Type aliases: storage, resolution, protocol unions (progress on #15127)#15366
Closed
cgarvis wants to merge 2 commits into
Closed
Type aliases: storage, resolution, protocol unions (progress on #15127)#15366cgarvis wants to merge 2 commits into
cgarvis wants to merge 2 commits into
Conversation
…lixir-lang#15127 This branch advances the Type aliases bullet of elixir-lang#15127. The work is split across four logical pieces; they're combined here as one PR since each piece depends on the previous one and they all share a test surface in lib/elixir/test/elixir/module/types/integration_test.exs. ## 1. Foundation — typed struct fields via @type t A new `Module.Types.Typespec` converter walks typespec ASTs and produces `Module.Types.Descr` values. Built-ins, literals, unions, struct literals (%__MODULE__{...} and %Mod{...}), local references, and self-qualified references are all handled. Memoized via a `:pending` sentinel so self-cycles surface as an error; cycles in real code degrade the whole alias to `dynamic()` rather than break the build. `Kernel.Typespec.translate_typespecs_for_module/2` invokes the converter and stores the result in the module's data tables under `{:elixir, :types_descr}`. `Module.ParallelChecker.cache_from_module_map` snapshots it into the long-lived checker ETS table during `spawn_parallel_checker`, so the async checker can read it after the per-module data tables are torn down. The chunk-write path in `elixir_erl:compile/2` adds an optional `:types` key to the ExCk chunk; older readers ignore it (additive, no version bump). `Module.Types.Of.struct_instance/7` consults `User.t/0` to type-check field values, emitting a new `:badstructfield` diagnostic on mismatch. Default values injected by defstruct are skipped to avoid noise. To make this available to stdlib, `module/types/typespec.ex` is added to bootstrap MAIN — without this, modules compiled before Module.Types.Typespec is loaded would never get `:types` in their chunks. ## 2. Function types and remote type references The converter is extended to handle two more typespec shapes: * Function types `(args -> result)` parse as a single-element list containing `{:->, _, [args, return]}`. The handler calls `Descr.fun/2` (or `Descr.fun/0` for variadic `... -> result`). * Remote type references `Mod.t()` are resolved by reading the target's ExCk chunk via `:beam_lib.chunks` on the in-memory binary (`:code.get_object_code/1`, not the disk path, which fails when the beam isn't written yet). `@opaque` is dynamic from outside the defining module; parametric arity > 0 degrades silently because the offending typespec lives in another module. ## 3. Same-compile-unit ordering for typespec refs When module B's typespec references `A.t()`, B needs A's beam to be loaded by the time the converter runs. Without help, the parallel compiler doesn't see typespec references as deps and may schedule B before A is loaded. `Kernel.Typespec.typespec/4`'s remote-call branch now calls `Kernel.ErrorHandler.ensure_compiled` to block until A's beam is available — the same mechanism struct expansion uses. A `:type_reference` trace event is also emitted via `:elixir_env.trace` so external tools (mix xref, custom tracers) can consume it; the lexical tracker itself is intentionally not wired, since existing tests assert that typespecs don't add compile/runtime deps. To make `fetch_remote_types` see the in-memory checker ETS from each spawn subprocess, `elixir_module.erl` propagates `:elixir_checker_info` into the subprocess's process dictionary. ## 4. Protocol t/0 union Protocol consolidation now writes the union of every implementing struct/built-in type into the consolidated chunk's `:types[{:t, 0}]`. `Protocol.consolidate/5` extracts the domain already computed by `new_signatures/5` (which uses `Module.Types.Of.impl/2` for each impl) and stores it. The result is that references like `Enumerable.t()` in downstream typespecs become semantically meaningful: a value typed as `Foo.t()` is statically known to be one of the implementing types. ### Trade-off rationale for the protocol approach Three alternatives were considered: * **Rust `dyn Trait`** — keep the type opaque even after link-time vtable construction. Gives up refinement entirely. * **Scala/Java/Dart/Swift sealed traits** — author opts into a closed-world definition. Cleanest answer overall but requires a language addition. * **Closed-world per-build-artifact** (this commit). Open by language design, closed-by-consolidation for the current artifact. The union is correct for this build and useful for downstream type-checking on subsequent compilations. For protocols with `@fallback_to_any true` (most stdlib protocols), `Of.impl(Any, :open)` is `term()` and the union collapses to `term()`. Technically correct — those protocols runtime-accept any value — but no useful refinement. Third-party protocols without the fallback get the full benefit. ## What this branch does NOT address * **Same-run downstream re-check after consolidation.** Consolidation runs after the rest of compile, so dependents type-checked during the same `mix compile` saw the unconsolidated chunk. Precision arrives on subsequent compiles. * **Parametric `t(element)` of protocols and aliases.** Consolidator writes only `{:t, 0}`. Parametric protocol types are a separate bullet in elixir-lang#15127. * **INIT bootstrap modules' missing `:types`.** Modules in the very early bootstrap (Macro, Range, Keyword, etc., before `kernel/typespec.ex`) still don't get `:types`. Moving `Module.Types.Descr` to INIT would unlock them but is a much larger reshuffling. ## Verification * 508 checker tests pass (+15 from new typed-struct, function-type, remote-ref, dep-tracking, and protocol-union tests). * 79 top-level typespec tests pass. * 26 Protocol tests pass. * 36 lexical tracker tests pass. * 55 Mix.Dep tests pass. * `make clean && make compile` produces no new stdlib warnings. Signed-off-by: Chris Garvis <cgarvis@gmail.com>
Follow-up commit on top of the type-aliases foundation. Three categories
of change, all stylistic or performance — no new behavior.
## Bug fix: false-positive "protocol for missing module" warnings
`Protocol.__impl__!/3`'s missing-module check was treating transient
parallel-compiler deadlock states (`Code.ensure_compiled` returns
`{:error, :unavailable}`) as genuinely-missing modules and emitting a
misleading warning. The new typespec `ensure_compiled` calls
introduced earlier in this series caused more of these transient
states, surfacing ~33 false positives across the top 100 Hex
packages.
Fix: distinguish `:unavailable` (transient — compiler may resolve)
from `:nofile`/`:badfile`/`:embedded` (genuinely missing). Only the
latter group warrants the warning.
## Refactors (no behavior change)
* `Module.Types.Of.struct_instance/7`: extract a small helper for
the typed-field-vs-expected fallback (`expected_field_type/2`).
* `Of.format_diagnostic({:badstructfield, ...})`: change message to
"but expected type:" to match the convention used by `:badmap`
and sibling diagnostics.
* `Module.Types.Typespec`: function-spec clause uses two pattern
heads instead of a single-branch `cond`; collapsed an intermediate
`case pairs` in `fetch_remote_types_from_checker`; reordered
`do_to_descr/2` clauses so the hot remote-call shape matches
earliest.
* `Kernel.Typespec.convert_type_to_descr/2`: inlined a bare `arity =`
binding from a `with` chain.
* `Protocol.consolidate/5`: replaced 3-line Map.get/put/put dance
with `update_in` matching the adjacent `update_in(checker.exports)`
style.
## Performance (measurable)
* `Module.Types.Typespec.fetch_remote_types/1`: memoize via process
dictionary, scoped per compile worker. Eliminates redundant
`:beam_lib.chunks` + `:erlang.binary_to_term` calls for the same
remote module within one typespec block. Highest-payoff
optimization for typespec-heavy modules (Ecto schemas reference
`String.t()` etc. 30+ times per module).
* `Kernel.Typespec.typespec/4` remote handler: dedupe
`Kernel.ErrorHandler.ensure_compiled` calls per worker process via
`:"$elixir_typespec_ensured"` process-dict set. Eliminates redundant
`:waiting` round-trips to the parallel compiler.
## Bootstrap safety
`Kernel.Typespec.collect_defined_type_pairs/1`'s use of `Enum.any?`
was crashing during the bootstrap of `Range` (Enum isn't compiled yet
at that point). Replaced with `:lists.any/2`, an Erlang BIF that's
always available. Pure bootstrap fix; no runtime behavior change.
## Tests
* Updated 6 `integration_test.exs` assertions to match the new
`:badstructfield` diagnostic wording.
* Full checker suite: 508/508 passes on clean rebuild.
* Protocol suite: 26/26.
* Top-level typespec: 79/79.
* `make clean && make compile` clean.
Signed-off-by: Chris Garvis <cgarvis@gmail.com>
This was referenced May 13, 2026
cgarvis
added a commit
to cgarvis/nimble_options
that referenced
this pull request
May 13, 2026
The struct already stores a list of atoms for the "unknown options" error (lib/nimble_options.ex:524 passes a list of keys; test "unknown options" asserts `key: [:not_an_option1, :not_an_option2]`). The @type for the field said `atom()` only, so the new set-theoretic checker in elixir-lang/elixir#15366 flags every caller of error_tuple at that site and cascades into every package that depends on nimble_options. Widening to `atom() | [atom()]` matches actual runtime behavior. Doc updated to call out the multi-key case. No code change, no test change. Signed-off-by: Chris Garvis <cgarvis@gmail.com>
Member
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Progress on #15127's "Type aliases" bullet.
What ships
@type t :: %__MODULE__{name: binary()}declarations resolve through a newModule.Types.TypespecAST-to-Descr converter and persist via the ExCk chunk. The checker reads them fromOf.struct_instance/7and emits a new:badstructfielddiagnostic when a value doesn't match its declared field type. Function-type fields, remote type references, and protocolt/0(union of impls, written at consolidation) all resolve precisely. Wasdynamic()before.The decision worth flagging: protocol
t/0as closed-world per build artifactElixir protocols are open-world by design. Consolidation closes the world for a given build. Three approaches considered:
dyn Trait: keep the type opaque. Gives up refinement entirely.@fallback_to_any trueprotocols (most stdlib) collapse toterm()becauseOf.impl(Any, :open)isterm(). Third-party protocols without the fallback get the full benefit.Not in scope (separate follow-ups)
t(element).:types.Verification
make test: 14/8362 failures, all OTP-28.0 stderr-banner contamination (lib/elixir/src/elixir.erl:134), identical onorigin/main. Not branch-related.make format && make clean compile: clean.Hex bench: top 100 packages
0 new compile errors. 14 distinct upstream typespec violations surfaced (Postgrex.Result, Absinthe.Resolution, nimble_options.error_tuple, Bandit.Adapter, Finch.HTTPError, etc.). First upstream fix already filed: dashbitco/nimble_options#151.
Performance: +1.53% wall-clock across all 100 packages. Concentrated in typespec-heavy libraries (ecto_sql +18%).
fetch_remote_typesmemoization in the second commit recovered about half the regression from the first.AI disclosure
Implementation assisted by Claude. Human review and architectural decisions throughout. Sign-off carries DCO responsibility.