[Bug target/104580] New: [nvptx] Use prevent_branch_around_nothing only when necessary

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104580

            Bug ID: 104580
           Summary: [nvptx] Use prevent_branch_around_nothing only when
                    necessary
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: vries at gcc dot gnu.org
  Target Milestone: ---

I had the following idea:

The prevent_branch_around_nothing workaround was added to force
a uniform warp after this:
...
            {
                    .reg.u32        %x;
                    mov.u32 %x, %tid.x;
                    setp.ne.u32     %r23, %x, 0;
            }
            @%r23   bra     $L2;
    $L2:
...

With TARGET_PTX_6_0, we enforce a uniform warp by using bar.warp.sync, and
consequently this workaround is not necessary anymore.
...

However, this introduces the following regression:
...
FAIL: libgomp.fortran/task-detach-8.f90 -Os execution test
FAIL: libgomp.oacc-c/../libgomp.oacc-c-c++-common/reduction-7.c
-DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -foffload=nvptx-none  -O0
-DGOMP_NVPTX_JIT=-O0 execution test
...

Needs investigation.  Either the idea is wrong, or the tests FAIL for some
other reason.