Improve heuristic for global-vs-local regs

[Michael Matz]

Hi,

On Tue, 24 Jan 2006, Roger Sayle wrote:

> To break the chicken-and-egg deadlock, I'll propose that I'll approve
> all three recommended changes provided they bootstrap and regression
> test well together on several platforms, and the impressive speed-ups
> seen on some platforms are not at the expense of others.

So, I've regtested the thing again on i686,x86-64,ppc,ppc64,ia64 with 
trunk without regressions.  I've also done some SPEC runs on ia64, ppc 
and x86-64 and got only 1% changes mostly (in fact most of the executables 
were just the same or had slightly different allocation, but no different 
spills).  I did get speedups on ppc (with -O2 on eon from 180 to 170 
seconds) and on x86-64 (with -O3 -march=k8 -funroll-all-loops 
-ftree-loop-linear -ffast-math -ftree-vectorize --param 
max-aliased-vops=800 on mesa (124 to 118 seconds) and facerec (208 to 202 
seconds)).

I've checked in the patch to trunk and would like to have it also in 4.1.1 
I.e. I'm not proposing to include it into 4.1 right now, but into the 
branch after 4.1 is released.  (as indicated I also regtested this with 
4.1 as of last week)

Thanks for the review.


Ciao,
Michael.