The startup time of hg increased during the 4.6 development cycle. A
cause of that was importing more modules and doing more work at module
import time.
The import of zope.interface and the declaring of various interfaces
is partially responsible for the startup time regression.
Our current usage of zope.interface doesn't do much at run time: we are
merely declaring interfaces and stating that certain types implement
various interfaces. Core Mercurial is not (yet) using of any of
zope.interface features that actually require that interface plumbing be
defined. The only place we actually need the interface metadata is in
test-check-interfaces.py.
This commit establishes a new interfaceutil module. It exposes the subset
of the zope.interface API that we currently use. By default, the APIs
no-op. But if an environment variable is set, we export the real
zope.interface APIs.
Existing importers of zope.interface have been converted to use the new
module. test-check-interfaces.py has been updated to define the
environment variable so the real zope.interface is used.
The net effect of this change is we stop importing 9 zope.interface.*
modules and we no longer perform interface bookkeeping when registering
interfaces.
On my i7-6700K on Linux, a shell loop that runs hg log -r . 300 times
on a repo with 1 commit shows a significant CPU time improvement
(average of 4 runs):
4.5: 14.814s
before: 19.028s
after: 16.945s
And with run-tests.py -j10 (single run):
4.5: ~3100s (~51.7m)
before: ~4450s (~74.2m)
after: ~3980s (~66.3m)
So this claws back about half of the regressions in 4.6.