Per-object flags

[markshannon]

This presumes a one-time ABI breakage, moving to abi4. Such a breakage seems likely in order to make no-gil feasible. If it doesn't happen, then the following is moot.

We change the object header from:

struct _object {
    intptr_t ob_refcount;
    PyTypeObject *ob_type;
};

to

struct _object {
    uint32_t ob_refcount;
    uint32_t ob_flags;
    PyTypeObject *ob_type;
};

which is the same size on 64 bit platforms, and one word larger on 32 bit platforms.

We can do a lot of performance optimizations with those bits. First of all, setting them during object creation should be free, or close to it, as we need to set the ref count at the same time.

Then we can make many improvements, such as:

• Free objects direct to the freelist by indexing on the size class in the header.
• Avoid expensive lookup via the class to check if objects should be checked for cycles
• Avoid allocating space for weakref pointer. Just use a single bit in the flags.
• Store the size of the preheader, to further speed deallocation
• Check for compact ints in a single operation
• Store more GC information, to allow a single GC pointer in the pre-header.

Possible flags:

• Size class (8 bits)
• Is GC
• Preheader size (2 bits)
• Needs finalizing
• GC state (5 or 6 bits)
• Has weakrefs
• Is compact int
• Statically allocated

Related issue on object layout: https://github.com/faster-cpython/ideas/issues/553

[gvanrossum]

I expect this will have to wait until we've stopped supporting abi3. During the time when abi3 and abi4 coexist, compatibility with abi3 will presumably require us to keep the layout of the ob_refcnt field unchanged. Let's discuss in Brno when we can drop abi3 -- we could conceivably do this overnight (i.e., in 3.13) as long as we preserve source compatibility. Or perhaps there are pressing reasons (like promises to the users of abi3? I have no idea what we've promised) to keep it around longer.