Re: [tree-ssa] LRS enabled.

["Vladimir N. Makarov" <vmakarov at redhat dot com>]

Andrew MacLeod wrote:

>  I've had this patch kicking around for a while, it makes a noticable
> difference in spec on my x86 box (~2% on my 1.8 Ghz 512MB P4). I've been
> getting spec marks pretty close to mainline's.
>  
Congratulation!

> I would be interested in anyone's spec results on other architectures.
> We should see a general improvement, the only downside I am aware of is
> that we sometimes use a little more stack space due to the additional
> variables cause by the LRS. Intense recursive functions will use more
> memory, and this can sometimes trigger more swapping.
>
>  
I've sent a patch to solve this problem in reload several days ago 
(coalescing stack slots allocated for pseudo-regsiters not get a hard 
register and spilled hard registers).   I found that it improves SPEC95 
go (which as I know uses recursion heavily)  even for architectures with 
many registers like ppc.

As for debugging problems, it has been solved too by a patch tracking 
variables and  the new version of gdb.  Paolo Bonzini is trying to get 
an approvement to switch renaming on mainline (-fweb -frename-registers) 
by default.  It probably will require a decision what code to use when 
ssa branch is merged.

Vlad