solidjs / solid / 39014688291

78%

main: 89%

push

github

perf(solid): a suspended read unwinds to its boundary through fewer handler frames (#4122)

* perf(solid): a suspended read unwinds to its boundary through fewer handler frames

Profiling the shell of a streamed page with ten pending boundaries
(pinned to a performance core, pre-warmed, 20 µs sampling) put a fifth of
the first chunk's CPU on the NotReadyError unwind itself, not on any work:
V8 runs a `finally` as a catch-and-rethrow, and between the suspended
read and the boundary's catch sat the sync memo's catch + finally, the
ssrScope closure's inner catch + outer finally, two runWithOwner frames
(catch for stampThrower + finally) and runWithBoundaryErrorContext's
finally — about eight re-raises per boundary at the cost of the throw
each. And every compiler hole memo was pulled twice: the eager pull at
creation met the NotReady (swallowed, "surface on read"), then ssr()'s
read pulled again and met it again.

Now:
- runWithOwner, createSyncMemo's pull, ssrScope's closure and
  runWithBoundaryErrorContext restore their state inline on the normal
  path and in one catch (which also stamps, where it did), no finally.
  runWithBoundaryErrorContext does its own owner swap instead of wrapping
  runWithOwner, and loses the render closure.
- A sync memo whose eager pull met a NotReadyError latches it for the rest
  of the synchronous turn: a read in that turn re-raises it instead of
  running the compute a second time. Nothing can settle the source inside
  a turn (promises settle on microtasks; server renders are pure), so the
  second pull could only reproduce the first's answer. One microtask,
  shared by every latch armed in the turn, clears them; the next turn —
  the boundary's retry, after the source may have settled — pulls as
  before. The latch is allocated only when a pull suspends.

No behavior or wire change; solid, web (default, server, hydrate) and
signals suites pass.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* perf(solid): ... (continued)

756 of 1050 branches covered (72.0%)

Branch coverage included in aggregate %.

1097 of 1339 relevant lines covered (81.93%)

27.83 hits per line