Bug 24431 Treat __ksymtab as int32_t for v4.19+ kernels
Calculating the relocation for values in __ksymtab with GElf_Addr (i.e.
uint64_t), makes the calculation rely on overflows for negative offsets.
Address that by treating these as 32bit signed values for the v4.19+
__ksymtabs and calculate the offset with them. This also allows, similar
an earlier commit, to drop the distinction between 64bit and 32bit
kernels.
* src/abg-dwarf-reader.cc (maybe_adjust_sym_address_from_v4_19_ksymtab):
treat passed addr as 32bit signed offset in case of v4.19+ __ksymtabs
Bug 24431 Read 32bit values when testing for the v4.19 symbol table format
Reading into uint64_t when reading the symbol table values drops the
sign and subsequently offset calculations will only be correct if either
the offset is positive or if the calculation overflows.
Read the relative value as signed int32_t (indepently of the target's
bitness) to allow negative offsets. That also allows to drop the code
that formerly handled the overflow.
That change fixes an assertion raised when dealing with aarch64 kernel
binaries that have a __ksymtab with 32bit relocations. i.e. Bug #24431
* src/abg-dwarf-reader.cc
(try_reading_first_ksymtab_entry_using_v4_19_format): attempt to
read first __ksymtab entry into int32_t to preserve sign
Dodji Seketeli [Thu, 9 May 2019 13:19:05 +0000 (15:19 +0200)]
Better handle several anonymous types of the same kind
This is a follow-up patch for the commit:
43d56de Handle several member anonymous types of the same kind
It allows support for severan anonymous types even when these are not
members of a class/unions.
The patch introduces the concept of a scoped name. It's a qualified
name for a decl made of the name of the decl appended to the
*unqualified* name of its scope. Unlike for qualified names, the
scoped name won't have a "__anonymous_*__" string in its name if its
directly containing scope is not anonymous; a qualified name might
still have that string in its name because the decl has a parent scope
(not necessarily its directly containing scope though) that is
anonymous.
The patch goes on to update the logic for comparison of decls that are
anonymous. For a decl which direct scope is *NOT* anonymous, the
scoped name is what's used in the comparison. Otherwise, only the
name of the decl is used.
The patch also updates how we detect changes in data members and
member types, in the comparison engine. It now uses the names of the
data members, rather than their qualified name. This is in the scope
of the current class/union anyway. The improvement is that the fact
that the class/union itself is anonymous (even if its anonymous name
changes to another anonymous name) won't have any spurious impact on
the detection of name change of the members.
The patch considers the change of an anonymous decl name which
anonymous name changes to another anonymous name as being harmless.
The patch updates the logic of category propagation in the comparison
engine. Although a public typedef to private underlying type needs to
stay public and thus not propagate the PRIVATE_TYPE_CATEGORY from its
child diff node to himself, it still needs to suppress the changes to
the private underlying diff node that were suppressed (because of the
private-ness), unless that typedef has local changes.
* include/abg-ir.h (decl_base::get_scoped_name): Declare new
member function.
(scope_decl::get_num_anonymous_member_{classes, unions, enums}):
Declare new virtual member functions.
(class_decl::get_num_anonymous_member_{classes, unions, enums}):
Adjust to make these virtual. It's not necessary but I feel
redundancy is a kind of self-documentation here.
* src/abg-comp-filter.cc (has_harmless_name_change): Consider
anonymous name changes as harmless.
* src/abg-comparison.cc
(class_or_union_diff::ensure_lookup_tables_populated): Consider
the names of the members rather than their qualified names.
(suppression_categorization_visitor::visit_end): Suppress the
changes to the private underlying diff node that were suppressed
because of the private-ness, unless that typedef has local
changes.
* src/abg-dwarf-reader.cc (build_enum_type)
(add_or_update_class_type, add_or_update_union_type): Handle
anonymous types in namespaces as well, not just in class/unions.
* src/abg-ir.cc (decl_base::priv::scoped_name_): Define new data
member.
(decl_base::get_scoped_name): Define new member function.
(equals): For the decl_base overload, use scoped name in the
comparison, unless the decl belongs to an anonymous type. For the
class_or_union_diff, only consider scoped_name during comparison.
Avoid name comparison between anonymous types.
(scope_decl::get_num_anonymous_member_{classes, unions, enums}):
Define new member functions.
(types_have_similar_structure): Do not compare names between
anonymous types.
(qualified_name_setter::do_update): Update scoped names too.
* tests/data/test-abidiff/test-PR18791-report0.txt: Adjust.
* tests/data/test-annotate/libtest23.so.abi: Likewise.
* tests/data/test-annotate/test13-pr18894.so.abi: Likewise.
* tests/data/test-annotate/test14-pr18893.so.abi: Likewise.
* tests/data/test-annotate/test15-pr18892.so.abi: Likewise.
* tests/data/test-annotate/test21-pr19092.so.abi: Likewise.
* tests/data/test-diff-dwarf/test43-PR22913-report-0.txt:
Likewise.
* tests/data/test-diff-dwarf/test46-rust-report-0.txt: Likewise.
* tests/data/test-diff-filter/test30-pr18904-rvalueref-report0.txt:
Likewise.
* tests/data/test-diff-filter/test30-pr18904-rvalueref-report1.txt:
Likewise.
* tests/data/test-diff-filter/test30-pr18904-rvalueref-report2.txt:
Likewise.
* tests/data/test-diff-filter/test31-pr18535-libstdc++-report-0.txt:
Likewise.
* tests/data/test-diff-filter/test31-pr18535-libstdc++-report-1.txt:
Likewise.
* tests/data/test-diff-filter/test33-report-0.txt: Likewise.
* tests/data/test-diff-filter/test35-pr18754-no-added-syms-report-0.txt:
Likewise.
* tests/data/test-diff-filter/test44-anonymous-data-member-report-0.txt:
Likewise.
* tests/data/test-diff-pkg/libsigc++-2.0-0c2a_2.4.0-1_amd64--libsigc++-2.0-0v5_2.4.1-1ubuntu2_amd64-report-0.txt:
Likewise.
* tests/data/test-diff-pkg/nss-3.23.0-1.0.fc23.x86_64-report-0.txt:
Likewise.
* tests/data/test-diff-pkg/spice-server-0.12.4-19.el7.x86_64-0.12.8-1.el7.x86_64-report-0.txt:
Likewise.
* tests/data/test-diff-pkg/spice-server-0.12.4-19.el7.x86_64-0.12.8-1.el7.x86_64-report-1.txt:
Likewise.
* tests/data/test-diff-pkg/spice-server-0.12.4-19.el7.x86_64-0.12.8-1.el7.x86_64-report-2.txt:
Likewise.
* tests/data/test-diff-pkg/spice-server-0.12.4-19.el7.x86_64-0.12.8-1.el7.x86_64-report-3.txt:
Likewise.
* tests/data/test-read-dwarf/PR22015-libboost_iostreams.so.abi:
Likewise.
* tests/data/test-read-dwarf/PR22122-libftdc.so.abi: Likewise.
* tests/data/test-read-dwarf/libtest23.so.abi: Likewise.
* tests/data/test-read-dwarf/test10-pr18818-gcc.so.abi: Likewise.
* tests/data/test-read-dwarf/test11-pr18828.so.abi: Likewise.
* tests/data/test-read-dwarf/test12-pr18844.so.abi: Likewise.
* tests/data/test-read-dwarf/test13-pr18894.so.abi: Likewise.
* tests/data/test-read-dwarf/test14-pr18893.so.abi: Likewise.
* tests/data/test-read-dwarf/test15-pr18892.so.abi: Likewise.
* tests/data/test-read-dwarf/test16-pr18904.so.abi: Likewise.
* tests/data/test-read-dwarf/test21-pr19092.so.abi: Likewise.
* tests/data/test-read-dwarf/test22-pr19097-libstdc++.so.6.0.17.so.abi:
Likewise.
* tests/data/test-read-dwarf/test9-pr18818-clang.so.abi: Likewise.
Dodji Seketeli [Tue, 7 May 2019 08:22:29 +0000 (10:22 +0200)]
Handle several member anonymous types of the same kind
When there are several anonymous types (e.g, anonymous classes, unions
or enums) in a given class or union, libabigail's internals do
struggle.
An anonymous class, for instance, is named __anonymous_struct__. When
there are more than one of these inside a given class, then we can't
name and look them up, because they all have the same name.
Incidentally, when add_or_update_class_type completes a class type
that was initially constructed before, it fails to determine that an
anonymous member type of that class was already present in that
context. It thus wrongly duplicates anonymous structs/unions/enums in
there and that leads to spurious textual (abixml) representation
differences later, where duplicated anonymous member types would
appear intermittently, depending on the order in which the class was
built.
This patch addresses this general issue by naming anonymous member
types in a way that allows several of them to exist. That is, if there
are two anonymous structs in a class, they are going to be named
__anonymous_struct__ and __anonymous_struct__1. We do follow a
similar scheme for anonymous unions and enums. This is handled by the
DWARF reader that builds the internal representation.
While looking at this issue, I also fixed a tangent bug; some DWARF
emitters wrongly *define* types in the scope of a
DW_TAG_subroutine_type or DW_TAG_array_type. We handle that by
actually defining those types in the scope of that subroutine or
array. But then it appears that if that scope itself a class and if
the type defined is an anonymous type, then putting that anonymous
type in the class scope might interfere with the *naming* of the
existing legit anonymous types of that scope. I decided to put those
anonymous types in the containing namespace instead. We'll see how
that goes in real time use.
The patch also updates lots of existing tests and adds a new one.
* include/abg-ir.h
(class_or_union::get_num_anonymous_member_{classes, unions,
enums}): Declare new member functions.
* src/abg-dwarf-reader.cc (get_internal_anonynous_die_base_name)
(build_internal_anonymous_die_name)
(get_internal_anonymous_die_name, is_anonymous_type_die): Define
new static functions.
(die_qualified_type_name): Use the new
get_internal_anonymous_die_name.
(get_scope_for_die): Fix this to put anonymous types that were
wrongly emitted into the scope of DW_TAG_subroutine_type or
DW_TAG_array_type by buggy DWARF emitters into the enclosing
namespace, rather than into the enclosing class/union.
(build_enum_type): Take the scope of the enum to have a chance to
properly name potential anonymous enums.
(lookup_class_typedef_or_enum_type_from_corpus): Take an anonymous
member type index for when the DIE we are lookup up represents an
anonymous type. Support proper building of the internal anonymous
name of the anonymous type we are lookup up.
(add_or_update_class_type): Use the new
get_internal_anonynous_die_base_name and
build_internal_anonymous_die_name functions. Support making sure
that the anonymous member type we are adding to the class wasn't
already there, especially for cases where we are updating a class
type.
(add_or_update_union_type): Use the new
get_internal_anonynous_die_base_name and
build_internal_anonymous_die_name functions.
(build_ir_node_from_die): Adjust the use of build_enum_type to
pass it the scope of the enum type we are building.
* src/abg-ir.cc (lookup_union_type): Add a new overload.
(lookup_class_or_typedef_type): Use the new overload of
lookup_union_type above to support looking up union types too.
(class_or_union::get_num_anonymous_member_{classes, unions,
enums}): Define new member functions.
* src/abg-reporter-priv.cc (represent): Detect when anonymous
types of anonymous data members have their internal names change,
probably because anonymous member types were inserted in the scope.
* tests/data/Makefile.am: Add the new test-anonymous-members-0.*
test input files to the source distribution.
* tests/data/test-annotate/test-anonymous-members-0.cc: New test
input file.
* tests/data/test-annotate/test-anonymous-members-0.o: Likewise.
* tests/data/test-annotate/test-anonymous-members-0.o.abi: Likewise.
* tests/data/test-annotate/test17-pr19027.so.abi: Adjust.
* tests/data/test-annotate/test18-pr19037-libvtkRenderingLIC-6.1.so.abi:
Likewise.
* tests/data/test-annotate/test19-pr19023-libtcmalloc_and_profiler.so.abi:
Likewise.
* tests/data/test-annotate/test21-pr19092.so.abi: Likewise.
* tests/data/test-diff-filter/test30-pr18904-rvalueref-report0.txt:
Likewise.
* tests/data/test-diff-filter/test30-pr18904-rvalueref-report1.txt
* tests/data/test-diff-filter/test30-pr18904-rvalueref-report2.txt:
Likewise.
* tests/data/test-diff-filter/test35-pr18754-no-added-syms-report-0.txt:
Likewise.
* tests/data/test-read-dwarf/PR22122-libftdc.so.abi: Likewise.
* tests/data/test-read-dwarf/test12-pr18844.so.abi: Likewise.
* tests/data/test-read-dwarf/test16-pr18904.so.abi: Likewise.
* tests/data/test-read-dwarf/test17-pr19027.so.abi: Likewise.
* tests/data/test-read-dwarf/test18-pr19037-libvtkRenderingLIC-6.1.so.abi:
Likewise.
* tests/data/test-read-dwarf/test19-pr19023-libtcmalloc_and_profiler.so.abi:
Likewise.
* tests/data/test-read-dwarf/test21-pr19092.so.abi: Likewise.
* tests/data/test-read-dwarf/test22-pr19097-libstdc++.so.6.0.17.so.abi:
Likewise.
* tests/test-annotate.cc (int_out_specs): Add the new test inputs
to this test harness.
Use canonical types hash maps for type IDs in abixml writer
This patch stores canonical types in the hash maps that are used in
the abixml writer to compute the type IDs. This limits the
possibility that two types that are equivalent (especially when one is
a declaration-only class) end-up in the same abixml.
* src/abg-writer.cc (write_context::{type_has_existing_id,
get_id_for_type}): Save the canonical type of the type in the map,
not the type itself.
(write_context::{type_is_emitted}): Use the canonical type rather
than the type itself.
* tests/data/test-read-dwarf/test9-pr18818-clang.so.abi: Adjust.
Don't try to de-duplicate all anonymous struct DIEs
Trying to de-duplicate anonymous struct DIEs can lead to subtle
issues, because there can be two different naming typedefs designating
two anonymous structs that are equivalent, in the same translation
unit. In that case, de-duplicating the two leaf anonymous structs
DIEs leads to non-resolvable conflict.
This patch avoids de-duplicating anonymous structs DIEs and rather
de-duplicates (naming) typedefs.
* include/abg-fwd.h (is_typedef): Remove the overloads for
type_base_sptr and decl_base_sptr. Replace those with an overload
for type_or_decl_base_sptr.
* src/abg-ir.cc (is_typedef): Do the same for the definitions.
* src/abg-dwarf-reader.cc (add_or_update_class_type)
(add_or_update_union_type): Do not de-duplicate anonymous
struct/union DIEs.
(build_typedef_type): Try to de-duplicate typedefs DIEs.
* tests/data/test-annotate/test17-pr19027.so.abi: Adjust.
* tests/data/test-annotate/test19-pr19023-libtcmalloc_and_profiler.so.abi:
Likewise
* tests/data/test-annotate/test21-pr19092.so.abi: Likewise.
* tests/data/test-read-dwarf/PR22015-libboost_iostreams.so.abi: Likewise.
* tests/data/test-read-dwarf/PR22122-libftdc.so.abi: Likewise.
* tests/data/test-read-dwarf/test16-pr18904.so.abi: Likewise.
* tests/data/test-read-dwarf/test17-pr19027.so.abi: Likewise.
* tests/data/test-read-dwarf/test19-pr19023-libtcmalloc_and_profiler.so.abi:
Likewise.
* tests/data/test-read-dwarf/test21-pr19092.so.abi: Likewise.
* tests/data/test-read-dwarf/test9-pr18818-clang.so.abi: Likewise.
During DWARF reading, it can happen that some types are created that
are not tied to a given DIE. This happens for instance when editing
the internal representation of types to better reflect the intent of
the code rather than the letter of the code; more precisely this
happens when a qualified array is edited to represent an array of
qualified elements, for instance. In that case, the qualified element
type is created by the DWARF analyser. And in the present incarnation
of the code, we forget to canonicalize that type.
The fact that that type is not canonicalized leads to cases where two
equivalent types can still be present in a translation unit in the
resulting abixml representation, without the abixml writer noticing
it. That duplication can lead to unnecessary cycles in the type graph
where passes can either take one side or the other of the cycle; and
that leads to subtle heisen-changes in the emitted abixml representation.
This patch thus canonicalizes those newly created types.
* src/abg-dwarf-reader.cc
(read_context::extra_types_to_canonicalize_): Add new data member.
(read_context::{initialize, clear_types_to_canonicalize}): Adjust.
(read_context::extra_types_to_canonicalize): Create new accessor.
(read_context::schedule_type_for_late_canonicalization): Add new
overload for type_base_sptr.
(read_context::perform_late_type_canonicalizing): Perform the
canonicalization of the types created by the DWARF analyzer, but
that are not tied to any DIE.
(maybe_strip_qualification): Take a read_context&. Schedule newly
created types (during type edition) for late canonicalization.
(build_ir_node_from_die): Adjust the call to
maybe_strip_qualification to pass a read_context.
* tests/data/test-annotate/test15-pr18892.so.abi: Adjust.
* tests/data/test-annotate/test17-pr19027.so.abi: Likewise.
* tests/data/test-annotate/test21-pr19092.so.abi: Likewise.
* tests/data/test-read-dwarf/PR22122-libftdc.so.abi: Likewise.
* tests/data/test-read-dwarf/test12-pr18844.so.abi: Likewise.
* tests/data/test-read-dwarf/test15-pr18892.so.abi: Likewise.
* tests/data/test-read-dwarf/test17-pr19027.so.abi: Likewise.
* tests/data/test-read-dwarf/test21-pr19092.so.abi: Likewise.
* tests/data/test-read-dwarf/test22-pr19097-libstdc++.so.6.0.17.so.abi:
Likewise.
Mark Wielaard [Thu, 18 Apr 2019 15:41:41 +0000 (17:41 +0200)]
Don't try to read a build_id as string in find_alt_debug_info_link.
The GCC8 address sanitizer found an issue in find_alt_debug_info_link.
It tried to convert a build-id byte sequence into a string. But the
build-id byte sequence is not a zero terminated sequence of chars.
So it could run off way past the section data.
The code never actually uses the build-id. It could use it to verify
the referenced alt-file is the correct one. But since it uses elfutils
to actually load the alt file it doesn't have to, since elfutils will
already check the build-id matches.
So just remove the build_id argument from find_alt_debug_info_link
and don't try to convert and return it as a string.
* src/abg-dwarf-reader.cc (find_alt_debug_info_link): Remove
build_id argument. Don't try to read the buildid chars as a
string.
(find_alt_debug_info): Don't call find_alt_debug_info_link
with a build_id string argument.
Signed-off-by: Mark Wielaard <mark@klomp.org> Signed-off-by: Dodji Seketeli <dodji@redhat.com>
dwarf-reader: fix undefined behaviour in get_binary_load_address
Within the loop, the call `gelf_getphdr(elf_handle, i, &ph_mem)` is
returning a pointer to `ph_mem` that is only valid in this loop
iteration. The later assignment to *lowest_program_header and its
eventual use to assign load_address leads to undefined behaviour.
* src/abg-dwarf-reader.cc (get_binary_load_address): Move the
ph_mem and program_header variables out of the inner for-loop.
Delay canonicalization for array and qualified types
Array and qualified types can be edited after they are built, e.g, to
fold the qualifiers of the array as those should always apply to the
element types of the array, at least in C and C++. So we need to
delay the canonicalization of these types until *after* all that type
editing is done. We were not doing that properly and I suspect this
is the cause of some "heisenregressions" we are seeing on some
platforms intermittently.
This patch does delay the type canonicalization and insures that we
don't edit types that are canonicalized already.
* src/abg-dwarf-reader.cc (maybe_canonicalize_type): Delay the
canonicalization of array and qualified types, just like what we
do for classes and function types already.
(maybe_strip_qualification): Do not
re-canonicalize array and qualified types here because it should
not be necessary anymore.
The problem is that the method used to determine if the format of the
__ksymtab section is pre or post v4.19. In that commit, we assumed
that a kernel without relocations for the __ksymtab section implies a
v4.19 format, because in that format, we don't need relocations for
that section. But then it can happen that the entire kernel be built
without relocations, in the case a non relocatable kernels. So just
looking a the presence of relocations is not a good enough
discriminant to determine the format of the __ksymtab section.
This patch rather tries to read the first entry of the __ksymtab
section assuming a pre v4.19 format, comes up with an address, and
then looks the .symtab section up with that address. If the lookup
succeeds and we find a symbol, then we can reasonably assume that the
__ksymtab section is in the v4.19 format. Otherwise, we do the same
assuming a v4.19 format, etc.
Tested on v4.9, v4.16 and v4.16 kernels in 32 and 64 bits.
* src/abg-dwarf-reader.cc
(read_context::{try_reading_first_ksymtab_entry_using_pre_v4_19_format,
try_reading_first_ksymtab_entry_using_v4_19_format}): Define new
member functions.
(read_context::maybe_adjust_sym_address_from_v4_19_ksymtab): Make
member function this const.
(read_context::get_ksymtab_format): Implement the new heuristic
here, using try_reading_first_ksymtab_entry_using_pre_v4_19_format
and try_reading_first_ksymtab_entry_using_v4_19_format, rather
than assuming that if we have no relocations, then we are in the
v4.19 format.
(maybe_adjust_sym_address_from_v4_19_ksymtab): When on a 64 bits
architecture, ensure the 32 bits address read from the v4.19
format __ksymtab section is converted into a 64 bits address.
dwarf-reader: fix recursion in expr_result::operator&
operator& was implemented in terms of itself, leading to infinite
recursion. Fix that by implementing it in terms of &='ing the const
value. That is consistent with all other ?= operators.
distinct_diff: avoid expression with side effects within typeid
When compiling with clang, the following warning is emitted:
abg-comparison.cc:2674:17: warning: expression with side effects will be
evaluated despite being used as an operand to 'typeid' [-Wpotentially-evaluated-expression]
return typeid(*first.get()) != typeid(*second.get());
^
Mitigate that warning by moving potential side effects out of typeid.
ir: drop unused data members from {environment,qualified_name}_setter
The data members environment_setter::artifact_ and
qualified_name_setter::node_ are not used and can therefore be dropped
along with all their references.
Several virtual desctructors were missing. Even though there might not
have been actual leaks or similar bugs, it is worth fixing these
locations as they might lead to bugs in the future.
Clang also warns at these locations:
warning: delete called on non-final 'abigail::ir::corpus' that has virtual
functions but non-virtual destructor [-Wdelete-non-virtual-dtor]
* include/abg-comparison.h: add virtual destructor for corpus_diff and diff_node_visitor
* include/abg-corpus.h: add virtual destructor for corpus
* include/abg-reporter.h: add virtual destructor for reporter_base
* include/abg-traverse.h: add virtual destructor for traversable_base
The postfix increment / decrement operators were implemented by calling
themselves recursively. Fix that by implementing these in terms of their
prefix counter parts.
When compiling with clang, the following warning is emitted:
abg-reader.cc:1981:15: warning: using the result of an assignment as a condition without parentheses [-Wparentheses]
while (corp = read_corpus_from_input(ctxt))
~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
That is certainly a common pitfall and can be mitigated by placing
parentheses around the assignment.
* src/abg-reader.cc: clarify boolean use of assignment
Since the commit
https://github.com/torvalds/linux/commit/7290d58095712a89f845e1bca05334796dd49ed2
of the kernel, the format of the ksymtab section changed, so the ELF
reader of Libabigail cannot get the list of kernel symbols that are
meaningful to analyze.
In the new format, each ksymtab entry is now 32 bits length,
independently of if the kernel was built in 32 or 64 bits mode. This
way, 64 bit kernels save 32 bits per entry, from the ksymtab section
alone.
Also, note that the symbol address in the ksymtab entry is stored in a
"place-relative address" format. That means for a given symbol value
'Sym_Addr', the value which is stored at a given offset 'Off' of the
ksymtab secthion which address is Ksymtab_Addr is:
Sym_Addr - Ksymtab_Addr - Off.
This is what is meant by storing Sym_Addr in a "place-relative" manner.
One advantage, of course, is to make Sym_Addr take 32 bits, even on a
64 bits arch. Another advantage is that the resulting value doesn't
need to be relocated! So we don't need any relocation section either!
So lots of space saving.
This patch teaches the ELF reader into this new format.
* src/abg-dwarf-reader.cc (enum kernel_symbol_table_kind): Move this
enum at the top.
(enum ksymtab_format): Define new enum.
(read_context::{ksymtab_format_, ksymtab_entry_size_,
nb_ksymtab_entries_, nb_ksymtab_gpl_entries_}): Define new data
members.
(read_context::initiliaze): Initialize the new data members above.
(read_context::{get_ksymtab_format, get_ksymtab_symbol_value_size,
get_ksymtab_entry_size, get_nb_ksymtab_entries,
get_nb_ksymtab_gpl_entries,
maybe_adjust_sym_address_from_v4_19_ksymtab}): Define new member
functions.
(read_context::load_kernel_symbol_table): Support loading from
both pre and post v4.19 linux kernels with their different ksymtab
formats. Add more comments.
In C and C++ the qualifiers of a qualified array type apply to the
element type of said array.
In this particular problem report, GCC and Clang emit DWARF that
represent the qualifiers either on the array type, or on the element
type. Quoting the test of the bug:
Given the following example:
struct test {
const char asdf[4];
};
void func(struct test arg) {}
abidiff says:
1 function with some indirect sub-type change:
[C]'function void func(test)' at test.c:5:1 has some indirect sub-type
changes:
parameter 1 of type 'struct test' has sub-type changes:
type size hasn't changed
1 data member change:
type of 'const char test::asdf[4]' changed:
entity changed from 'const char[4]' to 'const char[4] const'
This is because GCC represents the array as const array of const
signed char, whereas clang represents it as an array of const signed
char.
In this patch, libabigail's DWARF reader detects qualified array types
and appropriately qualifies the array element type, instead of
qualifying the array type. The patch accordingly adjusts the various
regression tests and adds a new test which comes from the problem
report.
* include/abg-fwd.h (is_array_of_qualified_element): Declare 2
overloads of this function.
(re_canonicalize): Declare a new function.
* include/abg-ir.h (class {decl_base, type_base}): Declare
re_canonicalize as a friend of these classes.
* src/abg-dwarf-reader.cc (maybe_strip_qualification): Detect
qualified array types and appropriately qualifies the array
element type, instead of qualifying the array type itself.
Re-canonicalize the resulting type if necessary.
* src/abg-ir.cc (is_array_of_qualified_element): Define 2
overloads of this function.
(re_canonicalize): Define new function.
* tests/data/Makefile.am: The two new test binary input files
PR24430-fold-qualified-array-clang and
PR24430-fold-qualified-array-gcc to source distribution, as well
as the expected reference output.
* tests/data/test-annotate/test15-pr18892.so.abi: Adjust.
* tests/data/test-annotate/test17-pr19027.so.abi: Likewise.
* tests/data/test-annotate/test19-pr19023-libtcmalloc_and_profiler.so.abi:
Likewise.
* tests/data/test-annotate/test21-pr19092.so.abi: Likewise.
* tests/data/test-diff-filter/PR24430-fold-qualified-array-clang:
New binary test input coming from the bug report.
* tests/data/test-diff-filter/PR24430-fold-qualified-array-gcc:
Likewise.
* tests/data/test-diff-filter/PR24430-fold-qualified-array-report-0.txt:
Expected reference abi difference.
* tests/data/test-diff-filter/test33-report-0.txt: Adjust.
* tests/data/test-diff-pkg/nss-3.23.0-1.0.fc23.x86_64-report-0.txt:
Likewise.
* tests/data/test-read-dwarf/PR22122-libftdc.so.abi: Likewise.
* tests/data/test-read-dwarf/test12-pr18844.so.abi: Likewise.
* tests/data/test-read-dwarf/test15-pr18892.so.abi: Likewise.
* tests/data/test-read-dwarf/test17-pr19027.so.abi: Likewise.
* tests/data/test-read-dwarf/test19-pr19023-libtcmalloc_and_profiler.so.abi:
Likewise.
* tests/data/test-read-dwarf/test21-pr19092.so.abi: Likewise.
* tests/data/test-read-dwarf/test22-pr19097-libstdc++.so.6.0.17.so.abi:
Likewise.
* tests/test-diff-filter.cc: Add the new binary test input to this
test harness.
1b83138 Propagate private type diff category through refs/qualified type diffs dc84fee Fix anonymous union constructed under the wrong context 522ac25 Internal pretty repr of union cannot be flat representation
But then I forgot to add a regression test for that issue.
This patch does that.
* tests/data/test-diff-pkg/PR24410-new/poppler-debuginfo-0.73.0-8.fc30.x86_64.rpm:
Add new test input.
* tests/data/test-diff-pkg/PR24410-new/poppler-qt5-0.73.0-8.fc30.x86_64.rpm:
Add new test input.
* tests/data/test-diff-pkg/PR24410-new/poppler-qt5-debuginfo-0.73.0-8.fc30.x86_64.rpm:
Add new test input.
* tests/data/test-diff-pkg/PR24410-new/poppler-qt5-devel-0.73.0-8.fc30.x86_64.rpm:
Add new test input.
* tests/data/test-diff-pkg/PR24410-old/poppler-debuginfo-0.73.0-4.fc30.x86_64.rpm:
Add new test input.
* tests/data/test-diff-pkg/PR24410-old/poppler-qt5-0.73.0-4.fc30.x86_64.rpm:
Add new test input.
* tests/data/test-diff-pkg/PR24410-old/poppler-qt5-debuginfo-0.73.0-4.fc30.x86_64.rpm:
Add new test input.
* tests/data/test-diff-pkg/PR24410-old/poppler-qt5-devel-0.73.0-4.fc30.x86_64.rpm:
Add new test input.
* tests/data/test-diff-pkg/PR24410-report-0.txt: Add new test
input.
* tests/data/Makefile.am: Add the test input above to source
distribution.
* tests/test-diff-pkg.cc: Make this test harness use the new input
rpms above.
Propagate private type diff category through refs/qualified type diffs
This patch is the third of the series:
Internal pretty repr of union cannot be flat representation
Fix anonymous union constructed under the wrong context
Propagate private type diff category through refs/qualified type diffs
The intent of this series is to fix the bug:
https://sourceware.org/bugzilla/show_bug.cgi?id=24410
"Empty change report emitted for libpoppler-qt5.so.1.18.0"
We (mistakenly) don't propagate private type diff categories through
reference and qualified type diffs. This leads to some diff nodes not
being suppressed just because they are private type diffs which
category weren't properly propagated.
This patch fixes this.
Note that the tests updated in this patch reflect the regression tests
changes needed for the entire set of 3 patches.
* src/abg-comparison.cc
(suppression_categorization_visitor::visit_end): Propagate
suppressed and private type diff categories for reference and
qualified types. For qualified types, make sure they don't have
local changes. Even when there are no local changes, do not
propagate private diff categories to typedefs.
* tests/data/test-annotate/test17-pr19027.so.abi: Adjust.
* tests/data/test-annotate/test21-pr19092.so.abi: Likewise.
* tests/data/test-read-dwarf/PR22122-libftdc.so.abi: Likewise.
* tests/data/test-read-dwarf/test10-pr18818-gcc.so.abi: Likewise.
* tests/data/test-read-dwarf/test11-pr18828.so.abi: Likewise.
* tests/data/test-read-dwarf/test12-pr18844.so.abi: Likewise.
* tests/data/test-read-dwarf/test16-pr18904.so.abi: Likewise.
* tests/data/test-read-dwarf/test17-pr19027.so.abi: Likewise.
* tests/data/test-read-dwarf/test21-pr19092.so.abi: Likewise.
* tests/data/test-read-dwarf/test22-pr19097-libstdc++.so.6.0.17.so.abi: Likewise.
* tests/data/test-read-dwarf/test9-pr18818-clang.so.abi: Likewise.
Fix anonymous union constructed under the wrong context
This patch is second of the series:
Internal pretty repr of union cannot be flat representation
Fix anonymous union constructed under the wrong context
Propagate private type diff category through refs/qualified type diffs
The intent of this series is to fix the bug:
https://sourceware.org/bugzilla/show_bug.cgi?id=24410
"Empty change report emitted for libpoppler-qt5.so.1.18.0"
What happens here is that when the DWARF reader sees an anonymous
union/struct, it can mistakenly think that it has seen it before
(because the comparison doesn't take the scope of the union/struct
into account), and thus mistakenly represent the union/struct.
The solution implemented by this patch is to take the scope of the
anonymous union/struct into account.
Note that regression tests are all updated in the last patch of the
series.
* src/abg-dwarf-reader.cc (add_or_update_class_type)
(add_or_update_union_type): Only reuse anonymous class/union types
which have the same scope as the current one.
Internal pretty repr of union cannot be flat representation
This is the first patch of this series:
Internal pretty repr of union cannot be flat representation
Fix anonymous union constructed under the wrong context
Propagate private type diff category through refs/qualified type diffs
The intent of this series is to fix the bug:
https://sourceware.org/bugzilla/show_bug.cgi?id=24410
"Empty change report emitted for libpoppler-qt5.so.1.18.0"
The internal pretty representation of a union must be its fully
qualified name, even when it's a anonymous union. It cannot be its
flat representation as for anonymous unions, that would lead to
confusion between anonymous unions that have the same flat
representation but are in different scopes.
Fixed thus.
Note that regression tests are all updated in the last patch of the
series
* src/abg-ir.cc (union_decl::get_pretty_representation):
Anonymous internal pretty representation of unin is its fully
qualified name.
Dodji Seketeli [Mon, 25 Mar 2019 13:46:31 +0000 (14:46 +0100)]
Bug 24378 - DW_TAG_subroutine_type as a DIE scope causes infinite loop
GCC 4.3.2 wrongly emits some type definition DIEs in the scope of a
DW_TAG_subroutine_type. Whenever the DWARF reader tries to get the
scope of a DIE during the computation of the pretty name of a type DIE
which scope is (wrongly) emitted as being a DW_TAG_subroutine_type
things end-up in an infinite loop.
This patch makes get_scope_die to look through the
DW_TAG_subroutine_type to return the proper scope instead, just like
what we already do for DW_TAG_subprogram and DW_TAG_array_type.
* src/abg-dwarf-reader.cc (get_scope_die): Look through
DW_TAG_subroutine_type to get the scope of a given DIE.
* tests/data/Makefile.am: Add the two new files below to source
distribution.
* tests/data/test-read-dwarf/PR24378-fn-is-not-scope.abi: New
reference test output.
* tests/data/test-read-dwarf/PR24378-fn-is-not-scope.o: New binary
test input.
* tests/test-read-dwarf.cc (in_out_specs): Add the new test input
to the test harness.
Dodji Seketeli [Thu, 21 Mar 2019 17:32:38 +0000 (18:32 +0100)]
Better pointer name equality optimization in DIE de-duplication code
During DIE de-duplication for the purpose of DIE canonicalization, we
use the so called "pointer name equality" optimization. That is, when
in a given translation unit, we have two different pointers named T*,
we assume they are the same type, without having to compare their
pointed-to T types recursively.
Unfortunately, some Ts can be different, especially in C. Think about
two typedefs with different underlying types, for instance. Or two
struct with data members which types are pointers to typefefs with
different underlying types. We are having the case in the libc binaries
provided in bug https://sourceware.org/bugzilla/show_bug.cgi?id=24257,
for instance.
In those case, we try not only to look at the name of the translation
unit associated to the pointer to T, but also, at the name of the
translation unit of the leaf node of T*. By leaf node, I mean the
resulting of stripping T* from all pointers and typedefs. This won't
solve the case of the Ts being structs with data members which types
are pointers to typedefs with different underlying types. But it
helps with the easier earlier cases.
* src/abg-dwarf-reader.cc
(die_is_pointer_reference_or_typedef_type)
(die_peel_pointer_and_typedef): Define new static functions.
(compare_dies_string_attribute_value): Turn this function into a
static one.
(compare_dies_cu_decl_file): Make this function compare the cu
decl file name of the leaf type of the pointer, not just the one
of the pointer itself.
(compare_as_decl_dies): Compare the DWARF tags too.
(compare_dies): Simplify logic.
While looking at something else, I figured it's useful, for debugging
purposes, to be able to lookup a given data member of a union/class by
name, as well as a function parameter by index.
Dodji Seketeli [Thu, 21 Mar 2019 17:08:59 +0000 (18:08 +0100)]
Better detection of void* to something* change
Whenever a void* pointer changes to a T* pointer, we already consider
that change to be ABI-compatible. The issue though is that we don't
detect the case of foo* changing into T* where foo is typedef void
foo. This patch fixes that.
* include/abg-ir.h (is_void_type): Add a new overload that takes
type_base*.
* src/abg-ir.cc (is_void_type): Define the new overload that takes
type_base*.
(is_void_pointer_type): Look through typedefs in
the pointed-to type.
Dodji Seketeli [Wed, 20 Feb 2019 10:25:39 +0000 (11:25 +0100)]
Do not build DIE -> parent map just because we see an asm TU
For languages like C that don't have namespaces, we don't need to
build the DIE -> parent map, and that speeds up the analysis of the
ELF binary in the DWARF reader.
Unfortunately, just because we are seeing some MIPS assembler
translation unit, dwarf_reader::read_context::build_die_parent_maps
unnecessarily builds the DIE -> parent map, making the ELF binary
analysis slower.
This patch introduces the new function
dwarf_reader::read_context::do_we_build_die_parent_maps to make the
decision about building the DIE -> parent map in a central place. And
it makes dwarf_reader::read_context::build_die_parent_maps use it.
With this, analysing the linux kernel becomes faster again.
Dodji Seketeli [Tue, 19 Feb 2019 15:34:45 +0000 (16:34 +0100)]
Avoid over-suppressing fns & vars when analysing the Kernel
When analyzing the Linux Kernel (with e.g abidw) and providing a
symbol whitelist the DWARF reader performs a speed optimization which
is that it does not read the symbol table of the kernel binary.
To know if a function is to be suppressed (because it doesn't belong
to the whitelist, for instance) dwarf_reader::function_is_suppressed
actually tries to see if the address of the function corresponds to
the address of an ELF symbol that is publicly exported, before looking
at if the function was suppressed by a suppression specification.
But then, when a whitelist is provided and the speed optimization of
not reading the symbol table is at play, there is no information about
ELF symbol addressed around. So dwarf_reader::function_is_suppressed
almost always supposes that the function is suppressed.
This patch fixes that by not trying to look at the address of the
function/variable when the symbol table optimization is on.
* include/abg-dwarf-reader.h (get_ignore_symbol_table): Take a
const read_context&.
* src/abg-dwarf-reader.cc (get_ignore_symbol_table): Likewise.
(function_is_suppressed): When the symbol table optimization is in
flight -- that is, when no symbol table has been loaded -- do not
try to see if a given function symbol was exported at the ELF
level or not. Just look at if the function was suppressed or not.
(variable_is_suppressed): Likewise for variables.
Dodji Seketeli [Fri, 8 Feb 2019 08:48:40 +0000 (09:48 +0100)]
Bug 24188 - Assertion failed while analysing a Fortran binary
While analysing a Fortran binary the DWARF reader performs DIE
de-duplication. During that process, the compare_dies function
stumbles accross the a DIE of the DW_TAG_string_type kind. And it
doesn't know how to compare those DIEs. And this leads to an abort of
the abipkgdiff program, in particular.
DW_TAG_string_type DIEs do have a DW_AT_string_length attribute which
value is a location expression that does not resolve to a constant
value. In general, we cannot evaluate those expressions in a static
context like in Libabigail because we lack things like register values
which are dynamic in nature.
So, I decided for now to consider that two DW_TAG_string_type seen at
different DWARF offsets are considered to be different for now. This
pessimises DIEs de-duplication for types that contain
DW_TAG_string_type as their subtypes, but at least this is a basic
support for DW_TAG_string_type.
Tested with the RPMs on which abipkgdiff was failing.
Signed-off-by: Dodji Seketeli <dodji@redhat.com>
* src/abg-dwarf-reader.cc (compare_as_type_dies): Handle
DW_TAG_string_type DIEs here.
(compare_dies): Handle DW_TAG_string_type DIEs by using
compare_as_type_dies.
* tests/data/test-diff-pkg/netcdf-fortran-debuginfo-4.4.4-10.fc29.x86_64.rpm:
New test RPM.
* tests/data/test-diff-pkg/netcdf-fortran-debuginfo-4.4.4-11.fc30.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/netcdf-fortran-mpich-4.4.4-10.fc29.x86_64-4.4.4-11.fc30.x86_64-report-0.txt:
New expected test reference output.
* tests/data/test-diff-pkg/netcdf-fortran-mpich-4.4.4-10.fc29.x86_64.rpm:
New test RPM.
* tests/data/test-diff-pkg/netcdf-fortran-mpich-4.4.4-11.fc30.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/netcdf-fortran-mpich-debuginfo-4.4.4-10.fc29.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/netcdf-fortran-mpich-debuginfo-4.4.4-11.fc30.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/netcdf-fortran-mpich-devel-4.4.4-10.fc29.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/netcdf-fortran-mpich-devel-4.4.4-11.fc30.x86_64.rpm:
Likewise.
* tests/data/Makefile.am: Add the new test input material above to
source distribution.
* tests/test-diff-pkg.cc (in_out_spec): Add the new test RPMs
above to the set of RPMs to use as test input.
Dodji Seketeli [Wed, 6 Feb 2019 12:16:44 +0000 (13:16 +0100)]
Bug 24157 - Wrong support of Ada ranges
In this problem report, the DWARF reader comes across an ADA subrange
which lower bound is a negative signed integer. The problem is that
the internal representation of libabigail expects the bounds of the
subrange to be unsigned integer. This incompatibility leads to an
assertion failure.
Further down the road I realized that the function
types_have_similar_structure doesn't support subrange types.
This patch thus introduces a new
array_type_def::subrange_type::bound_value class that captures the
signedness of the bound value. This allows the internal
representation to keep the sign information of the bound value, and
yet be able to treat the bound value as either a signed or unsigned
value depending on the contexts.
The patch also extends the function types_have_similar_structure to
make it support subrange types.
* include/abg-ir.h (array_type_def::subrange_type::bound_value):
Define new class.
(array_type_def::subrange_type::subrange_type): Adjust to use the
new bound_value type for bound values.
(array_type_def::subrange_type::{get_upper_bound, get_lower_bound,
set_upper_bound, set_lower_bound}): Return or take int64_t rather
than size_t.
(array_type_def::subrange_type::get_length): Return uint64_t
rather than size_t.
* src/abg-dwarf-reader.cc (die_signed_constant_attribute)
(die_constant_attribute, die_attribute_has_form)
(die_attribute_is_signed, die_attribute_is_unsigned)
(die_attribute_has_no_signedness): Define new static functions.
(get_default_array_lower_bound): Return uint64_t rather than int.
(build_subrange_type): Use the new
array_type_def::subrange_type::bound_value type for bound values.
Use the new die_constant_attribute function, rather than
die_unsigned_constant_attribute to fecth the bound values.
* src/abg-ir.cc
(array_type_def::subrange_type::bound_value::{bound_value,
get_signedness, set_signedness, get_signed_value,
get_unsigned_value, set_unsigned, set_signed}): Define new member
functions.
(array_type_def::subrange_type::priv::{lower_bound_,
upper_bound}): Use the new class bound_value.
(array_type_def::subrange_type::priv::priv): Adjust to use the new
bound_value class to hold bound values.
(array_type_def::subrange_type::subrange_type): Likewise.
(array_type_def::subrange_type::{get_upper_bound, get_lower_bound,
set_upper_bound, set_lower_bound}): Return or take int64_t rather
than size_t.
(array_type_def::subrange_type::get_length): Return uint64_t
rather than size_t.
(types_have_similar_structure): Handle array_type_def::subrange_type
* src/abg-reader.cc (build_subrange_type): Use the new
array_type_def::subrange_type::bound_value to hold bound values.
* tests/data/test-diff-pkg/GtkAda-debuginfo-2.24.2-29.fc29.x86_64.rpm:
New binary RPM as test input.
* tests/data/test-diff-pkg/GtkAda-debuginfo-2.24.2-30.fc30.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/GtkAda-devel-2.24.2-29.fc29.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/GtkAda-devel-2.24.2-30.fc30.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/GtkAda-gl-2.24.2-29.fc29.x86_64--2.24.2-30.fc30.x86_64-report-0.txt:
New expected test output.
* tests/data/test-diff-pkg/GtkAda-gl-2.24.2-29.fc29.x86_64.rpm:
New binary RPM as test input.
* tests/data/test-diff-pkg/GtkAda-gl-2.24.2-30.fc30.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/GtkAda-gl-debuginfo-2.24.2-29.fc29.x86_64.rpm:
Likewise.
* tests/data/test-diff-pkg/GtkAda-gl-debuginfo-2.24.2-30.fc30.x86_64.rpm:
Likewise.
* tests/data/Makefile.am: Add the new test material above to source
distribution.
* tests/test-diff-pkg.cc (in_out_specs): Add the new input testing
RPMs in here.
This property is taken into accound iff the current suppress_type does
have the 'type_kind = enum' property.
Changes to enum types that match the new 'changed_enumerators'
property are suppressed.
* doc/manuals/libabigail-concepts.rst: Document the new
'changed_enumerators' property.
* include/abg-suppression.h
(type_suppression::{g, s}et_changed_enumerator_names): Declare two
new member functions.
* src/abg-suppression-priv.h
(type_suppression::priv::changed_enumerator_names_): Add a new
data member.
* src/abg-suppression.cc
(type_suppression::{g,s}et_changed_enumerator_names): Define two
new member functions.
(type_suppression::suppresses_diff): Support evaluating the new
'changed_enumerators = <vector of changed enumerators>'.
(read_type_suppression): Read the new list
property'changed_enumerators" and store it into the
type_suppression using the new
type_suppression::set_changed_enumerator_names ().
* tests/data/test-diff-suppr/libtest4{0,1}-enumerator-changes-v{0,1}.so:
Add new test inpujts.
* tests/data/test-diff-suppr/test4{0,1}-enumerator-changes-0.suppr:
Add a new suppr spec for this new test.
* tests/data/test-diff-suppr/test4{0,1}-enumerator-changes-report-0.txt:
The default report.
* tests/data/test-diff-suppr/test4{0,1}-enumerator-changes-v{0,1}.cc:
Add Source code of libtest4{0,1}-enumerator-changes-v{0,1}.so.
* tests/data/Makefile.am: Add the test files above to source
distribution.
* tests/test-diff-suppr.cc: Add the test input files above to the
harness.
Dodji Seketeli [Fri, 25 Jan 2019 10:52:41 +0000 (11:52 +0100)]
Better comments in the comparison engine
* include/abg-comparison.h (enum diff_category): Add comments to
describe what to update when a new enumerator is added to this enum.
* src/abg-comp-filter.cc (has_fn_return_type_cv_qual_change): Fix
comment thinko here.
Dodji Seketeli [Fri, 25 Jan 2019 09:42:57 +0000 (10:42 +0100)]
Bug 20175 - Classify CV qual changes in variable type as harmless
When the type of a variable changes and that change is only a CV qual
change, we want libabigail to automatically classify that change as
being harmless.
This patch thus introduces a new VAR_TYPE_CV_CHANGE_CATEGORY change
category (enumerator of the abigail::comparison::diff_category enum)
for this kind of changes and classifies that category as being
harmless.
The patch then detects diff nodes that carry CV qualifiers change on
variable types and flags them as belonging to the new
VAR_TYPE_CV_CHANGE_CATEGORY.
* include/abg-comparison.h (VAR_TYPE_CV_CHANGE_CATEGORY): Add new
enumerator to diff_category enum.
(EVERYTHING_CATEGORY): Update this enumerator.
* src/abg-comp-filter.cc (type_diff_has_cv_qual_change_only):
Support array diff nodes carrying a cv qual change on the element
type.
(has_var_type_cv_qual_change): Define new static function.
(categorize_harmless_diff_node): Use the new
has_var_type_cv_qual_change to categorize variable diff node with
cv qual change on its type as harmless.
* src/abg-comparison.cc
(get_default_harmless_categories_bitmap): Update this.
(operator<<(ostream& o, diff_category c)): Likewise.
* include/abg-ir.h (equals_modulo_cv_qualifier): Declare new ...
* src/abg-ir.cc (equals_modulo_cv_qualifier): ... function.
* tests/data/test-diff-pkg/libICE-1.0.6-1.el6.x86_64.rpm--libICE-1.0.9-2.el7.x86_64.rpm-report-0.txt:
Update expected test output.
* tests/data/test-diff-pkg/spice-server-0.12.4-19.el7.x86_64-0.12.8-1.el7.x86_64-report-2.txt: Likewise.
* tests/data/Makefile.am: Add the new test material below to
source distribution.
* tests/data/test-diff-pkg/nss-3.23.0-1.0.fc23.x86_64-report-0.txt:
New expecte test output.
* tests/data/test-diff-pkg/nss-3.23.0-1.0.fc23.x86_64.rpm: New
test input.
* tests/data/test-diff-pkg/nss-3.24.0-1.0.fc23.x86_64.rpm: Likewise.
* tests/data/test-diff-pkg/nss-debuginfo-3.23.0-1.0.fc23.x86_64.rpm: Likewise.
* tests/data/test-diff-pkg/nss-debuginfo-3.24.0-1.0.fc23.x86_64.rpm: Likewise.
* tests/data/test-diff-pkg/nss-devel-3.23.0-1.0.fc23.x86_64.rpm: Likewise.
* tests/data/test-diff-pkg/nss-devel-3.24.0-1.0.fc23.x86_64.rpm: Likewise.
* tests/test-diff-pkg.cc (in_out_specs): Add the test input above
to the test harness.
Dodji Seketeli [Thu, 24 Jan 2019 09:31:57 +0000 (10:31 +0100)]
Fix a typo in the recent Rust support and update regression tests
* configure.ac: Fix the typo HAS_LANG_Rust into HAS_DW_LANG_Rust.
* tests/data/test-diff-dwarf/test46-readme.txt: Add new file to
the test suite.
* tests/data/test-diff-dwarf/test46-rust-libone.so: Likewise.
* tests/data/test-diff-dwarf/test46-rust-libtwo.so: Likewise.
* tests/data/test-diff-dwarf/test46-rust-report-0.txt: Likewise.
* tests/test-diff-dwarf.cc (in_out_specs): Update the tests array
to compare the two new binaries included above.
Dodji Seketeli [Fri, 18 Jan 2019 10:21:48 +0000 (11:21 +0100)]
Overhaul detection the DW_LANG_* enumerators from dwarf.h
elfutils regularly adds new members to the anonymous DWARF language
encodings enum which contains the DW_LANG_* enumerators, from the
dwarf.h header file.
If we want libabigail to keep compiling with older versions of
elfutils, we need to detect the absence of a given DW_LANG_*
enumerator and avoid using it in that case.
Until now, we were doing #ifdef DW_LANG_* for that purpose. But then
the DW_LANG_* are *enumerators*, not preprocessor macros. So that
preprocessor macro.
This patch detects the presence of each of the "newer" DW_LANG_*
enumerator using autoconf. And for each DW_LANG_xxx enumerator that
is present, autoconf defines the HAVE_DW_LANG_xxx_enumerator macro.
Libabigail source code can thus do #ifdef HAVE_DW_LANG_xxx_enumerator
to guard the use of DW_LANG_xxx.
Tested with the Rust binaries from
https://gitlab.gnome.org/federico/abi-rust.
* configure.ac: Detect the presence of DW_LANG_{UPC, D, Python,
Go, C11, C_plus_plus_03, C_plus_plus_11, C_plus_plus_14,
Mips_Assembler, Rust} and define the corresponding
HAVE_DW_LANG_*_enumerator macro accordingly.
* include/abg-ir.h (LANG_C_plus_plus_03): Define this new
enumerator in the translation_unit::language enum.
* src/abg-dwarf-reader.cc (dwarf_language_to_tu_language): Use the
new HAVE_DW_LANG_*_enumerator macros.
(get_default_array_lower_bound): Support the
translation_unit::LANG_C_plus_plus_03 enumerator.
* src/abg-ir.cc (is_cplus_plus_language): Support the
translation_unit::LANG_C_plus_plus_03 enumerator.
Mark Wielaard [Thu, 17 Jan 2019 10:19:02 +0000 (11:19 +0100)]
Conditionalize the use of DW_LANG_C_plus_plus_03 and DW_LANG_Rust
Older elfutils (pre-0.170) don't define these constants in dwarf.h so
don't use them in that case.
* include/abg-ir.h (LANG_C_plus_plus_03): Add this new language
enum to "enum translation_unit::language".
* src/abg-dwarf-reader.cc (dwarf_language_to_tu_language): Do not
use DW_LANG_Rust or DW_LANG_C_plus_plus_03 if these are not
defined.
(get_default_array_lower_bound): Handle the new
translation_unit::LANG_C_plus_plus_03 enumerator.
Signed-off-by: Mark Wielaard <mark@klomp.org> Signed-off-by: Dodji Seketeli <dodji@redhat.com>
[C]'function one::Foo one::foo(u32)' at lib.rs:8:1 has some indirect sub-type changes:
'function one::Foo one::foo(u32) {foo}' now becomes 'function two::Foo two::foo(u32, u32) {foo}'
return type changed:
type name changed from 'one::Foo' to 'two::Foo'
type size changed from 32 to 64 (in bits)
1 data member insertion:
'u32 two::Foo::b', at offset 32 (in bits)
no data member change (1 filtered);
parameter 2 of type 'u32' was added
* include/abg-ir.h (LANG_Rust): Add this new enumerator to the
"enum language" enum.
* src/abg-dwarf-reader.cc (dwarf_language_to_tu_language): Handle
the Rust language.
(get_default_array_lower_bound): Likewise.
Dodji Seketeli [Tue, 15 Jan 2019 17:33:12 +0000 (18:33 +0100)]
Separate public types of first binary from those of the second
In abidiff when the user uses --headers-dir{1,2}, she implicitly
indicates the public types of the first and the second binary. That
means that any type that is defined in a file that is not located
under the directory tree designated by --headers-dir{1,2} is
considered private and any change involving those private types will
be suppressed.
In practice, what happens is that libabigail constructs a suppression
specification which suppress all types that are defined in files that
are not under the directories located by --headers-dir{1,2}.
abidiff has a problem, though. It uses the same private type
suppressions (derived from --headers-dir1 and --headers-dir2) when
looking at the first binary *and* the second binary. It should rather
only use the private type suppression specifications derived from
--headers-dir1 when looking at the first binary, and use the private
type suppression specifications derived from --headers-dir2 when
looking at the second binary.
This problem leads to some false positives like the one reported at
https://gitlab.gnome.org/GNOME/pango/issues/343#note_397761.
This patch fixes this issue by using the private type suppression
specifications derived from --headers-dir1 only when looking at the
first binary, and using the private type suppression specifications
derived from --headers-dir2 only when looking at the second binary.
* include/abg-dwarf-reader.h (read_context_get_path): Declare new
function.
* include/abg-reader.h (read_context_get_path): Likewise.
* src/abg-dwarf-reader.cc (read_context_get_path): Define new function.
* src/abg-reader.cc (read_context_get_path): Likewise.
* tools/abidiff.cc (set_suppressions): Set the suppression
specification derived from the --headers-dir1 option only for the first
binary, and similarly, from the --headers-dir2 option only for the
second binary.
Dodji Seketeli [Wed, 9 Jan 2019 14:22:01 +0000 (15:22 +0100)]
Bug 23044 - Assertions with side effects
There are lots of spots in libabigail's source code where the argument
of the assert() call does have side effects. This is a problem
because when the code is compiled with the NDEBUG macro defined, the
assert call does nothing, so the side effects of its argument are then
suppressed, changing the behaviour of the program.
To handle this issue, this patch introduces the ABG_ASSERT macro which
is a wrapper around the assert call that enable the use of side
effects in its argument. The patch now uses that ABG_ASSERT macro
instead of using the assert call directly.
The patch also makes it so that the configure option accepts the
--disable-assert option so that the user can build libabigail with the
NDEBUG macro defined.
Tested by running the testsuite with and without the --disable-assert
option to configure.
* COMPILING: Add the pkg-config dependency.
* doc/manuals/kmidiff.rst: Replace the redundant --full-impact
documentation with the proper --impacted-interfaces one.
Dodji Seketeli [Thu, 29 Nov 2018 10:18:57 +0000 (11:18 +0100)]
Add basic support for Fortran binaries
Apparently, the only DWARF TAG that we needed to support to allow the
analysis of Fortran binaries is DW_TAG_string_type, at least for now.
The only place where we need to handle the DW_TAG_string_type is in
the DIE canonicalizer. Basically, with this patch we now consider all
the DW_TAG_string_type DIEs as being equal. This seems good enough for
ABIs purpose, unless proven otherwise.
Note that this fixes PR23492 and PR23905.
* src/abg-dwarf-reader.cc (die_pretty_print_type): Support
DW_TAG_string_type DIEs. They all have the same representation
for now.
Dodji Seketeli [Thu, 8 Nov 2018 08:04:53 +0000 (09:04 +0100)]
Add a --fail-no-debug-info to abidiff
Add an option to make abidiff to fail if doesn't find debug info.
Without this option, abidiff keeps going and works with only ELF
information, as it can't get the debug info.
* doc/manuals/abidiff.rst: Document the new --fail-no-debug-info
option.
* tools/abidiff.cc (options::fail_no_debug_info): Define new data
member.
(display_usage): Provide a help string for the new
--fail-no-debug-info option.
(parse_command_line): Parse the new option.
(main): If --fail-no-debug-info and no debug info was found, or
not alternate debuginfo file was found, bail out.
Dodji Seketeli [Thu, 8 Nov 2018 06:26:14 +0000 (07:26 +0100)]
Support having several debuginfo search dirs for a binary
There are use cases where the split debuginfo file of a given binary
is under a given root directory and that the split debuginfo file
itself depends on an alternate debuginfo file that is under another
unrelated root directory.
In that case, the dwarf reader must be able to look for the debuginfo
files under several unrelated root directories. The tools abidiff and
abidw must thus support having several occurences of the option
--debug-info-dir1 (or --debug-info-dir2) meaning that the debuginfo
files for the binary must be looked for under several root
directories.
This is what this patch does.
* doc/manuals/abidiff.rst: Adjust doc for the
--debug-info-dir{1,2} that can now be provided several times.
* include/abg-dwarf-reader.h ({create, reset}_read_context)
(read_corpus_from_elf): Take a vector of debug info root dirs.
* include/abg-tools-utils.h (trim_leading_string)
(find_file_under_dir, make_path_absolute_to_be_freed)
(convert_char_stars_to_char_star_stars): Declare new functions.
* src/abg-dwarf-reader.cc (find_alt_debug_info_link): Renamed
find_alt_debug_info_location into this.
(find_alt_debug_info_path): Define new static function.
(find_alt_debug_info): Take a vector of debug info root dirs. Use
the new find_alt_debug_info_path to look into the debug info root
dirs for the alt debug info.
(read_context::debug_info_root_paths_): Define new data member.
(read_context::read_context): Take a vector of debug info root
dirs and initialize the new read_context::debug_info_root_paths_.
(read_context::{initialize, create_default_dwfl}): Take a vector
of debug info root dirs and adjust.
(read_context::{add_debug_info_root_paths,
add_debug_info_root_path, find_alt_debug_info}): Define new member
functions.
(read_context::load_debug_info): Look into the debug info roots
for split debug info files.
(create_read_context, read_corpus_from_elf): Take a vector of
debug info root dirs and adjust.
(has_alt_debug_info): Adjust.
* src/abg-tools-utils.cc (trim_leading_string)
(make_path_absolute_to_be_freed, find_file_under_dir)
(convert_char_stars_to_char_star_stars): Define new functions.
(entry_of_file_with_name): Define new static function.
(build_corpus_group_from_kernel_dist_under): Adjust.
* tests/print-diff-tree.cc (main): Adjust.
* tests/test-diff-dwarf.cc (main): Adjust.
* tests/test-ir-walker.cc (main): Adjust.
* tests/test-read-dwarf.cc (main): Adjust.
* tools/abicompat.cc (main): Adjust.
* tools/abidiff.cc (options::di_root_paths{1,2}): Changed
di_root_path{1,2} into this, change their types into vectors of
allocated char*.
(options::prepared_di_root_paths{1,2}): Define new data members.
(options::~options): Define new destructor.
(parse_command_line): Adjust.
(prepare_di_root_paths): Define new static function.
(handle_error): Remove arguments input_file_name,
debug_info_dir{1,2}. Now just take an instance of options
instead. Adjust.
(main): Adjust.
* tools/abidw.cc (options::dir_root_paths): Renamed dir_root_path
into this and make it be a vector of allocated char*.
(options::prepared_di_root_paths): Define new data member.
(options::~options): Free the allocated char* in
options::dir_root_paths.
(parse_command_line): Support several --debug-info-dir.
(load_corpus_and_write_abixml): Adjust.
(prepare_di_root_paths): Define static function.
(main): Adjust.
* tools/abilint.cc (main): Adjust.
* tools/abipkgdiff.cc (compare): Adjust.
Dodji Seketeli [Tue, 16 Oct 2018 08:59:37 +0000 (10:59 +0200)]
Better support array with unknown upper bound
It can happen that arrays are described in DWARF with an unknown upper
bound. In those cases, if the array is the type of a global variable,
the ELF object corresponding to that global variable should have a
size. It's that ELF object size that should be taken into account
when comparing versions of that global variable and its type.
This patch fixes issues in the detection and representation of array
subranges with unknown size so that we behave like presented above.
* include/abg-comparison.h
(BENIGN_INFINITE_ARRAY_CHANGE_CATEGORY): A new enumerator in the
diff_category enum.
(EVERYTHING_CATEGORY): Adjust.
* src/abg-comparison.cc (get_default_harmless_categories_bitmap):
Adjust.
(operator<<(ostream& o, diff_category)): Likewise.
* include/abg-ir.h (array_type_def::subrange_type::is_infinite):
Declare new member function.
* src/abg-ir.cc (array_type_def::subrange_type::priv::infinite_):
New data member.
(array_type_def::subrange_type::priv::priv): Initialize it.
(array_type_def::subrange_type::get_length): Better support
unknown sized subrange.
(array_type_def::subrange_type::is_infinite): Define new member
function.
* src/abg-comp-filter.cc (has_benign_infinite_array_change):
Define new static function.
(categorize_harmless_diff_node): Use the new
has_benign_infinite_array_change above.
* src/abg-dwarf-reader.cc (build_subrange_type): Better recognize a
subrange type with unknown upper bound. Represent that with the
new array_type_def::subrange_type::is_infinite member property.
* src/abg-reader.cc (build_subrange_type): Likewise.
* tests/data/test-abidiff/test-PR18166-libtirpc.so.abi: Adjust.
* tests/data/test-annotate/libtest23.so.abi: Likewise.
* tests/data/test-annotate/libtest24-drop-fns-2.so.abi: Likewise.
* tests/data/test-annotate/libtest24-drop-fns.so.abi: Likewise.
* tests/data/test-annotate/test14-pr18893.so.abi: Likewise.
* tests/data/test-annotate/test19-pr19023-libtcmalloc_and_profiler.so.abi:
Likewise.
* tests/data/test-annotate/test7.so.abi: Likewise.
* tests/data/test-read-dwarf/libtest23.so.abi: Likewise.
* tests/data/test-read-dwarf/libtest24-drop-fns-2.so.abi: Likewise.
* tests/data/test-read-dwarf/libtest24-drop-fns.so.abi: Likewise.
* tests/data/test-read-dwarf/test10-pr18818-gcc.so.abi: Likewise.
* tests/data/test-read-dwarf/test11-pr18828.so.abi: Likewise.
* tests/data/test-read-dwarf/test14-pr18893.so.abi: Likewise.
* tests/data/test-read-dwarf/test19-pr19023-libtcmalloc_and_profiler.so.abi: Likewise.
* tests/data/test-read-dwarf/test22-pr19097-libstdc++.so.6.0.17.so.abi: Likewise.
* tests/data/test-read-dwarf/test9-pr18818-clang.so.abi: Likewise.
Dodji Seketeli [Wed, 3 Oct 2018 09:24:42 +0000 (11:24 +0200)]
Bug 23708 - categorize void* to pointer change as harmless
Changing a void* pointer into another pointer of the same size is a
change that is harmless in terms of data layout.
This commit thus categorizes such a change as harmless.
* include/abg-comparison.h (VOID_PTR_TO_PTR_CHANGE_CATEGORY): New
enumerator in the diff_category enum. Also, adjust the
EVERYTHING_CATEGORY enumerator.
* include/abg-fwd.h (is_void_pointer_type): Declare new function.
* src/abg-comp-filter.cc (has_void_ptr_to_ptr_change): Define new
static function and ...
(categorize_harmless_diff_node): ... use it here.
* src/abg-comparison.cc (get_default_harmless_categories_bitmap):
Add the new abigail::comparison::VOID_PTR_TO_PTR_CHANGE_CATEGORY
category in here.
(operator<<(ostream& o, diff_category c)): Add support for the new
VOID_PTR_TO_PTR_CHANGE_CATEGORY.
* src/abg-ir.cc (is_void_pointer_type): Define new function.
* tests/data/Makefile.am: Add the new test material below to source distribution.
* tests/data/test-diff-filter/test47-filter-void-ptr-change-report-0.txt:
New test reference output.
* tests/data/test-diff-filter/test47-filter-void-ptr-change-v{0,1}.c:
Source code of the new binary test input below.
* tests/data/test-diff-filter/test47-filter-void-ptr-change-v{0,1}.o:
New binary test input.
* tests/test-diff-filter.cc: Add the test input/output above to
test harness.
Dodji Seketeli [Tue, 2 Oct 2018 11:49:55 +0000 (13:49 +0200)]
Categorize CV qualifier changes on fn return types as harmless
This partially fixes PR23700.
A change in the CV qualifiers of a function return value type should
be categorized as harmless. And this is what this patch does.
* include/abg-comparison.h (FN_RETURN_TYPE_CV_CHANGE_CATEGORY):
New enumerator for diff_category.
(EVERYTHING_CATEGORY): Update.
* src/abg-comp-filter.cc (type_diff_has_cv_qual_change_only):
Factorize this function out of ...
(has_fn_parm_type_cv_qual_change): ... this one.
(has_fn_return_type_cv_qual_change): Define new static function.
(categorize_harmless_diff_node): Use the new
has_fn_return_type_cv_qual_change.
* src/abg-comparison.cc (get_default_harmless_categories_bitmap):
Adjust to add the new FN_RETURN_TYPE_CV_CHANGE_CATEGORY category.
(operator<<(ostream& o, diff_category c)): Support the new
FN_RETURN_TYPE_CV_CHANGE_CATEGORY.
* tests/data/Makefile.am: Add the new test material below to
source distribution.
* tests/data/test-diff-filter/test46-fn-return-qual-change-report-0.txt:
New reference output for the new input test.
* tests/data/test-diff-filter/test46-fn-return-qual-change-v{0,1}.c:
New source code for the new binary test input.
* tests/data/test-diff-filter/test46-fn-return-qual-change-v{0,1}.o:
New binary test input files.
* tests/test-diff-filter.cc: Add the new test input above to test
harness.
Dodji Seketeli [Mon, 1 Oct 2018 14:14:50 +0000 (16:14 +0200)]
Fix propagation of private type suppression category
This is a partial fix of PR23700.
Conceptually, there are two kinds of type suppression specifications:
1/ a generic user-provided suppression specification that is meant to
suppress changes on types specified by the user
2/ a private type suppression specification that is automatically
generated from the path to public header files provided by the user.
Technically, one difference between 1 and 2 lays in the way we
propagate categories of changes matched by those suppression
specifications.
If a class type change of category SUPPRESSED_CATEGORY is referenced
in a typedef change, then the typedef change is also considered to be
of category SUPPRESSED_CATEGORY. In other words, the
SUPPRESSED_CATEGORY category is propagated to the typedef change.
That means that if a change to a class type is suppressed, a (changed)
typedef to that class is considered to be suppressed too.
But then that is not true if the class type was changed because it's
private. In that, a typedef to that class can be *public*, because
the said typedef is defined in a public header. In that case the
typedef change should *NOT* be considered suppressed just because the
class type change was suppressed.
The problem we have here is that we don't make any difference between
1/ and 2/. So we need to introduce different propagation rules for 1/
and 2/.
So this patch introduces a new PRIVATE_TYPE_CATEGORY category for
types suppression specification that are automatically generated for
private types. That new category has its own propagation rule which
is basically "no propagation"; every type must be matched by the
private type suppression specification to be considered as private.
* include/abg-comp-filter.h (has_harmful_name_change): Declare new
function overloads.
* include/abg-comparison.h (PRIVATE_TYPE_CATEGORY): New enumerator
for diff_category;
(EVERYTHING_CATEGORY): Adjust this enumerator in diff_category;
(is_suppressed): Take an output parameter to say if the
suppression is a private type suppression.
* include/abg-suppression.h (is_private_type_suppr_spec): Take a
const reference parameter and add an overload for a shared
pointer.
* src/abg-comp-filter.cc (has_harmful_name_change): Define new
function.
* src/abg-comparison-priv.h (diff::priv::is_filtered_out): Diffs
of category PRIVATE_TYPE_CATEGORY are also considered filtered
out.
* src/abg-comparison.cc (diff::is_filtered_out): Adjust to account
for canonical diffs of category PRIVATE_TYPE_CATEGORY.
(diff::is_suppressed): Add an overload that takes a
is_private_type output parameter. Re-write the old overload in
terms of the new one.
(operator<<(ostream& o, diff_category c)): Handle
PRIVATE_TYPE_CATEGORY.
(category_propagation_visitor::visit_end): Do not propagate
PRIVATE_TYPE_CATEGORY here. Do not propagate
HARMLESS_DECL_NAME_CHANGE_CATEGORY either, when the class does
have a harmful decl name change.
(suppression_categorization_visitor::visit_begin): Set the new
PRIVATE_TYPE_CATEGORY category but do not propagate it.
(suppression_categorization_visitor::visit_end): Add some
comments.
* src/abg-default-reporter.cc (default_reporter::report): Avoid
reporting typedef underlying types that are in the
PRIVATE_TYPE_CATEGORY category.
* src/abg-suppression.cc (type_suppression::suppresses_diff): Do
not peel typedefs if we are a private type suppression.
(is_private_type_suppr_spec): Take a const reference.
* tests/data/Makefile.am: Add the new test material below to
source distribution.
* tests/test-diff-suppr.cc: Use new test binary input.
* tests/data/test-diff-filter/test7-report.txt: Adjust.
* tests/data/test-diff-suppr/test39-opaque-type-report-0.txt: New
test reference output.
* tests/data/test-diff-suppr/test39-opaque-type-v{0,1}.c: Source
code of new test binary input.
* tests/data/test-diff-suppr/test39-opaque-type-v{0,1}.o: New test
binary input.
* tests/data/test-diff-suppr/test39-public-headers-dir/test39-header-v{0,1}.h:
Source code of new test binary input.
PR23641 - confusion when a type definition DIE is matched by a supprspec and its decl DIEs aren't
This is a followup of the patch with commit hash 90b4e76.
It turns out that in some cases, when we see a suppressed private
class definition, we haven't *previously* seen any declaration-only
instance of it. So we cannot use any pre-existing declaration-only
instance as an opaque type for the private type that is suppressed.
In that case, this patch creates (and uses) a declaration-only
instance for the suppressed private type definition.
* src/abg-dwarf-reader.cc (get_opaque_version_of_type): If no
pre-existing opaque version was found, a new one is created and
returned. Take a needed "where_offset" parameter.
(build_ir_node_from_die): Adjust the call to
get_opaque_version_of_type.
PR23641 - Type definition DIE matched by a supprspec but not its decl
Suppose we have two versions of a library which are almost identical.
Suppose the difference between the two binaries is a slight difference of
ordering in how the linker put together what the DWARF describes.
Then, while processing the debug info of the first library, libabigail
first comes across the forward declaration of a type T. Suppose that
declaration is not matched by any private type suppression
specification. Libabigail is going to keep the declaration of T and
build an internal representation (IR) for it. It's also going to
build an IR for types and decls using T like "T*" or "T* var".
Now suppose that while processing the debug info of the second
library, libabigail first comes across the *definition* of T --
because in this second library, that definition comes first. Suppose
that the definition is matched by a private type suppression
specification, unlike the declaration in the first library. In this
case, T is going to be dropped, and no IR is going to be built for it.
If other types or decl like 'T*' or 'T *var" refer to this
*definition* of T, they will be dropped too.
We'll end up with those two libraries that are identical (modulo the
order in which libabigail sees type declarations and their
definitions) and are considered different when a suppression
specification makes us drop T: the second library appears to
libabigail as if T was removed from it.
This is the problem addressed by this patch.
When the definition of a type T is suppressed because it's considered
private then we look if there was a forward declaration for it
elsewhere, that is not matched by the private type suppression
specification. If we encountered such a type declaration then it
means that declaration is in effect an "opaque" version of T. So
rather than just dropping T altogether, we keep (and build an IR) for
its opaque version only. And we drop the definition of T.
This seems to fix the issue.
I can't seem to reproduce the slight re-ordering of DIEs descriptions that
uncover the issue so I'll rely on integration tests to catch future
regressions on this issue, rather than on unit tests. Sigh.
* include/abg-tools-utils.h (PRIVATE_TYPES_SUPPR_SPEC_NAME):
Remove this extern constant definition.
* src/abg-dwarf-reader.cc (type_is_suppressed): Add an overload
that takes an additional type_is_private output parameter.
(get_opaque_version_of_type): New static function.
(build_ir_node_from_die): For class types, get the opaque version
for suppressed private types rather than dropping them altogether.
* src/abg-reader.cc (type_is_suppressed): Adjust.
* src/abg-suppression-priv.h (type_is_suppressed): Add an overload
that takes a type_is_private output parameter.
* include/abg-suppression.h (get_private_types_suppr_spec_label)
(is_private_type_suppr_spec): Declare new functions.
* src/abg-suppression.cc
(get_private_types_suppr_spec_label, is_private_type_suppr_spec):
Define new functions.
(suppression_matches_type_name_or_location): Use the new
get_private_types_suppr_spec_label rather than a global extern
variable.
* src/abg-tools-utils.cc (handle_fts_entry): Adjust to use the new
get_private_types_suppr_spec_label.
(gen_suppr_spec_from_headers): Handle the case or an empty headers
root dir.
Add option to avoid walking abigail::ir nodes twice
When the ir_traversable_base::traverse() walks the IR graph, it
happens that it can visit a type node that was already visited before.
For instance, it visits the 'struct S' once and later, as part of its
visit of struct S*, it can visit struct S again.
There are use cases where we want the walker to avoid visiting a given
type node again. This patch adds the option to do so.
Basically the ir_node_visitor class can now be configured to tell the
walker to avoid re-visiting a node.
The test-ir-walker.cc example is amended to avoid re-visiting type nodes
as well.
* include/abg-ir.h (struct ir_node_visitor): Make this be a class.
Add a private data member to it, following the 'pimpl' idiom.
(ir_node_visitor::{allow_visiting_already_visited_type_node,
mark_type_node_as_visited, forget_visited_type_nodes,
type_node_has_been_visited}): Declare new member functions.
* src/abg-ir.cc ({type_base, type_decl, scope_type_decl,
qualified_type_decl, pointer_type_def, reference_type_def,
array_type_def, enum_type_decl, typedef_decl, class_or_union,
class_decl, union_decl}::traverse): Avoid re-visiting the type
node if the visitor was configured as such.
(struct ir_node_visitor::priv): Define new struct.
(ir_node_visitor::{allow_visiting_already_visited_type_node,
mark_type_node_as_visited, forget_visited_type_nodes,
type_node_has_been_visited}): Define new member functions.
* tests/test-ir-walker.cc
(name_printing_visitor::name_printing_visitor): Avoid visiting a
type node twice.
This is because the ini parser won't accept the '=' character in the
URL as a valid character for ini property values.
So the entire [suppress_file] section is ignored by the suppression
specification engine.
This patch fixes that by making the equal character valid in property
values.
* src/abg-ini.cc (char_is_delimiter): Take a new include_equal
flag to control is the equal character should be considered as a
delimiter or not.
(char_is_property_value_char): Accept the equal character as a
valid property value character.
* tests/Makefile.am: Build a new runtestini test from the new
tests/test-ini.cc source file.
* tests/data/Makefile.am: Add the two new test inputs below to
source distribution.
* tests/data/test-ini/test01-equal-in-property-string.{abignore,
abignore.expected}: New test inputs.
* tests/test-ini.cc: New test harness.
Dodji Seketeli [Mon, 6 Aug 2018 08:28:06 +0000 (10:28 +0200)]
Make test-ir-walker work on ELF binaries directly
The tests/test-ir-walker.cc example was working on abixml files
resulting from abidw. This commit changes that to make it work on
ELF input files directly. The commit also adds some comments to make
it easier to understand the concepts.
* test-ir-walker.cc (main): Load an ABI corpus from an elf file
and walk its translation units.
Allow use of python even when fedabipkgdiff is disabled
It turns out that when we disable fedabipkgdiff, we also disable
detection of python and the setting of the proper environment
variables that are expected by some makefiles down the road. The
leads to the runtestdefaultsupprs.py test to be blocked at runtime
because the environement variable PYTHON is not defined. This patch
fixes that.
* configure.ac: Detect the presence and version of python even
when fedabipkgdiff is disabled.
Sometimes, one wants to be able to write suppression specifications
like:
[suppress_function]
filename_regexp = ^test38-char-class-in-ini-v[[:digit:]].*
symbol_name_regexp = bar
change_kind = added-function
without having to escape the square brackets in the regexp. Normally,
one has to escape the '[' and the ']' because these characters are
used to define ini section names (e.g, [suppress_function]).
This patch allows the presence of the square bracket characters in a
property value, making the suppression specification above valid.
* src/abg-ini.cc (char_is_delimiter): Possibly disallow square
bracket characters into the set of delimiters.
* tests/data/test-diff-suppr/test38-char-class-in-ini-report-0.txt:
New reference output.
* tests/data/test-diff-suppr/test38-char-class-in-ini-v{0,1}.c:
Source code new test binaries.
* tests/data/test-diff-suppr/test38-char-class-in-ini-v{0,1}.o:
New test binaries.
* tests/data/test-diff-suppr/test38-char-class-in-ini.abignore:
New test abi suppression file.
* tests/data/Makefile.am: Add the new test materials above to
source distribution.
* tests/test-diff-suppr.cc: Add the test materials above to the
set of tests to run.
runtestdefaultsupprspy3.sh just runs runtestdefaultsupprs.py using
python3. So when both run in parallel, it can happen that they step
on their toes as they emit content into some files.
This patch commits makes it so that runtestdefaultsupprspy3.sh is
executed only if we are in the python3 mode, otherwise we just run
runtestdefaultsupprs.py, fixing the occasional failure that we see on
either runtestdefaultsupprs.py or runtestdefaultsupprspy3.sh.
* tests/Makefile.am: Run runtestdefaultsupprspy3.sh if we are in
python3 mode otherwise run runtestdefaultsupprs.py.
Dodji Seketeli [Fri, 29 Jun 2018 10:32:02 +0000 (12:32 +0200)]
Ensure die_function_type_is_method_type returns a class type die
When die_function_type_is_method returns true, it meanins that the
function die it's looking at is a member function. In that case,
die_function_type_is_method must also return the class of which the
function is a member. It appears that there are cases where this
die_function_type_is_method returns a *typedef* DIE (to a class DIE),
rather than a class DIE.
This results in the assertion violation below when loading the file
/usr/lib/libreoffice/program/libanalysislo.so from, e.g, the
libreoffice-calc-5.3.6.1-10.el7.i686 package:
This patch fixes that by peeling off the potential typedefs that might
be there.
* src/abg-dwarf-reader.cc (die_peel_typedef): Define new static
function.
(die_function_type_is_method_type): Use the function above to peel
the class die from potential typedefs wrapping it.
git://sourceware.org / libabigail.git / log