http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47530
--- Comment #5 from Luke Dalessandro <luked at cs dot rochester.edu> 2012-05-02 16:50:41 UTC --- Created attachment 27288 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=27288 Preprocessed RBTree microbenchmark from RSTM that triggers this bug. I'm still seeing this bug with gcc-4.8 (svn r187050). I've attached a preprocessed source file from the RSTM microbenchmark suite (for fc13, x86_64) that manifests the problem. It simply takes an integer set, implemented as an RBTree, and randomly performs insert/delete/lookup operations as transactions. The "p" command line parameter sets the number of threads to use. The rest of the command line parameters aren't relevant. Unoptimized compilation produces expected output: luke> g++ -fgnu-tm -o TreeBench TreeBench.i -lrt luke> ./TreeBench -p2 Verification: Passed csv, ALG=gcc-tm, B=RBTree, R=34, d=1, p=2, X=0, m=256, S=1, O=1, txns=1231838, time=1000001329, throughput=1231836 Optimized compilation shows stack corruption ("time" output is computed using a stack location that is trashed when _ITM_commitTransaction longjmps from the tail call). luke> g++ -O3 -fgnu-tm -o TreeBench TreeBench.i -lrt luke> ./TreeBench -p2 Verification: Passed csv, ALG=gcc-tm, B=RBTree, R=34, d=1, p=2, X=0, m=256, S=1, O=1, txns=1301244673, time=1335976459721524108, throughput=0 Tail (sibling?) calls to _ITM_commitTransaction are identified by 'jmp' instructions. luke> g++ -O3 -fgnu-tm -o /dev/stdout -S TreeBench.i | grep _ITM_commitTransaction call _ITM_commitTransaction jmp _ITM_commitTransaction jmp _ITM_commitTransaction I expect that this is also related to http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51771--and in fact may be the same bug, but it's manifesting as the specific tail-call problem that was noted here as fixed. The returns_twice "hack" that suppresses pr51771 in gcc-4.7 also suppresses this bug (pr47530) in gcc-4.7, so it's gcc-4.8-only at this point. I will reopen pr47530, but if someone can confirm it as really being pr51771 I'm happy to have it closed again.
