[PATCH 58/62] AVX512FP16: Optimize for code like (_Float16) __builtin_ceif ((float) f16).

[Richard Biener]

On Thu, Jul 1, 2021 at 11:26 PM Joseph Myers <joseph@codesourcery.com> wrote:
>
> On Thu, 1 Jul 2021, liuhongt via Gcc-patches wrote:
>
> > +/* Optimize for code like (_Float16) __builtin_ceif ((float) f16)
> > +   since it's not handled in frontend.  */
>
> Much the same comments apply as for sqrt.  But in this case, the
> conversion code is in match.pd - right now, specific to pairs of types
> (float, double) and (float, long double).  And it's logically valid for
> any pair of same-radix floating-point types, the values of one of which
> are a subset of the values of the other (a strict subset, for it actually
> to be an interesting optimization).  (So when making it apply to more
> general types, take care that it does *not* apply to the __ibm128 /
> _Float128 pair on powerpc64le, in either order, because neither of those
> types has values a subset of the other.)
>
> (Also, the match.pd code isn't handling roundeven at present, but that
> should be a trivial addition to it.)

Note that for non-"standard" float formats emitting builtins would require
guaranteed libm support (IIRC we don't have rounding functions in
libgcc), so guarding on availability of a machine instruction (optab)
for the operation should be tested instead - see my proposed match
rule for sqrt and using internal-functions and the direct optab machinery.

Richard.

> --
> Joseph S. Myers
> joseph@codesourcery.com