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

[rguenth at gcc dot gnu.org]

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

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #43287|0                           |1
        is obsolete|                            |

--- Comment #24 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 46044
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=46044&action=edit
updated vectorizer patch

I've updated the vectorizer patch that was posted to the mailing list last
year.
I changed it to also apply to the non-SLP case and instead of looking for a
non-vector-mode return RTX (which would pessimize word_mode vectorization)
look at hard_regno_nregs and only pessimize multi-reg return locations.

I've removed CONCAT and PARALLEL handling with PARALLEL left as todo
(not sure what to do there, I'd need to have a target / testcase combo
to look at - I suppose if any of the return locations is "compatible"
with the vector result we want to not pessimize).