Add INLINABLE to AST traversals — full builds -9%

[kozak]

Summary

Two-line change: add {-# INLINABLE #-} to everywhereOnValuesTopDownM and everywhereOnValuesM in src/Language/PureScript/AST/Traversals.hs.

+ {-# INLINABLE everywhereOnValuesTopDownM #-}
  everywhereOnValuesTopDownM f g h = (f' <=< f, g' <=< g, h' <=< h)
  ...
+ {-# INLINABLE everywhereOnValuesM #-}
  everywhereOnValuesM f g h = (f', g', h')

Why

The hot caller is Entailment.replaceTypeClassDictionaries (src/Language/PureScript/TypeChecker/Entailment.hs:118-139). It runs once per typechecked declaration with monad m = WriterT (Any, [...]) (StateT InstanceContext TypeCheckM) — four transformer layers, each with its own >>= per recursive descent over Expr trees.

The traversal helpers were polymorphic over Monad m =>. Without INLINABLE, GHC can specialise within the defining module but cannot see the unfolding when compiling other modules — so callers go through the runtime Monad dictionary on every recursive bind. Adding INLINABLE exposes the unfolding in the .hi file; GHC's specialiser at each call site then collapses the polymorphic helper into a flat, monad-specialised loop.

Measurements

Two interleaved-measurement runs (exp run, median-of-4 after warm-up) against the pre-change baseline (c84101d8) on the pr-admin workload (1758 modules):

Scenario	Run 1 (load 4–7)	Run 2 (load 1.85)
full	−9.6%	−8.9%
nochange	+8.4%	−3.0%
prelude	−0.5%	+1.6%
leaf	−4.3%	−1.5%

Per-round full deltas (Run 2): −10.1%, −15.2%, −9.6%, −8.9% — even the worst round is −8.9%, so the full-build win is robust. nochange/prelude/leaf are within noise floor (Run 1's nochange +8.4% was a load-contamination artefact on a 643 ms wallclock; Run 2's −3.0% confirms it).

Cost

• Diff size: 2 lines.
• Binary size: 48,625,952 → 48,839,456 bytes (+213 KB / +0.4%) — GHC emits a few specialised copies at heavy callers.
• stack build time: unchanged within noise.
• Tests: stack test --fast passes 1340 examples.

Test plan

• stack build (optimised) — green
• stack test --fast — 1340 examples, 0 failures
• Full pr-admin compile via exp run --scenarios all --runs 5 — robust ~−9% on full builds across two runs

🤖 Generated with Claude Code

[zyla]

LG

I was super confused at first when I saw this - my thought was "Wait, didn't we add INLINE to the traversals a long time ago?". The answer is: yes, but to the type traversals, not full AST.

[seastian]

Hi! the work you are doing here is very good! Does it make sense to port this to the official one?