Package: libvorbisidec-dev On armel, when i build ivorbisfile_example.c against the stock libvorbisidec package, it triggers a series of CPU alignment faults, and consequently produces bad data. Interestingly, this seems to be because gcc is happy to align an array of chars on an odd address under armel (on i386, gcc appears to always place char arrays on 4-byte boundaries). since ov_read asks for a char* buffer, but then casts it to shorts, if the char* buffer is not short-aligned, the CPU sees an alignment fault. What the CPU does in the case of an alignment fault is configurable by echoing the decimal representation of a bitmask into /proc/cpu/alignment: 1 -> warn 2 -> fixup (software workaround to correct the alignment error) 4 -> signal (send a SIGBUS to the child process) What to do on a user fault defaults to 0 (none of the above), but if you set it to 4 or 5, the process will receive a SIGBUS. with the alignment user fault set to 0, the ivorbisfile_example.c simply produces bad data. I reported a few things that should probably be fixed to upstream, and have a new version of the debian package ready which fixes the example to work properly; i'll upload that version shortly. further reference: I originally documented the problem upstream in april: http://lists.xiph.org/pipermail/tremor/2009-April/001564.html and followed up again earlier this month (the proposed fixes follow this thread): http://lists.xiph.org/pipermail/tremor/2009-September/001587.html Martin Guy pointed me in the right direction on http://bugs.debian.org/547525 --dkg
Attachment:
signature.asc
Description: OpenPGP digital signature