Files
AILang/crates/ail/tests/mono_xmod_ctor_pattern.rs
T
Brummel e580f75adf test: red for mono pass mis-resolving cross-module ctor patterns
`monomorphise_workspace` re-runs `synth` on every fn body to recover
residual class constraints; the env it uses is built by
`mono::build_workspace_env`, which delegates to `build_check_env` and
produces a workspace-flat `ctor_index` (every Def::Type ctor across
every module, keyed by bare ctor name → bare type name).

`check_in_workspace` (lib.rs:1247-1258) explicitly clears that flat
index after `build_check_env` and rebuilds it per-module so that
`Pattern::Ctor`'s local-first / imports-fallback resolution at
lib.rs:2486-2521 keeps the qualified-type-name comparison at
lib.rs:2526 intact: imports-fallback yields a qualified
`resolved_type_name` (`Mod.Type`), local-hit yields a bare one.

The mono pass never does the per-module overlay, so when a body in
module B pattern-matches a ctor `C` whose Def::Type lives in imported
module A, the workspace-flat index resolves `C` locally and yields the
bare type name. The scrutinee, however, was typed against the
qualified name, and the comparison fails with
`PatternTypeMismatch { ctor: "Cons", ty: "A.Type<...>" }`.

Surfaced by iter 23.2 Task 3, which adds `class Eq a` + Eq Int/Bool/Str
instances to the prelude. Pre-Task-3 the prelude is class-free, so
`workspace_has_typeclasses(ws) == false` and the mono pass early-outs
at `mono.rs:73`. Task 3 flips the gate; every workspace now traverses
bodies, which brings the latent bug to the surface for the
pre-existing `nested_ctor_pattern_first_two_sum` and
`std_either_list_demo` E2E tests.

Sibling regressions in the same family: `mono_xmod_qualified_ref.rs`
(env.imports not seeded), commit 13b36cc (env.globals not seeded for
self-recursive fns), commit 5c5180f (env.types / env.ctor_index not
seeded for user ADTs — the original "flat ctor_index" decision that
this bug now exposes as wrong-by-construction for cross-module ctor
pattern resolution).

The minimal fixture is two modules with five defs total:
test_mono_ctor_listmod (data List a = Nil | Cons a (List a)) and
test_mono_ctor_main (class Trivial a + instance Trivial Int + a fn
that pattern-matches Cons against the imported List<Int> + a main).
The test pins the inner cause: matches against the specific
`CheckError::PatternTypeMismatch` variant with the qualified
scrutinee type and bare ctor name, so it stays RED regardless of the
prelude's typeclass content.
2026-05-10 22:16:51 +02:00

123 lines
5.5 KiB
Rust

//! Bug regression: the workspace-monomorphisation pass must not mis-resolve
//! a cross-module constructor pattern. The mono pass re-runs `synth` on
//! every fn body to recover residual class constraints; that env is built
//! by `mono::build_workspace_env`, which delegates to `crate::build_check_env`
//! and produces a **workspace-flat** `ctor_index` (every Def::Type ctor
//! across every module, keyed by bare ctor name → bare type name).
//!
//! In contrast, `check_in_workspace` (lib.rs:1247-1258) explicitly clears
//! that flat index after `build_check_env` and rebuilds it per-module so
//! that `Pattern::Ctor`'s local-first / imports-fallback resolution
//! (lib.rs:2486-2521) can keep the qualified-type-name comparison at
//! lib.rs:2526 intact: the imports-fallback branch produces a qualified
//! `resolved_type_name` (`Mod.Type`), the local-hit branch produces a bare
//! one (`Type`).
//!
//! The mono pass never does the per-module overlay. So when a body in
//! module B pattern-matches a constructor `C` whose Def::Type lives in
//! imported module A, the workspace-flat ctor_index resolves `C` locally
//! (in fact: anywhere in the workspace) and yields the bare type name
//! `"Type"`. The scrutinee, however, was typed by the main pass against
//! the qualified name `"A.Type"` (per the imports-fallback path), so the
//! comparison
//!
//! ```ignore
//! Type::Con { name, args } if name == &resolved_type_name => args.clone(),
//! _ => return Err(CheckError::PatternTypeMismatch { ... }),
//! ```
//!
//! at lib.rs:2526 fails with `cannot match constructor pattern C against
//! type A.Type<...>`.
//!
//! Sibling regressions in the same family:
//! * `mono_xmod_qualified_ref.rs` — `env.imports` not seeded.
//! * commit 13b36cc — `env.globals` not seeded for self-recursive fns.
//! * commit 5c5180f — `env.types` / `env.ctor_index` not seeded for
//! user ADTs (the original "flat ctor_index" decision that this
//! bug now exposes as wrong-by-construction for cross-module
//! pattern-match resolution).
//!
//! Surfaced by iter 23.2 Task 3, which adds `class Eq a` + Eq Int/Bool/Str
//! instances to `examples/prelude.ail.json`. Before Task 3 the prelude has
//! only `Def::Type Ordering`, so `workspace_has_typeclasses(ws) == false`
//! and the mono pass early-outs at `mono.rs:73`. Task 3 flips the gate;
//! every workspace now traverses bodies, which brings the latent bug to
//! the surface.
//!
//! Without the prelude change, two pre-existing E2E tests already
//! exercise the bug shape (`nested_ctor_pattern_first_two_sum`,
//! `std_either_list_demo`) and silently pass because the mono pass is a
//! no-op on a class-free workspace. This test pins the inner cause
//! against a minimal 2-module fixture so it stays RED regardless of the
//! prelude's typeclass content.
use std::path::PathBuf;
fn examples_dir() -> PathBuf {
let mut d = PathBuf::from(env!("CARGO_MANIFEST_DIR"));
d.pop();
d.pop();
d.join("examples")
}
/// Property: `monomorphise_workspace` must succeed on a workspace that
/// (a) flips the typeclass gate (any `Def::Class` + `Def::Instance` is
/// enough), and (b) contains a fn whose body pattern-matches a
/// constructor whose `Def::Type` lives in an imported module.
///
/// The minimal fixture is two modules:
///
/// * `test_mono_ctor_listmod` — declares `data List a = Nil | Cons a (List a)`.
/// * `test_mono_ctor_main` — imports the listmod, declares
/// `class Trivial a` + `instance Trivial Int` (to flip the
/// `workspace_has_typeclasses` gate), and defines
/// `head_or_zero : test_mono_ctor_listmod.List<Int> -> Int` that
/// matches on `Cons h _ | Nil`.
///
/// Pre-fix repro: `monomorphise_workspace` returns
/// `Err(CheckError::PatternTypeMismatch { ctor: "Cons", ty:
/// "test_mono_ctor_listmod.List<Int>" })`. The same workspace
/// typechecks cleanly because `check_in_workspace` clears the flat
/// `ctor_index` before running `Pattern::Ctor` resolution.
///
/// Post-fix expectation: `monomorphise_workspace` returns `Ok` and the
/// `head_or_zero` fn body's match arm types correctly.
#[test]
fn mono_pass_handles_xmod_ctor_pattern() {
let entry = examples_dir().join("test_mono_ctor_main.ail.json");
let ws = ailang_core::load_workspace(&entry).expect("load");
// Pre-condition: the workspace typechecks cleanly. The bug is
// localised to the mono pass; the typecheck path does not have it.
let diags = ailang_check::check_workspace(&ws);
assert!(
diags.is_empty(),
"fixture must typecheck before mono runs: {:?}",
diags
);
// Pin the symptom to the inner cause. Pre-fix this returns Err
// with the qualified scrutinee type and the bare `Cons` ctor.
let result = ailang_check::monomorphise_workspace(&ws);
match result {
Ok(_) => { /* post-fix: pass */ }
Err(ailang_check::CheckError::PatternTypeMismatch { ctor, ty }) => {
panic!(
"mono pass mis-resolves cross-module ctor pattern: \
ctor=`{}` ty=`{}` (expected mono to succeed; the bare \
ctor_index in build_workspace_env resolves `Cons` to \
the local bare `List` instead of qualified \
`test_mono_ctor_listmod.List`)",
ctor, ty
);
}
Err(other) => {
panic!(
"mono pass failed with unexpected error variant: {:?} \
(expected PatternTypeMismatch on cross-module Cons)",
other
);
}
}
}