[cfe-commits] r142885 - in /cfe/trunk: include/clang/Analysis/Analyses/Dominators.h lib/Analysis/CMakeLists.txt lib/Analysis/Dominators.cpp lib/StaticAnalyzer/Checkers/Checkers.td lib/StaticAnalyzer/Checkers/DebugCheckers.cpp test/Analysis/domtes

Cameron Zwarich zwarich at apple.com
Mon Oct 24 21:32:15 PDT 2011


The performance results of that paper are dubious. There is a followup paper by Georgiadis, Tarjan, and Werneck, entitled "Finding Dominators in Practice", which presents results to the contrary and quotes private communication from Cooper saying that he retested with a more careful implementation of Lengauer-Tarjan and found the same.

I went and found all of the tricks and optimizations I could find for LLVM's implementation of Lengauer-Tarjan. There are still some things that could probably be improved for a few percentage points of extra performance.

Cameron

On Oct 24, 2011, at 8:49 PM, Guoping Long wrote:

> Thanks for pointing out this. I shall do the refactoring.
> 
> Chris, why do you think this implementation is slower than the version in llvm/Analysis (based on the classical Lengauer-Tarjan algorithm)? My patch was based on this 2001 paper: K. D. Cooper "A Simple, Fast Dominance Algorithm", which actually included an interesting analysis and comparison with the classical Lengauer-Tarjan algorithm. I believe the implementation in core LLVM shall be very stable and efficient. I am just interested in know why choosing this.




More information about the cfe-commits mailing list