[PATCH 5/7] x86: re-work insn/suffix recognition

[H.J. Lu]

On Fri, Aug 19, 2022 at 1:28 AM Jan Beulich <jbeulich@suse.com> wrote:
>
> On 18.08.2022 17:14, H.J. Lu wrote:
> > On Wed, Aug 17, 2022 at 11:24 PM Jan Beulich <jbeulich@suse.com> wrote:
> >>
> >> On 17.08.2022 22:29, H.J. Lu wrote:
> >>> On Tue, Aug 16, 2022 at 12:32 AM Jan Beulich <jbeulich@suse.com> wrote:
> >>>>
> >>>> x86: re-work insn/suffix recognition
> >>>>
> >>>> Having templates with a suffix explicitly present has always been
> >>>> quirky. Introduce a 2nd matching pass in case the 1st one couldn't find
> >>>
> >>> I don't like the second pass.   What problem does it solve?
> >>
> >> It addresses the reasons we have various pretty odd (and confusing by
> >> their mere presence) insn templates which better would never have been
> >> there. If you have a better suggestion to eliminate those, I'm all ears.
> >>
> >> You can also easily see the issues this solves by looking at the
> >> testsuite changes. Among other things this once again is a matter of
> >> providing consistent and hence predictable behavior.
> >
> > Did you mean the error reporting behavior?  I don't think we should add
> > a second pass just for it.
>
> No. Certain insns simply were not accepted previously (this is actually
> what finally made me think of a solution here; prior observations
> weren't severe enough to try to get past your possible opposition which
> was to be expected based on past discussions). And certain other ones
> were wrongly accepted.

Please open bug reports for these cases.

> >> Further this sets the stage for the subsequent two changes, which I
> >> don't think are easily possible without this 2nd pass.
> >
> > Does it indicate that the second pass is used quite often?
>
> No, what I did say ...
>
> >> And finally you've likely spotted that this is actually a reduction in
> >> code size, first and foremost because the odd maybe_adjust_templates()
> >> can now go away. Plus I think you realize that the 2nd pass wouldn't
> >> be engaged in many cases - it requires a template match failure in the
> >> 1st pass, after all, which isn't going to happen very often.

There is a fixed cost to prepare for the second pass.

> ... here will continue to be the case with those later changes.
>
> Jan



-- 
H.J.