https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80286
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |jakub at gcc dot gnu.org
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 41111
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41111&action=edit
gcc7-pr80286.patch
Untested fix for the correctness part. Now if the shift count comes from
memory or XMM register, it would be certainly faster and shorter to emit
v?pmovzxdq
before the shift (for SSE4.1+) or pxor; punpck* for SSE2, wonder how to do that
best though; first of all it would be nice to transform sign-extension followed
by shift using the sign-extended count into zero-extension followed by shift,
because if the shift count is negative, the insn will handle it as very large
count no matter if it is zero or sign extended. The second question is if
zero_extendsidi2 etc. instruction should have a =v/vm alternative at least for
sse4 isa, but it already has some alternative with x and uses ? to make it not
likely to be used.
