Richard Biener <richard.guent...@gmail.com> writes:
> On Thu, May 21, 2020 at 10:17 PM Segher Boessenkool
> <seg...@kernel.crashing.org> wrote:
>>
>> Hi!
>>
>> On Thu, May 21, 2020 at 03:29:49PM +0200, Martin Liška wrote:
>> > Adding Segher to CC, he can help us.
>>
>> Oh dear. Are you sure?
>>
>> > On 5/21/20 2:51 PM, Martin Liška wrote:
>> > >Back to this I noticed that ppc64le target build is broken due to:
>>
>> > >insn-emit.o -MMD -MP -MF ./.deps/insn-emit.TPo insn-emit.c
>> > >/home/marxin/Programming/gcc/gcc/config/rs6000/vector.md:357:11: error:
>> > >vcondv4sfv4sf cannot FAIL
>> > > 357 | FAIL;
>>
>> Is it new that vcond cannot FAIL? Because we have done that for years.
>>
>> Since this breaks bootstrap on a primary target, please revert the patch
>> until it is sorted.
>>
>> > >which is caused by the 4 added optabs:
>> > >
>> > >+DEF_INTERNAL_OPTAB_FN (VCOND, 0, vcond, vec_cond)
>> > >+DEF_INTERNAL_OPTAB_FN (VCONDU, 0, vcondu, vec_condu)
>> > >+DEF_INTERNAL_OPTAB_FN (VCONDEQ, 0, vcondeq, vec_condeq)
>> > >+DEF_INTERNAL_OPTAB_FN (VCOND_MASK, 0, vcond_mask, vec_cond_mask)
>>
>> > >looking at the generator:
>>
>> > >I get there due to:
>> > >
>> > >B- │516 if (find_optab (&p, XSTR (expand,
>> > >0)))│
>> > > │517
>> > > { │
>> > > │518 gcc_assert (p.op <
>> > > NUM_OPTABS); │
>> > > │519 if
>> > > (nofail_optabs[p.op]) │
>> > > │520 can_fail_p =
>> > > false; │
>> > > │521
>> > > } │
>
> OK, so this is an "artifact" of direct internal functions. We do check that
> expansion does not actually FAIL before emitting calls to those IFNs.
>
> I guess this simply makes direct internal functions not a 100% match for
> our use and the way out is to add regular internal functions mapping to
> the optabs. That is, I guess, for direct-internal functions it should be
> enough to check direct_internal_function_supported_p which it is not
> for the case of vcond*.
>
> Richard, do you agree?
Sorry for the late reply, been off for a few days.
I guess that would be OK for VCOND(U) as an intermediate step,
but long term, I think we should try to make all VCOND* directly-mapped.
If we're doing instruction selection on gimple (a good thing IMO)
we need to know before expand whether an operation is supported.
So longer-term, I think we should replace VCOND(U) with individual ifns,
like for VCONDEQ. We could reduce the number of optabs needed by
canonicalising greater-based tests to lesser-based tests.
Thanks,
Richard
