Re: rearrange -mcpu= code for rs6000

[Geoffrey Keating <geoffk at apple dot com>]

On Wednesday, September 24, 2003, at 09:53  PM, David Edelsohn wrote:

> > > > > > Geoffrey Keating writes:
>
> Geoff> I'll wait a bit before committing this, in the hope that 
> someone can
> Geoff> help me with the FIXME.  In particular:
>
> Geoff> - Does POWER3 have the GPOPT extensions?
> Geoff> - What about POWER4?  620?  630?
> Geoff> - Does the 8540 have Altivec?
>
> 	Just because a processor supports GPOPT extensions, does not mean
> that it is a performance benefit for the compiler to generate those
> instructions.

Should we just pretend that the processor does not have them?  Is this 
true for all the four processors listed?  Is there ever a circumstance 
under which they should be generated?

> 	Motorola does not mention Altivec support for the 8540 processor
> in their documentation.

OK.  That answers that question.

> 	Also, processors with Altivec extensions are not always used in
> settings with Altivec support enabled.  MASK_ALTIVEC changes the 
> behavior
> of the GCC, especially on targets other than Darwin.  I am concerned 
> about
> enabling MASK_ALTIVEC based solely on a processor choice.

MASK_ALTIVEC is supposed to only control the emission of Altivec 
instructions, not any other behaviour.  Presumably if you're using the 
processor in an environment where Altivec is unavailable (or where the 
FPU is unavailable, or whatever) you'd use -mcpu=xxx -mno-altivec, or 
-mcpu=xxx -msoft-float.