On 2015-02-27 16:52, Jacco Ligthart wrote:
On 02/27/2015 02:31 PM, Gordan Bobic wrote:
On 2015-02-26 21:30, Jacco Ligthart wrote:
On 02/26/2015 01:03 PM, Gordan Bobic wrote:
Thanks for the detailed analysis.
Would it be worth rebuilding pixman with 4.4.5 to see if that makes
the problem not manifest? That is not to say it's a fix if it works
- it
could just be masking the issue, but at least it gives a narrower
starting point for bisection.
Just tested, gcc 4.4.5 does not produce a working pixman
Hmm... Off the wall idea, but if you rebuild the old, known
working version of pixman on a fully up to date build box,
does that pixman then work? Or does that end up being broken
as well?
I'll try to do some more investigation on my side this weekend,
once I have my chromebook working again.
OK, new workaround :)
if you add "--disable-arm-simd --disable-arm-neon" to the configure
flags you'll end with a working version.
Do the binaries actually include SIMD/NEON instructions? No way should
gcc be generating those for armv5tel. Is there (new) assembly code for
ARM in the newer version?
Having said that, IIRC invalid instructions would cause a different
error (not a segfault).
If it isn't invalid instructions that are causing the problem, there
is a reasonably good chance that all this does is introduces a small
memory hole somewhere that is masking a buffer overrun. Because the
overrun doesn't step outside the program's malloc()-ed area, it
doesn't segfault. Whether the overrun ends up trampling actual data
or code or falls harmlessly into a memory hole, is a different
question.
Do you happen to have the debug symbols installed? It might be
worth getting a core and stacktracing it. There could be a bona
fide bug that's been there for ages on different platforms which
just hasn't been tripped in previous builds.
But then, I still do not understand why building without these flags on
EL7 results in a working pixman ???
Just guessing, but probably different gcc version producing different
binary code with slightly different memory occupancy details.
Gordan
_______________________________________________
users mailing list
[email protected]
http://lists.redsleeve.org/mailman/listinfo/users
