Re: 4 GCC regressions, 2 new, with your patch on 2002-07-18T09:56:17Z.

[Jan Hubicka <jh at suse dot cz>]

> On Fri, Aug 09, 2002 at 11:11:37AM +0200, Jan Hubicka wrote:
> > Really? I thought that the SSA on lowlevel RTL is no-no becuase of
> > subregs and strict low parts etc.  We need some CSE-like pass to cleanup
> > after last code lowering stage in any stage and it should be global.
> 
> I think the argument goes that the bulk of the user code will
> already have had global optimizations applied and that, hopefully,
> there not have been hordes of code generated from the tree->rtl
> lowering process that could benefit from global opts, only local.
> 
> It's something we'll have to experiment with, for sure.

Yes, we need to dicuss it in more detail too.
If I think about it, the midrtl->lowrtl lowering pass seems to be
perfrect source of redundancies both local and global for instance from
constant expansion on RISCy chips. We want CSE on these as well as loop
invariant motion, so PRE looks like one of the "a must" passes for the
final version of RTL.

It can be interesting to prepare list of passes we would like to
implement in some horisont and at what level we do want them.

I would like GCSE to be broken up into separate passes.  I like how LCM
is implemented separately and I believe number of other analysis and
functions can be reused by other passes as well (midlevel or lowlevel,
SSA or not), so thats the purpose what Zdenek and I are trying to break
GCSE up and get the dead bits working.

Honza
> 
> 
> r~