[LLVMbugs] [Bug 15000] New: Dead Store Elimination takes a very long time (and has no effect)
bugzilla-daemon at llvm.org
bugzilla-daemon at llvm.org
Fri Jan 18 21:56:30 PST 2013
http://llvm.org/bugs/show_bug.cgi?id=15000
Bug #: 15000
Summary: Dead Store Elimination takes a very long time (and has
no effect)
Product: libraries
Version: 3.2
Platform: PC
OS/Version: All
Status: NEW
Severity: enhancement
Priority: P
Component: Scalar Optimizations
AssignedTo: unassignedbugs at nondot.org
ReportedBy: caustin at gmail.com
CC: llvmbugs at cs.uiuc.edu
Classification: Unclassified
I suspect the Dead Store Elimination pass has algorithmic deficiencies. While
investigating why my Emscripten build was so slow, I ran opt with -time-passes
and witnessed:
third-party/mac/clang+llvm-3.1-x86_64-apple-darwin11/bin/opt -time-passes
-std-compile-opts -disable-inlining -std-link-opts -internalize -globaldce
-mergefunc -constmerge -always-inline -o=build/opt.bc build/total.bc
===-------------------------------------------------------------------------===
... Pass execution timing report ...
===-------------------------------------------------------------------------===
Total Execution Time: 53.2006 seconds (53.2081 wall clock)
---User Time--- --System Time-- --User+System-- ---Wall Time--- ---
Name ---
22.9070 ( 43.5%) 0.0270 ( 5.5%) 22.9340 ( 43.1%) 22.9402 ( 43.1%) Dead
Store Elimination
22.6525 ( 43.0%) 0.0315 ( 6.4%) 22.6840 ( 42.6%) 22.6900 ( 42.6%) Dead
Store Elimination
0.8339 ( 1.6%) 0.0083 ( 1.7%) 0.8422 ( 1.6%) 0.8417 ( 1.6%)
Global Value Numbering
...
Manually specifying all passes in -std-compile-opts and -std-link-opts but
excluding -dse had literally no effect on the resulting opt.bc file.
Why is dead store elimination so slow?
This occurs in both LLVM 3.1 and LLVM 3.2.
I have attached total.bc so you can reproduce the slowdown with the above
command.
--
Configure bugmail: http://llvm.org/bugs/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.
More information about the llvm-bugs
mailing list