https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128722

Jerry DeLisle <jvdelisle at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
                 CC|                            |jvdelisle at gcc dot gnu.org
   Last reconfirmed|                            |2026-11-16
             Status|UNCONFIRMED                 |NEW

--- Comment #2 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> ---
Confirmed. I ran a benchmark here: a 5M-element pointer array through a dummy,
sync and async, built with trunk and with the patched compiler, plus a
non-pointer control that both pass by descriptor.


┌──────────────────────────────────────┬─────────┬─────────┬─────────┐
│                 Case                 │  Trunk  │ Patched │ Speedup │
├──────────────────────────────────────┼─────────┼─────────┼─────────┤
│ Pointer to d%x (32-byte span), sync  │ 0.073 s │ 0.053 s │ 1.4×    │
├──────────────────────────────────────┼─────────┼─────────┼─────────┤
│ Pointer to d%x (32-byte span), async │ 1.269 s │ 0.053 s │ 24×     │
├──────────────────────────────────────┼─────────┼─────────┼─────────┤
│ Contiguous pointer, sync             │ 0.073 s │ 0.008 s │ 9×      │
├──────────────────────────────────────┼─────────┼─────────┼─────────┤
│ Contiguous pointer, async            │ 1.268 s │ 0.008 s │ 160×    │
├──────────────────────────────────────┼─────────┼─────────┼─────────┤
│ Control (non-pointer dummy)          │ 0.008 s │ 0.008 s │ —       │
└──────────────────────────────────────┴─────────┴─────────┴─────────┘

- Async is where the big gain is. Trunk queues one item per element: a calloc
plus a lock and signal, 5 million times.
- Sync with a non-contiguous span improves only 1.4×. The library still copies
a strided array element by element; it just no longer goes through a library
call per element.

Reply via email to