Fix PR681 by using the standard Lengauer and Tarjan algorithm for dominator

set construction, rather than intersecting various std::sets.  This reduces
the memory usage for the testcase in PR681 from 496 to 26MB of ram on my
darwin system, and reduces the runtime from 32.8 to 0.8 seconds on a
2.5GHz G5.  This also enables future code sharing between Dom and PostDom
now that they share near-identical implementations.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26707 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Nate Begeman 2006-03-11 02:20:46 +00:00
parent 78167faa18
commit 442b32b5c5
2 changed files with 238 additions and 172 deletions

View File

@ -18,6 +18,42 @@
namespace llvm {
//===-------------------------------------
/// ImmediatePostDominators Class - Concrete subclass of ImmediateDominatorsBase
/// that is used to compute a normal immediate dominator set.
///
struct ImmediatePostDominators : public ImmediateDominatorsBase {
ImmediatePostDominators() : ImmediateDominatorsBase(false) {}
virtual bool runOnFunction(Function &F);
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
}
private:
struct InfoRec {
unsigned Semi;
unsigned Size;
BasicBlock *Label, *Parent, *Child, *Ancestor;
std::vector<BasicBlock*> Bucket;
InfoRec() : Semi(0), Size(0), Label(0), Parent(0), Child(0), Ancestor(0){}
};
// Vertex - Map the DFS number to the BasicBlock*
std::vector<BasicBlock*> Vertex;
// Info - Collection of information used during the computation of idoms.
std::map<BasicBlock*, InfoRec> Info;
unsigned DFSPass(BasicBlock *V, InfoRec &VInfo, unsigned N);
void Compress(BasicBlock *V, InfoRec &VInfo);
BasicBlock *Eval(BasicBlock *v);
void Link(BasicBlock *V, BasicBlock *W, InfoRec &WInfo);
};
/// PostDominatorSet Class - Concrete subclass of DominatorSetBase that is used
/// to compute the post-dominator set. Because there can be multiple exit nodes
/// in an LLVM function, we calculate post dominators with a special null block
@ -27,40 +63,20 @@ namespace llvm {
///
struct PostDominatorSet : public DominatorSetBase {
PostDominatorSet() : DominatorSetBase(true) {}
virtual bool runOnFunction(Function &F);
/// getAnalysisUsage - This pass does not modify the function at all.
/// getAnalysisUsage - This simply provides a dominator set
///
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<ImmediatePostDominators>();
AU.setPreservesAll();
}
// stub - dummy function, just ignore it
static void stub();
};
/// ImmediatePostDominators Class - Concrete subclass of ImmediateDominatorsBase
/// that is used to compute the immediate post-dominators.
///
struct ImmediatePostDominators : public ImmediateDominatorsBase {
ImmediatePostDominators() : ImmediateDominatorsBase(true) {}
virtual bool runOnFunction(Function &F) {
IDoms.clear(); // Reset from the last time we were run...
PostDominatorSet &DS = getAnalysis<PostDominatorSet>();
Roots = DS.getRoots();
calcIDoms(DS);
return false;
}
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequired<PostDominatorSet>();
}
private:
void calcIDoms(const DominatorSetBase &DS);
};
/// PostDominatorTree Class - Concrete subclass of DominatorTree that is used to
/// compute the a post-dominator tree.
///
@ -69,18 +85,19 @@ struct PostDominatorTree : public DominatorTreeBase {
virtual bool runOnFunction(Function &F) {
reset(); // Reset from the last time we were run...
PostDominatorSet &DS = getAnalysis<PostDominatorSet>();
Roots = DS.getRoots();
calculate(DS);
ImmediatePostDominators &IPD = getAnalysis<ImmediatePostDominators>();
Roots = IPD.getRoots();
calculate(IPD);
return false;
}
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequired<PostDominatorSet>();
AU.addRequired<ImmediatePostDominators>();
}
private:
void calculate(const PostDominatorSet &DS);
void calculate(const ImmediatePostDominators &IPD);
Node *getNodeForBlock(BasicBlock *BB);
};

View File

@ -16,8 +16,131 @@
#include "llvm/Support/CFG.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/SetOperations.h"
#include <iostream>
using namespace llvm;
//===----------------------------------------------------------------------===//
// ImmediatePostDominators Implementation
//===----------------------------------------------------------------------===//
static RegisterAnalysis<ImmediatePostDominators>
D("postidom", "Immediate Post-Dominators Construction", true);
unsigned ImmediatePostDominators::DFSPass(BasicBlock *V, InfoRec &VInfo,
unsigned N) {
VInfo.Semi = ++N;
VInfo.Label = V;
Vertex.push_back(V); // Vertex[n] = V;
//Info[V].Ancestor = 0; // Ancestor[n] = 0
//Child[V] = 0; // Child[v] = 0
VInfo.Size = 1; // Size[v] = 1
// For PostDominators, we want to walk predecessors rather than successors
// as we do in forward Dominators.
for (pred_iterator PI = pred_begin(V), PE = pred_end(V); PI != PE; ++PI) {
InfoRec &SuccVInfo = Info[*PI];
if (SuccVInfo.Semi == 0) {
SuccVInfo.Parent = V;
N = DFSPass(*PI, SuccVInfo, N);
}
}
return N;
}
void ImmediatePostDominators::Compress(BasicBlock *V, InfoRec &VInfo) {
BasicBlock *VAncestor = VInfo.Ancestor;
InfoRec &VAInfo = Info[VAncestor];
if (VAInfo.Ancestor == 0)
return;
Compress(VAncestor, VAInfo);
BasicBlock *VAncestorLabel = VAInfo.Label;
BasicBlock *VLabel = VInfo.Label;
if (Info[VAncestorLabel].Semi < Info[VLabel].Semi)
VInfo.Label = VAncestorLabel;
VInfo.Ancestor = VAInfo.Ancestor;
}
BasicBlock *ImmediatePostDominators::Eval(BasicBlock *V) {
InfoRec &VInfo = Info[V];
// Higher-complexity but faster implementation
if (VInfo.Ancestor == 0)
return V;
Compress(V, VInfo);
return VInfo.Label;
}
void ImmediatePostDominators::Link(BasicBlock *V, BasicBlock *W,
InfoRec &WInfo) {
// Higher-complexity but faster implementation
WInfo.Ancestor = V;
}
bool ImmediatePostDominators::runOnFunction(Function &F) {
IDoms.clear(); // Reset from the last time we were run...
Roots.clear();
// Step #0: Scan the function looking for the root nodes of the post-dominance
// relationships. These blocks, which have no successors, end with return and
// unwind instructions.
for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
if (succ_begin(I) == succ_end(I))
Roots.push_back(I);
Vertex.push_back(0);
// Step #1: Number blocks in depth-first order and initialize variables used
// in later stages of the algorithm.
unsigned N = 0;
for (unsigned i = 0, e = Roots.size(); i != e; ++i)
N = DFSPass(Roots[i], Info[Roots[i]], N);
for (unsigned i = N; i >= 2; --i) {
BasicBlock *W = Vertex[i];
InfoRec &WInfo = Info[W];
// Step #2: Calculate the semidominators of all vertices
for (succ_iterator SI = succ_begin(W), SE = succ_end(W); SI != SE; ++SI)
if (Info.count(*SI)) { // Only if this predecessor is reachable!
unsigned SemiU = Info[Eval(*SI)].Semi;
if (SemiU < WInfo.Semi)
WInfo.Semi = SemiU;
}
Info[Vertex[WInfo.Semi]].Bucket.push_back(W);
BasicBlock *WParent = WInfo.Parent;
Link(WParent, W, WInfo);
// Step #3: Implicitly define the immediate dominator of vertices
std::vector<BasicBlock*> &WParentBucket = Info[WParent].Bucket;
while (!WParentBucket.empty()) {
BasicBlock *V = WParentBucket.back();
WParentBucket.pop_back();
BasicBlock *U = Eval(V);
IDoms[V] = Info[U].Semi < Info[V].Semi ? U : WParent;
}
}
// Step #4: Explicitly define the immediate dominator of each vertex
for (unsigned i = 2; i <= N; ++i) {
BasicBlock *W = Vertex[i];
BasicBlock *&WIDom = IDoms[W];
if (WIDom != Vertex[Info[W].Semi])
WIDom = IDoms[WIDom];
}
// Free temporary memory used to construct idom's
Info.clear();
std::vector<BasicBlock*>().swap(Vertex);
return false;
}
//===----------------------------------------------------------------------===//
// PostDominatorSet Implementation
//===----------------------------------------------------------------------===//
@ -30,119 +153,59 @@ B("postdomset", "Post-Dominator Set Construction", true);
// sets for the function.
//
bool PostDominatorSet::runOnFunction(Function &F) {
Doms.clear(); // Reset from the last time we were run...
// Scan the function looking for the root nodes of the post-dominance
// relationships. These blocks end with return and unwind instructions.
// While we are iterating over the function, we also initialize all of the
// domsets to empty.
Roots.clear();
for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
Doms[I]; // Initialize to empty
for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
if (succ_begin(I) == succ_end(I))
Roots.push_back(I);
}
// If there are no exit nodes for the function, postdomsets are all empty.
// This can happen if the function just contains an infinite loop, for
// example.
ImmediatePostDominators &IPD = getAnalysis<ImmediatePostDominators>();
Doms.clear(); // Reset from the last time we were run...
if (Roots.empty()) return false;
// If we have more than one root, we insert an artificial "null" exit, which
// has "virtual edges" to each of the real exit nodes.
if (Roots.size() > 1)
Doms[0].insert(0);
//if (Roots.size() > 1)
// Doms[0].insert(0);
bool Changed;
do {
Changed = false;
// Root nodes only dominate themselves.
for (unsigned i = 0, e = Roots.size(); i != e; ++i)
Doms[Roots[i]].insert(Roots[i]);
// Loop over all of the blocks in the function, calculating dominator sets for
// each function.
for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
if (BasicBlock *IPDom = IPD[I]) { // Get idom if block is reachable
DomSetType &DS = Doms[I];
assert(DS.empty() && "PostDomset already filled in for this block?");
DS.insert(I); // Blocks always dominate themselves
std::set<BasicBlock*> Visited;
DomSetType WorkingSet;
for (unsigned i = 0, e = Roots.size(); i != e; ++i)
for (idf_ext_iterator<BasicBlock*> It = idf_ext_begin(Roots[i], Visited),
E = idf_ext_end(Roots[i], Visited); It != E; ++It) {
BasicBlock *BB = *It;
succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
if (SI != SE) { // Is there SOME successor?
// Loop until we get to a successor that has had it's dom set filled
// in at least once. We are guaranteed to have this because we are
// traversing the graph in DFO and have handled start nodes specially.
//
while (Doms[*SI].size() == 0) ++SI;
WorkingSet = Doms[*SI];
for (++SI; SI != SE; ++SI) { // Intersect all of the successor sets
DomSetType &SuccSet = Doms[*SI];
if (SuccSet.size())
set_intersect(WorkingSet, SuccSet);
}
// Insert all dominators into the set...
while (IPDom) {
// If we have already computed the dominator sets for our immediate post
// dominator, just use it instead of walking all the way up to the root.
DomSetType &IPDS = Doms[IPDom];
if (!IPDS.empty()) {
DS.insert(IPDS.begin(), IPDS.end());
break;
} else {
// If this node has no successors, it must be one of the root nodes.
// We will already take care of the notion that the node
// post-dominates itself. The only thing we have to add is that if
// there are multiple root nodes, we want to insert a special "null"
// exit node which dominates the roots as well.
if (Roots.size() > 1)
WorkingSet.insert(0);
DS.insert(IPDom);
IPDom = IPD[IPDom];
}
WorkingSet.insert(BB); // A block always dominates itself
DomSetType &BBSet = Doms[BB];
if (BBSet != WorkingSet) {
BBSet.swap(WorkingSet); // Constant time operation!
Changed = true; // The sets changed.
}
WorkingSet.clear(); // Clear out the set for next iteration
}
} while (Changed);
return false;
}
//===----------------------------------------------------------------------===//
// ImmediatePostDominators Implementation
//===----------------------------------------------------------------------===//
static RegisterAnalysis<ImmediatePostDominators>
D("postidom", "Immediate Post-Dominators Construction", true);
// calcIDoms - Calculate the immediate dominator mapping, given a set of
// dominators for every basic block.
void ImmediatePostDominators::calcIDoms(const DominatorSetBase &DS) {
// Loop over all of the nodes that have dominators... figuring out the IDOM
// for each node...
//
for (DominatorSet::const_iterator DI = DS.begin(), DEnd = DS.end();
DI != DEnd; ++DI) {
BasicBlock *BB = DI->first;
const DominatorSet::DomSetType &Dominators = DI->second;
unsigned DomSetSize = Dominators.size();
if (DomSetSize == 1) continue; // Root node... IDom = null
// Loop over all dominators of this node. This corresponds to looping over
// nodes in the dominator chain, looking for a node whose dominator set is
// equal to the current nodes, except that the current node does not exist
// in it. This means that it is one level higher in the dom chain than the
// current node, and it is our idom!
//
DominatorSet::DomSetType::const_iterator I = Dominators.begin();
DominatorSet::DomSetType::const_iterator End = Dominators.end();
for (; I != End; ++I) { // Iterate over dominators...
// All of our dominators should form a chain, where the number of elements
// in the dominator set indicates what level the node is at in the chain.
// We want the node immediately above us, so it will have an identical
// dominator set, except that BB will not dominate it... therefore it's
// dominator set size will be one less than BB's...
//
if (DS.getDominators(*I).size() == DomSetSize - 1) {
IDoms[BB] = *I;
break;
}
} else {
// Ensure that every basic block has at least an empty set of nodes. This
// is important for the case when there is unreachable blocks.
Doms[I];
}
}
return false;
}
//===----------------------------------------------------------------------===//
@ -152,59 +215,45 @@ void ImmediatePostDominators::calcIDoms(const DominatorSetBase &DS) {
static RegisterAnalysis<PostDominatorTree>
F("postdomtree", "Post-Dominator Tree Construction", true);
void PostDominatorTree::calculate(const PostDominatorSet &DS) {
DominatorTreeBase::Node *PostDominatorTree::getNodeForBlock(BasicBlock *BB) {
Node *&BBNode = Nodes[BB];
if (BBNode) return BBNode;
// Haven't calculated this node yet? Get or calculate the node for the
// immediate postdominator.
BasicBlock *IPDom = getAnalysis<ImmediatePostDominators>()[BB];
Node *IPDomNode = getNodeForBlock(IPDom);
// Add a new tree node for this BasicBlock, and link it as a child of
// IDomNode
return BBNode = IPDomNode->addChild(new Node(BB, IPDomNode));
}
void PostDominatorTree::calculate(const ImmediatePostDominators &IPD) {
if (Roots.empty()) return;
// Add a node for the root. This node might be the actual root, if there is
// one exit block, or it may be the virtual exit (denoted by (BasicBlock *)0)
// which postdominates all real exits if there are multiple exit blocks.
BasicBlock *Root = Roots.size() == 1 ? Roots[0] : 0;
Nodes[Root] = RootNode = new Node(Root, 0); // Add a node for the root...
// Iterate over all nodes in depth first order...
for (unsigned i = 0, e = Roots.size(); i != e; ++i)
for (idf_iterator<BasicBlock*> I = idf_begin(Roots[i]),
E = idf_end(Roots[i]); I != E; ++I) {
BasicBlock *BB = *I;
const DominatorSet::DomSetType &Dominators = DS.getDominators(BB);
unsigned DomSetSize = Dominators.size();
if (DomSetSize == 1) continue; // Root node... IDom = null
// If we have already computed the immediate dominator for this node,
// don't revisit. This can happen due to nodes reachable from multiple
// roots, but which the idf_iterator doesn't know about.
if (Nodes.find(BB) != Nodes.end()) continue;
// Loop over all dominators of this node. This corresponds to looping
// over nodes in the dominator chain, looking for a node whose dominator
// set is equal to the current nodes, except that the current node does
// not exist in it. This means that it is one level higher in the dom
// chain than the current node, and it is our idom! We know that we have
// already added a DominatorTree node for our idom, because the idom must
// be a predecessor in the depth first order that we are iterating through
// the function.
//
for (DominatorSet::DomSetType::const_iterator I = Dominators.begin(),
E = Dominators.end(); I != E; ++I) { // Iterate over dominators.
// All of our dominators should form a chain, where the number
// of elements in the dominator set indicates what level the
// node is at in the chain. We want the node immediately
// above us, so it will have an identical dominator set,
// except that BB will not dominate it... therefore it's
// dominator set size will be one less than BB's...
//
if (DS.getDominators(*I).size() == DomSetSize - 1) {
// We know that the immediate dominator should already have a node,
// because we are traversing the CFG in depth first order!
//
Node *IDomNode = Nodes[*I];
assert(IDomNode && "No node for IDOM?");
// Add a new tree node for this BasicBlock, and link it as a child of
// IDomNode
Nodes[BB] = IDomNode->addChild(new Node(BB, IDomNode));
break;
}
Nodes[Root] = RootNode = new Node(Root, 0);
Function *F = Roots[0]->getParent();
// Loop over all of the reachable blocks in the function...
for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
if (BasicBlock *ImmPostDom = IPD.get(I)) { // Reachable block.
Node *&BBNode = Nodes[I];
if (!BBNode) { // Haven't calculated this node yet?
// Get or calculate the node for the immediate dominator
Node *IPDomNode = getNodeForBlock(ImmPostDom);
// Add a new tree node for this BasicBlock, and link it as a child of
// IDomNode
BBNode = IPDomNode->addChild(new Node(I, IPDomNode));
}
}
}
//===----------------------------------------------------------------------===//
// PostETForest Implementation
//===----------------------------------------------------------------------===//