v3: use emitted fixed-array names in function pointers

[antono3]

Summary

V3 emits fixed-array typedefs using the backed enum's C name, while several function-pointer paths use a different TypeChecker encoding and refer to an undeclared Array_fixed_int_2. Alias-only callback declarations can also omit their required fixed-array typedef.

Use the emitted fixed-array name consistently in function-pointer keys and collect arrays reachable through type aliases. Resolve named callback aliases in sizeof through the same key.

This supersedes #28867: it covers that PR's imported sizeof(Callback) case while retaining the backed-enum typedef name, instead of changing fixed-array naming globally. It is independent of #28868 (imported constants in array lengths).

Verification against current master

Rebased onto 6c359c228241b44f9d9a942696e83b513931e532 on 2026-09-26.

• Both imported callback regressions fail on unmodified master.
• Both pass together with this patch under strict V3 (-new-compiler), including native compilation, execution, and emitted typedef checks.
• ./v self, formatting verification, and git diff --check pass.
• The full fixed_array_typedef_codegen_test.v file passes (all seven test functions).
• VJOBS=4 ./v -new-compiler test vlib/v/gen/c/ passes: 28/28 test files.
• Both focused callback regressions pass again after the final pointer-preservation refinement. A separate callback-pointer alias compiles and runs, and its generated sizeof expression retains the extra *.

Latest tested head: 45ad1ea2fae51aa54e32545fa024aeba96dccdb1.

The previously failing twelve formatter files now pass on master. The unrelated anonymous-struct compiler regression still fails on unmodified master in its genuinely anonymous parameter case.

[medvednikov]

@codex review

[medvednikov]

@codex review

[medvednikov]

Reviewed b10ec41, including typedef collection, alias emission, return-wrapper registration, and the function-parameter emitter. The new alias-only and sizeof(Callback) cases are useful, but the emitted-name normalization is not yet used by every callback-signature path; see the inline comment.

Please extend the regression to pass and invoke a callback, in addition to declaring/measuring its type. That exercises the parameter declaration and the call ABI rather than only the alias's typedef.

This is a source review; I have not built V3 or executed these tests in this environment.

[medvednikov]

@codex review

[medvednikov]

@codex review

[medvednikov]

@codex review

[medvednikov]

Callback parameters now use normalized type names, including nested signatures. The added regressions actually invoke V and C callbacks and cover aliases, fixed-array pointers, and Option/Result callback returns.

[medvednikov]

Reviewed head 9897bbaf6ccce9da3b48164d198647c174c553b2, ignoring CI. The earlier function-parameter normalization finding is addressed. Since the PR is now merged, here is one remaining in-scope follow-up from source inspection:

[P2] Normalize callback signatures when emitting struct fields too

preseed_fn_ptr_type now registers g.fn_ptr_type_key(typ), but write_struct_field in the same file still uses g.tc.c_type(raw_field_type) for ordinary function-valued fields and g.tc.c_type(field_type.base_type) for pointer-to-callback fields. soa_field_c_type has the same fallback. The C-ABI override lookup does not normalize an ordinary V callback field.

Consequently, putting the callback from the new regression into a struct still requests a different typedef key from the one preseeded here: the field-emission path encodes [2]fixture.Mode as Array_fixed_int_2, while fixed-array collection emits Array_fixed_fixture__Mode_2. resolve_fn_ptr_type registers the supplied encoding without repairing it, and emit_fn_ptr_typedef uses that array name unchanged. The generated declarations therefore still have an undeclared-type problem for this use of the callback.

Please route those ordinary field fallbacks through the normalized function-type conversion too, preserving any explicit C-ABI override and pointer depth. Extend test_backed_enum_callbacks_are_passed_and_invoked with a stored callback, for example:

struct Holder {
    cb Callback
}

and, inside its generated main:

holder := Holder{cb: first}
assert holder.cb([fixture.Mode.one, fixture.Mode.two]!) == 26

The existing generated-C assertion rejecting Array_fixed_int_2 should cover this path as well. A pointer-to-callback field is worth covering alongside it.

This is a remaining normalization gap, not a claim that the unchanged field-emission code was introduced by this PR. I have not built V3 or executed the regressions locally: this environment has no V compiler and the repository clone failed because GitHub could not be resolved.