https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127100
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Assignee|unassigned at gcc dot gnu.org |jakub at gcc dot gnu.org
Status|NEW |ASSIGNED
--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 65432
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=65432&action=edit
gcc17-pr127100.patch
This is what I came up with. It doesn't help that I think only s390 ever has
non-zero bias and only on len_load/len_store, while most of the other named
patterns are riscv only and always 0 bias, for which + vs. - doesn't really
matter.
The patch doesn't include the testcase, unsure what to do about that.
On one side, it is already tested whenever one configures s390x gcc with
--with-arch=z13 or --with-arch=z14 (or perhaps higher). If we wanted to test
it even in other configurations, there is currently no gfortran.target/s390 and
using s390_useable_hw outside of s390 specific directories (e.g. in
gfortran.dg) is not possible because the tcl function is only defined in
s390.exp.
Plus, I can test this in 16 branch in Fedora build system, but it takes there ~
2 days, or I can test on cfarm192 which could be somewhat faster, but still I
think Stefan has faster boxes around.
