Avoid CPS effect type instantiation

[plajjan]

Summary

• Read CPS call effects through schemaOf instead of forcing the full callee type with typeOf.
• Dropped the typed-local initializer QuickType skip after review. That path can lose argument coercions because genExp/qType also rewrites call arguments with qMatch.

Performance

Measured on the 61 MB concurrent_typecheck_class_heavy fixture with out cleared before each run:

compiler-bench --pipeline dist/base/out/types test/compiler/concurrent_typecheck_class_heavy/src/concurrent_typecheck_class_heavy.act +RTS -T -A64M -RTS

The earlier combined numbers included the dropped QuickType commit and no longer apply. This is the current one-commit branch result compared with the previous baseline run from the compiler-bench head now merged to main.

Baseline:
parse 10.655304s
front 347.252679s
front_timing env 0.0s kinds 5.467403s types 341.778869s
back 153.737079s
back_stats alloc 371524555088 copied 90198456432 max_live 12653155120 max_mem 37528535040 gc_elapsed 35213930000 gcs 424
back_timing normalize 0.0s deactorize 3.600145s cps 1.0e-6s llift 37.588621s boxing 0.0s codegen 1.0e-6s render 107.705677s write 4.842622s

CPS-only branch:
parse 10.01912s
front 342.031121s
front_timing env 1.0e-6s kinds 5.433949s types 336.579875s
back 168.719363s
back_stats alloc 369063478000 copied 90455088080 max_live 13566249416 max_mem 36836474880 gc_elapsed 41838148000 gcs 422
back_timing normalize 0.0s deactorize 3.987647s cps 1.0e-6s llift 40.53988s boxing 0.0s codegen 1.0e-6s render 119.530711s write 4.661106s

This run does not show a wall-time improvement for the remaining CPS-only change. Back allocation is slightly lower, but back wall time is worse in this sample.

Check

stack --stack-yaml compiler/stack.yaml build libacton:exe:compiler-bench

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d8461bf2f6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve argument coercions for unboxed initializers

For unboxable local declarations this now bypasses genExp, but genExp does more than the final result cast: it runs qType, whose Call case applies qMatch to coerce each argument to the callee's parameter types. When an initializer such as x: int = f(v) calls a function whose argument needs one of those non-unboxable casts (for example an optional or superclass parameter), the generated call now uses raw gen env e and omits the required argument coercion, while the old path still avoided any final cast because adjust is a no-op for unboxable result types. This can produce invalid or mis-typed C for typed local initializers even though the target itself is unboxable.

Useful? React with 👍 / 👎.