xref: /third_party/libabigail/ABIXML-FORMAT-VERSIONS

2.1 version
===========

This occurred in libabigail 2.1.  The IR doesn't contain no-op
qualified types anymore.  These were the ancient way to represent
qualified type that were edited by libabigail to avoid spurious
diagnostics. E.g, const references or const void type.  In those
cases, we were replacing these by qualified types that were no-ops.
Those no-ops qualified types further needed to be stripped off in the
diff IR to avoid emitting spurious changes.  It turned out stripping
out those no-op qualified types was causing some difficulties in the
algorithms that propagate type diff node categorization.

For instance, variable diff type node would have the categorization of
its diff->type_diff() be different from what we could see from the
change observed between its diff->first_subject()->get_type() and
diff->second_subject()->get_type().  This is because diff->type_diff()
strips the no-op qualified types to create a new diff IR node, whereas
looking at the difference between diff->first_subject->get_type() and
diff->second_subject()->get_type() gives a different perspective.

So looking at the "changes" carried by 'diff' could yield a result
suggesting that there was a change to be diagnosed, carried by the
difference between diff->first_subject()->get_type() and
diff->second_subject()->get_type() -- a difference between two no-ops
qualified types -- whereas actually looking at diff->type_diff() (that
should be equivalent) yields a different information because it's
stripped off of the no-op qualified type.

Just removing the no-op qualified types removed all those issues
altogether.  So in concrete terms, when we see a const reference, it's
replaced by a reference (because a reference is always const) and
when we see a const void, it's replaced by void.  As simple as that.

But then the emitted ABIXML won't carry any no-op qualified type
anymore. Older versions of ABIXML might still carry those, and so
abidiff-ing those against their original binary might yield some
spurious diagnostics.  Hence the minor version bump.

relevant commit:
----------------
commit 8a4f72005bc96fecf63414784cf94857eddbd124
Author: Dodji Seketeli <dodji@redhat.com>
Date:   Thu Nov 4 10:59:32 2021 +0100

    Bug 28450 - Fix cloned member function handling in DWARF
2.0 version
===========

This occured in libabigail 2.0.  The interpretation of
DW_AT_bit_offset (DWARF 5) changed so the value of the
'layout-offset-in-bits' property of the 'data-member' element might
now have changed in an incompatible way.

relevant commit:
----------------
commit 23046152e0843bc2f141e37dcce047305c3f4379
Author: Dodji Seketeli <dodji@redhat.com>
Date:   Thu Oct 22 16:04:08 2020 +0200

    Bump ABIXML format version to 2.0


1.0 version
===========

This was the initial version of abixml.