At some point, both dmd and Phobos1 added a new member to TypeInfo_Struct: m_flags.
The way dmd handles this causes an ABI breakage: dmd will use the _calculated_ size of the struct to calculate the offset of the name string. After that, dmd _unconditionally_ writes the new m_flags field, and writes it at the same offset where the name string is supposed to start.
As a result, TypeInfo names for structs are broken. The first 4 items of typeid(somestruct).toString() will contain garbage. I'm not really sure why after these 4 bytes, the expected string starts. Either I got it wrong, or the backend data segment writer does strange things?
The problem doesn't appear if the object.d[i] contains the proper m_align. E.g. compare:
http://www.dsource.org/projects/phobos/browser/branches/phobos-1.x/phobos/object.d#L141http://www.dsource.org/projects/tango/browser/trunk/object.di#L163 (WARNING: BSD LICENSED CODE)
Comment #1 by nfxjfg — 2010-11-19T08:20:24Z
(In reply to comment #0)
> At some point, both dmd and Phobos1 added a new member to TypeInfo_Struct:
> m_flags.
Damn, that should be m_align.