[Bug target/73350] AVX512: GCC optimizes away rounding flags

[glisse at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=73350

--- Comment #7 from Marc Glisse <glisse at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #5)
> The "simplest" way would be to have separate UNSPECs for all ops (but that
> explodes the number of unspecs I guess).

Would that be so bad? As far as gcc understands addition, this is not an
addition, representing it with 'plus' is a lie.

> So another way would be sth like
> 
> (insn (set (reg:V16SF 110)
>                 (vec_merge:V16SF (plus:V16SF (reg:V16SF 111)
>                         (reg:V16SF 112))
>                     (reg:V16SF 113)
>                     (reg:HI 114))))
> 
> (insn (set (reg:V16SF 110) (unspec [ (reg:V16SF 110) (const_int 9) ]
> UNSPEC_EMBEDDED_ROUNDING) ])
> 
> or merging this "rounding operation" into the operation (around the
> plus, one of its operands or around the merge).

I guess the unspec alone wouldn't match anything, which prevents constant
propagation, but I would still fear combine merging a regular addition with a
rounded one (when they simplify), producing a single rounded operation. If you
want to wrap plus with unspecs, you may have to put them on both operands and
the result to be safe...