An interface name in a var initialiser stores the GUID's ADDRESS, not the GUID

An interface type name used as a value means its GUID — defs.inc:965, AN_GUIDCONST, "which in Pascal means exactly one thing, its GUID". That works in a statement and is silently wrong in a var initialiser.

Measured 2026-09-06 at compiler cda68a91bec5, {$mode objfpc}, on ICom = interface ['{04B6AB72-8F86-45F8-8D49-393E799F51A8}'], printing the first eight bytes of the destination:

114 171 182 4 is 04B6AB72 little-endian — the right answer, which the statement path produces. 248 27 65 0 0 0 0 0 is 0x00411BF8: an address. AN_GUIDCONST's own doc says lowering "yields that blob's data address", and the var-initialiser path stores that pointer into the record instead of copying the 16 bytes behind it.

Why it ranks at 60 rather than as a missing arm

It is a silent wrong value in interface identity. TObject.GetInterface looks an interface up BY GUID at runtime (defs.inc:447), so a TGUID global seeded this way holds an address where 16 bytes of identity belong, and every comparison against it fails — or, worse, succeeds against whatever else happens to sit at that address pattern. Nothing diagnoses it: the declaration compiles, the record is the right size, and the program runs.

The refusal on the const line is the better of the two failures, and note which one a user hits: const is the spelling this construct is normally written in, so the shape that stays quiet is the one people reach for second, after the diagnostic pushes them off the first.

The pattern this completes, which is the reason to fix both together

Four cells: two value forms crossed with two initialiser paths. Three are wrong, in three different ways, and the two paths fail in OPPOSITE directions:

The @ row was fixed this morning by making the var path call the same helper the const path already used. This row wants the mirror, and the strongest argument for one shared path is that nobody could have predicted which direction each asymmetry ran — they are not a systematic "const is ahead of var", they are two independent omissions that happen to point opposite ways.

Also absent, found in the same file and not fixed here

IsEqualGUID is not in lib/rtl/sysutils.pas (zero hits). That is Track B and a separate row; tinterface6.pp needs it as well as the two arms above, which is why that conformance row takes three changes across two lanes and is not a single burn.

Repro

{$mode objfpc}
type ICom = interface ['{04B6AB72-8F86-45F8-8D49-393E799F51A8}'] end;
var g: TGUID = ICom; i: Integer;
begin for i := 0 to 3 do Write(PByte(PtrUInt(@g) + i)^, ' '); end.
{ fpc: 114 171 182 4     pxx: 248 27 65 0 }

RESOLVED — all four cells, and the fix found a fifth the ticket did not list

The ticket's table has four cells (two value forms x two initialiser paths). Working it turned up a fifth axis it does not mention: global vs routine-local, which is a genuinely separate emitter, and the first version of this fix compiled clean and SIGSEGV'd on it.

Cause, and it is the one the statement path already fixed

Not an "address instead of a copy" in the GUID lowering. The var-section initialiser's METACLASS arm claimed the name. Its guard is tkIdent + IsClassType + not-a-symbol, an interface satisfies all three, and the arm records init kind 5 — an AN_CLASSREF VMT address. 0x00410BD8 was never a GUID pointer; it was a class reference.

That is precisely bug-a-an-interface-name-as-a-guid-value-copies-rtti, one construct over: there the statement path fell through to AN_CLASSREF and "made the record copy read 16 bytes of the RTTI blob... with no diagnostic and a plausible-looking result". Fixed the same way, by asking about interfaces FIRST — now through a named IsInterfaceTypeName, because it is two lookups and the second is the easy half to forget at the next site.

Kind 10, not a new baked shape

PendingInitKind 10 already means "Val IS AN AST NODE INDEX", and its own declaration says why it exists: the other kinds name a constant the emitter rebuilds in Data[], and this one keeps the AST so the initialiser runs at program entry "like the assignment it is". g := ICom as a statement has always produced the right bytes — an ordinary 16-byte record copy from the interned blob — so the fix hands that working path the same node rather than teaching a second one to bake a GUID. Both sections inherit it because both call TryParseInitValForm, which is exactly how the @ half of this table was fixed at 21ac9e7bc.

The fifth cell, which is the part worth reading

FlushLocalInits is a separate emitter and knew kinds 1/2/4/5/9. Its else is a silent catch-all that builds an AN_INT_LIT from whatever Val held — so a routine-local var g: TGUID = ICom assigned a node index into a TGuid and the program segfaulted, while the compile printed ok. Measured, not predicted: the fixture's local rows are what caught it.

A kind the global emitter knows and the local one does not is not a missing feature there, it is a wrong value. Kind 10 added to the local emitter with that written beside it.

The const branch is dedicated rather than a widening of the general const path, and deliberately: letting a record-typed const fall through would hand LOCAL consts to TryParseInitValForm for the first time, whose string arm keys on the destination type — so const S: ShortString = 'x' inside a routine would move. That is a separate extent to enumerate.

Verification

test/test_interface_name_as_guid_initialiser.pas, wired into test-core, .expected generated from fpc 3.2.2. Five rows, all byte-identical: statement, global var, global const, local var, local const.

The statement row is the control — it was always correct, so a fixture without it would pass if every declaration broke the same way. Eight bytes are the assertion, not a <> nil or a size: 114 171 182 4 is 04B6AB72 little-endian; an address shows as a small number followed by zeros, which is what made the original defect invisible to any check that only asked whether something was stored.

Control run, not assumed: with all four compiler files reverted, the fixture does not compile (expected '(' before 'ICom' at the global const). Self-host fixedpoint converged (a7b03135f504).

Log

• 2026-09-08 — resolved; this names the commit that carried the resolve, which is not always the one that carried the change — commit 2cf3b41c0.