[Bug rtl-optimization/70542] [6 Regression] Wrong code with -O3 -mavx2.

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70542

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 38191
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38191&action=edit
gcc6-pr70542.patch

Perhaps better fix.  Looking at PR64286, the comment is right, if we change
somehow the definition for VECTOR_MODE_P, we need to make sure all uses are
modified, or none.  But there is tons of reasons why some of them could fail,
e.g. for the copy_needed case there is lots of tests that can fail, but even
for the !copy_needed case it could fail to get recognized etc., and we are
unable to handle all the uses as a single transaction.