Pointer cast on custom Char pointer aliases fails to skip string length prefix

• Type: bug
• Status: DONE 2026-06-20
• Owner: —
• Opened: 2026-06-20
• Relation: surfaced in gtk3widgets.pas when using typed pointer casts of custom aliases (like PC = ^Char; PC(s)) instead of literally PChar(s) or PAnsiChar(s).

Problem

When a custom alias to ^Char is defined (e.g. type PC = ^Char;), casting an inline string (tyString) using this alias (e.g. PC(someString)) does not add the +8 offset required to skip the 8-byte length prefix. As a result, the cast yields a pointer to the length prefix itself rather than to the NUL-terminated character data. This causes external C/GTK functions to receive control/length bytes instead of the actual string contents.

Reproduction

program repro;
type
  PC = ^Char;
var
  s: string;
  p1: PChar;
  p2: PC;
begin
  s := 'Hello';
  p1 := PChar(s); // correctly points to 'H' (adds +8 offset)
  p2 := PC(s);    // BUG: points to the 8-byte length prefix (first byte contains length 5)
end.

Root cause

In compiler/parser.inc line 3256:

if CaseEqual(name, 'pchar') or CaseEqual(name, 'pansichar') then

The compiler checks literally for 'pchar' and 'pansichar' to trigger the ASTIVal[node] := -2 sentinel (which instructs the code generator to add the +8 offset for string types). For custom aliases or any other pointer type names, it performs a raw pointer reinterpret cast, which skips the +8 adjustment.

Workaround

Use the built-in PChar type name for all casts of string values to C-style strings.

Fix direction

The compiler should check if the target cast type resolved is a pointer to Char/AnsiChar (i.e. checking LastTypePointerElemTk = tyChar), rather than matching the type name literally against 'pchar' or 'pansichar'.

Log

• 2026-06-20 — opened. Discovered in gtk3widgets.pas.

DONE 2026-06-20

Fixed in parser.inc: the type-alias cast path now uses the -2 PChar adapter (skip the 8-byte length prefix for a string operand; plain reinterpret otherwise) when the alias resolves to ^Char (AliasElemTk = tyChar), instead of a raw pointer reinterpret. PC(s) with type PC = ^Char now points at the char data like PChar(s). Non-Char aliases unchanged. Validated; self-host byte-identical, make test green.

(Aside, pre-existing & unrelated: writeln(p^) on an alias-cast pointer whose pointee is dereferenced directly can segfault — PI(@n)^ with type PI=^Integer crashes too, independent of this fix. Worth a separate ticket if it bites.)

Resolved-in: 71036f6 (finalizing commit)