[PATCH, i386]: Add SSE4.2 support - pcmpstr part

[Uros Bizjak]

On 6/5/07, Jan Hubicka <hubicka@ucw.cz> wrote:

> Considering Uros' comment on shanky ground, I think it does not make
> much difference, reload should abadon the SSE alternative if XMM0 is
> occupied and ECX is free.  So the ordering should be based on 1) CPU
> preferrence (i.e latency and resource usage) and 2) instruction encoding
> length. If both are same, than it is all good (and I will update my
> knowledge of Core's instruction timmings).
>
> Considering 'z0', I would still preffer 'Y0' (with possible rename of
> the internal Y2).  I would a lot more preffer have all XMM constraints
> starting by 'x', but we can't do that for backward compatibility. Having
> 'x' for basic SSE register file and 'Y' for various alternatives works
> very well for me.  I don't see use of 'x', 'Y' and 'z' more consistent.

I'm OK with Y0, too. Probably, you should resolve Y2 rename with H.J.,
as he cares about this.

Regarding _cconly - according to published documentation, the length
of _i and _m insn is the same, but timings are nowhere to be found.
IIUC, it is always better to stay in the same register set, so I guess
that preferences should be switched.

Uros.