Passing the state around as a parameter is one way to get rid of ET_STATE_GLOBAL. It results in a larger diff (more churn), but IMO it is an ok approach. Alternatively, we can store state as a member in the type contexts, which will result in a smaller diff.

(I will rebase this onto main after #101187 has landed.)

cc. @arhadthedev

After this, the remaining parts are:

• use PyModule_GetState where possible (1 small PR)
• use PyType_GetModuleState where possible (1 slightly larger PR w/clinic changes)
• use PyType_GetModuleByDef+PyModule_GetState where possible
• adapt to multi-phase init
• reapply the pyexpat-c-api-on-the-heap PR

use PyModule_GetState where possible (1 small PR)
use PyType_GetModuleState where possible (1 slightly larger PR w/clinic changes)
use PyType_GetModuleByDef+PyModule_GetState where possible
adapt to multi-phase init

Up to you but I can review of all this in one go, they are relatively small changes.

Side thought: sizeof(ElementObject) is 56 which rounds off to 64 for alignment so we can add a state pointer and the memory usage will be same.

Side thought: sizeof(ElementObject) is 56 which rounds off to 64 for alignment so we can add a state pointer and the memory usage will be same.

Yes; well observed.