Opened 10 years ago
Closed 10 years ago
#687 closed defect (fixed)
Segmentation fault in latest 4.8 and 4.9 gfortran (?)
| Reported by: | Juergen Reuter | Owned by: | kilian |
|---|---|---|---|
| Priority: | P0 | Milestone: | v2.2.3 |
| Component: | core | Version: | 2.2.2 |
| Severity: | blocker | Keywords: | |
| Cc: |
Description
Seems to happen only for 4.8-LATEST and 4.9-LATEST, and for extended precision only.
Change History (9)
comment:1 follow-up: 2 Changed 10 years ago by
comment:2 follow-up: 3 Changed 10 years ago by
Replying to jr_reuter:
Seems like this is an artifact of the OpenMP ... Question is whether we want to further investigate this or not. It might be a problem in the OpenMP v4.0 implementation in gcc 4.8/4.9.
Depends on whether this occurs in less special circumstances. Extended precision may be required by some applications, but definitely not in context with NLO. Gosam doesn't support extended precision, neither does it explicitly support OpenMP. (Nevertheless, CS has applied the correct type conversion to Gosam calls, as far as I can see.)
comment:3 Changed 10 years ago by
Replying to kilian:
(Nevertheless, CS has applied the correct type conversion to Gosam calls, as far as I can see.)
CW, not CS, this time :-)
comment:4 Changed 10 years ago by
We definitely have to solve this problem, as I fear that 4.8.4 and 4.9.2 might be coming up very soon.
comment:5 Changed 10 years ago by
Damn, I don't get the seg fault with 4.8.4 prerelease on MAC OS X with --enable-fc-openmp --enable-fc-extended ....
comment:6 Changed 10 years ago by
Used the setup of the trunk-dist-weekly (which has a strange script BTW) to reproduce this thingy. It seems to only appear with flags -g -O2 and when using option -r. WTF?
comment:7 Changed 10 years ago by
Can again not reproduce on MAC OS X (where I do have a bit more modern svn version of gcc 4.8-LATEST, but don't know of any changes)
comment:8 Changed 10 years ago by
One thing that is strange: for CW's test there is only one single VAMP grid file: smtest_16_p1_m2.vg. While for multi_comp tests there is one grid file per subprocess. Could this be a source of the problem? (connected with the fact that -r does the problem, reusing stuff does not... (and generating the grids the first place without -r option).
comment:9 Changed 10 years ago by
| Resolution: | → fixed |
|---|---|
| Status: | new → closed |
Quoting CW: was caused by uninitialized variable, fixed in 6215. Checks work now, closing.
Seems like this is an artifact of the OpenMP ... Question is whether we want to further investigate this or not. It might be a problem in the OpenMP v4.0 implementation in gcc 4.8/4.9.