On 5/14/07, Serge Belyshev <[EMAIL PROTECTED]> wrote:
"Richard Guenther" <[EMAIL PROTECTED]> writes:
> It was a patch to enable more optimization. Reverting it should be as safe
> or unsafe as exchanging forwprop and dce passes. And I have no idea
> as how to quantify safeness of either ;)
>
> I'd say we better analyze what goes wrong (as the problem is possibly
> latent on the mainline) and fix it properly. If we paper over it we will not
> fix it at all.
>
I vote for reverting that change (r111639) on the branch before release.
We would gain nothing by having bug 30252, however latent,
uncovered in a released compiler.
The patch only changes the heuristic where we create SFTs for. The bug is in
handling aliasing with SFTs. So you'll fix this particluar case --
but I have no idea
on what other testcases you re-expose the bug. So I'd rather stay with a few
now known cases than shipping with other unknown ones (it maybe glibc which
we miscompile after reverting the change).
Richard.