bpo-46329: Avoid PUSH_NULL when followed by LOAD_GLOBAL

[markshannon]

Uses the low bit of LOAD_GLOBAL's oparg to indicate whether it should push a NULL first.

len(alist)

now compiles to

LOAD_GLOBAL              1 (NULL + len)
LOAD_FAST                0 (alist)
PRECALL                  1
CALL                     1

This reduces the total number of VM instructions executed by ~2%.

The benchmark results show an unlikely 3% speedup, probably just one of those random build variations.

https://bugs.python.org/issue46329

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @markshannon for commit 198517e 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[markshannon]

Buildbot failures seem unrelated to this PR: 2 pre-existing failures and one asyncio timeout.

[MatthieuDartiailh]

I may be missing something but shouldn't the computation for the stack effect of LOAD_GLOBAL be also adjusted since it can now push 2 elements on the stack ?