gh-145665: Configurable tos cache

[aisk]

• Issue: Make size of TOS cache configurable #145665

[markshannon]

Thanks for giving this a go. I've had a quick look.

Rather than take up memory for tables and stencils for all permutations up to MAX_GENERATED_CACHED_REGISTER we only want them up to MAX_CACHED_REGISTER.

We still want to generate the code for all possible cases, but have it turned off by the preprocessor.
For example, in executor_cases.c.h you have:

        case _NOP_r55: {
            #if MAX_CACHED_REGISTER >= 5

but it will need to be:

    #if MAX_CACHED_REGISTER >= 5
        case _NOP_r55: {

Likewise in pycore_uop_ids.h, the names will need to be guarded:

#if MAX_CACHED_REGISTER >= 5
    #define _NOP_r55
#endif

To avoid gaps, the ids will need to be sorted by the largest cached register used.

[aisk]

Sorry, I don't have free time in the next few weeks, so I will close this temporarily incase of someone want work on this.

Will re-open and continue the work if no one work on this when I have time.

[aisk]

Hi @markshannon, with different MAX_CACHED_REGISTER, get_uop_cache_depths will generate different rXY combinations, so I changed to generate completely different cases / metadata variants by different MAX_CACHED_REGISTER, guarded by a C macro, and using the C macro to select them when building.

But I have a few questions:

1. Currently, only 3–5 variants are generated, but what's the recommended range for this?
2. What's the right way to configure MAX_CACHED_REGISTER? Currently, it's hardcoded in pycore_uop.h and _targets.py.