co_names per code object

[iritkatriel]

I'm working on the pyco part now.

[iritkatriel]

Almost all tests are passing now. Only two of the values from test_basic are still failing, the ones that contain strings - "abc" and "(0, 1, 'aa', 'bb')". Looks like something is going wrong with STORE that deal with strings. I'll debug.

[iritkatriel]

That was actually easy to fix. All tests in test_new_pyc are passing now.

[gvanrossum]

I'll take a look at the code soon, but first a suggestion. A test that has e.g. 100 functions using 100 distinct globals each (so 10,000 globals in total) should work, since the co_names of each function would have only 100 items, right?

FWIW One of my goals is to eventually be able to compile every stdlib module correctly -- then we should be able to run the full test suite. We're not there yet, and I would focus first on quick wins that help us demonstrate that this is faster first. And a big blocker (for completeness, not for speed) is that we need to be able to hydrate any constant by calling a C function, so we can make the co_consts attribute appear fully hydrated to Python code. (In an earlier version of my branch I had such a function and called it from LAZY_LOAD_CONSTANT, but it was too slow and I rewrote everything. But I now realize that we still need such a function -- and for this use case it doesn't need to be as fast as LAZY_LOAD_CONSTANT.)

[gvanrossum]

This is exciting! Here are a few suggestions.

[gvanrossum]

PS. I discovered a logic bug in helper() in do_test_speed(). I've pushed a commit that fixes it (93e91c2).

[iritkatriel]

Yes, I know we still need co_names. It's easier to debug when I break one thing at a time. I had to hard-reboot my windows machine twice today. I felt sorry for it so I switched to the mac which doesn't get offended so easily.

The test with 100 functions with 100 globals each works if I define the globals inside the function, but if I define them in the module then we go over 256 there.

[gvanrossum]

LGTM. Maybe we can merge this and then continue iterating?

[iritkatriel]

LGTM. Maybe we can merge this and then continue iterating?

Sure. My next thing is EXTENDED_ARGs for strings, then to move on to consts.