[llvm-commits] [llvm] r122684 - /llvm/trunk/include/llvm/Analysis/DominatorInternals.h
Cameron Zwarich
zwarich at apple.com
Sat Jan 1 23:53:49 PST 2011
Author: zwarich
Date: Sun Jan 2 01:53:49 2011
New Revision: 122684
URL: http://llvm.org/viewvc/llvm-project?rev=122684&view=rev
Log:
Remove the #ifdef'd code for balancing the eval-link data structure. It doesn't
compile, and everyone's tests have shown it to be slower in practice, even for
quite large graphs.
I also hope to do an optimization that is only correct with the simpler data
structure, which would break this even further.
Modified:
llvm/trunk/include/llvm/Analysis/DominatorInternals.h
Modified: llvm/trunk/include/llvm/Analysis/DominatorInternals.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Analysis/DominatorInternals.h?rev=122684&r1=122683&r2=122684&view=diff
==============================================================================
--- llvm/trunk/include/llvm/Analysis/DominatorInternals.h (original)
+++ llvm/trunk/include/llvm/Analysis/DominatorInternals.h Sun Jan 2 01:53:49 2011
@@ -22,13 +22,9 @@
// A Fast Algorithm for Finding Dominators in a Flowgraph
// T. Lengauer & R. Tarjan, ACM TOPLAS July 1979, pgs 121-141.
//
-// This implements both the O(n*ack(n)) and the O(n*log(n)) versions of EVAL and
-// LINK, but it turns out that the theoretically slower O(n*log(n))
-// implementation is actually faster than the "efficient" algorithm (even for
-// large CFGs) because the constant overheads are substantially smaller. The
-// lower-complexity version can be enabled with the following #define:
-//
-#define BALANCE_IDOM_TREE 0
+// This implements the O(n*log(n)) versions of EVAL and LINK, because it turns
+// out that the theoretically slower O(n*log(n)) implementation is actually
+// faster than the almost-linear O(n*alpha(n)) version, even for large CFGs.
//
//===----------------------------------------------------------------------===//
@@ -157,75 +153,17 @@
typename GraphT::NodeType *V) {
typename DominatorTreeBase<typename GraphT::NodeType>::InfoRec &VInfo =
DT.Info[V];
-#if !BALANCE_IDOM_TREE
- // Higher-complexity but faster implementation
if (VInfo.Ancestor == 0)
return V;
Compress<GraphT>(DT, V);
return VInfo.Label;
-#else
- // Lower-complexity but slower implementation
- if (VInfo.Ancestor == 0)
- return VInfo.Label;
- Compress<GraphT>(DT, V);
- GraphT::NodeType* VLabel = VInfo.Label;
-
- GraphT::NodeType* VAncestorLabel = DT.Info[VInfo.Ancestor].Label;
- if (DT.Info[VAncestorLabel].Semi >= DT.Info[VLabel].Semi)
- return VLabel;
- else
- return VAncestorLabel;
-#endif
}
template<class GraphT>
void Link(DominatorTreeBase<typename GraphT::NodeType>& DT,
unsigned DFSNumV, typename GraphT::NodeType* W,
typename DominatorTreeBase<typename GraphT::NodeType>::InfoRec &WInfo) {
-#if !BALANCE_IDOM_TREE
- // Higher-complexity but faster implementation
WInfo.Ancestor = DFSNumV;
-#else
- // Lower-complexity but slower implementation
- GraphT::NodeType* WLabel = WInfo.Label;
- unsigned WLabelSemi = DT.Info[WLabel].Semi;
- GraphT::NodeType* S = W;
- InfoRec *SInfo = &DT.Info[S];
-
- GraphT::NodeType* SChild = SInfo->Child;
- InfoRec *SChildInfo = &DT.Info[SChild];
-
- while (WLabelSemi < DT.Info[SChildInfo->Label].Semi) {
- GraphT::NodeType* SChildChild = SChildInfo->Child;
- if (SInfo->Size+DT.Info[SChildChild].Size >= 2*SChildInfo->Size) {
- SChildInfo->Ancestor = S;
- SInfo->Child = SChild = SChildChild;
- SChildInfo = &DT.Info[SChild];
- } else {
- SChildInfo->Size = SInfo->Size;
- S = SInfo->Ancestor = SChild;
- SInfo = SChildInfo;
- SChild = SChildChild;
- SChildInfo = &DT.Info[SChild];
- }
- }
-
- DominatorTreeBase::InfoRec &VInfo = DT.Info[V];
- SInfo->Label = WLabel;
-
- assert(V != W && "The optimization here will not work in this case!");
- unsigned WSize = WInfo.Size;
- unsigned VSize = (VInfo.Size += WSize);
-
- if (VSize < 2*WSize)
- std::swap(S, VInfo.Child);
-
- while (S) {
- SInfo = &DT.Info[S];
- SInfo->Ancestor = V;
- S = SInfo->Child;
- }
-#endif
}
template<class FuncT, class NodeT>
More information about the llvm-commits
mailing list