https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97291
Bug ID: 97291
Summary: [SIMT] Move SIMT_XCHG_* out of non-uniform execution
region
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libgomp
Assignee: unassigned at gcc dot gnu.org
Reporter: vries at gcc dot gnu.org
CC: jakub at gcc dot gnu.org
Target Milestone: ---
We have:
...
/* Allocate per-lane storage and begin non-uniform execution region. */
static void
expand_GOMP_SIMT_ENTER_ALLOC (internal_fn, gcall *stmt)
...
and:
...
/* Deallocate per-lane storage and leave non-uniform execution region. */
static void
expand_GOMP_SIMT_EXIT (internal_fn, gcall *stmt)
...
So, if the SIMT_ENTER_ALLOC and the SIMT_EXIT mark the start and end of a
region of non-uniform execution, it's strange that such a region can contain
SIMT_XCHG_*, which on nvptx requires uniform execution.
Moving SIMT_VOTE_ANY/SIMT_LAST_LANE/SIMT_XCHG_* as a whole after SIMT_EXIT is
not possible given that VOTE_ANY may have data dependencies to storage that is
deallocated by SIMT_EXIT (as Alexander mentioned here:
https://gcc.gnu.org/pipermail/gcc-patches/2020-October/555475.html).
A possible solution would be to split the SIMT_EXIT into separate bits for
exiting non-uniform execution and deallocation, and have:
- SIMT_ENTER_ALLOC
- SIMT_EXIT_UNI
- SIMT_VOTE_ANY/SIMT_LAST_LANE/SIMT_XCHG_*
- SIMT_EXIT_DEALLOC
Also I've wondered if we could do:
- SIMT_ENTER_ALLOC
- SIMT_VOTE_ANY
- SIMT_EXIT
- SIMT_LAST_LANE/SIMT_XCHG_*
but perhaps there are again data dependency problems.
