Repository navigation
TypedPython: compiled code never checks the eval breaker (signals, GC, GIL hand-off) #141
Description
Since N-8, compiled loops do not check the eval breaker, so during a long compiled loop four things are deferred until it returns: signals (KeyboardInterrupt), scheduled garbage collections, pending calls, and dropping the GIL for other threads. The first three only move in time. The last starves other Python threads, which is a behaviour difference, not only a timing one.
Proposal: check the eval breaker at loop back-edges and at calls, with a cost measurement (nbody, spectral-norm, fannkuch) before and after. Design note: docs/design/typedpython.md §4.3.4 마.2 (PR for N-12).
Done when: a test shows a second Python thread makes progress while a compiled loop runs, KeyboardInterrupt reaches a compiled loop, and the measured cost is recorded.
Activity
Perf: Poll loops through a function-local countdown (#141)
Docs: Bring the TypedPython design up to date (N-11 landed, #141 phas…
b-re-w commentedon Oct 3, 2026
Phase 1 (
feat/tp-eval-breaker) cost, interleaved A/B against develop in a quiet window (load 2.2 to 3.0), 11 pairs:whileloops. Next: poll only at the back-edges of loops that are not innermost, plus innermost loops by strip-mining (a chunked inner loop with one poll per chunk), then re-measure.Follow-ups:
feat/tp-free-threaded) is committed on top of phase 1: 974 passed, 3 skipped on 3.15t.Perf: Keep the eval-breaker slow path out of line and the loop tick i…
Docs: Record the eval-breaker polling cost (#141)
Merge pull request #207 from thisisthepy/feat/tp-eval-breaker
b-re-w commentedon Oct 17, 2026
Landed in develop via #207. Measured cost: fannkuch +8.2% (accepted, see the PR); chunked range-loop polling is #206; phase 2 is in #159.