FlexiBLAS: Workaround miscompilation on PPC #15867
Conversation
Flamefire
force-pushed
the
20220714172540_new_pr_FlexiBLAS304
branch
from
3bb6721 to
3323ad1
Compare
|
Test report by @Flamefire |
|
Test report by @Flamefire |
| @@ -11,6 +11,13 @@ toolchain = {'name': 'GCC', 'version': '10.3.0'} | |||
| local_extra_flags = "-fstack-protector-strong -fstack-clash-protection" | |||
| toolchainopts = {'pic': True, 'extra_cflags': local_extra_flags, 'extra_fflags': local_extra_flags} | |||
|
|
|||
| if ARCH == 'ppc64le': | |||
There was a problem hiding this comment.
I don't want to stall this, but in general I think it is better if we avoid logic in easyconfigs if possible (since they are not supposed to contain any). Instead you could create a local_configopts dict and just add local_configopts[ARCH] to the component.
I say this because this easyconfig will now fall over with some of the try-* options. (Well, not fall over but it will evaluate that logic and it will be lost)
There was a problem hiding this comment.
BTW ARCH is not defined is it? Shouldn't it be %(arch)s?
Show's how much I know! List of current known archs is at https://github.com/easybuilders/easybuild-framework/blob/develop/easybuild/tools/systemtools.py#L93
|
The fix needs to go into OpenBLAS anyway. Closing this. |
(created using
eb --new-pr)Addresses #12968 and alternative to easybuilders/easybuild-framework#3809 / #13836
I found that although in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100799 I was able to fix the issue with
-OgI wasn't able to do reproduce that now. I.e.SciPy-bundle-2021.05-foss-2021a.ebwould always segfault or produce wrong results whenever anything but-O0was used for the C-compilation. I played around with various combinations of the C_FLAGS & Fortran_FLAGS but nothing worked but setting C_FLAGS to-O0You also have to be careful as
C_FLAGS_RELEASE(orC_FLAGS_DEBUGetc.) is appended to the value ofC_FLAGS