https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95435
Alexander Monakov <amonakov at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amonakov at gcc dot gnu.org
--- Comment #5 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
Ugh. Stringop tuning for Ryzens is terribly anachronistic, all AMD processors
since K8 (!!) use the exact same tables, and 32-bit memset/memcpy don't use
libcall for large sizes:
static stringop_algs znver2_memcpy[2] = {
{libcall, {{6, loop, false}, {14, unrolled_loop, false},
{-1, rep_prefix_4_byte, false}}},
{libcall, {{16, loop, false}, {64, rep_prefix_4_byte, false},
{-1, libcall, false}}}};
(first subarray is 32-bit tuning, the second is for 64-bit)
Using test_stringop microbenchmark from PR43052 it's easy to see that library
memset/memcpy are fastest on sizes 256 and above. Below that, the result from
the microbenchmark may be debatable, I think we should prefer the libcall
almost always except for tiniest sizes for I-cache locality reasons.
But anyway, current tuning is completely inappropriate.
