It has been observed that the changes added by
gdb-libexec-add-index.patch will result in GDB testing hanging when
the tests are being run using an in-tree GDB; that is when using 'make
check'. One test that is known to fail is gdb.base/with-mf.exp,
though any test that calls the gdb-add-index.sh script will also hang.
The problem is that when the gdb-add-index.sh script is run, the GDB
testsuite passes the GDB command to use within the GDB environment
variable. For in-tree testing this will be something like:
GDB="/path/to/gdb -data-directory /path/to/data-directory"
Notice that the environment variable contains both an executable and
an argument.
Our changes to gdb-add-index.sh add this:
GDB2=/usr/libexec/gdb
if test -x $GDB2 && ! which $GDB &>/dev/null; then
GDB=$GDB2
fi
The problem then is that '-data-directory' is treated as a set of
options to 'which'. Many of these options are not known to 'which',
but the '-i' option is known. The documentation of '-i' says:
--read-alias, -i
Read aliases from stdin, reporting matching ones on
stdout. This is useful in combination with using an alias for
which itself. For example
alias which=´alias | which -i´.
And here's the problem; this option causes 'which' to read from
stdin. As the GDB testsuite doesn't send any additional input on
stdin then the which command will never complete, and the test will
hang.
The solution I think is to avoid calling 'which' like this on a user
supplied GDB environment variable.
The changes in the gdb-libexec-add-index.patch were really about what
the _default_ GDB executable should be. The upstream version of this
script does this:
GDB=${GDB:=gdb}
That is, the default is just 'gdb'. However, for RH this is not good
enough. We want to handle two additional cases, first, when only the
gdb-minimal package is installed, in which case the default should be
/usr/bin/gdb.minimal. Then we also want to handle the case where the
user doesn't have 'gdb' itself in their $PATH, but does have the 'gdb'
executable installed in /usr/libexec/gdb.
The code as it currently stands also has a problem where, if
gdb.minimal is installed on the machine this will _always_ be used in
preference to the user supplied GDB value (assuming the code worked at
all) this means that when doing in-tree testing we wouldn't actually
be using the in-tree GDB to build the index, which isn't ideal.
So in this commit I propose that we rework our gdb-add-index.sh
changes. Now, we only use the RH special values in the case that
there is no GDB environment variable set. I believe this handles all
the required use cases:
1. When doing in-tree testing GDB environment variable will be set,
and this will always be used as is, with no special processing,
2. When gdb-add-index.sh is used and GDB environment variable is not
set then we will use the first of the following as the default:
(a) /usr/bin/gdb.minimal if this file exists and is executable,
(b) The first gdb executable that can be found in the $PATH,
(c) /usr/libexec/gdb if this file exists and is executable.
While I was changing this patch anyway I've removed the libexec part
of the patch name -- this no longer seemed relevant, I suspect this
related to an older version of this patch.
Remove gdb-6.6-bz230000-power6-disassembly-test.patch as this is now
covered by upstream tests. This patch added some tests of the Power6
disassembler. Disassembler tests are better carried in the gas tree.
Upstream commit:
commit 0fcf99b8ab5ccbde30fa7d36742e670cd4df48ef
Date: Mon Jan 30 11:47:31 2023 +0000
gas/ppc: Additional tests for DFP instructions
Added some new assembler/disassembler tests to the gas tree that
covered all the instructions that were covered by our local patch.
...so that $_gdb_major and $_gdb_minor will be obtained correctly.
Also, change define of 'use_guile' from empty (which rpmbuild
complained about) to a value of 1. I encountered this problem
while attempting to build on F37.
This upstream commit:
commit 24f3aded1d42f515527e2de7e8e9e26f0b77c932
Date: Tue Dec 20 15:01:29 2022 +0000
gdb/testsuite/tui: more testing of the 'focus' command
replaces the tests added in gdb-6.3-focus-cmd-prev-test.patch. As
this patch was only adding a test I see no need to back-port the
upstream commit, we'll pick this up in a later rebase.
- Backport fix for problems associated with GCC 13's self-move warning.
- Tweak gdb-6.6-buildid-locate-rpm.patch so that running GDB's configure
script will not error out due to GCC 13's warnings.
- Backport replace deprecated distutils.sysconfig in python-config.
Fixes RHBZ 2152431, an issue where reading a label symbol's value
would cause GDB to crash. This pulls in two upstream commits
38665d717a3 and c3efaf0afd9 both of which should be in GDB 13 when it
is released.
Changes in this commit:
- Remove gdb-bz601887-dwarf4-rh-test.patch
This adds a very simple test to check whether DWARF4
is implemented. GDB has been using DWARF4 for several
years, and the testsuite coverage supercedes this patch.
- Remove gdb-6.6-buildid-locate-rpm-scl.patch
This patch is applicable only on RHEL6 where .gdb_index
version 5 is used. In recent GDBs, i.e., all DTS releases,
this version could only be read with the option
"use-deprecated-index-sections". RHEL7+ use .gdb_index version 8.
Since RHEL6 is no longer supported, this is unnecessary.
Also remove GDB_INDEX_VERIFY_VENDOR references/code from
gdb-6.6-buildid-locate.patch and gdb.spec
Remove the 'gdb-fortran-frame-string.patch' patch from Fedora GDB. I
have pushed an upstream test that covers everything that this test
covered, see upstream GDB commit:
commit 1c01b23603766fbca4ed4dd12fdd710860e6038e
Date: Fri Nov 4 14:51:11 2022 +0000
gdb/fortran/testsuite: print values and types of string variables
As this is just a testsuite patch, I'm not going to backport this to
Fedora GDB, but when we next rebase, we will pick up this test.
The patch 'gdb-bfd-aliasing.patch' is no longer applied as part of the
Fedora GDB build, since this commit to the Fedora GDB repository:
commit c0cbb1aae4
Date: Sat Mar 19 21:26:41 2016 +0100
.spec cleanup: Drop strict-aliasing GCC bug workaround (from RH BZ 1315191).
I propose that we remove this file from the repository.
The three patches:
gdb-entryval-crash-1of3.patch
gdb-entryval-crash-2of3.patch
gdb-entryval-crash-3of3.patch
are not being applied to the Fedora build of GDB, since this commit to
the Fedora GDB repository:
commit 902c8e5abc
Date: Thu Jul 24 20:04:47 2014 +0200
Rebase to FSF GDB 7.7.91.20140724 (pre-7.8 snapshot).
The contents of the first two were reverting an upstream commit,
which, based on the dialog in the patch files, was done to mask a bug
in a different part of GDB, which I assume is now fixed.
The third patch adds a couple of tests, which are now all upstream,
and passing, I assume these tests covered the issue that was causing
the problem with the first two patches.
Given these patches are not being applied, and the relevant work seems
to be in upstream GDB, I propose we drop these three files from the
repository.
- gdb-physname-pr11734-test.patch
- gdb-physname-pr12273-test.patch
These two tests were merged upstream under a different file name.
- gdb-runtest-pie-override.patch
This patch appears unnecessary today. I have tested on all supported
host architectures and detected no regressions. I presume this was
fixed in a different manner through the years.
- gdb-test-expr-cumulative-archer.patch
The patch contains two tests. Both are upstream already.
While one is a little more involved than what landed upstream,
both tests are equivalent. The other test, gdb.cp/namespace-no-imports.exp,
is identical to upstream's gdb.cp/nsnoimports.exp with whitespace
and other trivial changes.
The test added by this patch from 2005 is a tweaked verion of the upstream
test gdb.threads/watchthreads.exp from 2004. (The current version of
gdb.threads/watchthreads.exp has changed somewhat since 2004.)
The tweaks added to the circa 2004 upstream test consist of setting
a breakpoint on thread_function and then continuing to it. At
that point, initial / runtime values of args[2] and args[3] are
fetched and saved in some TCL variables (which are renamed versions
of args_0 and args_1 in watchthreads.exp). The rest of the changes
to the tweaked script are mostly related to the renaming related to
args_0 and args_1. The other major difference is that watchthreads.exp
places hardware watchpoints on args[0] and args[1] whereas the
tweaked test instead watches args[2] and args[3]. However, all
of these initialized to 0. In the tweaked test, fetching the
initial values when the thread function is first hit should result
in zero values being fetched for args[2] and args[3].
So, in conclusion, the tweaked version of the test doesn't actually
test anything new. I took a look at old posts / threads from the
gdb-patches mailing list, but I couldn't find a post where Jeff
Johnston included this patch (and an explanation of what it does over
the other patch.)
Since this test doesn't offer anything new/better over what's already
upstream, I'm deleting it.
In my .gitconfig I have this:
[format]
useAutoBase = true
thread = shallow
As a consequence, when I run generate-patches-from-git-repo.sh I see
two problems, first, I see lots of errors like this:
fatal: failed to get upstream, if you want to record base commit automatically,
please use git branch --set-upstream-to to track a remote branch.
Or you could specify base commit by --base=<base-commit-id> manually
These are caused by the use of the useAutoBase option.
If I disable useAutoBase, and regenerate the patches, I see that every
patch changes, with the addition of a line like this:
Message-Id: <some message id string here...>
This is caused by the use of the thread option.
Rather than forcing me to not have these settings in my .gitconfig, I
propose that we add --no-base and --no-thread to our invocation of
'git format-patch' in generate-patches-from-git-repo.sh, this fixes
both of the above problems.
With these new options the 'git format-patch' line was getting pretty
long, so I've wrapped it to make it more readable.
This patch disables the gdb-minimal package on GTS for RHEL9. Since
the -minimal package only exists to shrink the buildroot size and
we don't use GTS in the buildroot, there's no reason to build it.
Defining this option will iimit the tests performed with the "--with testsuite"
option:
$ rpmbuild -bb --with testsuite --define 'tests "gdb.base/break.exp gdb.dwarf2/*.exp"' \
gdb.spec
The above will build the RPM with testing, but it will only test the listed tests
(just as if they were passed to TESTS="gdb.base/break.exp gdb.dwarf2/*.exp").
David Abdurachmanov
Kevin Buettner
Alexandra Hájková
Andrew Burgess
Keith Seitz
Bruno Larsen
Fedora Release Engineering
Michal Kolar
Amit Shah