purescript-lua · Unisay · Jul 13, 2026 · Jul 13, 2026 · Jul 13, 2026 · Jul 13, 2026
diff --git a/bench/goldens/fnew_Bench.EffectStep.txt b/bench/goldens/fnew_Bench.EffectStep.txt
@@ -0,0 +1,20 @@
+chunk: Bench.EffectStep.lua
+runtime: LuaJIT 2.1.1741730670
+main-chunk FNEW: 11
+function-body FNEW: 13
+total FNEW: 24
+prototypes: 25
+function-body FNEW sites:
+  Bench.EffectStep.lua:4
+  Bench.EffectStep.lua:6
+  Bench.EffectStep.lua:6
+  Bench.EffectStep.lua:9
+  Bench.EffectStep.lua:10
+  Bench.EffectStep.lua:12
+  Bench.EffectStep.lua:12
+  Bench.EffectStep.lua:33
+  Bench.EffectStep.lua:54
+  Bench.EffectStep.lua:53
+  Bench.EffectStep.lua:48
+  Bench.EffectStep.lua:57
+  Bench.EffectStep.lua:66
diff --git a/bench/goldens/trace_effect_step.txt b/bench/goldens/trace_effect_step.txt
@@ -0,0 +1,28 @@
+spec: effect_step
+runtime: LuaJIT 2.1.1741730670
+workload: n=100000 reps=4 result=1500000
+aborts (distinct site -- reason):
+  Bench.EffectStep.lua:10 -- NYI: bytecode FNEW
+  Bench.EffectStep.lua:12 -- NYI: bytecode FNEW
+  Bench.EffectStep.lua:48 -- NYI: bytecode FNEW
+  Bench.EffectStep.lua:49 -- NYI: bytecode UCLO
+  Bench.EffectStep.lua:53 -- NYI: bytecode FNEW
+  Bench.EffectStep.lua:54 -- NYI: bytecode FNEW
+  Bench.EffectStep.lua:66 -- NYI: bytecode FNEW
+  Bench.EffectStep.lua:9 -- NYI: bytecode FNEW
+bytecode end state (J*=compiled, I*=blacklisted):
+  Bench.EffectStep.lua:10 IFUNCF
+  Bench.EffectStep.lua:10 JFUNCF
+  Bench.EffectStep.lua:11 IFUNCF
+  Bench.EffectStep.lua:12 IFUNCF
+  Bench.EffectStep.lua:12 JFUNCF
+  Bench.EffectStep.lua:17 IFUNCF
+  Bench.EffectStep.lua:36 IFUNCF
+  Bench.EffectStep.lua:37 IFUNCF
+  Bench.EffectStep.lua:39 IFUNCF
+  Bench.EffectStep.lua:40 JFUNCF
+  Bench.EffectStep.lua:9 JFUNCF
+  effect_step.lua:11 JFUNCF
+  effect_step.lua:13 JFORI
+  effect_step.lua:13 JFORL
+counts: aborts=8 compiled=7 blacklisted=7
diff --git a/bench/macro/effect_step.lua b/bench/macro/effect_step.lua
@@ -0,0 +1,16 @@
+-- A hot ST loop whose per-iteration step is a unary effect action, always
+-- fully applied and immediately run. Uncurried late (after magicDo), the
+-- step is one n-ary worker call per iteration; curried, every iteration
+-- allocates the thunk closure and pays a second call to force it.
+return {
+  artifact = "Bench.EffectStep",
+  n = 100000,
+  drive = function(mod, n)
+    return mod.run(n)
+  end,
+  ideal = function(n)
+    local acc = 0
+    for _ = 1, n do acc = acc + 15 end
+    return acc
+  end,
+}
diff --git a/changelog.d/20260713_160000_unisay_late_uncurry_rerun.md b/changelog.d/20260713_160000_unisay_late_uncurry_rerun.md
@@ -0,0 +1,13 @@
+### Changed
+
+- The uncurrying worker/wrapper split (#24) now runs a second time at the end
+  of the IR pipeline, catching the two saturated-call families the early run
+  cannot see: effect functions whose thunk-forcing application only appears
+  once magic-do has run — the worker absorbs the thunk parameter, so a fully
+  applied effect statement compiles to one n-ary call instead of a curried
+  call plus a thunk allocation and force — and the `$kont` continuation
+  helpers minted by `flattenDeepBinds`, saturated by construction. The pass is
+  rerun-safe (wrappers built by the early run are recognised and reused, never
+  split again), effect statements keep their dead-code protection through the
+  rewrite (`isEffectRun` recognises the absorbed trailing marker), and a
+  closing dce pass drops the wrappers whose sites were all saturated (#200).
diff --git a/lib/Language/PureScript/Backend/IR/Optimizer.hs b/lib/Language/PureScript/Backend/IR/Optimizer.hs
@@ -112,8 +112,9 @@ optimizerPipeline policy =
   , RunFixpoint "optimize+dce-post-merge" (optimizePass :| [dcePass])
   , -- Split curried bindings into n-ary workers and curried wrappers
     -- and rewrite the saturated call sites to direct worker calls
-    -- (issue #24). Runs after the post-merge fixpoint, so manifest
-    -- arities are measured once inlining has settled. See
+    -- (issue #24) — the early of the pass's two runs (the late one
+    -- closes the pipeline). Runs after the post-merge fixpoint, so
+    -- manifest arities are measured once inlining has settled. See
     -- Language.PureScript.Backend.IR.Uncurry.
     RunPass uncurryPass
   , -- The post-uncurry fixpoint dead-code-eliminates wrappers with no
@@ -174,6 +175,21 @@ optimizerPipeline policy =
     -- likewise consumes and preserves the unique naming.
     -- See Language.PureScript.Backend.IR.FlattenDeepBinds.
     RunPass flattenDeepBindsPass
+  , -- The late uncurry run (issue #200): the same pass again, now that
+    -- the two saturated-site families invisible to the early run exist —
+    -- the effect-run spines magicDo completed (f(a)(b)(run), saturating
+    -- the effect function's manifest chain, thunk parameter included)
+    -- and the saturated-by-construction $kont helpers minted by the
+    -- flattening above. Bindings the early run split are recognised and
+    -- left split (see the Rerun section in
+    -- Language.PureScript.Backend.IR.Uncurry).
+    RunPass uncurryLatePass
+  , -- Drop the wrappers the late run left unreferenced. A single dce
+    -- pass, deliberately not an optimize+dce fixpoint like the early
+    -- run's: optimize's use-once inlining would paste the $kont workers
+    -- (each is called exactly once) back into their call sites round by
+    -- round, re-nesting exactly what flattenDeepBinds just flattened.
+    RunPass dcePass
   ]
  where
   uniquifyPass =
@@ -222,6 +238,7 @@ optimizerPipeline policy =
       , passRequires = guc
       , passEnsures = guc
       }
+  uncurryLatePass = uncurryPass {passName = "uncurry-late"}
   floatInPass = gucPass "float-in" floatIn
   shareAccessorsPass =
     gucPass "share-accessors" (shareForeignAccessors policy)

diff --git a/lib/Language/PureScript/Backend/IR/Types.hs b/lib/Language/PureScript/Backend/IR/Types.hs
@@ -445,9 +445,13 @@ pattern EffectRunArg ∷ ann → RawExp ann
 pattern EffectRunArg ann =
   Ref ann (Imported (ModuleName "Prim") (Name "$magicDoRun"))
 
-{- | Recognise an Effect/ST statement as magic-do emits it: running a thunk, the
-application @m EffectRunArg@ (see 'Language.PureScript.Backend.IR.MagicDo'). Its
-side effect is observable and its statement sequencing is size-managed by
+{- | Recognise an Effect/ST statement by its trailing 'EffectRunArg': running a
+thunk, the application @m EffectRunArg@ as magic-do emits it (see
+'Language.PureScript.Backend.IR.MagicDo'), or the n-ary worker call that
+absorbed the run — the late uncurry rerun rewrites the saturated spine
+@f(a)(b)(EffectRunArg)@ to @f$w(a, b, EffectRunArg)@ (issue #200). The marker
+never occurs in a non-trailing position, so a trailing one is precise. The
+statement's side effect is observable and its sequencing is size-managed by
 magic-do's chunking, so passes that run after magic-do must leave it alone:
 dead-code elimination keeps it even when its binder is unreferenced, beta
 reduction does not reduce through it (which would merge a chunk into its parent
@@ -457,7 +461,7 @@ copy would execute the effect twice).
 -}
 isEffectRun ∷ RawExp ann → Bool
 isEffectRun = \case
-  AppN _ _ (EffectRunArg _ :| []) → True
+  AppN _ _ args | EffectRunArg _ ← last args → True
   _ → False
 
 ctorId ∷ ModuleName → TyName → CtorName → Text