mirror of
https://github.com/HeyPuter/gex.git
synced 2026-11-13 18:34:21 -10:00
velzieandClaude Opus 5
df848cbfc6
`region.rs` motivates regions partly with the observation that V8 sees
hundreds of thousands of tiny functions and never optimises most of them --
76% of SQLite's execution staying in the baseline compiler. Read forward,
that suggests a large win for anything that makes functions fewer and
bigger, and it has been the argument for chasing more region coverage.
It does not survive measurement. CPython statically recompiled, the same
module both sides, one with V8 told to compile everything with TurboFan up
front, five alternating rounds, minimums:
fannkuch default 7.103s TurboFan-only 6.629s 0.933x
nbody default 7.568s TurboFan-only 8.218s 1.086x
dicts default 2.258s TurboFan-only 2.172s 0.962x
Two marginally faster, one marginally slower, all near the six per cent this
harness can resolve. Removing the baseline compiler entirely -- the whole of
the effect the statistic describes -- is worth nothing measurable.
That is not an argument against regions, which are measured on their own
terms further down the same file. It is an argument against expecting a
large win from function count, and it is written where the claim that
suggests otherwise lives.
`AB_FLAGS_B` in the harness runs the B side with extra node flags, which is
how an engine setting gets A/B'd without changing the program.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01KcidnrZ5aLqKoBJBqdjx9d