On Wed, Nov 23, 2022 at 08:28:20PM +0800, liuhongt via Gcc-patches wrote:
> After supporting real __bf16 type, implementation of mm_cvtsbh_ss went wrong.
> The patch supports extendbfsf2/truncsfbf2 with pslld/psrld,
> and then refined the intrinsic with implicit conversion.
This is not correct.
While using such code for _mm_cvtsbh_ss is fine if it is documented not to
raise exceptions and turn a sNaN into a qNaN, it is not fine for HONOR_NANS
(i.e. when -ffast-math is not on), because a __bf16 -> float conversion
on sNaN should raise invalid exception and turn it into a qNaN.
We could have extendbfsf2 expander that would FAIL; if HONOR_NANS and
emit extendbfsf2_1 otherwise.
And the truncsfbf2 case isn't correct IMHO even for -ffast-math.
float -> __bf16 conversion should be properly rounding depending on the
current rounding mode, while {,v}psrld will always round toward zero.
Jakub
