This is based on the ARCv2 32-bit little-endian hard-float ISA.

Closes: #980963
Based-on-patch-by: Alexey Brodkin <Alexey.Brodkin@synopsys.com>

Running code in the module scope means that anything importing the
module will execute that code, which is undesirable. Move the
initialization into a _select_objdump() sub, which gets assigned into
a state variable on demand.

The spawn() call was removed but not the matching import.

Fixes: commit 8ae966ae

An ELF executable is defined by whether the object has the EXEC_P flag
defined or an interpreter in the program header. The former applies to
statically linked programs, the latter to dynamically linked programs
and possibly some shared libraries that can be executed, such as the
ones provided by glibc.

This is now more relevant as PIE makes normal executables show up as
shared objects, so they do not contain the EXEC_P flag, and were not
being detected as executables.

If the ELF class or endianness are unknown or bogus, ignore the file.

Reported-by: Niels Thykier <niels@thykier.net>

Emit an explicit warning whenever we cannot detect the format for
an executable object, instead of delegating this to the subsequent
objdump, and letting it die, which ca be canfusing and is not
future-proof.

Closes: #854536

The rest of the code handles non-binary files (ELF in this case)
gracefully and ignores them, even though not very explicitly, as
objdump will emit a warning that might be difficult to decrypt.

We will still fail for other read failures that are not just
short-reads, as those imply some actual problem with the passed files.

Closes: #854536

The affected code in NetBSD was bogus, and has been removed now. So
there is no point in trying to special case the EM_SPARC32PLUS ELF
machine ID depending on the ELF class, for something that should never
happen.

Ref: https//gnats.netbsd.org/51925

These are too unreliable for exact matches. There are objects with
EABIv4 in the wild, even when the current is EABIv5. The soft and
hard float flags are not always set on armel and armhf respectively,
although the Tag_ABI_VFP_args attribute in the the ARM attribute
section should always be present on armhf. And there are cases were
both soft and hard float flags are set at the same time(!).

Mask all flags on ARM, so that we get back to the previous behavior
with objdump. We can try to revisit this for a better matching during
the dpkg 1.19.x cycle.

Closes: #853793

Some ELF binaries contain alternative or old ELF machine types, which
should match with their canonical forms. Map those before encoding the
ABI.

We ignore some mappings for things that should certainly never appear
in current systems. Of note are EM_PPC_OLD (17) that conflicts with
EM_VPP550 (17), and EM_PJ_OLD (99) that conflicts with EM_SNP1K (99).

Warned-by: rebootstrap

This way when unpacking for output, the result gives meaningful results.

These do not define the ABI, and seem to be set depending on the ISA
used. Mask them for now, and postpone possibly making more fine-grained
matching in the future.

The previous ELF ABI mismatch detector was very naïve, as the string
returned by «objdump -a» is a very simplistic representation of the
ELF ABI used.

Switch to our own ELF header parser, so that we can distinguish based
on the fields that define the object ABI.

This is still not enough, and we will have collisions with things such
as linux-i386 and hurd-i386, but most of the previously colliding
objects are now filtered.

This also makes us independent of objdump not supporting any unknown
ELF object ABI.

Closes: #849913

This should make the regular expressions easier to read and understand,
and allows to add comments describing specific parts.

Addresses RegularExpressions::RequireExtendedFormatting.

Warned-by: perlcritic

The ELF spec does not disallow symbols with embedded spaces, so we
should really be supporting those. This is required by Go shared
libraries.

Closes: #785344
Based-on-patch-by: Michael Hudson-Doyle <michael.hudson@canonical.com>
Signed-off-by: Guillem Jover <guillem@debian.org>

This moves the variables closer to the block they are being used in. And
protects them from external access.

Place 'use' strict and warnings first, then Exporter 'our' declarations,
then Test module imports, then system module imports, then Dpkg module
imports, then 'use' parent and overload pragmas, separated by a blank
line for each block.

Split each exported symbol declaration into its own line to ease
modifications.

The old function name was inconsistent with the other gettext family
of short aliases which has already caused some code typos, and functions
starting with underscore are considered by convention private in Perl.

Fixes BuiltinFunctions::ProhibitUselessTopic
and RegularExpressions::ProhibitUselessTopic.

Warned-by: perlcritic

This reverts some of the unsafe assignments, as the variable might have
contained a value evaluating to false, and we are assigning non-false
values.

This specifically fixes dpkg-source --commit and -b not working any
longer.

Regression introduced in commit b0337f00.

Closes: #758426

Reported-by: Neil Williams <codehelp@debian.org>

Replace only safe usages, i.e. those that fallback on initialization
values that evaluate to false anyway. Or when the API is explicit about
the variable being undefined.

Replace only safe usages, i.e. those that fallback on initialization
values that evaluate to false anyway. Or when the API is explicit about
the variable being undefined.

These are false positives, as when there's no match perl will leave
the subpattern variables untouched. We'll just use the opportunity to
unconfuse perlcritic, making some of those variable usages more clear
by using intermediate variables.

Fixes RegularExpressions::ProhibitUnusedCapture.

Warned-by: perlcritic

This way we make sure we don't affect the $_ value for the caller,
in case it was being used.

This will make it safe to remove parenthesis around builtins.

This gets rid of the localized $OBJDUMP variable, by explicitly passing
the overriden value to the recursive call.

Fixes Variables::ProhibitLocalVars.

Using double-quotes imposes a small performance penalty as the perl
parser needs to check if any interpolation is needed. Use double-quotes
only when the string contains single-quotes. Ideally we'd use
double-quotes too for escaped meta-characters that might otherwise be
confusing to immediately see if they need interpolation or not, but the
policy does not (currently) allow to ignore these.

Fixes ValuesAndExpressions::ProhibitInterpolationOfLiterals.

Warned-by: perlcritic

Addresses ValuesAndExpressions::ProhibitInterpolationOfLiterals.

Warned-by: perlcritic

Do not quote the keys for any hash which will always use simple
identifiers, as that is taken care of by the => operator. So this does
not apply to overload pragmas, version operators, field names, substvars,
command line options or filenames.

Addresses ValuesAndExpressions::ProhibitInterpolationOfLiterals.

Warned-by: perlcritic

Fixes InputOutput::ProhibitBarewordFileHandles.

Warned-by: perlcritic

This makes sure the functions always return the appropriate value
independent of the context.

Fixes Subroutines::ProhibitExplicitReturnUndef.

Warned-by: perlcritic

In many cases the cross objdump is not able to analyze the binaries from
the build host and yet dpkg-shlibdeps need to be able to verify that
the various binaries use the same ELF format. To achieve this get_format()
is enhanced to fallback on the objdump command when the cross one failed.

When <cross-prefix>-objdump is available and when we're cross-compiling
let's use the cross objdump in preference over the standard objdump.

Based-on-patch-by: Loïc Minier <lool@debian.org>

Modules whose API should be stable have a version 1.00.
The modules with version 0.01 are expected to have further API
changes in the (near) future or are deemed to be useful mostly
for dpkg-dev's internal usage.