Richard Earnshaw (lists) wrote:
> On 05/05/17 17:10, Wilco Dijkstra wrote:
> > However I think all uses of '*' in md files are incorrect and the
> > feature should
> > be removed. '?' already exists for cases where the alternative may be
> > expensive.
> >
>
> It's not quite as simple as that. It may be, however, that we should
> only use it for restricting subclasses (eg generally avoiding high
> registers on Thumb1).
We needed to add TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS to ensure
register preferencing works correctly, so this can be used to choose low/high
registers on Thumb-1.
> However, things have changed somewhat since the move to LRA and what was
> once true might be quite different now.
It's possible that existing uses of '*' were trying to hack around the
preferencing
issues, but with this callback I don't think they are needed. We will still
require '?'
and '!' to set the relative cost of alternatives - if you have a negate and one
of the
operands is a SIMD register (so you need one int<->FP move either way), and SIMD
negate has a higher latency, you need '?' to say that it is more expensive than
the
integer negate.
Wilco
