Build fix for Tcl 9.0 (while remaining compatible with Tcl 8.4-8.6)

[devrimgunduz]

Problem

pgtcl fails to compile against Tcl 9.0 headers (e.g. building against PostgreSQL 19 / recent Fedora-Rawhide-style toolchains that ship Tcl 9). The build aborts with dozens of

error: expected ';', ',' or ')' before 'objv'

and, once that is fixed, a second wave of

error: passing argument N of '...' from incompatible pointer type
note: expected 'Tcl_Size *' but argument is of type 'int *'

and one instance of

error: storage size of 'cmd_info' isn't known

After those are resolved, the build succeeds but leaves a small number of warnings, also fixed by this patch (see item 6 below).

Root cause

Tcl 9.0 removed two long-deprecated Tcl-8.0-era compatibility shims that pgtcl's sources still relied on:

1. The CONST / CONST84 / CONST86 macros are gone. Every Tcl_Obj *CONST objv[] (and one CONST char *) in pgtcl no longer parses, since CONST is now just an unknown identifier sitting where the compiler expects a parameter name. Plain const has been the correct, portable spelling since Tcl 8.4, so this patch simply spells it that way everywhere.

2. String/list lengths and element counts that used to be returned via "int *" out-parameters (Tcl_GetStringFromObj, Tcl_GetByteArrayFromObj, Tcl_ListObjGetElements, Tcl_ListObjLength, ...) now use the new, pointer-width Tcl_Size type. Every local variable, struct field and helper-function parameter in pgtcl that receives one of these lengths has been changed to Tcl_Size, with a small compatibility header (generic/pgtclCompat.h) providing a typedef int Tcl_Size; fallback for Tcl 8.x, which doesn't define this type at all. This mirrors the migration path documented by the Tcl maintainers: https://www.tcl-lang.org/doc/howto/size_t.html

3. struct Tcl_CmdInfo is used in two places in pgtclSqlite.c, but Tcl_CmdInfo has always actually been a typedef, not a tagged struct; older Tcl headers happened to also declare a struct tag of the same name; the anonymous-struct-only header in newer Tcl no longer does, so the struct keyword must be dropped.

4. One additional, previously-hidden bug surfaces once the above parse errors are fixed: Pg_lo_read() passed the long-obsolete TCL_PARSE_PART1 flag to Tcl_ObjSetVar2(). That flag was removed in Tcl 9.0. Tcl_ObjSetVar2() has parsed array syntax ("name(index)") in part1 automatically for a very long time regardless of this flag, so it is simply dropped; behaviour is unchanged on every supported Tcl version.

5. A handful of ckfree(some_const_char_ptr) calls relied on the old ckfree(x) macro casting its argument to (char *) internally. Tcl 9.0's non-debug ckfree is a bare #define ckfree Tcl_Free with no cast, so these now need an explicit (void *) cast at the call site to avoid a "discards const qualifier" warning.

6. Three warnings remained after the above (all pre-existing, just never previously reached because the build died earlier):

   • pgtclId.c: getresid() and PgGetConnByResultId() take a
     const char * handle string, then temporarily NUL the last '.'
     in place (and restore it) to split the connection id from the
     result id. mark = strrchr(id, '.') assigned the result
     straight into a char *, which some toolchains flag as
     discarding the const qualifier since id is const. The
     temporary in-place mutation itself is unchanged (it's the
     existing, intentional behaviour); the cast is now explicit:
     mark = (char *) strrchr(id, '.');

   • pgtclCmds.c: Pg_select() declared int tuplesProcessed = 0;
     and incremented it once per row, but never read it anywhere.
     The -count var feature that this looks like it was meant to
     feed is actually driven by a separate numTuples/tuplesVarObj
     pair a few lines away, so tuplesProcessed was simply dead
     code left over from an earlier version. It has been removed
     (declaration + increment); this does not change any observable
     behaviour, since its value was never used.

None of this changes pgtcl's behaviour on any currently supported Tcl version -- it only replaces removed/renamed compatibility spellings with their modern equivalents, and drops one genuinely-unused variable.

Hacked by Claude, tested by me.

Per: pgdg-packaging/pgdg-rpms#213