General TypeName(expr) reinterpret cast (named record/class/pointer)

• Type: feature
• Status: backlog
• Owner: —
• Opened: 2026-06-18
• Split from: [[feature-explicit-typecasts]] (scope §3). The cast-token allowlist gap (Char/Boolean/String) is closed there; this is the larger, separate piece deferred at resolution.

Motivation

PXX casts are still a fixed token allowlist in the factor parser (Integer/LongWord/Char/Boolean/String/Pointer/Ord/Chr/PChar). There is no general TypeName(expr) form for a user-named type — i.e. a reinterpret/checked cast to a named record, class, pointer-to-T, or type alias:

TFoo(p)^           { pointer reinterpret to a named record }
TDerived(baseObj)  { downcast a class reference (unchecked here; `as` is checked) }
PMyRec(addr)       { typed-pointer reinterpret }

Today these only work for the hardcoded builtins; a TypeName(expr) with a user type name falls through to "expected expression" or is misparsed.

Scope

• In the factor parser, when an identifier in value position is a known type name (record/class/alias/typed-pointer) and is immediately followed by (expr), build a reinterpret-cast node carrying the target type (reuse the existing AN_PTR_CAST machinery / type-alias AliasElemTk path where it fits). No value conversion — same bits, new static type — except where a checked form is wanted.
• Disambiguate from a call: a type name followed by ( is a cast; a routine name is a call. Constructor calls (TFoo.Create) are unaffected.
• Decide interaction with is/as ([[feature-class-is-as]]): as is the checked class cast; TClass(obj) is the unchecked reinterpret. Keep them distinct (matches FPC/Delphi).

Acceptance

• TRec(ptr)^.field, PRec(addr), and TClass(obj).method compile and reinterpret correctly on all targets; make test + make cross-bootstrap stay byte-identical. Add an oracle test (output-equality vs FPC).

Notes

• Lower-value than is/as (which is the real OOP gap); schedule after the class-cast work so the checked/unchecked split is designed together.

Log

• 2026-06-18 — surveyed during the autonomous run. Pointer-type-alias casts already work end to end: PFoo(expr), PFoo(expr)^.field, PFoo(addr)[i] via the AN_PTR_CAST branch in ParseFactor (with its own ^/./[] suffix loop). So the record-via-pointer reinterpret case (PRec(addr)^) is covered today. Still open: the bare unchecked class reinterpret TClass(obj) with a trailing .member — needs class-cast detection in ParseFactor plus postfix member/method integration on the cast result. Low value: the checked form (obj as TClass).member already works (feature-class-is-as), and unchecked reinterpret is mainly for type-punning. Deferred behind the float-correctness bugs (feature-int-to-float-assign, feature-result-in-loop).

RESOLVED 2026-06-20 (Track A)

All the acceptance forms now work:

• PRec(addr) typed-pointer reinterpret — already worked (pointer-alias cast).
• TClass(obj).method / .field — already worked (AN_CLASS_CAST).
• TRec(ptr)^.field record-name reinterpret — ADDED, read and write. A record name followed by ( is reinterpreted as a pointer to that record; the pointee rec id is carried on the AN_DEREF node (ResolveNodeRec AN_DEREF ASTIVal>0 path) so trailing .field offsets resolve. Added in both the factor parser (expression/read) and the statement parser (lvalue/write), mirroring the existing class- and pointer-alias-cast branches.

Verified: TRec(p)^.a/.b read and TRec(p)^.a := v write hit the right field offsets. make test + cross-bootstrap (i386/aarch64/arm32) byte-identical; test/test_record_typecast.pas added. (A bare value reinterpret TRec(x) without ^ — record value cast — is not part of this; the ^.field form covers the typed-pointer use the ticket and Track B reflection/binding code want.)

Resolved-in: 453bbc5 (finalizing commit)