On 10/30/19 2:12 AM, Richard Biener wrote:
> On Tue, Oct 29, 2019 at 8:34 PM Jeff Law <l...@redhat.com> wrote:
>
> I think the wiki has better examples. That said, I wonder how much can
> be automated here, for example when just considering CCmode (m68k has
> setcc IIRC) then for example all define_insns like
>
> (define_insn "*cmpsi_cf"
> [(set (cc0)
> (compare (match_operand:SI 0 "nonimmediate_operand" "mrKs,r")
> (match_operand:SI 1 "general_operand" "r,mrKs")))]
> "TARGET_COLDFIRE"
> {...}
The compare/test insns are relatively straightforward as they're a
fairly simple substitution. It's all the other insns that implicitly
set cc0 that are so painful.
>
> The various expanders need to be handled manually I guess, though somehow
> automagically generating the define_insn_and_split might be possible
> as well, no?
I was looking at define_subst to help here, but it didn't look like it
was really going to simplify things in any significant way. Maybe
someone with more experience with define_subst would see something that
would dramtically simplify this process.
>
> That is, the work is quite repetitive (looking at m68k.md) and for this reason
> error-prone. But a case#2 conversion could essentially be automated?
It's a case2. ANd yes, it's highly repetitive. That's the part of the
v850 conversion my son did. Once he had all the mechanical stuff in
place I started banging it into shape.
>
> That is, do you see it as impossible to implement a case#2 cc0 "conversion"
> in gen*?
I'd be surprised if it could be that simple.
>
>> Some have claimed this might be easier with define_subst, but I wandered
>> it a bit and it didn't seem necessarily simpler in the end. But maybe
>> someone with more experience with define_subst could make the transition
>> simpler.
>>
>> Anyway, that has to be done for each and every pattern that modifies the
>> condition codes. It's a ton of mechanical work. The m68k has the
>> additional complication that condition codes can be either in cc0 or
>> fcc0. So that has to be taken into account as well.
>
> I don't see that it does this currently though (cc0 doesn't allow this).
You're right. fcmp seems to just set the normal condition codes, just
using a different insn. So the FP unit shouldn't be a major problem.
>
>> You also have to rewrite the tst/cmp and conditional branching patterns
>> to use the new style.
>>
>> Sadly until you do *all* of that you don't have anything you can really
>> test. And even once you've done all that, the code quality is going to
>> suffer because we're not doing any compare/tst elimination. To enable
>> that you have to go back and start adding patterns like this:
>
> For m68k what helps is that you could concentrate on TARGET_COLDFIRE
> (or !TARGET_COLDFIRE). Maybe even split m68k.md along this.
>
> A lot of the define_insns also look like optimization (so could be disabled
> for the moment).
That's certainly how I'd approach it. Disable everything not strictly
needed, convert what has to be converted, then bring back the patterns
incrementally.
Jeff
