Switch the internal "Info" map from an std::map to a DenseMap. This
speeds up idom by about 45% and postidom by about 33%.
Some extra precautions must be taken not to invalidate densemap iterators.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40827 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Analysis/PostDominators.cpp b/lib/Analysis/PostDominators.cpp
index e3169a5..244f8cd 100644
--- a/lib/Analysis/PostDominators.cpp
+++ b/lib/Analysis/PostDominators.cpp
@@ -27,28 +27,25 @@
static RegisterPass<PostDominatorTree>
F("postdomtree", "Post-Dominator Tree Construction", true);
-unsigned PostDominatorTree::DFSPass(BasicBlock *V, InfoRec &VInfo,
- unsigned N) {
- std::vector<std::pair<BasicBlock *, InfoRec *> > workStack;
+unsigned PostDominatorTree::DFSPass(BasicBlock *V, unsigned N) {
+ std::vector<BasicBlock *> workStack;
std::set<BasicBlock *> visited;
- workStack.push_back(std::make_pair(V, &VInfo));
+ workStack.push_back(V);
do {
- BasicBlock *currentBB = workStack.back().first;
- InfoRec *currentVInfo = workStack.back().second;
+ BasicBlock *currentBB = workStack.back();
+ InfoRec &CurVInfo = Info[currentBB];
// Visit each block only once.
- if (visited.count(currentBB) == 0) {
-
- visited.insert(currentBB);
- currentVInfo->Semi = ++N;
- currentVInfo->Label = currentBB;
+ if (visited.insert(currentBB).second) {
+ CurVInfo.Semi = ++N;
+ CurVInfo.Label = currentBB;
Vertex.push_back(currentBB); // Vertex[n] = current;
// Info[currentBB].Ancestor = 0;
// Ancestor[n] = 0
// Child[currentBB] = 0;
- currentVInfo->Size = 1; // Size[currentBB] = 1
+ CurVInfo.Size = 1; // Size[currentBB] = 1
}
// Visit children
@@ -58,8 +55,8 @@
InfoRec &SuccVInfo = Info[*PI];
if (SuccVInfo.Semi == 0) {
SuccVInfo.Parent = currentBB;
- if (visited.count (*PI) == 0) {
- workStack.push_back(std::make_pair(*PI, &SuccVInfo));
+ if (!visited.count(*PI)) {
+ workStack.push_back(*PI);
visitChild = true;
}
}
@@ -130,7 +127,7 @@
// 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);
+ N = DFSPass(Roots[i], N);
for (unsigned i = N; i >= 2; --i) {
BasicBlock *W = Vertex[i];
diff --git a/lib/VMCore/Dominators.cpp b/lib/VMCore/Dominators.cpp
index 9b36e1d..bb780c5 100644
--- a/lib/VMCore/Dominators.cpp
+++ b/lib/VMCore/Dominators.cpp
@@ -146,12 +146,12 @@
}
}
-unsigned DominatorTree::DFSPass(BasicBlock *V, InfoRec &VInfo,
- unsigned N) {
+unsigned DominatorTree::DFSPass(BasicBlock *V, unsigned N) {
// This is more understandable as a recursive algorithm, but we can't use the
// recursive algorithm due to stack depth issues. Keep it here for
// documentation purposes.
#if 0
+ InfoRec &VInfo = Info[Roots[i]];
VInfo.Semi = ++N;
VInfo.Label = V;
@@ -164,7 +164,7 @@
InfoRec &SuccVInfo = Info[*SI];
if (SuccVInfo.Semi == 0) {
SuccVInfo.Parent = V;
- N = DFSPass(*SI, SuccVInfo, N);
+ N = DFSPass(*SI, N);
}
}
#else
@@ -313,7 +313,7 @@
#endif
}
-void DominatorTree::calculate(Function& F) {
+void DominatorTree::calculate(Function &F) {
BasicBlock* Root = Roots[0];
// Add a node for the root...
@@ -323,9 +323,7 @@
// 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]], 0);
+ unsigned N = DFSPass(Root, 0);
for (unsigned i = N; i >= 2; --i) {
BasicBlock *W = Vertex[i];