SPECfp95: Huge performance loss with -fnew-ra?

[Michael Matz]

Hi,

On Mon, 22 Jul 2002, Daniel Berlin wrote:

> > I am not sure at the moment if it's the new register allocator by
> > itself or some combination of -fnew-ra and some other change.
>
> No, it's the new-ra, i'm sure.

Me too ;)

> In particular, Michael, the GENERAL_REGS change you mad (instead
> of the old ALL_REGS) guarantees we'll always spill float_regs,

Yes, unfortunately.  I think for the time being I'll just introduce a new
class of regs of potential spill temporaries.  These often would be
ALL_REGS, but e.g. on x86 would exclude MMX_REGS.

The real fix (webclass) is in the works, that's why I ignored the possible
decrease in performance for now.  But I'm not comfortable with the current
SPECfp numbers, so ... ;)


Ciao,
Michael.