perf: avoid `call_indirect` for capturing closures with statically-known fi by ggreif · Pull Request #5964 · caffeinelabs/motoko

ggreif · 2026-03-31T21:07:45Z

Summary

Introduces SR.StaticClosure of int32 in both compile_classical.ml and compile_enhanced.ml to carry the statically-known Wasm function index from FuncDec.closure to the CallPrim call site
A special case in compile_dec eagerly compiles LetD (VarP v, FuncE …) non-const expressions to extract fi, binding v as VarEnv.Local (SR.StaticClosure fi, …) so the call site sees the correct SR
CallPrim matches SR.StaticClosure fi, Type.Local before the generic _, Type.Local arm, emitting Call fi directly instead of loading funptr_field and doing call_indirect — the closure is still allocated and passed as $clos
test/run/fmodf-forward.mo: $bar.1 now emits call $quux.1 instead of call_indirect
Guards prevent the optimisation from firing unsafely: AllocHow.(LocalImmut _) (skips variables needing heap indirection for forward references) and VarEnv.all_in_scope (skips mutually-recursive groups where captures aren't yet in scope)

Resolves the optimization opportunity sketched in #5961.

What is NOT done

Subcase 3 (function passed as argument to higher-order callee) — requires devirtualization, out of scope

Test plan

test/run/fmodf-forward.mo passes all phases: [tc] [run] [comp] [valid] [FileCheck] [wasm-run]
tests.mo-idl clean (mutually-recursive guard)
tests.run-eop-release clean including quicksort (AllocHow guard)

🤖 Generated with Claude Code

ggreif · 2026-03-31T21:18:14Z

+         (match fi_opt with
+          | Some fi ->


Why not

Suggested change

(match fi_opt with

| Some fi ->

(match fun_sr with

| SR.StaticClosure fi ->

ggreif · 2026-03-31T21:23:13Z

+        (fun _ae -> fun_code ^^ G.i (LocalSet (nr local_i))),
+        unmodified )
+    | _ ->
+      let (pre_ae1, alloc_code, pre_code, sr, fill_code) = compile_unboxed_pat env pre_ae how (Source.({it = VarP v; at = e.at; note = typ})) in


Suggested change

let (pre_ae1, alloc_code, pre_code, sr, fill_code) = compile_unboxed_pat env pre_ae how (Source.({it = VarP v; at = e.at; note = typ})) in

let (pre_ae1, alloc_code, pre_code, sr, fill_code) = compile_unboxed_pat env pre_ae how Source.{e with it = VarP v; note = typ} in

ggreif · 2026-03-31T21:38:33Z

+    let fun_sr, fun_code = compile_exp env pre_ae e in
+    (match fun_sr with
+    | SR.StaticClosure fi ->
+      let (pre_ae1, local_i) = VarEnv.add_direct_local env pre_ae v (SR.StaticClosure fi) typ in


Suggested change

let (pre_ae1, local_i) = VarEnv.add_direct_local env pre_ae v (SR.StaticClosure fi) typ in

let pre_ae1, local_i = VarEnv.add_direct_local env pre_ae v fun_sr typ in

ggreif · 2026-03-31T21:39:25Z

+        (fun _ae -> fun_code ^^ G.i (LocalSet (nr local_i))),
+        unmodified )
+    | _ ->
+      let (pre_ae1, alloc_code, pre_code, sr, fill_code) = compile_unboxed_pat env pre_ae how pat in


Suggested change

let (pre_ae1, alloc_code, pre_code, sr, fill_code) = compile_unboxed_pat env pre_ae how pat in

let pre_ae1, alloc_code, pre_code, sr, fill_code = compile_unboxed_pat env pre_ae how pat in

github-actions · 2026-03-31T23:11:22Z

Comparing from 4dfc891 to f6f4f1e:
In terms of gas, 4 tests improved and the mean change is -0.0%.
In terms of size, 5 tests improved and the mean change is -0.0%.

…cally-known fi Introduces SR.StaticClosure (fi : int32) to compile_enhanced.ml. FuncDec.closure now returns SR.StaticClosure fi instead of SR.Vanilla, threading the statically-known function index to the call site. A special case in compile_dec eagerly compiles LetD (VarP v, FuncE …) non-const expressions with pre_ae, extracts fi from SR.StaticClosure fi, and binds v as VarEnv.Local (SR.StaticClosure fi, local_i). This is safe for non-recursive LetD since the FuncE does not reference v and all its captures are already in pre_ae. In CallPrim's _, Type.Local branch, fi_opt = Some fi causes a direct G.i (Call (nr fi)) instead of loading funptr_field and calling call_indirect. The closure is still allocated and passed as $clos (captures are needed). StackRep.adjust (SR.StaticClosure _) SR.Vanilla is a no-op (same i64 on stack). SR.StaticClosure is wired into to_block_type, to_var_type, to_string, drop, join, and adjust. test/run/fmodf-forward.mo: $bar.1 now emits call $quux.1 instead of call_indirect; CHECK updated accordingly. Classical backend (compile_classical.ml) not yet ported. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Documents the three subcases of the missed optimisation, the SR.StaticClosure approach (Option B), and current status. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Port the capturing-closure direct-call optimization (subcase 2) from compile_enhanced.ml to compile_classical.ml: - Add SR.StaticClosure of int32 variant (closure ptr on stack, fi known) - Wire into to_var_type/to_block_type (I32Type), to_string, drop, join, adjust - FuncDec.closure returns SR.StaticClosure fi instead of SR.Vanilla - compile_dec special-cases non-const LetD(VarP v, FuncE) to eagerly compile the FuncE and bind v as Local(SR.StaticClosure fi, …) - CallPrim's _, Type.Local branch matches fun_sr directly for SR.StaticClosure fi and emits Call fi + FakeMultiVal.load instead of call_indirect Also refactors compile_enhanced.ml to match fun_sr directly (dropping the intermediate fi_opt) as suggested in PR review. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Split the `_, Type.Local` arm into `SR.StaticClosure fi, Type.Local` (direct call, fi bound from pattern) and `_, Type.Local` (call_indirect fallback), eliminating the inner `match fun_sr` in both backends. Also bind the LetD pattern as `pat` in compile_dec's FuncE special case to pass it directly to compile_unboxed_pat. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Per review: use `let a, b = ...` instead of `let (a, b) = ...`, and pass fun_sr to add_direct_local instead of re-wrapping as SR.StaticClosure fi. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

In mutually-recursive local function groups AllocHow may promote a variable to StoreHeap (MutBox) during its fixed-point iteration so that forward references get a stable heap location. The SR.StaticClosure fast path in compile_dec bypasses compile_unboxed_pat, so it must not fire when AllocHow has decided the variable needs heap indirection. Add an AllocHow.M.find_opt guard (Some (AllocHow.LocalImmut _)) to the special case in both compile_classical.ml and compile_enhanced.ml. Without this the quicksort test aborts at runtime because the captured closure ends up being an unset local (0) instead of a valid MutBox pointer. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…sistency Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Verifies that anonymous capturing lambdas also get direct calls through both closure levels ($foo_lam → $@anon-func-N → $@anon-func-M directly). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…n details Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ggreif requested a review from a team as a code owner March 31, 2026 21:07

ggreif self-assigned this Mar 31, 2026

ggreif marked this pull request as draft March 31, 2026 21:13

ggreif changed the title ~~feat(enhanced): avoid call_indirect for capturing closures with statically-known fi~~ perf: avoid call_indirect for capturing closures with statically-known fi Mar 31, 2026

ggreif commented Mar 31, 2026

View reviewed changes

ggreif force-pushed the gabor/static-closure branch from 6d59b28 to f6f4f1e Compare April 1, 2026 01:19

ggreif mentioned this pull request Apr 1, 2026

perf: precisely short-circuit exact forwarders (fmodf et al. libm calls) and zero-forwarders (Motoko-specific) #5961

Draft

3 tasks

ggreif added the performance Affects only gas usage or code size label Apr 1, 2026

ggreif force-pushed the gabor/forward branch from 8d46414 to 809e739 Compare April 19, 2026 01:21

ggreif and others added 13 commits April 19, 2026 03:28

docs: add plan for avoiding call_indirect for statically-known functions

9b1a5fd

Documents the three subcases of the missed optimisation, the SR.StaticClosure approach (Option B), and current status. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cleanup: remove stale //MOxC-FLAG debug line from fmodf-forward.mo

e8350a7

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cleanup: remove spurious blank line in SR.StaticClosure call arm

5ba64a8

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cleanup: drop tuple parens in let bindings; use fun_sr directly

57a933e

Per review: use `let a, b = ...` instead of `let (a, b) = ...`, and pass fun_sr to add_direct_local instead of re-wrapping as SR.StaticClosure fi. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cleanup: use local open AllocHow.(...) to avoid double qualification

4d26c40

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cleanup: drop redundant parens around first LetD argument pattern

c8e82d2

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cleanup: rename pat to p in SR.StaticClosure special case for con…

c121cf8

…sistency Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

test: add foo_lam lambda analog to fmodf-forward.mo with FileCheck

e597bed

Verifies that anonymous capturing lambdas also get direct calls through both closure levels ($foo_lam → $@anon-func-N → $@anon-func-M directly). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docs: update avoid-static-indirect-call plan with final implementatio…

5fac707

…n details Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ggreif force-pushed the gabor/static-closure branch from 5d3cb5f to 5fac707 Compare April 19, 2026 01:29

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

perf: avoid `call_indirect` for capturing closures with statically-known fi#5964

perf: avoid `call_indirect` for capturing closures with statically-known fi#5964
ggreif wants to merge 13 commits intogabor/forwardfrom
gabor/static-closure

ggreif commented Mar 31, 2026 •

edited

Loading

Uh oh!

ggreif Mar 31, 2026

Uh oh!

ggreif Mar 31, 2026 •

edited

Loading

Uh oh!

ggreif Mar 31, 2026

Uh oh!

ggreif Mar 31, 2026

Uh oh!

github-actions Bot commented Mar 31, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

	let (pre_ae1, alloc_code, pre_code, sr, fill_code) = compile_unboxed_pat env pre_ae how (Source.({it = VarP v; at = e.at; note = typ})) in
	let (pre_ae1, alloc_code, pre_code, sr, fill_code) = compile_unboxed_pat env pre_ae how Source.{e with it = VarP v; note = typ} in

	let (pre_ae1, local_i) = VarEnv.add_direct_local env pre_ae v (SR.StaticClosure fi) typ in
	let pre_ae1, local_i = VarEnv.add_direct_local env pre_ae v fun_sr typ in

	let (pre_ae1, alloc_code, pre_code, sr, fill_code) = compile_unboxed_pat env pre_ae how pat in
	let pre_ae1, alloc_code, pre_code, sr, fill_code = compile_unboxed_pat env pre_ae how pat in

Conversation

ggreif commented Mar 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

What is NOT done

Test plan

Uh oh!

ggreif Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

ggreif Mar 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ggreif Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

ggreif Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented Mar 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

ggreif commented Mar 31, 2026 •

edited

Loading

ggreif Mar 31, 2026 •

edited

Loading

github-actions Bot commented Mar 31, 2026 •

edited

Loading