Eventually I'd like to automate this so that one or the other of these
files is canonical and the other is generated from it. Until then, add
this check to the release process to avoid a skewed release being
shipped.
On Linux, a C program that depends on a C library which in turn
depends on a C++ can be linked with the C compiler, (avoiding a direct
link from the program to the C++ runtime libraries).
Other platforms with less fancy linkers need to use the C++ compiler
for this linking.
Useful for verifying that our tar-file creation works. The tar-file
name can't easily be used as a target directly since it depends on the
current git revision.
Theese were previously pointing to "make VERSION=X.Y release", but
we've recently changed to an alternate scheme involving the updated
version in a file named "version".
We do this so that "git archive" produces a usable tar file without us
having to post-modify it, (since tools like git-buildpackage might not
give us an easy way to hook into the tar-file-creation step).
To support this we also have to change our preference to prefer the
git-described-based version (if available) and only if not available
do we fallback to using what's in the "version" file. Finally, we also
ovverride this preference when releasing, (where what's in the
"version" file wins).
Note that using our Makefile's rule to create a tar file still will
insert the git-based version into the tar file. This is useful for
creating snapshots which will correctly report the git version from
which they were created.
David Bremner informs me that shoving everything from the notmuch "git
log" into the debian/changelog is a bit excessive. Instead, we'll
start manually updating this file, (which feels a bit redundant with
NEWS, but perhaps makes us a better Debian-comunity member).
On Bdale Garbee's recommendation I'm switching from gitpkg, (which
constructed a source tree but still required me to go run debuild), to
git-buildpackage. I hadn't originally used git-buildpackage because it
didn't seem to work without a configuration file, (where gitpkg was
fine).
Bdale was kind enough to point me to his fw/altos source at
git.gag.com where I found an example gpb.conf file as well as a target
in debian/rules to automatically update debian/changelog with the new
version number.
This reverts commit fbec989fe3.
I only pushed this accidentally. See message
id:871ver6u9r.fsf@yoom.home.cworth.org for the various reasons I
didn't like this patch, (mostly I think the association of 'F' is
wrong).
We previously output "notmuch version 0.1" as response to notmuch --version.
Shorten this to "notmuch 0.1" as we know that we will receive a version
number when we explicitely ask for it.
Signed-off-by: Sebastian Spaeth <Sebastian@SSpaeth.de>
With the recent addition of "*" being a special case for a search
matching all messages, we have to take care when doing a filter
operation. In this case it's not legal to simply append and get:
* and <some-new-search-terms>
Instead we carefully construct a new search string of only:
<some-new-search-terms>
This could all be avoided if we had a parser that could understand
"*" with the meaning we want.
It was noted today in IRC that libnotmuch is not yet careful about
wrapping all Xapian calls with try/catch blocks to print nicer error
messages. It seems it would be natural to audit that at the same time
as doing the symbol-hiding work.
We actually want this version to be incremented by the commits that
extend the interface. So the release process really is not to just
verify two things (NEWS and libnotmuch version), then run "make
VERSION=x.y release", and send the mail. Quite nice.
The entire "make sure the code you want is in place" thing is part of
a larger release process that we don't document here at all. Instead,
we just focus here on the mechanics of pushing things out once the
larger process has determined the code is ready.
And the fewer steps there are, the better, (for making the
release-process as painless as possible and for avoiding any
mistakes).
Before and after the assignment operator, no spaces are allowed.
I don't know if there are any /bin/sh which allow spaces, but at least
in bash, csh and zsh, the former code was no valid assigment.
We put verify-version as a dependency, not a recursive action to keep
its output clean, (I know that I will always type "make release"
instead of "make VERSION=X.Y release" so I want a nice, neat
reminder).
Also, put the various ssh-based commands together, and after the
build, (so that it doesn't ask for a password/passphrase both before
and after building).
Previously, we had a separate release-upload target that a user might
mistake as something useful to call directly, (which would have the
undesired effect or uploading a new package, but without first making
all the checks that we want).
So we eliminate that target, (folding its actions into "make
release"), and we also rename the several release-verify-foo targets
to simply verify-foo. This leaves as the only targets with "release"
in the name as "release" and "release-message". Both of these are
intended for the user to call directly.
We've now changed to using "git describe" to automatically report a
version number that changes with every git commit. So we no longer
need to manually update anything in the Makefile during the release
process.
Hurrah---no more manual verification of that PASS column.
This means that "make test" can actually be a useful part of the
release process now, (since it will exit with non-zero status if there
are any failures).
Previously some tests (dump/restore) were doing ad-hoc verification of
values and their own printing of PASS/FAIL, etc. This made it
impossible to count test pass/fail rates in a single place.
The only reason these tests were written that way was because the old
execute_expecting function only worked if one could directly test the
stdout output of a notmuch command. The recent switch to pass_if_equal
means that all tests can use it.
I just wasted far too much time looking for a bug that wasn't actually
there only because I hadn't recompiled before running the test
suite. Now we can take advantage of actual dependency information to
force a rebuild for "make test".
When constructing a thread, we usually run a nested query to find all
messages in the thread that match the original search string. However,
we need to have special-case handling of an original search string of
"*" now that that is a supported means of specifying all messages.
The special-case ends up bein quite simple---we do less work, (just
skipping the nested search since we know that all messages must
match). I had been wanting to write this identical code to more
efficiently handle "notmuch search thread:<foo>" which was previously
running two identical searches. So that case is now more efficient as
well.
This feature was added recently and should have gotten a new test at
the time.
As this test demonstrates, the code is broken, ("notmuch search '*'
returns bogus dates of the Unix epoch for any threads where the
term "and" does not appear in any messages).
Using a date in the current year makes the test suite fragile since
the search output will include a date of "January 05" for now, but
will start doing "2010-01-05" in the future.
The filenames aren't predictable (including the current directory) nor
stable from one run to the next (including the PID). This makes it
hard to predict the output from a search command that returns such a
message (such as "*").
The original goal was simply to ensure that each generated message was
distinguishable somehow. So just use the message counter instead.
The old execute_expecting function was doing far too much for its own
good. One of the worst aspects of this was that it introduced
shell-quoting challengers where the caller could not easily control
the precise invocation of the command to be executed.
I personally couldn't find a way to test "notmuch search '*'" without
the shell expanding * against files in the current directory, or
having bogus quotation marks appearing in the search string,
(defeating the recognition of "*" as a special search term).
Hopefully this aspect of the test suite will be much easier to maintain now.
One of these is a bad bug I noticed this morning, (archiving messages I had
never read when going through a search of "tag:inbox and tag:to-me" and
hitting space bar).
The other ideas came from recent conversations with Dirk and Eric.
Recent coding around the "*" feature suggests some improvements that
we could make, (some of which might push us into writing a custom
query parser rather than using the one that exists in Xapian).
The recent fix to properly decode encoded headers made the expected
output of "notmuch reply" differ by a single space, (previously, there
were two spaces before the References: value and now there is just
one).
Fix the test suite so that these are all noted as correct results
again.
Apparently the OS X linker can't resolve symbols when linking a
program (notmuch) against a library (libnotmuch) when the library
depends on another library (libgmime) that the program doesn't depend
on directly.
For this case, we need to link the program directly against both
libraries, but we don't want to do this on Linux, where the linker can
do this on its own and the explicit, unneeded link would cause
problems.
This encodes the library version into the library, where the linking
binary can pick it up, and the linker can even enforce mismatches in
the minor release, (such as linking a binary against version 1.2 and
then attempting to run it against version 1.1).
I'm not sure which system Aaron used, but on the machine I have access
to, (Darwin 8.11.0), the -shared and -dylib_install_name options are
not recognized. Instead I use -dynamic_lib and -install_name as
documented here:
http://www.finkproject.org/doc/porting/shared.php
