Comment # 28 on bug 1207989 from Egbert Eich

(In reply to Richard Biener from comment #26)
> (In reply to Richard Biener from comment #25)
> > Since the lapacke TMG interfaces are part of the API (they are declared in
> > the
> > header) I would argue we should include them, thus define LAPACKE_WITH_TMG
> > even in Factory.
> 
> Note the original issue was likely that liblapacke.so.3 contained the testing
> functions but they refer to LAPACK_zlatms and friends which were not included
> and the fix was to include all of TMGLIB into liblapack.so.3 (instead of
> just the strictly required subset to make --no-undefined work).

LAPACKE_WITH_TMG=yes will do exactly that: build the TMGLIB C-Bindings into
LAPACKE. It will not build ltmglib - nor include it into the lapack library.

Looking at all this, it seems like the c-bindings for tmglib have become part
of the LAPACKE API (by virtue of not cleanly separated headers), these test
functions require tmglib (the Fortran library). Thus, by shipping LAPACKE,
there is probably no choice but to ship tmglib as well (probably separately,
not as part of liblapack).