Explicit type-casts (Char/Boolean/String and a general TypeName(expr))

• Type: feature
• Status: done
• Resolved: 2026-06-18 (commits a7e8dd4, 9ff4ac1) — Char/Boolean/String casts wired; general TypeName() reinterpret split to a separate future ticket.
• Owner: —
• Opened: 2026-06-18

Motivation

Type-casting in PXX is ad-hoc: the factor parser (parser.inc ~2372) has a hardcoded allowlist of cast tokens, not a general TypeName(expr) mechanism. Surveyed 2026-06-18 (surfaced while wiring managed strings on ESP, but target-independent — same char/shortstring/ansistring typing family as the literal-concat fold fix):

Not urgent: Boolean and Char are easily worked around today (Ord/Chr, or arithmetic), and String(x) has little use until external-library interop grows (PChar->string already works for imports). Filed so the gap is tracked.

Scope

• Add tkChar_T, tkBoolean_T, tkString_T to the factor parser's cast cases (mirror tkInteger_T): build an AN_CALL with ASTIVal = -Ord(tk*_T) and the right ASTTk.
• Codegen passthrough: add those ids to each backend's type-pun passthrough list (Ord/Chr/tkInteger_T/tkLongWord_T already passthrough on x86-64 line ~1300 and the ESP backends). Char(x) = low-byte (mask/passthrough on the value model), Boolean(x) = passthrough (nonzero semantics already hold), String(x) = char/PChar -> managed string (route to PXXStrFromLit / PCharToString as the existing inline-string store does).
• Consider a real general TypeName(expr) cast (named record/class/pointer reinterpret) instead of the token allowlist — larger, separate sub-task.

Acceptance

• Char(i), Boolean(i), String(c) compile on all targets and produce the FPC-equivalent value; make test + make cross-bootstrap stay byte-identical.

Notes

• Additive / target-independent, so byte-identity-safe.
• Related: the ESP literal-string-concat fold (commit e1c9198) and the managed-string typing work — same family of char/string conversions.

Log

• 2026-06-18 — Char(x) + Boolean(x) landed (commit a7e8dd4). Factor-parser cases tkChar_T/tkBoolean_T → AN_CALL value-pun (result tyChar/tyBoolean); passthrough id added to all backends (x86-64/i386/aarch64/arm32/riscv32/xtensa). Test test_cast_char_bool.pas byte-identical to FPC; make test + cross-bootstrap byte-identical. String(x) still open: unlike the int/char puns, it must materialise a string value from an arbitrary expression. The frozen-string model only builds a string at assignment store sites (s := c works); there is no helper turning a Char/ordinal into a string rvalue (PCharToString/PXXStrFromLit cover PChar/literal only). Belongs with the managed-string right-sizing arc ([[project_managed_string_f2_direction]]). General TypeName(expr) reinterpret cast also still open (separate larger sub-task).
• 2026-06-18 — String(x) landed (commit 9ff4ac1). Solved the rvalue- materialisation: new AN_STR_FROM_CHAR node desugars String(c) (parse time) to a hidden tyString temp + __t := c store + temp read; IR lowers Left then yields Right (no backend code). String(s) identity = value-pun passthrough (tkString_T added to all six backends). Temp is tyString (matches the String keyword, always frozen even under managed mode — a managed temp mismatched on copy). All three table gaps (Char/Boolean/String) now closed; FPC byte-identical; cross-bootstrap byte-identical i386+aarch64+arm32. Resolved. The general TypeName(expr) named-record/class/pointer reinterpret cast (scope §3) is split out as its own future ticket — the cast token allowlist gap (the actual motivation) is fully closed.