Re: RFA: GCC 4.2.1: Stabalizing coalesce_list's qsort

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 8/20/07, Nick Clifton <nickc@redhat.com> wrote:
> Hi Guys,
>
>   I am not sure if this patch counts as fixing a regression, but I am
>   going to submit it anyway.
>
>   The problem is that gcc appears to rely upon the
>   sort_coalesce_list() function in tree-ssa-live.c performing a stable
>   sort of the coalesce list.  (ie one where items with the same key
>   value retain the same relative ordering after the sort).
>   Unfortunately the C library function qsort() does not guarantee
>   this, and while it just so happens that under Linux the qsort is
>   stable, under other operating systems and execution environments
>   this is not the case.  (The problem was actually detected when
>   running gcc under Cygwin).
>
>   The proposed patch fixes the problem by extending the comparison
>   function passed to the invocation of qsort so that when two
>   partition pairs do have the same cost they are compared based on
>   their location in the partition arrays.  This has the effect of
>   stabalizing the sort, allowing the Cygwin hosted toolchain to behave
>   in exactly the same way as the Linux hosted toolchain, and so
>   preventing Cygwin-only code generation bugs.
>
>   Is this patch suitable for application to the 4.2 branch ?

No, but for trunk if you bootstrapped/tested this.  That is, this patch
is ok for mainline if you did.

Thanks,
Richard.

> Cheers
>   Nick
>
> gcc/ChangeLog
> 2007-08-20  Nick Clifton  <nickc@redhat.com>
>
>         * tree-ssa-live.c (compare_pairs): Stabalize the comparison by
>         checking the partition indicies if the costs are the same.