Bug#325131: gcc-4.0: regression on hppa and arm, possibly wrong code

[Max Vozeler <xam@debian.org>]

Package: gcc-4.0
Version: 4.0.1-6

Hi gcc maintainers,

I'm seeing an interesting failure of the aespipe testsuite on hppa 
when built with gcc-4.0. I can't say for sure that gcc-4.0 is at fault,
but it does look like it and a hppa porter suggested it likely is.

The problem shows up in aespipe's C implementation of SHA256/384/512
(sha512.c). SHA256 works as expected, but SHA384 and -512 give wrong
results when built with gcc-4.0 and -O2, but only on hppa and arm.

This is what I've found out so far:
 - No problem with gcc-3.3 or gcc-3.4
 - No problem with -O0 or -O1
 - No runtime problem; old binary passes the testsuite
 - No difference in preprocessed sha512.c between gcc-3.4 and gcc-4.0

I can reproduce the problem in paer's sid chroot (hppa) and have put
gcc-3.4 and gcc-4.0 builds with -save-temps on [1]. If there is other
information that could be useful to you, please let me know.
<http://people.debian.org/~xam/aespipe/hppa/>

cheers,
Max