[Bug bootstrap/42157] [4.5 regression] ICE building stage 1 libgcc on IRIX 5.3: SEGV in compare_access_positions

[ro at CeBiTec dot Uni-Bielefeld dot DE]

------- Comment #2 from ro at CeBiTec dot Uni-Bielefeld dot DE  2009-11-24 18:42 -------
Subject: Re:  [4.5 regression] ICE building stage 1 libgcc on IRIX 5.3: SEGV in
compare_access_positions

> ------- Comment #1 from jamborm at gcc dot gnu dot org  2009-11-24 18:16 -------
> Can you please add a check before the qsort call (tree-sra.c, line
> 1407) whether all pointers in the access_vec seem OK or whether it is
> SRA that passes the invalid pointer to qsort?  Nothing fancy,
> something as stupid as the loop below should do:
>
> for (i =0; i < access_count; i++)
>   gcc_assert (((unsigned) VEC_index (access_p, access_vec, i)) > 0x1000);
>
> If it passes I can't see any other explanation but a bad qsort.  If it
> doesn't we'll need to figure out how the bad pointers get there.

It seems that the qsort is broken (although I haven't found any report
about qsort on IRIX 5.3 yet).  Anyway, I've added a slightly modified
qsort from glibc to libiberty (which won't work due to license
differences, I suppose, but maybe newlib is an option), which allowed
the bootstrap to continue.

I'll later check (either with the code above or with gdb watchpoints)
what is causing the corruption.

     Rainer


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42157