The last leg of the #63 drop-soundness cluster: an (own ...) heap param consumed on one match-arm / if-branch but live on a sibling was never dropped on the live path, leaking one slab per call. This was the residual live=2 leak in series_sma that kept the series headline from being leak-clean. Root: the per-fn aggregate consume_count (the worst-case max over all branches from uniqueness::merge_states) was codegen's only consume signal, and every owned-param drop site gated on it. A param consumed on SOME branch makes the aggregate >= 1, so every site skipped it — including the branch where it stays live. match and if share this root (if is first-class MTerm::If, not desugared to match). Mechanism (spec 0068): - CAPTURE: uniqueness retains the per-branch consume snapshots it already computes per branch and discarded at the max merge — new BranchConsume + infer_module_with_cross_branches (the aggregate-only infer_module_with_cross now delegates and drops the channel). - CARRY: additive MArm.consume / MTerm::If.{then,else}_consume, attached by lower_to_mir via a traversal-order cursor (post-order pop, kind- and exhaustion-asserts make a desync a loud panic; neutral for const bodies which uniqueness does not walk). The AST carries no node id, so the correspondence is the structural pre-order both walks share — pinned by branch_consume_maps_attach_to_matching_arms. - GATE: emit_leakclass_branch_param_drops fires a fall-through drop only for the leak class (branch_consume==0 AND aggregate>=1), in match arms + both if branches; the pre-tail-call dec switches its gate source from the aggregate to per-arm. The fn-return dec and arm-close pattern-binder dec are UNCHANGED. Safety: - Double-free: the drop sites partition by aggregate (==0 -> existing fn-return/pre-tail-call; >=1 -> new per-branch). Disjoint, so no param is dropped twice — no fn-return disable, no tail-position analysis. - Use-after-free: the checker rejects use-after-consume, so an aggregate>=1 param is provably dead past the construct (path-terminal drop, tail position irrelevant). - Type eligibility (found by the full-suite gate during implement; spec 0068 refined): the new agg>=1 site is the FIRST drop site that can reach a STATIC closure-pair param (a top-level fn ref like inc in Either.either, consumed on one arm, live on the other) — a .rodata constant with no rc-header. field_drop_call routes Type::Fn / static-Str / Type::Var to the bare ailang_rc_dec, which underflowed on it. The helper now drops only params with a real per-type heap-ADT drop fn (field_drop_call != "ailang_rc_dec"). Sound (no underflow) and leak-correct (a static closure/Str allocates nothing). Witnessed green by std_either_demo / std_list_demo / poly_rec_capture_demo. Verification: full workspace suite green (116 binaries); both new leak pins (if + match) and series_sma_no_leak_pin green at live=0; legs 1/2 pins still green; INTERCEPTS<->(intrinsic) bijection intact; no Pattern::Lit reject path added. The leg-3 helper's type precondition was applied inline by the orchestrator (a narrowing guard clause in an already-reviewed Task-4 helper, full context loaded) after the implement-orchestrator correctly surfaced the spec gap rather than papering over the regression. This clears the series_sma leak tail; the series milestone (#61) close stays a separate deliberate step (its end-to-end milestone fieldtest). closes #63
38 KiB
Branch-Consume-Split Drop Soundness (#63 leg 3) — Implementation Plan
Parent spec:
docs/specs/0068-branch-consume-split-drop.mdFor agentic workers: REQUIRED SUB-SKILL: use the
implementskill to run this plan. Steps use- [ ]checkboxes for tracking.
Goal: Drop an (own ...) heap param exactly once per control-flow
path when it is consumed on one match-arm / if-branch but stays live
on a sibling — closing the last leg of #63 and the residual live=2
leak in series_sma (#61).
Architecture: The uniqueness pass already computes a per-branch
consume snapshot for every if/match branch and discards it at the
max merge in merge_states. Retain those snapshots, carry them
additively on MArm.consume / MTerm::If.{then,else}_consume
(attached by lower_to_mir in pre-order traversal-order lock-step with
the uniqueness walk, since the AST carries no node id), and let codegen
emit a NEW fall-through drop for the leak class — branch_consume[p] == 0 && aggregate >= 1. The drop sites partition by aggregate consume
(== 0 → existing fn-return/pre-tail-call machinery; >= 1 →
per-branch), so no param is dropped twice and no tail-position analysis
is needed; the checker's use-after-consume rejection guarantees an
aggregate >= 1 param is dead past the construct, so the per-branch
drop is path-terminal.
Tech Stack: ailang-check (uniqueness capture + MIR lowering
cursor), ailang-mir (additive branch-consume fields), ailang-codegen
(leak-class drop helper + gate-source switch), ail integration pins.
Files this plan creates or modifies:
- Create:
examples/if_branch_consume_split_no_leak_pin.ail— new RED fixture (if-branch split). - Create:
crates/ail/tests/if_branch_consume_split_no_leak_pin.rs— new RED pin,live==0. - Create:
crates/ail/tests/series_sma_no_leak_pin.rs— #61 close pin,series_smalive==0. - Modify:
crates/ail/tests/match_arm_consume_split_no_leak_pin.rs:74— remove#[ignore]. - Modify:
crates/ailang-check/src/uniqueness.rs—BranchConsumeenum,Walker.branch_consumes, capture at If/Match,UniquenessResult+infer_module_with_cross_branches. - Modify:
crates/ailang-mir/src/lib.rs:91,116— additive fields onMTerm::IfandMArm. - Modify:
crates/ailang-check/src/lower_to_mir.rs—Ctxcursor fields, attach at If/Match, switch to the branches API, exhaustion guard. - Modify:
crates/ailang-codegen/src/escape.rs:612,737,747— thread the newMArmfield in test-mod constructions. - Modify:
crates/ailang-codegen/src/match_lower.rs:823,~965,~1009—:823gate-source switch; new fall-through drop in ctor-arms loop + default/open arm. - Modify:
crates/ailang-codegen/src/lib.rs:1906—MTerm::Ifdestructure + per-branch drop in both branches; add theemit_leakclass_branch_param_dropshelper. - Test:
crates/ailang-check/tests/lower_to_mir_ty.rs— lock-step correspondence test.
Task 1: RED — if-branch fixture + pin, un-ignore match pin
Files:
-
Create:
examples/if_branch_consume_split_no_leak_pin.ail -
Create:
crates/ail/tests/if_branch_consume_split_no_leak_pin.rs -
Modify:
crates/ail/tests/match_arm_consume_split_no_leak_pin.rs:74 -
Step 1: Write the if-branch RED fixture
Create examples/if_branch_consume_split_no_leak_pin.ail verbatim (this is
the spec's Concrete-code block 2; ail check passes today with only an
over-strict-mode warning on sink):
(module if_branch_consume_split_no_leak_pin
(data Box (ctor B (con Int)))
(fn sink
(type (fn-type (params (own (con Box))) (ret (own (con Unit))) (effects IO)))
(params b)
(body (match b (case (pat-ctor B n) (do io/print_str "x")))))
(fn run_if
(type (fn-type (params (own (con Box)) (own (con Int))) (ret (own (con Unit))) (effects IO)))
(params b k)
(body
(if (app ge k 0)
(do io/print_str "")
(app sink b))))
(fn main
(type (fn-type (params) (ret (own (con Unit))) (effects IO)))
(params)
(body
(let b (term-ctor Box B 7)
(seq
(app run_if b 0)
(do io/print_str "done\n"))))))
- Step 2: Write the if-branch RED pin
Create crates/ail/tests/if_branch_consume_split_no_leak_pin.rs. This is a
structural copy of the committed match_arm_consume_split_no_leak_pin.rs
(same build/run/parse-stats body), with the fixture name swapped and NO
#[ignore]. Header doc explains: b is consumed on the else branch
(sink b), live on the then branch (ge 0 0 true → then runs), so
b stays live and must be dropped on the then branch:
//! RED-pin for leg 3 of bug #63 — the per-branch-vs-per-fn consume-
//! accounting leak at an `if`-branch fall-through (the `if` half of the
//! branch-consume-split class; the `match` half is
//! match_arm_consume_split_no_leak_pin.rs).
//!
//! Property protected: under `--alloc=rc`, when an `(own ...)` heap
//! param is consumed on ONE `if` branch but is LIVE on the sibling
//! branch, it MUST be dropped on the live branch before the fn returns.
//! `run_if b 0` takes the `then` branch (`ge 0 0` is true) where `b` is
//! live; the `else` branch (`sink b`) consumes it. As of HEAD this fails:
//! the binary reports `live=1` because the per-fn aggregate consume of
//! `b` is `>= 1` (the `else` branch consumes it) so the fn-return dec
//! skips `b` and nothing drops it on the `then` path.
//!
//! `if` is a first-class `MTerm::If` (it does NOT desugar to `match`),
//! so it needs the per-branch drop in its own codegen node. Goes green
//! with the #63-leg-3 fix.
use std::path::Path;
use std::process::Command;
fn ail_bin() -> &'static str {
env!("CARGO_BIN_EXE_ail")
}
#[test]
fn alloc_rc_owned_param_consumed_on_one_if_branch_live_on_sibling_does_not_leak() {
let manifest_dir = env!("CARGO_MANIFEST_DIR");
let workspace = Path::new(manifest_dir).parent().unwrap().parent().unwrap();
let src = workspace
.join("examples")
.join("if_branch_consume_split_no_leak_pin.ail");
let tmp = std::env::temp_dir().join(format!(
"ailang_if_branch_consume_split_no_leak_pin_{}",
std::process::id()
));
std::fs::create_dir_all(&tmp).unwrap();
let out = tmp.join("bin");
let status = Command::new(ail_bin())
.args(["build", src.to_str().unwrap(), "--alloc=rc", "-o"])
.arg(&out)
.status()
.expect("ail build failed to run");
assert!(
status.success(),
"ail build --alloc=rc failed for if_branch_consume_split_no_leak_pin.ail"
);
let output = Command::new(&out)
.env("AILANG_RC_STATS", "1")
.output()
.expect("execute binary");
assert!(
output.status.success(),
"binary exited non-zero: status {:?}",
output.status
);
let stderr = String::from_utf8(output.stderr).expect("stderr utf8");
let stats_line = stderr
.lines()
.find(|l| l.starts_with("ailang_rc_stats:"))
.unwrap_or_else(|| {
panic!("missing ailang_rc_stats line in stderr; stderr was:\n{stderr}")
});
let mut allocs: Option<u64> = None;
let mut frees: Option<u64> = None;
let mut live: Option<i64> = None;
for tok in stats_line.split_whitespace() {
if let Some(v) = tok.strip_prefix("allocs=") {
allocs = v.parse().ok();
} else if let Some(v) = tok.strip_prefix("frees=") {
frees = v.parse().ok();
} else if let Some(v) = tok.strip_prefix("live=") {
live = v.parse().ok();
}
}
let allocs = allocs.expect("missing allocs= field");
let frees = frees.expect("missing frees= field");
let live = live.expect("missing live= field");
assert_eq!(
live, 0,
"an (own) heap param consumed on one if branch but live on the \
sibling branch leaks {live} slab(s) (allocs={allocs} frees={frees}): \
the per-branch live param is never dropped because the fn-return \
dec gates on the per-fn aggregate consume (>= 1 from the consuming \
branch) instead of the per-branch consume."
);
assert_eq!(
allocs, frees,
"alloc/free mismatch (allocs={allocs} frees={frees} live={live})"
);
}
- Step 3: Un-ignore the committed match pin
In crates/ail/tests/match_arm_consume_split_no_leak_pin.rs, delete line 74:
#[ignore = "leg 3 of #63 — design fork (per-arm consume dataflow); un-ignore when fixed"]
(Leave the #[test] and everything else intact.)
- Step 4: Run the two leak pins — expect RED
Run: cargo test -p ail --test if_branch_consume_split_no_leak_pin --test match_arm_consume_split_no_leak_pin
Expected: BOTH FAIL with live=1 (assertion left: 1, right: 0). This is
the RED baseline — the leak is reproduced for both constructs.
- Step 5: Confirm legs 1/2 still pass (regression baseline)
Run: cargo test -p ail --test tail_recur_owned_param_no_leak_pin --test tail_recur_let_wrapped_no_leak_pin
Expected: BOTH PASS (these are unaffected by Task 1; they are the
behaviour-preservation baseline for Task 4).
Task 2: CAPTURE — retain per-branch consume in the uniqueness pass
Files:
-
Modify:
crates/ailang-check/src/uniqueness.rs -
Step 1: Add the
BranchConsumetype and a snapshot helper
After the BinderState struct (around line 248) add:
/// Per-branch consume snapshots retained for codegen's leak-class drop
/// siting (#63 leg 3). Each map is the absolute per-binder
/// `consume_count` at a branch's end-state — the value the `max` merge
/// in `merge_states` would otherwise collapse. The pass emits one entry
/// per `if`/`match` construct in PRE-ORDER of branch constructs (the
/// same order `lower_to_mir` lowers them), so `lower_to_mir` correlates
/// entries to MIR branch nodes by a traversal-order cursor (the AST
/// carries no node id).
#[derive(Debug, Clone)]
pub enum BranchConsume {
If {
then: BTreeMap<String, u32>,
else_: BTreeMap<String, u32>,
},
Match {
arms: Vec<BTreeMap<String, u32>>,
},
}
fn snapshot_consume(binders: &BTreeMap<String, BinderState>) -> BTreeMap<String, u32> {
binders
.iter()
.map(|(k, v)| (k.clone(), v.consume_count))
.collect()
}
- Step 2: Add the accumulator field to
Walker
In struct Walker<'a> (around line 250) add a field after out:
/// Per-branch consume snapshots, accumulated in pre-order during the
/// walk (#63 leg 3). Drained per fn by `infer_fn`.
branch_consumes: Vec<BranchConsume>,
- Step 3: Capture at
Term::If
Replace the Term::If arm (currently lines 297-304) with:
Term::If { cond, then, else_ } => {
self.walk(cond, Position::Consume);
let saved = self.binders.clone();
self.walk(then, pos);
let after_then = std::mem::replace(&mut self.binders, saved);
self.walk(else_, pos);
// #63 leg 3: retain both branch end-states (pre-order push
// AFTER walking both children, so nested branches inside
// then/else are already pushed — matches lower_to_mir's
// post-order pop).
self.branch_consumes.push(BranchConsume::If {
then: snapshot_consume(&after_then),
else_: snapshot_consume(&self.binders),
});
merge_states(&mut self.binders, &after_then);
}
- Step 4: Capture at
Term::Match
Replace the Term::Match arm (currently lines 305-322) with:
Term::Match { scrutinee, arms } => {
self.walk(scrutinee, Position::Borrow);
let saved = self.binders.clone();
let mut acc: Option<BTreeMap<String, BinderState>> = None;
let mut arm_maps: Vec<BTreeMap<String, u32>> =
Vec::with_capacity(arms.len());
for arm in arms {
self.binders = saved.clone();
self.walk_arm(arm, pos);
// #63 leg 3: snapshot this arm's end-state (pattern
// binders already popped by walk_arm) before the merge.
arm_maps.push(snapshot_consume(&self.binders));
match acc.as_mut() {
None => acc = Some(self.binders.clone()),
Some(m) => merge_states(m, &self.binders),
}
}
// pre-order push AFTER all arms (nested arm-body branches
// already pushed) — matches lower_to_mir's post-order pop.
self.branch_consumes
.push(BranchConsume::Match { arms: arm_maps });
if let Some(merged) = acc {
self.binders = merged;
} else {
self.binders = saved;
}
}
- Step 5: Drain the accumulator per fn in
infer_fn
infer_fn gains a branches_out param and drains walker.branch_consumes.
Change the signature (line 181-186) to:
fn infer_fn(
f: &FnDef,
globals: &BTreeMap<String, Type>,
cross_module_types: &BTreeMap<String, BTreeMap<String, Type>>,
out: &mut UniquenessTable,
branches_out: &mut BTreeMap<String, Vec<BranchConsume>>,
) {
Replace the param_states block (lines 192-214) so it also returns the
drained branch vec, and init the new Walker field:
let (param_states, branch_consumes) = {
let mut walker = Walker {
globals,
cross_module_types,
binders: BTreeMap::new(),
def_name: f.name.clone(),
out,
branch_consumes: Vec::new(),
};
// Install fn parameters. Their consume counter starts at 0;
// the walk over the body increments it.
for name in f.params.iter() {
walker.binders.insert(name.clone(), BinderState::default());
}
walker.walk(&f.body, Position::Consume);
// Snapshot params before dropping the walker.
let ps = f
.params
.iter()
.filter_map(|name| walker.binders.get(name).map(|s| (name.clone(), s.clone())))
.collect::<Vec<_>>();
(ps, walker.branch_consumes)
};
Then after the existing for (name, s) in param_states { ... } loop (ends
line 223), append:
branches_out.insert(f.name.clone(), branch_consumes);
- Step 6: Add
UniquenessResult+infer_module_with_cross_branches; delegate the old API
Add the result type near UniquenessTable (after line 100):
/// The full uniqueness output: the per-`(def, binder)` aggregate table
/// (codegen's classification + scope-close consume) AND the per-fn
/// ordered list of per-branch consume snapshots (#63 leg 3).
pub struct UniquenessResult {
pub table: UniquenessTable,
pub branches: BTreeMap<String, Vec<BranchConsume>>,
}
Rename the current infer_module_with_cross body to
infer_module_with_cross_branches returning UniquenessResult, and make
infer_module_with_cross a thin delegate. Concretely, change the
signature line (133-136) to:
pub fn infer_module_with_cross_branches(
m: &Module,
cross_module_types: &BTreeMap<String, BTreeMap<String, Type>>,
) -> UniquenessResult {
Replace the table-build tail (lines 170-177) with:
let mut table: UniquenessTable = BTreeMap::new();
let mut branches: BTreeMap<String, Vec<BranchConsume>> = BTreeMap::new();
for def in &m.defs {
if let Def::Fn(f) = def {
infer_fn(f, &globals, cross_module_types, &mut table, &mut branches);
}
}
UniquenessResult { table, branches }
}
/// Aggregate-only convenience used by existing callers (codegen tests,
/// etc.). Delegates to [`infer_module_with_cross_branches`] and drops the
/// per-branch channel.
pub fn infer_module_with_cross(
m: &Module,
cross_module_types: &BTreeMap<String, BTreeMap<String, Type>>,
) -> UniquenessTable {
infer_module_with_cross_branches(m, cross_module_types).table
}
- Step 7: Build the crate, confirm existing uniqueness tests stay green
Run: cargo build -p ailang-check && cargo test -p ailang-check
Expected: PASS (the change is purely additive — the new channel is
produced but not yet consumed; the aggregate table is unchanged, so every
existing test is unaffected).
Task 3: CARRY — additive MIR fields + lower_to_mir traversal-order cursor
Files:
-
Modify:
crates/ailang-mir/src/lib.rs:91,116 -
Modify:
crates/ailang-codegen/src/escape.rs:612,737,747 -
Modify:
crates/ailang-check/src/lower_to_mir.rs -
Test:
crates/ailang-check/tests/lower_to_mir_ty.rs -
Step 1: Add the additive fields on
MArmandMTerm::If
In crates/ailang-mir/src/lib.rs, change MTerm::If (line 91) to:
If {
cond: Box<MTerm>,
then: Box<MTerm>,
else_: Box<MTerm>,
/// #63 leg 3: per-branch consume of each owned binder on the
/// `then` / `else` branch (absolute count at the branch end).
/// Codegen drops a param live here (`0`) but consumed on the
/// sibling (`MirDef.consume >= 1`). Empty for const bodies.
then_consume: BTreeMap<String, u32>,
else_consume: BTreeMap<String, u32>,
ty: Type,
},
Change struct MArm (lines 116-119) to:
pub struct MArm {
pub pat: Pattern,
pub body: MTerm,
/// #63 leg 3: per-arm consume of each owned binder (absolute count at
/// this arm's end). See `MTerm::If.then_consume`. Empty for const bodies.
pub consume: BTreeMap<String, u32>,
}
(BTreeMap is already imported in this file — MirDef.consume uses it.)
- Step 2: Build the workspace to enumerate the break sites
Run: cargo build --workspace 2>&1 | grep -E "error\[|missing field" | head -30
Expected: errors at the literal-construction sites only — lower_to_mir.rs
(the MTerm::If at ~326 and MArm at ~381) and the three test-mod MArm
literals in escape.rs (~612, ~737, ~747). All ..-pattern read sites
compile clean. (This step just confirms the break set before threading.)
- Step 3: Thread the
escape.rstest-modMArmconstructions
In crates/ailang-codegen/src/escape.rs, the three MArm { pat, body }
literals (around lines 612, 737, 747) gain an empty consume map. For each,
add the field:
consume: std::collections::BTreeMap::new(),
(These are #[cfg(test)] fixtures; an empty map is correct — the escape
tests do not exercise branch-consume drops.)
- Step 4: Add the cursor fields to
Ctxand a pop helper
In crates/ailang-check/src/lower_to_mir.rs, add to the Ctx struct (the
struct holding env, in_def, locals, loop_stack) these fields:
/// #63 leg 3: per-branch consume snapshots for the current def, in the
/// uniqueness pass's pre-order. Consumed by a traversal-order cursor
/// as `lower_term` lowers each `if`/`match` (post-order pop). Empty +
/// `track_branches=false` for const bodies (uniqueness walks only fns).
branch_consumes: Vec<crate::uniqueness::BranchConsume>,
branch_idx: usize,
track_branches: bool,
Add an impl block with the two pop helpers (place near the other Ctx
methods):
impl Ctx<'_> {
/// Pop the next `If` branch-consume entry (post-order: called AFTER
/// lowering cond/then/else). Returns empty maps for const bodies. A
/// kind/cursor mismatch is a compiler-internal desync — panic loudly
/// rather than silently mis-attach a drop map.
fn next_branch_if(
&mut self,
) -> (std::collections::BTreeMap<String, u32>, std::collections::BTreeMap<String, u32>) {
use crate::uniqueness::BranchConsume;
if !self.track_branches {
return (std::collections::BTreeMap::new(), std::collections::BTreeMap::new());
}
match self.branch_consumes.get(self.branch_idx) {
Some(BranchConsume::If { then, else_ }) => {
let r = (then.clone(), else_.clone());
self.branch_idx += 1;
r
}
other => panic!(
"lower_to_mir branch-consume desync in `{}` at idx {}: expected If, got {}",
self.in_def,
self.branch_idx,
match other {
Some(BranchConsume::Match { .. }) => "Match",
Some(BranchConsume::If { .. }) => "If",
None => "end-of-list",
}
),
}
}
/// Pop the next `Match` branch-consume entry (post-order: called AFTER
/// lowering scrutinee + all arm bodies). Returns `n_arms` empty maps
/// for const bodies. Arm-count / kind mismatch is a desync — panic.
fn next_branch_match(
&mut self,
n_arms: usize,
) -> Vec<std::collections::BTreeMap<String, u32>> {
use crate::uniqueness::BranchConsume;
if !self.track_branches {
return vec![std::collections::BTreeMap::new(); n_arms];
}
match self.branch_consumes.get(self.branch_idx) {
Some(BranchConsume::Match { arms }) if arms.len() == n_arms => {
let r = arms.clone();
self.branch_idx += 1;
r
}
other => panic!(
"lower_to_mir branch-consume desync in `{}` at idx {}: expected Match[{}], got {}",
self.in_def,
self.branch_idx,
n_arms,
match other {
Some(BranchConsume::Match { arms }) => format!("Match[{}]", arms.len()),
Some(BranchConsume::If { .. }) => "If".into(),
None => "end-of-list".into(),
}
),
}
}
}
- Step 5: Attach per-branch maps at
Term::Iflowering
Replace the Term::If lowering (lines 326-331) with (lower children
first, then pop — post-order, matching the uniqueness push order):
Term::If { cond, then, else_ } => {
let m_cond = lower_term(ctx, cond)?;
let m_then = lower_term(ctx, then)?;
let m_else = lower_term(ctx, else_)?;
let (then_consume, else_consume) = ctx.next_branch_if();
MTerm::If {
cond: Box::new(m_cond),
then: Box::new(m_then),
else_: Box::new(m_else),
then_consume,
else_consume,
ty,
}
}
- Step 6: Attach per-arm maps at
Term::Matchlowering
In the Term::Match lowering (lines 359-383), build each MArm with an
empty consume, then fill from the popped entry after the arm loop. Change
the m_arms.push(...) line (381) to:
m_arms.push(MArm {
pat: a.pat.clone(),
body: m_body,
consume: BTreeMap::new(),
});
and insert, between the for a in arms { ... } loop close (382) and the
MTerm::Match { ... } construction (383):
let arm_maps = ctx.next_branch_match(arms.len());
for (marm, cmap) in m_arms.iter_mut().zip(arm_maps) {
marm.consume = cmap;
}
(Ensure use std::collections::BTreeMap; is in scope in this file; if not,
qualify as std::collections::BTreeMap::new().)
- Step 7: Switch
lower_moduleto the branches API + seed the cursor + exhaustion guard
In lower_module, change the uniqueness call (lines 602-603) to:
let result =
crate::uniqueness::infer_module_with_cross_branches(module, cross_module_types);
In the const-lowering Ctx construction (around line 612-617) add the new
fields (const bodies are not walked by uniqueness → neutral cursor):
let mut ctx = Ctx {
env: &env,
in_def: &c.name,
locals: IndexMap::new(),
loop_stack: Vec::new(),
branch_consumes: Vec::new(),
branch_idx: 0,
track_branches: false,
};
In the fn-lowering Ctx construction (around line 655-660) add:
let mut ctx = Ctx {
env: &env,
in_def: &f.name,
locals,
loop_stack: Vec::new(),
branch_consumes: result.branches.get(&f.name).cloned().unwrap_or_default(),
branch_idx: 0,
track_branches: true,
};
let body = lower_term(&mut ctx, &f.body)?;
assert!(
ctx.branch_idx == ctx.branch_consumes.len(),
"lower_to_mir branch-consume cursor desync for `{}`: consumed {} of {} entries",
f.name,
ctx.branch_idx,
ctx.branch_consumes.len()
);
(The let body = lower_term(&mut ctx, &f.body)?; already exists at line
661 — replace it with the version above that adds the trailing assert!.)
Finally, change the MirDef.consume fill (lines 667-671) to read
result.table instead of the old uniqueness binding:
consume: result
.table
.iter()
.filter(|((d, _), _)| d == &f.name)
.map(|((_, b), info)| (b.clone(), info.consume_count))
.collect(),
- Step 8: Build the workspace
Run: cargo build --workspace
Expected: 0 errors. (Codegen still reads the aggregate MirDef.consume;
the new per-branch fields are populated but not yet consumed by codegen —
so behaviour is unchanged and the leak pins stay RED.)
- Step 9: Add the lock-step correspondence test
Append to crates/ailang-check/tests/lower_to_mir_ty.rs a test that
elaborates the existing match_arm_consume_split_no_leak_pin fixture and
asserts the per-arm maps attach to the correct arms (guards the
traversal-order cursor). Use the file's existing elaborate_fixture and
body helpers — do NOT invent a new API. The fixture's run body is
directly a (match f ...) with arms in source order: Off (b live) at
arms[0], On (consumes b via sink) at arms[1].
/// #63 leg 3 lock-step: the per-arm consume map attaches to the matching
/// MArm in traversal order. In the `match_arm_consume_split_no_leak_pin`
/// fixture's `run`, `b` is consumed on the `On` arm (forwarded into
/// `sink`) and live on the `Off` arm — so traversal-order attachment must
/// land MArm.consume[b] == 0 on `Off` (source-order arms[0]) and >= 1 on
/// `On` (arms[1]). A cursor desync would mis-attach and flip these.
#[test]
fn branch_consume_maps_attach_to_matching_arms() {
let m = "match_arm_consume_split_no_leak_pin";
let mir = elaborate_fixture(m);
let arms = match body(&mir, m, "run") {
MTerm::Match { arms, .. } => arms,
other => panic!("run body is not a Match: {other:?}"),
};
assert_eq!(
arms[0].consume.get("b").copied().unwrap_or(99),
0,
"Off arm (arms[0]) must leave `b` live (consume 0); got {:?}",
arms[0].consume.get("b")
);
assert!(
arms[1].consume.get("b").copied().unwrap_or(0) >= 1,
"On arm (arms[1]) must consume `b` (>= 1); got {:?}",
arms[1].consume.get("b")
);
}
- Step 10: Run the correspondence test
Run: cargo test -p ailang-check --test lower_to_mir_ty branch_consume_maps_attach_to_matching_arms
Expected: PASS.
Task 4: GATE — codegen leak-class drop (turns the RED pins GREEN)
Files:
-
Modify:
crates/ailang-codegen/src/lib.rs(helper +MTerm::Ifbranches) -
Modify:
crates/ailang-codegen/src/match_lower.rs(:823switch + fall-through drops) -
Step 1: Add the
emit_leakclass_branch_param_dropshelper
Add this method to the codegen impl in crates/ailang-codegen/src/lib.rs
(near the other drop-emitting helpers — it reuses field_drop_call,
partial_drop_symbol_for_type, build_moved_mask, moved_slots,
current_param_modes, current_def, consume, locals, alloc):
/// #63 leg 3: drop owned heap params that are LIVE on this branch
/// (`branch_consume[p] == 0`) but CONSUMED on a sibling branch (per-fn
/// aggregate `self.consume[p] >= 1`) — the leak class. The
/// `aggregate >= 1` guard keeps this disjoint from the fn-return dec
/// (which owns the `aggregate == 0` class), so no param is double-
/// dropped — no fn-return disable / tail-position analysis is needed.
/// And because the checker rejects use-after-consume, an
/// `aggregate >= 1` param is provably dead past the branch construct,
/// so dropping it at the branch close is safe regardless of tail
/// position. `tail_val` is the branch's own tail SSA (a param returned
/// by the branch transfers ownership to the join consumer — skip it);
/// `scrutinee` is the match scrutinee binder to skip (None for `if`).
fn emit_leakclass_branch_param_drops(
&mut self,
branch_consume: &std::collections::BTreeMap<String, u32>,
tail_val: &str,
scrutinee: Option<&str>,
) {
if !matches!(self.alloc, AllocStrategy::Rc) {
return;
}
let owned_params: Vec<(String, ParamMode)> = self
.current_param_modes
.iter()
.map(|(n, m)| (n.clone(), *m))
.collect();
for (pname, mode) in &owned_params {
if !matches!(mode, ParamMode::Own) {
continue;
}
if scrutinee == Some(pname.as_str()) {
continue;
}
let branch_cc = branch_consume.get(pname).copied().unwrap_or(u32::MAX);
if branch_cc != 0 {
continue; // consumed on this branch — ownership transferred
}
let agg = self
.consume
.get(&(self.current_def.clone(), pname.clone()))
.copied()
.unwrap_or(0);
if agg == 0 {
continue; // live on every path — the fn-return dec owns this
}
let p_ssa = format!("%arg_{}", pname);
if p_ssa == tail_val {
continue; // param is this branch's return value
}
let p_ail = self
.locals
.iter()
.find(|(_, ssa, _, _)| ssa == &p_ssa)
.map(|(_, _, lty, ail)| (lty.clone(), ail.clone()));
let (p_lty, p_ail) = match p_ail {
Some(t) => t,
None => continue,
};
if p_lty != "ptr" {
continue;
}
// #63 leg 3 TYPE precondition: only drop params with a real
// per-type heap-ADT drop fn. `field_drop_call` routes
// `Type::Fn` (static OR heap closures), static-`Str`, and
// `Type::Var` to the bare `ailang_rc_dec`, which reads a
// non-existent rc-header on a `.rodata` static closure/Str
// and underflows. This `agg >= 1` site is the FIRST drop site
// that can reach a static closure param (a called closure has
// `agg >= 1`; every existing site needs `agg == 0`), so it
// must gate on a real drop fn. Sound + leak-correct: a static
// closure / static `Str` allocates nothing to drop.
let drop_call = self.field_drop_call(&p_ail);
if drop_call == "ailang_rc_dec" {
continue;
}
let moves = self.moved_slots.get(pname).cloned().unwrap_or_default();
if moves.is_empty() {
self.body
.push_str(&format!(" call void @{drop_call}(ptr {p_ssa})\n"));
} else {
let sym = self.partial_drop_symbol_for_type(&p_ail);
let mask = Self::build_moved_mask(&moves);
if let (Some(sym), Some(mask)) = (sym, mask) {
self.body
.push_str(&format!(" call void @{sym}(ptr {p_ssa}, i64 {mask})\n"));
} else {
self.body
.push_str(&format!(" call void @ailang_rc_dec(ptr {p_ssa})\n"));
}
}
}
}
- Step 2: Switch the pre-tail-call dec gate source to per-arm
In crates/ailang-codegen/src/match_lower.rs, replace the aggregate lookup
in the pre-tail-call Own-param dec (lines 823-827):
let consume_count = self
.consume
.get(&(self.current_def.clone(), pname.clone()))
.copied()
.unwrap_or(u32::MAX);
with the per-arm lookup (arm is the ctor-arms loop variable, in scope
here):
let consume_count = arm.consume.get(pname).copied().unwrap_or(u32::MAX);
(For an aggregate == 0 param — legs 1/2 — arm.consume[p] is 0 on
every arm, so this is the identical decision. It additionally drops a
param live on this tail-call arm but consumed on a sibling.)
- Step 3: Add the fall-through drop in the ctor-arms loop
In lower_match, after the arm-close pattern-binder dec block closes
(immediately before the // pop bindings comment / the for _ in 0..pushed { self.locals.pop(); } loop, around line 966), insert:
// #63 leg 3: a fall-through (non-tail-call) arm that left an
// owned param live (arm.consume == 0) but consumed it on a
// sibling arm (aggregate >= 1) must drop it here, before the
// br to the join. The aggregate >= 1 guard inside the helper
// keeps this disjoint from the fn-return dec.
if !self.block_terminated {
self.emit_leakclass_branch_param_drops(
&arm.consume,
&val,
scrutinee_binder.as_deref(),
);
}
- Step 4: Add the fall-through drop in the default/open arm
In the default-block handling, inside if let Some(arm) = open_arm { ... },
after let (val, vty) = self.lower_term(&arm.body)?; and the
for _ in 0..pushed { self.locals.pop(); } (around line 1009), before the
if !self.block_terminated { phi_inputs.push(...) ... } block (line 1010),
insert:
// #63 leg 3: same leak-class drop for a Wild/Var default arm.
if !self.block_terminated {
self.emit_leakclass_branch_param_drops(
&arm.consume,
&val,
scrutinee_binder.as_deref(),
);
}
- Step 5: Add the per-branch drops to
MTerm::Ifcodegen
In crates/ailang-codegen/src/lib.rs, change the MTerm::If destructure
(line 1906) to bind the new fields:
MTerm::If { cond, then, else_, then_consume, else_consume, .. } => {
In the then-branch emission (lines 1928-1930), emit the drop before the
br:
if !then_terminated {
self.emit_leakclass_branch_param_drops(then_consume, &then_v, None);
self.body.push_str(&format!(" br label %{join_lbl}\n"));
}
In the else-branch emission (lines 1941-1943):
if !else_terminated {
self.emit_leakclass_branch_param_drops(else_consume, &else_v, None);
self.body.push_str(&format!(" br label %{join_lbl}\n"));
}
- Step 6: Run the two leak pins — expect GREEN
Run: cargo test -p ail --test if_branch_consume_split_no_leak_pin --test match_arm_consume_split_no_leak_pin
Expected: BOTH PASS (live=0, allocs==frees).
- Step 7: Run legs 1/2 — expect still GREEN (behaviour-preservation)
Run: cargo test -p ail --test tail_recur_owned_param_no_leak_pin --test tail_recur_let_wrapped_no_leak_pin
Expected: BOTH PASS (the :823 gate-source switch is identical for
aggregate == 0 params; the new fall-through drop never fires for them).
Task 5: Verify — series_sma close + full workspace suite
Files:
-
Create:
crates/ail/tests/series_sma_no_leak_pin.rs -
Step 1: Write the series_sma RC-stats close pin
Create crates/ail/tests/series_sma_no_leak_pin.rs — structural copy of
the if/match leak pins, building examples/series_sma.ail and asserting
live==0. (series_sma_pin.rs already pins stdout; this pin is the RC
ledger, distinct.) Header doc: "Pins that the #61 headline series_sma
example is leak-clean under --alloc=rc. The residual live=2 was #63
leg 3 (branch-consume-split); this pin closes that leak tail." Body
identical to the if-pin's build/run/parse-stats, with:
let src = workspace.join("examples").join("series_sma.ail");
and the temp-dir / test-fn names adjusted (series_sma_no_leak_pin,
series_sma_is_leak_clean), asserting assert_eq!(live, 0, ...) and
assert_eq!(allocs, frees, ...).
- Step 2: Run the series_sma pin
Run: cargo test -p ail --test series_sma_no_leak_pin
Expected: PASS (live=0). This clears the #61 leak tail.
- Step 3: Run the full workspace suite
Run: cargo test --workspace
Expected: ALL PASS. (The typed-MIR re-synth strictness family means a
MIR-node change can ripple — this is the gate that catches a regression
the targeted pins miss, including any double-free/leak from the new drop
interacting with an existing branch-tailed fn in prelude/kernel/examples.)
In particular the std_either_demo, std_list_demo, and
poly_rec_capture_demo e2e tests are the closure-param regression
witnesses: they pass top-level fn refs (static closure-pairs) through a
match that consumes them on one arm and leaves them live on another —
exactly the leak class. Without the helper's field_drop_call == "ailang_rc_dec" type precondition (Step 1) they abort with
ailang_rc_dec: refcount underflow; with it they stay green.
- Step 4: Confirm no lockstep-pair drift
Run: cargo test -p ailang-codegen intercepts_bijection_with_intrinsic_markers
Expected: PASS. (Confirms the INTERCEPTS ↔ (intrinsic) bijection is
intact — this change adds no intrinsic. The Pattern::Lit ↔
pre_desugar_validation pair is likewise untouched: no Pattern::Lit
reject arm is added.)
- Step 5: Spot-check the diff for the double-free invariant
Read the codegen diff and confirm, by eye, that every new drop call sits
inside a branch_cc == 0 && agg >= 1 gate (the helper) and that the
fn-return dec (lib.rs:1489) and arm-close pattern-binder dec
(match_lower.rs:915) were NOT modified. This is the human check that the
disjointness-by-aggregate invariant holds in the shipped code.