Illegal WG barrier usage in prm/core/parallel/fbarrier/wait_race/1_1024x1x1_256x1x1

[pjaaskel]

This test case has a WG barrier that is entered only by a single WI in the whole grid. Others exit the kernel early. The behavior of the last WG that contains that WI (1023) is undefined due to the WG barrier's all-or-none semantics.

module &sample:1:0:$base:$large:$near;

prog kernel &test_kernel(kernarg_u64 %output)
{
    fbarrier %fb;
    workitemflatabsid_u64   $d0;
    cmp_ge_b1_u64   $c0, $d0, 1023;
  ; Only the last WI of the last WG in the 1024x1x1 grid proceeds...
    cbr_b1  $c0, @then;
  ; ...the rest WIs jump to the end which exits the kernel.
    br  @endif;

@then:
    sub_u64 $d1, $d0, 1023;
    workitemflatid_u32  $s0;
    cmp_ne_b1_u32   $c1, $s0, 0;
    cbr_b1  $c1, @L;
  ; Is any WI going here in case only 1023 executes
  ; the kernel? It's 255 in the last WG?
    initfbar    %fb;

@L:
  ; ...only WI 1023 reaches this barrier, thus
  ; it's undefined what happens because the others
  ; have exited.
    barrier;
....

[SamWot]

Yes, you are right - this is error in test.
I'm working on fixing it.