[Bug target/70321] [6 Regression] STV generates less optimized code

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70321

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 38057
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38057&action=edit
gcc6-pr70321.patch

I'm afraid for GCC 6.x we can't do much more than attached patch, which
improves the generated code, but there is no combiner post RA that would
improve it even more.  For GCC 6.x, I think it is out of question to move STV
pass somewhere else.

For stage1, I wonder if it can't move earlier, say before the combiner.  If it
could, then we could split the TARGET_STV patterns that weren't changed into
vector insns, and let the combiner deal with that.  Is there a reason why it is
done after the combiner now?