness backwards to operands.
while (!worklist.empty()) {
Instruction* curr = worklist.back();
worklist.pop_back();
for (BasicBlock::iterator II = BB->begin(), EI = BB->end(); II != EI; ) {
Instruction *I = II++;
if (CallInst *CI = dyn_cast<CallInst>(I)) {
if (AA.onlyReadsMemory(CI)) {
if (CI->use_empty()) {
BB->getInstList().erase(CI);
++NumCallRemoved;
}
} else {
markInstructionLive(I);
}
} else if (I->mayWriteToMemory() || isa<ReturnInst>(I) ||
isa<UnwindInst>(I) || isa<UnreachableInst>(I)) {
// FIXME: Unreachable instructions should not be marked intrinsically
// live here.
markInstructionLive(I);
} else if (isInstructionTriviallyDead(I)) {
// Remove the instruction from it's basic block...
BB->getInstList().erase(I);
++NumInstRemoved;
}
for (Instruction::op_iterator OI = curr->op_begin(), OE = curr->op_end();
OI != OE; ++OI)
if (Instruction* Inst = dyn_cast<Instruction>(OI))
if (alive.insert(Inst))
worklist.push_back(Inst);
}
// The inverse of the live set is the dead set. These are those instructions
// which have no side effects and do not influence the control flow or return
// value of the function, and may therefore be deleted safely.
SmallPtrSet<Instruction*, 32> dead;
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
if (!alive.count(I.getInstructionIterator())) {
dead.insert(I.getInstructionIterator());
I->dropAllReferences();
}
for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) {
// Back edges (as opposed to cross edges) indicate loops, so implicitly
// mark them live.
if (std::find(Stack.begin(), Stack.end(), *SI) != Stack.end())
markInstructionLive(BB->getTerminator());
if (!ReachableBBs.count(*SI))
Stack.push_back(*SI);
}
for (SmallPtrSet<Instruction*, 32>::iterator I = dead.begin(),
E = dead.end(); I != E; ++I) {
NumRemoved++;
(*I)->eraseFromParent();
}
// 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 && ReachableBBs.count(I))
for (pred_iterator PI = pred_begin(I), E = pred_end(I); PI != E; ++PI)
markInstructionLive((*PI)->getTerminator());
DOUT << "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 necessary as
// well... make them live as well.
//
while (!WorkList.empty()) {
Instruction *I = WorkList.back(); // Get an instruction that became live...
WorkList.pop_back();
BasicBlock *BB = I->getParent();
if (!ReachableBBs.count(BB)) continue;
if (AliveBlocks.insert(BB).second) // Basic block not alive yet.
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 (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
// If the incoming edge is clearly dead, it won't have control
// dependence information. Do not mark it live.
BasicBlock *PredBB = PN->getIncomingBlock(i);
if (ReachableBBs.count(PredBB)) {
// FIXME: This should mark the control dependent edge as live, not
// necessarily the predecessor itself!
if (AliveBlocks.insert(PredBB).second)
markBlockAlive(PN->getIncomingBlock(i)); // Block is newly ALIVE!
if (Instruction *Op = dyn_cast<Instruction>(PN->getIncomingValue(i)))
markInstructionLive(Op);
}
}
} else {
// 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);
}
}
DEBUG(
DOUT << "Current Function: X = Live\n";
for (Function::iterator I = Func->begin(), E = Func->end(); I != E; ++I){
DOUT << 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)) DOUT << "X ";
DOUT << *BI;
}
});
// All blocks being live is a common case, handle it specially.
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 deleting instructions
// to drop their references.
deleteDeadInstructionsInLiveBlock(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 an unconditional branch.
//
TerminatorInst *TI = I->getTerminator();
if (!LiveSet.count(TI))
convertToUnconditionalBranch(TI);
}
return MadeChanges;
}
// 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 = BasicBlock::Create();
BranchInst::Create(&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...
//
DomTreeNode *LastNode = DT[TI->getSuccessor(i)];
DomTreeNode *NextNode = 0;
if (LastNode) {
NextNode = LastNode->getIDom();
while (!AliveBlocks.count(NextNode->getBlock())) {
LastNode = NextNode;
NextNode = NextNode->getIDom();
if (NextNode == 0) {
LastNode = 0;
break;
}
}
}
// There is a special case here... if there IS no post-dominator for
// the block we have nowhere 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!
if (!isa<InvokeInst>(TI)) {
// 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.
//
TI = BB->getTerminator();
// Rescan this successor...
--i;
} else {
}
} else {
// 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();
isa<PHINode>(II); ++II) {
PHINode *PN = cast<PHINode>(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, deleting
// dead instructions. This is so that the next sweep over the program
// can safely delete dead instructions without other dead instructions
// still referring to them.
//
deleteDeadInstructionsInLiveBlock(BB);
}
// 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...
//
std::vector<BasicBlock*> DeadBlocks;
for (Function::iterator BB = Func->begin(), E = Func->end(); BB != E; ++BB)
if (!AliveBlocks.count(BB)) {
// Remove PHI node entries for this block in live successor blocks.
for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
if (!SI->empty() && isa<PHINode>(SI->front()) && AliveBlocks.count(*SI))
(*SI)->removePredecessor(BB);
BB->dropAllReferences();
MadeChanges = true;
DeadBlocks.push_back(BB);
}
NumBlockRemoved += DeadBlocks.size();
// 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).
for (std::vector<BasicBlock*>::iterator I = DeadBlocks.begin(),
E = DeadBlocks.end(); I != E; ++I)
Func->getBasicBlockList().erase(*I);
return MadeChanges;
return !dead.empty();
}
FunctionPass *llvm::createAggressiveDCEPass() {
return new ADCE();
}

View File

@ -1,27 +0,0 @@
; This testcase was failing because without merging the return blocks, ADCE
; didn't know that it could get rid of the then.0 block.
; RUN: llvm-as < %s | opt -adce | llvm-dis | not grep load
define void @main(i32 %argc, i8** %argv) {
entry:
call void @__main( )
%tmp.1 = icmp ule i32 %argc, 5 ; <i1> [#uses=1]
br i1 %tmp.1, label %then.0, label %return
then.0: ; preds = %entry
%tmp.8 = load i8** %argv ; <i8*> [#uses=1]
%tmp.10 = load i8* %tmp.8 ; <i8> [#uses=1]
%tmp.11 = icmp eq i8 %tmp.10, 98 ; <i1> [#uses=1]
br i1 %tmp.11, label %then.1, label %return
then.1: ; preds = %then.0
ret void
return: ; preds = %then.0, %entry
ret void
}
declare void @__main()

View File

@ -1,17 +0,0 @@
; RUN: llvm-as < %s | opt -adce | llvm-dis | not grep call
; The call is not live just because the PHI uses the call retval!
define i32 @test(i32 %X) {
; <label>:0
br label %Done
DeadBlock: ; No predecessors!
%Y = call i32 @test( i32 0 ) ; <i32> [#uses=1]
br label %Done
Done: ; preds = %DeadBlock, %0
%Z = phi i32 [ %X, %0 ], [ %Y, %DeadBlock ] ; <i32> [#uses=1]
ret i32 %Z
}