Jakub Jelinek <ja...@redhat.com> writes:
> On Fri, Mar 10, 2023 at 11:50:39AM +0000, Richard Sandiford wrote:
>> > Will test it momentarily (including the patch it depends on):
>
> Note, testing still pending, I'm testing in a Fedora scratch build
> and that is quite slow (lto bootstrap and the like).
>
>> A naive question:
>>
>> > --- libgcc/config/aarch64/t-softfp.jj 2022-11-14 13:35:34.527155682
>> > +0100
>> > +++ libgcc/config/aarch64/t-softfp 2023-03-10 12:19:58.668882041 +0100
>> > @@ -1,9 +1,10 @@
>> > softfp_float_modes := tf
>> > softfp_int_modes := si di ti
>> > -softfp_extensions := sftf dftf hftf
>> > -softfp_truncations := tfsf tfdf tfhf
>> > +softfp_extensions := sftf dftf hftf bfsf
>> > +softfp_truncations := tfsf tfdf tfhf tfbf dfbf sfbf hfbf
>>
>> Is bfsf used for conversions in which sf is the ultimate target,
>> as opposed to operations that convert bf to sf and then do something
>> with the sf? And so the libfunc is needed to raise exceptions, which in
>> more complex operations can be left to the following sf operation?
>>
>> Do we still optimise to a shift for -ffinite-math-only?
>
> Reminds me I should have added testcase coverage for PR107703, will post
> it momentarily.
>
> But, consider say:
> template <typename T, typename F>
> [[gnu::noipa]] T cvt (F f)
> {
> return T (F (f));
> }
>
> void
> foo ()
> {
> cvt <_Float32, __bf16> (0.0bf16);
> cvt <_Float64, __bf16> (0.0bf16);
> cvt <_Float128, __bf16> (0.0bf16);
> cvt <signed char, __bf16> (0.0bf16);
> cvt <signed short, __bf16> (0.0bf16);
> cvt <int, __bf16> (0.0bf16);
> cvt <long long, __bf16> (0.0bf16);
> cvt <__int128, __bf16> (0.0bf16);
> }
>
> This emits on x86_64 -O2:
> /usr/src/gcc/obj/gcc/cc1plus -quiet -O2 1111.C; grep call.*__ 1111.s
> call __extendbfsf2
> call __extendbfsf2
> call __extendbfsf2
> call __extendsftf2
> call __fixsfti
> where the first call is in cvt <_Float32, __bf16> is really needed,
> admittedly the second 2 calls could be replaced by shifts but aren't right
> now (we expand BF -> DF as BF -> SF -> DF and because sNaN would be already
> diagnosed on the SF -> DF conversion if BF -> SF is done with shift, I think
> it would be ok; similarly for BF -> TF). All the others (BF -> ?I) are
> expanded as BF -> SF using shift and then SF -> ?I. With -O2 -ffast-math
> /usr/src/gcc/obj/gcc/cc1plus -quiet -O2 -ffast-math 1111.C; grep call.*__
> 1111.s
> call __extendsftf2
> call __fixsfti
> so all the BF -> SF conversions are then done using shifts.
> And aarch64 is exactly the same:
> ./cc1plus -quiet -nostdinc -O2 1111.C; grep bl.*__[ef] 1111.s
> bl __extendbfsf2
> bl __extendbfsf2
> bl __extendbfsf2
> bl __extendsftf2
> bl __fixsfti
> ./cc1plus -quiet -nostdinc -O2 -ffast-math 1111.C; grep bl.*__[ef] 1111.s
> bl __extendsftf2
> bl __fixsfti
Thanks, sounds good. In some ways it's ironic that, in a bf->df
conversion, it's the bf->sf that needs a call, and the sf->df can
be done inline, given that one of the purposes of bf16 was to provide
cheap conversions to float. And similarly that bf->sf is more expensive
than sf->df. But that's not the patch's fault.
Rather than have an out-of-line call, would it be possible to synthesise
the checking inline by making bf->sf do a following sf->df conversion,
even when the df result is not used? It would obviously need to be kept
alive somehow (not sure how).
Richard
