Do you have a benchmark showing the effect of this change?

_dec_mpd_radix() is added (818c24f) so as not to be passed a different object, not for performance.

My different PGO builds says 3-5% faster on thetelco benchmark.

My different PGO builds says 3-5% faster on the telco benchmark.

Would you mind to share your results (before/after, difference)?

On release builds, the result was as fast as the global state access:#123100 (comment) (PR with alternatives: 1.08x )

Picked a fewtelco results on Windows PGO builds:

a) main withd69529d restored (hard to evaluate main as-is)

9.68 ms +- 0.17 ms -> 8.94 ms +- 0.15 ms: 1.08x faster  # PR (ctx object)9.68 ms +- 0.17 ms -> 8.76 ms +- 0.11 ms: 1.10x faster  # global state access

b) (a) with some patches

8.99 ms +- 0.11 ms -> 8.49 ms +- 0.16 ms: 1.06x faster  # PR (ctx object)8.99 ms +- 0.11 ms -> 8.44 ms +- 0.20 ms: 1.07x faster  # global state access

c) (b) with 1 diff inPyType_GetBaseByToken()

8.91 ms +- 0.21 ms -> 8.78 ms +- 0.16 ms: 1.02x faster  # PR (ctx object)8.91 ms +- 0.21 ms -> 8.25 ms +- 0.17 ms: 1.08x faster  # global state access

Sorry, I don't know how to read your 6 benchmarks. I don't understand what you compared. I don't know what are "some patches" nor why you mention PyType_GetBaseByToken(), it doesn't help me to comparethis PR.

So I just compared the main branch to this PR:

Mean +- std dev: [ref] 8.18 ms +- 0.28 ms -> [change] 7.53 ms +- 0.18 ms: 1.09x faster

Nice! 1.09x faster is worth it.

I added the skip news since Victor enabled auto-merge. If a NEWS entry is really needed, we can add it later in a separate commit though.

Merged, thanks@neonene!