This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Reload patch to improve 386 code
- To: mark at markmitchell dot com
- Subject: Re: Reload patch to improve 386 code
- From: Jeffrey A Law <law at cygnus dot com>
- Date: Sat, 05 Sep 1998 23:07:19 -0600
- cc: amylaar at cygnus dot co dot uk, crux at pool dot informatik dot rwth-aachen dot de, meissner at cygnus dot com, toon at moene dot indiv dot nluug dot nl, egcs at cygnus dot com
- Reply-To: law at cygnus dot com
In message <199809060428.VAA02085@smtp.earthlink.net>you write:
> I don't see why the base address tracking code should have to do
> anything. My plan was just to use different alias sets for each spill
> register (we have 2^32 of them, after all!). If you like, you could
> reuse these from function to function, so the maximum number of alias
> sets used up this was would be the number of spills in any one
> function. My guess is that if that gets close to 4 billion, we have
> worse problems.
>
> I'd be happy (delighted, even) to implement this, but I'd like a test
> case that someone things will benefit. (I am in *no* way doubting the
> existence of such a thing, but having one would allow me to verify my
> work.)
Just seemed easier to have a single alias set for spills. I'm not
particularly partial to either solution. They should both work.
As for a testcase. Who knows. I would think it would be reasonably
straightforward to build one.
ANy benefit right now would be in sched2 and possibly reload_cse since
we don't have generalized spill code motion yet.
Presumably we'd need a loop and something in the loop needs to be
spilled to the stack and we need unrelated memory accesses throuh a
pointer which we know nothing about.
jeff