Re: Limits of stage3 changes

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

On Nov 16, 2007 11:43 PM, Steven Bosscher <stevenb.gcc@gmail.com> wrote:
> Hello,
>
> The amount of duplicate work done on RTL is sometimes really amazing,
> especially since the merge of the dataflow branch.  Some of the people
> who have worked on the dataflow branch had hoped that other developers
> would help with the follow-up actions to actually *use* all the
> information that, for example, the df-scan problem collects.
>
> My favorite example of this lack of follow-through is gcse.c.  It
> computes reg-def chains and monotonic insn IDs. Guess what df-scan
> provides?
>
> Attached is an example of the cleanups I would have liked to see
> during stage 2 of GCC 4.3.  The patch is not finished, but it gives
> you an idea of what is possible.  The patch avoids computation of
> redundant information by making gcse use the df-scan information,
> instead of computing everything for itself.  In theory, this saves
> memory and it reduces the number of passes over all RTX'en in memory.
> So, if finished, this should give compile time improvements and a
> slightly smaller memory footprint.
>
> Question is, whether this kind of rather large changes is acceptable
> for stage 3 or not.  Me, I call it a "regression fix" if it reduces
> compile time.  But I will not work on it now (or ask help from others)
> if it is a priori not acceptable for stage 3.
>
> Thus, thoughts please! :-)

I think we should be able to rely on DF enough to make mostly mechanical
changes like this ok for stage3.  I don't know where we are in terms of
compile-time compared to 4.2 or 4.1, but certainly if you go back further
compile-time is always a "regression".

Thanks,
Richard.