On Thu, 07 Nov 2019 22:36:44 +0100, Jan Kratochvil wrote:
On Thu, 07 Nov 2019 15:59:41 +0100, Victor Stinner wrote:
> I cannot explain why inlining cannot be done more often in libpython.
>
> I cannot explain why PLT is needed when a libpython function calls a libpython
function.
Could you re-run the benchmark with shared library but with
-fno-semantic-interposition? I have run it locally but it takes a lot of time.
nbody python3-3.7.5-1.fc30.x86_64: Mean +- std dev: 217 ms +- 2 ms
nbody -fno-semantic-interposition: Mean +- std dev: 203 ms +- 3 ms - 6.9%
nbody static linkage claim: -27%
So -fno-semantic-interposition does help but it is not the whole static gain.
Jan