This code runs in under 1 second:

import scipy.stats
import numpy as np

scipy.stats.skellam.sf(np.arange(100), 3.752321119795838e-06, 0.000001)

but if you import gevent

import gevent
scipy.stats.skellam.sf(np.arange(100), 3.752321119795838e-06, 0.000001)

then it takes many seconds to complete.

I don't entirely understand the culprit, i.e. how it affects the execution of scipy's Fortran code. From randomly breaking with gdb, in the 2nd case, the scipy code doesn't seem to use libm. I've narrowed it down to gevent being compiled with -ffast-math (due to -Ofast, introduced in 59478eb, first released in 20.6) and thus changing process-wide floating point flags. The change happens immediately when gevent's shared libraries are loaded, since the C runtime's set_fast_math code is set to run as a library constructor.

I've checked it by running the following function after import gevent to "fix" the problem:

void set_slow_math()
{
  unsigned int mxcsr = __builtin_ia32_stmxcsr ();
  // inverse of https://github.com/gcc-mirror/gcc/blob/master/libgcc/config/i386/crtfastmath.c#L94-L96
  mxcsr &= ~(MXCSR_DAZ | MXCSR_FTZ);
  __builtin_ia32_ldmxcsr (mxcsr);
}