Make uop executor objects variable length

[gvanrossum]

Currently _PyUOpExecutorObject contains a fixed-length (_Py_UOP_MAX_TRACE_LENGTH, 32) array of micro instructions. We should switch to the usual variable-length array (making it a PyVarObject) so that we can make the largest superblock larger without paying the memory overhead cost for the full array.

Linked PRs

• [WIP] gh-106608: Have _PyUOpExecutorObject use a variable-length array #107411
• gh-106608: make uop trace variable length #107531

[gvanrossum]

This should be easy for someone with C experience. There are many examples in the CPython code base that do a similar thing. Basically the struct should be declared to end with _PyUOpInstruction trace[1]; and be given a size field, and various uses of the struct should be updated. I do expect a contributor taking this up to do their own research on how to build CPython, etc.

[dalalvarun]

I want to give it a try @gvanrossum

[gvanrossum]

@dalalvarun Go for it. Let me know if you get stuck.

[kgdiem]

Based on your description above I started off trying to base the implementation of trace off of a PyListObject / PySetObject;

From your description though,PyExecutorArray seemed to fit your description of the implementation better:

Basically the struct should be declared to end with _PyUOpInstruction trace[1]; and be given a size field

typedef struct {
    int size;
    int capacity;
    struct _PyExecutorObject *executors[1];
} _PyExecutorArray;

But this seems to be in conflict with the aim of avoiding having to copy everything into a new, larger array.

My questions:

1. Is something similar to a PyListObject but without all the casts the aim?
2. Should there be a new _Py_UOP_MAX_TRACE_LENGTH or should the size of the trace be SSIZE_MAX?

[gvanrossum]

@kgdiem Good questions. What I had in mind is not like PyListObject (which merely has a pointer to a separately malloc'ed array). It's more like a PyTupleObject. The struct you show is closer to what I have in mind, but we don't need a capacity field, and it should be an object. The idea is that the trace is initially written to a temporary buffer (_PyUOpInstruction trace[_Py_UOP_MAX_TRACE_LENGTH] in uop_optimize), and then a _PyUOpExecutorObject is allocated of the right size, into which the trace is copied. The role of _Py_UOP_MAX_TRACE_LENGTH will be reduced to the size of the temporary buffer.

[gvanrossum]

Looks like @rdrf2838 won the race.