https://bugs.kde.org/show_bug.cgi?id=523843

--- Comment #7 from Martin Cermak <[email protected]> ---
I've been measuring `perf stat -r 20  memcheck/memcheck-x86-linux
none/tests/x86/sse4-x86` and comparing this with and without
-mstack-alignment=16 -mstackrealign.  Although the cpu_core/instructions is
clearly slightly higher with realign in place (roughly + 5%), the IPC
(instructions per cycle count) is higher with alignment (roughly +7%) which
means less cycles overall with realignment in place. Total elapsed time dropped
from 1.990835071 to 1.970732172, which actually means that with the realignment
in place, none/tests/x86/sse4-x86 is slightly (1%) faster.

For a non-sse testcase, I've picked none/tests/x86/lzcnt32. 
cpu_core/instructions raise +2.7%, IPC: no change, total elapsed time raised
from 0.489375488 sec to 0.496833509, which seems like a minor (1.5%)
regression.

Of course, we have more non-sse tests than sse tests .  But the impact of the
stack realignment per patch from comment #4 seems negligible to me.

-- 
You are receiving this mail because:
You are watching all bug changes.

Reply via email to