The current DSO dependency sorting tests are for a limited number of
specific cases, including some from particular bug reports.
Add tests that systematically cover all possible DAGs for an
executable and the shared libraries it depends on, directly or
indirectly, up to four objects (an executable and three shared
libraries). (For this kind of DAG - ones with a single source vertex
from which all others are reachable, and an ordering on the edges from
each vertex - there are 57 DAGs on four vertices, 3399 on five
vertices and 1026944 on six vertices; see
https://arxiv.org/pdf/2303.14710 for more details on this enumeration.
I've tested that the 3399 cases with five vertices do all pass if
enabled.)
These tests are replicating the sorting logic from the dynamic linker
(thereby, for example, asserting that it doesn't accidentally change);
I'm not claiming that the logic in the dynamic linker is in some
abstract sense optimal. Note that these tests do illustrate how in
some cases the two sorting algorithms produce different results for a
DAG (I think all the existing tests for such differences are ones
involving cycles, and the motivation for the new algorithm was also to
improve the handling of cycles):
tst-dso-ordering-all4-44: a->[bc];{}->[cba]
output(glibc.rtld.dynamic_sort=1): c>b>a>{}<a<b<c
output(glibc.rtld.dynamic_sort=2): b>c>a>{}<a<c<b
They also illustrate that sometimes the sorting algorithms do not
follow the order in which dependencies are listed in DT_NEEDED even
though there is a valid topological sort that does follow that, which
might be counterintuitive considering that the DT_NEEDED ordering is
followed in the simplest cases:
tst-dso-ordering-all4-56: {}->[abc]
output: c>b>a>{}<a<b<c
shows such a simple case following DT_NEEDED order for destructor
execution (the reverse of it for constructor execution), but
tst-dso-ordering-all4-41: a->[cb];{}->[cba]
output: c>b>a>{}<a<b<c
shows that c and b are in the opposite order to what might be expected
from the simplest case, though there is no dependency requiring such
an opposite order to be used.
(I'm not asserting that either of those things is a problem, simply
observing them as less obvious properties of the sorting algorithms
shown up by these tests.)
Tested for x86_64.
