| Age | Commit message (Collapse) | Author |
|---|
|
Replace several expansions of
"${stamp_type:+, ${stamp_type}ing..}."
with only one, using an intermediate variable.
This makes sure we always use the same message.
Change-Id: Ib33c75a1fbb63ccadce04dfd2ead7acb7e6b4ee5
|
|
and enable it with binutils master config
Change-Id: I5c51cb539d01ebd5adcc77cd1ab715e4bfe6cc48
|
|
Make --enable make_docs behave as before a recent patch: install html
and info formats by default.
--enable full_docs adds pdf and dvi formats, which are not supported
by some of the older toolchain components versions.
This way, ABE is still able to build successfully toolchains of older
versions, and --enable full_docs enables more document formats if
desired.
Change-Id: I358f913970dd8c9505362afa6e929aeadaf3bca9
|
|
The extra options are the ones usually used by distributions. Although
all are security oriented, it improves the test coverage.
Change-Id: I4a93a3a9280111ff7cf8436990ac00503133ee13
|
|
Also use the same scheme to report errors when building docs for other components.
Change-Id: I6f1404d3d58ac8eeab2c7b56bd91a741ee4d74d3
|
|
Build PDF docs for binutils, gdb and gcc.
Build man pages for binutils, gcc and libstdc++.
For libstdc++, this requires 'doxygen' and 'dot' to be present (they
are detected at configure time, and this is a no-op if they are
absent).
Note that gdbserver documentation is part of the gdb component.
Change-Id: I94579cf085bb2ff78f45c34b93015100ceb8ae0a
|
|
ABE has a dedicated component for gdbserver, so do not build it as
part of binutils.
Change-Id: I077d3195513c4348479fefbe73d26106f8ef5406
|
|
The glibc tests are built in two different modes (enabled at configure
time):
1. No RPATH defined, tests will run by issuing the loader along with
the required library paths. It is the default option.
2. RPATH defined, tests will have RPATH pointing to the build
directory, so the kernel will be responsible for running the
loader. It enables with --enable-hardcoded-path-in-tests.
In both cases, the loader consults the system cache (ld.so.cache), so
any compiler dynamic library (libgcc_s.so for instance) is supplied by
the system.
However, this might fail for container tests, where even test-container
(the binary that setups and runs the tests in a container mode) creates
a loader cache that might not contain the required libraries. The glibc
build system tries to check and provide the required libraries by
tracing the DT_NEEDED one from a test binary, but the tracing inhibits
the system loader, and thus it will not copy the required system
libraries. It is expected the user to copy the required libraries to
build a directory.
This works for case 2., however, test-container does not issue the
loader for case 1.: it will parse the command arguments and issue the
binary and the kernel will be responsible for starting the loader. This
has the side-effect of ignoring all the defined --library-path and thus
making the compiler libraries copy trick ineffective.
So we copy the compiler libraries on both the build directory and on the
sysroot used by the test-container to set each test. The former handles
the case 1., while the latter will place the libraries on the default
system path, which will be included in the loader cache as well (and
handling case 2.).
Change-Id: I5b90c733a92a5980a8f33e5b6a5c67f768f91906
|
|
Change-Id: Ide588436ab5c3f5f5a4d945cabb5170752ff9ae5
|
|
It seems 'test -nt' does not check the same dates as reported by 'stat -c %Y'.
Change-Id: Ia403b4b23b6768cf3ef3623559b79f758d34b9cb
|
|
... after configure and after build, to help understand why a
subsequent call to ABE decides it's worth rebuilding.
Change-Id: I27ed7103231769280e635ffe9cc76014ef7f3ae4
|
|
Change-Id: Ic0040c14b36af622b8e45cf13b3a616538478219
|
|
Under --enable maintainer_mode, we probably want to enable make_docs,
unless disabled later on the command line.
This ensures that when we activate maintainer_mode, we also make sure
that the (possibly updated) docs can be built.
Change-Id: I78e57eb0fc53ca37e9d72360e4031b59dd612d8a
|
|
Change-Id: I880b743bef731d569386a896f4381ee50170bbd9
|
|
The loop which iterates over gcc, binutils and ${clibrary} expects the
sources of all these packages to be available so that we can copy
information into the manifest. However, this is not true when building
a single package (eg with --build gcc), leading to spurious error
messages in the logs, such as:
fatal: cannot change to 'default': No such file or directory
fatal: cannot change to 'log': No such file or directory
which happen because 'binutils' is defined to 'default' and 'clibrary'
to 'auto' in global.sh, leading to incorrect definition of srcdir.
This patch explicitly skips 'default' and '' values for srcdir.
Change-Id: I7caaf52f0b727e10fc15a322389062999b9b755f
|
|
Merge 'all', 'make_docs' and 'install' logs into a single file, to
make information easier to find/manage in CI.
Insert a "NOTE:" before each fragment to make identification of each
step easier.
Change-Id: I4a90d28b40b6c93ef9a88d43d306df2a478feca0
|
|
GCC's top-level Makefile does not propagate BUILD_INFO to sub-makes,
instead BUILD_INFO is defined at configure time depending on the
availability of a suitable version of makeinfo (so overriding
MAKEINFO=echo like we used to do is sufficient).
However, it seems cleaner to define a fake makeinfo wrapper outside
ABE (like we do in CI), and don't try to play tricks with GCC's
Makefiles.
With this patch, make_docs only enables/disables building the optional
versions of the documentations (eg HTML).
Change-Id: I56652e5048e2898a3f8e74af3cbde96a1a471aa7
|
|
We generate 'make all' logs as
${builddir}/make-${component}${2:+-$2}.log but 'make install' and
'make doc' do not follow the same pattern:
- we store 'make install' logs in ${builddir}/..
- we store 'make doc' logs as makedoc.log (without component name)
This patch makes all of them follow the same pattern:
${builddir}/{make,makedoc,install}-${component}${2:+-$2}.log
Change-Id: If3438fd63ce35e449ca6ea606d313f384efea094
|
|
The default buffer size Expect uses to hold the tool's output is 2000
bytes. This is too low for some GDB tests which run commands generating a
lot of output — e.g., when they turn on GDB debug output or when testing
GDB's Machine Interface which is verbose.
As a result those GDB tests become flaky because depending on the testsuite
run, the expected GDB output may have already been removed from the buffer
by the time Expect tries to match it. Well-written tests shouldn't have
that problem because they are supposed to frequently match GDB's output
(which removes it from the buffer), but in practice a number of tests don't
yet do that. By increasing the buffer size we can reduce the severity of
this problem, increasing the stability of testsuite results.
The Expect man page notes that "excessively large values can slow down the
pattern matcher". With the default buffer size, a tcwg_gdb_check job on
aarch64-linux takes about 115 minutes. With this patch, it takes about 145
minutes. I think the extra time is worth it though.
I also tried buffer sizes of 65k and 200k, but with the former the results
aren't as good as with 100k, and with 200k the results aren't better.
Change-Id: I0466c6f19658f7b638cc0e3dcf5023614de8fa41
|
|
Change-Id: I58ced760f4e45bbc98c18b889ac03c3354bccfca
|
|
For components binutils, gcc, gdb, gdbserver, glibc, newlib this flags
adds --enable-maintainer-mode to the configure flags.
Note that this expacts the appropriate versions of autoconf and
automake to be in $PATH (autoconf-2.69 and automake-1.15.1 at the
moment).
Tested by applying a binutils patch series which requires some files
to be regenerated, otherwise there are many failures when running the
testsuite: we do see the failures without this flag, and they
disappear with this flag.
Change-Id: Iadac4c1905761857f976c7b0aa228c4ac1e8c7d5
|
|
Unlike what we do for GDB, gcc-read1 builds the read1 library in
${abe_top} and makes EXPECT point to the expect-read1 wrapper.
We rely on ABE's copy of read1, at some point we can probably do the
same for GDB for consistency.
Change-Id: I43f7d08961f75968900bfa595d3e15977e5530aa
|
|
Instead of relying on each of the the projects to include support for
READ1, it seems simpler to build it once with ABE, then point to it
when running each project's testsuite. We could also think about
adding to Dejagnu instead, but adding it to ABE seems easier for the
time being.
This patch mimics what GDB does, and actually copies read1.c from GDB.
We now rely on Makefile.in to expand the appropriate paths at
configure time, so that the expect-read1 wrapper is able to find the
read1.so shared lib in ${abe_top}.
Change-Id: I3500eb714e1a05f5ff99bd73b241812e8879ef26
|
|
A recent glibc change (1d5024f4f05) made libsupport (which is used
by all tests) to be built with ‘-fexceptions' and '
-fasynchronous-unwind-tables’. Different than aarch64 and x86,
these options make arm tests require libgcc_so symbols to enable
unwind support ((__aeabi_unwind_cpp_pr0/__aeabi_unwind_cpp_pr1).
This extra required library is handled by glibc test-container,
which is responsible to setup the container tests (all containers
tests now fail due to the missing library).
The easier solution would be to copy both libgcc_s.so.1 and libstdc++.so.6
to the top-level build folder. It also makes the glibc testing less
dependent on the system configuration (for both libgcc_s.so, which is
used on cancellation and backtrace tests, and libstdc++.so).
Change-Id: I53fb3cfb49fdcdfd629075026219221290abbf60
|
|
Similar to GDB's "READ1" mode, I experimented with several buffering
workarounds for GCC.
Since GCC's Makefiles have many check-* targets, it is not practicable
to add a single check-read1 like GDB does. Instead, we could create a
'read1' rule in the top-level Makefile to build read1.so and
expect-read1, like what GDB can do. Then, the usual check-* targets
can be used, and we can override the READ1 environment variable as
needed (as is already supported by GDB). Note that this requires a
GCC patch which has not been submitted yet (and may never be, not sure
how useful/usable it would be).
This patch partially reverts a previous patch which added READ1
support for GDB, and adds the preliminary call to 'make read1' as
needed.
In addition to "READ!" more, this patch adds support for coreutils'
stdbuf-based wrappers, provided as helper scripts. They can be used
by overloading the EXPECT variable when calling 'make check'. A GCC
patch to fully allow that will be submitted.
The advantage of such wrappers is that the machinery is kept part of
ABE.
Change-Id: I184a486c7101550ad3cd0de12ddb7fd26a8deb7f
|
|
Maxim spotted that we are using "for_output" fails/passes to construct the
list of .exp files to re-test. So that even we "decide" that we need to
re-test only a.exp and b.exp, but "for_output" we'll have "a.exp b.exp
c.exp ... z.exp", then "c.exp ... z.exp" will also get re-tested.
Fix by using $new_fails_for_deciding and $new_passes_for_deciding to
generate the $failed_exps list.
Suggested-by: Maxim Kuvyrkov <maxim.kuvyrkov@linaro.org>
Change-Id: Id1aab541e681a2cf7718529b57baa304e1f1013c
|
|
We remove old results from the database, which also removes the flaky
entries they provided. This causes our CI to forget previously detected
flaky tests.
Fix this by making each job record all flaky tests that it detects. So
even if it is later removed from the results database, other results will
also have the same flaky entries. This has the additional advantage of
recording how often each flaky entry occurs.
We do this by not using the baseline flaky results when generating the
lists of flaky test failures and flaky test passes detected in each
testsuite run.
We still use the baseline flaky results to decide whether to run another
try of the testsuite, so this change won't impact runtime cost.
Change-Id: Idf2ec69af9d613511b743f871dc9abfe32e7b977
|
|
Since we often build withouth --send-results-to, and use that option
only when running the tests, don't rely on it when updating
gcc/REVISION which is used when building GCC.
Change-Id: Ia89c864329a5d355811ff414ebdd8977b8f1e96b
|
|
These files are used when sending test results summaries.
Since the CI uses Jenkins rather ABE to checkout the sources, we want
to update LAST_UPDATED and gcc/REVISION later than the checkout phase.
Since we want to include $ABE_TARGET_BOARD_OPTIONS early in the
summaries to make them easier to read/more informative, we have to
update these two files in two different locations:
- gcc/REVISION is used to build the output of 'gcc --version' and has
to be updated before starting the build, so we create it before
running the configure step. The drawback is that at this stage we
don't know about $ABE_TARGET_BOARD_OPTIONS in case the user invokes
ABE once to build the toolchain (without $ABE_TARGET_BOARD_OPTIONS),
and once to run the tests (with $ABE_TARGET_BOARD_OPTIONS, but too
late to update gcc/REVISION). (This is what we do in CI)
- LAST_UPDATED is used only by contrib/test_summary, and can be
updated when running the tests.
Change-Id: I5e50072fc189b7ae41a0829b83155d21294aa282
|
|
because Jenkins creates an error when it is present.
Change-Id: I46580b5aeac077b5ae225de19167a61b03b38dc1
|
|
from mail-recipient.txt to mail-recipients.txt for consistency with
our other scripts.
Change-Id: I44f509cdb350183065ad2c7923da817c7274abcb
|
|
Move this script from the jenkins-scripts repository.
Change-Id: I47ef3ffddf6a711fc9c2526efd6127915039c7dd
|
|
Rename current "default" value to "abe_default". This allows us
to build GCC's upstream default languages.
Change-Id: I7e3b573092d2747ffd3dd300cfe7a464357668a2
|
|
Change-Id: Iff6617aba91c82d6f006dbc28a0c5fae86c362ae
|
|
This option is meant to be used with GCC's contrib/test_summary to
send validation results to gcc-testresults mailing-list.
Instead of emailing the results directly from ABE, we want to let
Jenkins handle this step, so this option provides a way of "parsing"
the email to transform it into a format suitable for our CI.
The CI scripts can provide a filter whose interface would be:
cat BODY | filter -s SUBJECT RECIPIENT
For instance, it can store BODY, SUBJECT and RECIPIENT in different
files, and make them available to Jenkins.
Change-Id: I204900945433dac046815267e6993e61fa5ef39b
|
|
GDB has a test debugging library which is ld-preloaded into expect and
causes it to read one byte at a time from the tested tool. This allows
finding test races (see gdb/testsuite/README), and also makes the testsuite
run more deterministically.
Some testcases are buggy and have additional failures when running under
read1. In particular, gdb.linespec/cpcompletion.exp has 422 FAILs and
gdb.mi/mi2-var-child.exp has 425 FAILs. A couple of tests in gdb.mi are
still as flaky (and verbose) as before though.
The testsuite unexpected failures go from 56 to 1411. Still, the increased
stability of the test results makes this change worth it.
With this change, the check_gdb step goes from about 1h to 5h30m. I'll
explore enabling running the tests in parallel to compensate the increase
in run time.
Change-Id: I1ab812a6ba2492778bca5c87e4480499f76ad6cf
|
|
Newlib installs its header files under
$(exec_prefix)/$(host_alias)/include, but by default GCC expects them
under $(exec_prefix)/$(host_alias)/usr/include.
Before this patch, GCC's configure sets inhibit_libc to 'true', thus
preventing libgcc2 from containing any code which requires libc
support. For instance, this means that libgcov.a is essentially
empty.
This patch adds --with-headers=yes --with-native-system-header-dir=/include,
so that GCC's configure detects the presence of the needed headers.
--with-headers=yes is to be safe, as we don't want it to be set to no,
see this part of GCC's configure:
--------------------------------------------------
: ${inhibit_libc=false}
if { { test x$host != x$target && test "x$with_sysroot" = x ; } ||
test x$with_newlib = xyes ; } &&
{ test "x$with_headers" = xno || test ! -f "$target_header_dir/stdio.h"; } ; then
inhibit_libc=true
fi
--------------------------------------------------
Change-Id: I4521aea24509dbf5021e844e4a9647df24221966
|
|
The default timeout value in the GDB testsuite is supposed to be 1 min (see
gdb/testsuite/config/unix.exp), but is actually 10 seconds because the code
in lib/gdb.exp which sets 'gdb_test_timeout' from DejaGNU's 'timeout'
variable doesn't actually run.
Change-Id: I86c38239d79f7dadb5ee5e8077a74108b814e3b8
|
|
GDB's default timeout value is 1 minute. Setting it to 5 minutes leads to
very long GDB check jobs.
Change-Id: Ie798eeb12c628a55af24a243245d90601e3b0060
|
|
The GDB testsuite uses its own timeout variable.
Change-Id: I1c133cc8846e5d7263af4cc70fc14b25013d0965
|
|
The glibc testsuite supports the TIMEOUTFACTOR environment variable, so
extend that support to the DejaGNU testsuites we run too.
Change-Id: Ica4c086f22e14bd177bc444308c69ecebc0c4e9b
|
|
As an help to read/parse the build/check logs, display the value of
ABE_TARGET_BOARD_OPTIONS before running the tests.
Change-Id: Ia6031a09bb9d2f19362900fa25b7ec4113acd90e
|
|
When producing a manifest file, validate_failures.py also prints it to
stdout. This is a bit confusing in ABE's output so add a notice explaining
what it is.
Change-Id: If8602cf30ae54b534e04d0fe05c54548fb7e0a42
|
|
Change-Id: Icfcc045df67ba30bdb0601551ea97a9a2154b5d5
|
|
Change-Id: I1808a546dd30da503dc37c9b5880a5ea4dc885b6
|
|
... to create tests.log from output of failed glibc tests.
Change-Id: Idabcd45cfb1715d9933e76a71a4fbe678531553f
|
|
... which we no longer generate due to glibc's xcheck testsuite
being poorly supported.
Change-Id: I1a7d7bb571f488c63de68758dc04751a854fbd99
|
|
Change-Id: Ibd843c61fabde084654169624f5dd734745b5229
|
|
Add configure option "--with-additional-debug-dirs=/usr/lib/debug" so that
GDB can find the distro's debug info in /usr/lib/debug. This fixes Ada
tests for tcwg_gnu_native_check_gdb on arm, which use the distro's compiler
and thus need to find the distro's ld.so debug symbols.
At the moment this option is implemented by a patch applied locally, but
it has been submitted upstream.
Change-Id: I08470dd5253083642b64d4355d1d49b94af8d49c
|
|
Dejagnu supports overriding options when running the testsuite, by
using --target_board as part of RUNTESTFLAGS.
However, since we override RUNTESTFLAGS during automatic bisection, it
would be troublesome to maintain the --target_board part. Instead,
append these options to $target_list, which is then parsed by runtest.
For instance:
--set target_board_options="/-O3/-fmerge-constants"
--set target_board_options="{-mcpu=cortex-a53/-O2,-mcpu=cortex-a72}"
--set target_board_options="{-mlittle-endian,-mbig-endian}{-O1,-O2,-O3,}"
See also https://gcc.gnu.org/install/test.html
Change-Id: I99b50e35126bdcba9c05560f5213f4dd4482c96e
|
