Ah, that's interesting. I thought that would break my GIL profiling project that uses LD_PRELOAD (shameless plug: https://github.com/chrisjbillington/gil_load), but since I think I'm only overriding libc functions, it should be fine. I'm sure there are other things it will break (I could have overridden libpython functions instead - I wonder why I didn't, it seems simpler but there was probably a reason), but if someone wants to use code that hacks on the interpreter itself, having to install a custom python to do so is not so unreasonable. Those speedups are nothing to scoff at. -Chris On Tue, Nov 26, 2019 at 5:07 AM Ralph Corderoy <ralph@inputplus.co.uk> wrote:
Hi,
I thought this might be of interest. https://fedoraproject.org/wiki/Changes/PythonNoSemanticInterpositionSpeedup
By building with -fno-semantic-interposition they remove the PLT that provides a level of indirection when calling a libpython function. libpython often calls itself and the PLT adds L1-cache pressure plus prevents inlining. Gives gains of 25% on some workloads.
-- Cheers, Ralph.