[Bug rtl-optimization/84101] [7/8/9 Regression] -O3 and -ftree-vectorize trying too hard for function returning trivial pair-of-uint64_t-structure

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84101

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Segher Boessenkool from comment #6)
> I think subreg should deal with this.  What do you mean "there aren't
> any half of TImode subregs"?  insn 10 has it split at the start, and
> insn 18 at the end wants it split, too.

By that I mean that there aren't any (subreg:DI (reg:TI) [08]) destinations or
similar.  Admittedly lower-subreg also handles shifts, and a few others, but
not vec_concat etc.
Plus
V2DI move: original cost 4, split cost 4 * 2
so it decides not to do anything about V2DImode (generally the right thing).

So, we'd need to pattern recognize a vec_concat from scalar modes (integral
only) and the result used (single use) in a subreg to twice that big integral
mode and
handle it as if it was setting the two halves of the wider reg.