[JGamboa]

Summary

Four checker-driver fixes found while checking a large Composer project (a Laravel 13
application, ~3400 classes). They share the checker driver, pipeline and source-identity infrastructure. The shared
source-identity fix has also landed on main through #2038; this PR preserves it while
extending entry-file diagnostics, prelude discovery and checker optimizations.

Diagnostic format change. Every diagnostic that has a source position now names its
file: error[main.php:3:6]: Undefined class: Missing instead of
error[3:6]: Undefined class: Missing, and the same for warnings. This affects every
checker and later diagnostic, including the ones in the entry file. Parse errors already
printed error[file:line:col]; now all diagnostics do. The entry file is named as it was
given on the command line; included and autoloaded files by the path they were loaded
from. Diagnostics with no position still print error: message. README's "Error
messages" section is updated. Tests that match the message text (error_tests) are
unaffected; anything that parses error[line:col] literally would need updating.

1. file:line:col diagnostics and distinct span identities for autoloaded files.
   Autoloaded files were lexed with source id 0 (the entry file's), and the resolver of their
   includes rewound the id counter, so spans of different files collided as map keys. Each
   autoloaded file now gets its own source id (resolve_collecting_includes_continuing), and
   an id -> path registry (Span::register_source_path / Span::source_path) lets
   print_error / print_warning name the file when the error itself carries none. Class
   schema errors (circular inheritance, final class, duplicate class, unknown parent) point at
   the class declaration instead of Span::dummy().
   Visible effect besides the format: two autoloaded class files with a retyping assignment at
   the same line and column no longer fail with "Cannot re-bind $x here" (Give Span a file identity so cross-file same-position collisions stop being compile errors #778).

2. Post-autoload preludes. The pdo, mysqli, tz, image, hash, curl, xml and var_export
   preludes were decided on the entry program before name resolution, but the Composer
   autoload loads vendor classes later: an autoloaded class typed against PDO or
   CurlHandle failed with Unknown type: PDO unless --with-* was passed.
   autoload::prelude_probe resolves names and runs the autoload on a copy of the program,
   runs each prelude's detector over the result, and the pipeline injects the groups it finds
   as if the entry used them. The probe restores thread-local source IDs and anonymous-class numbering. Process-global
   synthetic spans and value-include temporaries stay monotonic, so a discarded probe cannot
   recycle identities allocated by another compiler thread.

3. Monotonic method variadics. A first-class callable ($r->is(...)) promotes an untyped
   method variadic's collector to array<mixed>; a later direct call with a string narrowed
   it back to array<string>, because wider_type_syntactic lets Str absorb Mixed. The
   callable path then stored mixed values into a string collector, and the program aborted
   with "heap memory exhausted". widen_method_variadic_element only widens.

4. Quadratic work removed (profiled on the 3400-class program; release user time
   290 s -> 62 s, same result except union member order, which HashMap order already varied):

   • update_method_return_type and update_method_callable_return_metadata walked every
     class for each method body: an index (implementing class, method, static) -> classes that
     dispatch to it;
   • propagate_constructor_arg_type walked every class per argument of each new: an index
     (property, declaring class) -> classes sharing it;
   • loop_storage_types was cleared with a retain over the whole map per method, function
     and closure: a scope -> loop spans index;
   • check_builtin scanned every declared function on each builtin call: the cheap condition
     goes first, and has_function_decl_folded / folded_map_key try the exact key before
     the scan;
   • validate_late_bound_constructor_targets walked the parent chain from every class at each
     new static: a direct-children index (class_and_descendants).

   The indexes are keyed by classes.len(): implementation and property maps are only written
   while classes are registered, which also changes the table size.

Closes #1445
Fixes the autoload part of #778 (autoloaded files no longer share span identities). The rest
of #778 (a retyping file required twice, file identity in Span equality) is not addressed
here, so #778 should stay open.

Left out

• The descendants index for widen_override_untyped_method_params. This PR is independent of
  the chained-call PR (fix(checker): widen untyped params reached through call-result receivers #1919), which adds that function; once both land, a follow-up switches
  it to class_and_descendants too.
• The opt-in fixpoint tracer (ELEPHC_CHECKER_TRACE) of the original branch.
• core and bcmath preludes are not probed through the autoload closure (they postdate the
  original change); an autoloaded file that is the only user of WeakMap or BcMath\Number
  is still not covered.
• Known remaining gap: two autoloaded files whose retype warnings sit at the same line and
  column still print only one of the two "changes type" warnings (the program compiles and
  runs correctly).

Tests

After merging main 30f1305e23, 14 exact local unit, CLI and codegen tests pass:

• Four unit checks cover concurrent synthetic-span identity, discarded-probe include
  temporaries, preserved source/anonymous-class numbering, span-to-file registration, and
  the source file taking precedence over an entry-file diagnostic tag.
• Six CLI checks cover entry/autoload diagnostic paths, missing includes, early rejected
  autoload-closure warnings, PDO/curl prelude detection and same-position retypes.
• Four codegen checks cover promoted method variadics, included-file lowering refusals,
  anonymous classes in includes and the original readonly-write span collision from Give autoloaded files distinct span identities #1445.

The three new regressions for global counter rollback and early warnings were verified
failing on 4666dae39 before the fixes. All focused commands ran through the shared
build queue, with one test matched per filter. The library, CLI and selected test binaries
compile without warnings; git diff --check passes.

Refs #1886

AI assistance

This change was written with substantial help from Claude Code (Anthropic). I reviewed it,
checked the expected outputs against php 8.5, and remain responsible for it.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no new blocking issue was established.

Summary

The PR extends file-qualified diagnostics, probes autoloaded files for needed preludes, preserves promoted method variadics, and replaces repeated checker scans with indexes.

• Positioned checker messages name the file they come from.
• Autoloaded files can bring in the preludes their code uses.
• Method variadic collectors keep the widest type seen at each call.
• Large class checks use indexes instead of repeated full-table scans.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Parse entry and resolve includes] --> B[Probe autoloaded files on a copy]
  B --> C[Restore thread-local counters]
  C --> D[Inject detected preludes]
  D --> E[Resolve names and load real files]
  E --> F[Check types using lookup indexes]

Loading

_{Reviews (6) · Last reviewed commit: "chore: merge main into pr-1921" · Reviewed by Greptile}