Profiling showed that we were spending 25s inside os.path.realpath()
on Windows while generating compile lines for build.ninja, inside
NinjaBackend.generate()
The real path for these will not (should not) change during a single
meson invocation, so cache all these.
Brings build.ninja generation from 73s to 47s on my machine.
The documentation for build_target(...) does not list file or str as
the possible types for the "objects" keyword argument, even though in
theory the argument is meant for prebuild object files that are part
of the sources.
Of course that is only the theory, because an ExtractedObjects object
is probably used a lot more than a file in the source tree. But
at least make the reference manual's typing information accurate.
TAP 14 states:
> Harnesses may treat any TAP stream lacking a version as a failed test.
TAP 13 states:
> In the absence of any version line version 12 is assumed. It is an
> error to explicitly specify any version lower than 13.
So, modern TAP is saying that we should treat a missing version as a
test definition bug, it's no longer okay to use a missing version as
saying "let's use TAP 12". But, we can choose whether to treat it that
way or error out.
Let's do a diagnostic, as we do elsewhere. But allow TAP streams that
aren't well defined, if they used to be well defined (back in TAP 12).
Darwin-based systems, at least macOS, provide various JDK executable
stubs in
/System/Library/Frameworks/JavaVM.framework/Versions/*/Commands.
These stubs are placed in such a way that they break the heuristics of
the JNI system dependency. If a javac being analyzed to find a Java home
is a stub, use /usr/libexec/java_home.
See https://stackoverflow.com/a/15133344/7572728 for more details.
Closes#11173
ctypes uses FFI, and surprisingly often people's Python installations
will be broken because ctypes is broken (e.g. the system libffi has been
updated and Python needs to be recompiled). That is not our fault, but
it does manifest as Meson failing to run. It turns out we aren't even
using it though. At least, pretty often.
We have two uses of ctypes, and both of them are for Windows. One of
them is already conditionally imported in the function that uses it, but
the other is imported at startup. Move this down into the invoking
function.
On non-Windows systems, it is now impossible for Meson to fail to run
when ctypes is broken, because we don't use it. Anecdotally, this issue
tends to come up on Linux systems primarily.
Fixes#11111Closes#11112
Link to feature options consistently, and point out that it controls
"whether" the function finds what it's trying to find. This clues people
in to the fact that disabled features exist.
fix_jar() tries to remove an existing Class-Path entry from the jar
manifest by postprocessing the manifest and passing it to `jar -um`.
However, `jar -um` can only add/replace manifest entries, not remove
them, and it also complains loudly when replacing an entry:
Dec 13, 2022 7:11:19 PM java.util.jar.Attributes read
WARNING: Duplicate name in Manifest: Manifest-Version.
Ensure that the manifest does not have duplicate entries, and
that blank lines separate individual sections in both your
manifest and in the META-INF/MANIFEST.MF entry in the jar file.
Thus fix_jar() produces one such warning for each entry in the manifest
and accomplishes nothing else.
Use jar -uM instead. This completely removes the manifest from the jar
and allows adding it back as a normal zip member, fixing fix_jar() and
avoiding the warnings.
Fixes: https://github.com/mesonbuild/meson/issues/10491
Fixes: c70a051e93 ("java: remove manifest classpath from installed jar")
Rather than passing arguments directly to less, set the LESS environment
variable to contain the desired arguments instead. This allows passing
arguments in case the user has PAGER=less set in their environment.
At least, if you tried to use it when passing an install_dir. Because
T.Sequence is horrible and we should never use it, and the annotations
are a lie that produces bugs.
So, fix the annotations on CustomTarget to never allow this to happen
again, and also fix the function too. Move some definitions elsewhere
inline to satisfy the linter.
Fixes#11157
T.Sequence is a questionable concept. The idea is to hammer out generic,
maximally forgiving APIs that operate on protocols, which is a fancy way
of saying "I don't care if you use tuples or lists". This is rarely
needed, actually, and in exchange for this fancy behavior you get free
bugs.
Specifically, `somestr` is of type `T.Sequence[str]`, and also
`somestr[0]` is another string of type you guessed it. It's ~~turtles~~
strings all the way down.
It's worth noting that trying to code for "protocols" is a broken
concept if the contents have semantic meaning, e.g. it operates on
"the install tags of this object" rather than "an iterable that supports
efficient element access".
The other way to use T.Sequence is "I don't like that T.List is
invariant, but also I don't like that T.Tuple makes you specify exact
ordering". This sort of works. In fact it probably does work as long as
you don't allow str in your sequences, which of course everyone allows
anyway.
Use of Sequence has cute side effects, such as actually passing lists
around, knowing that you are going to get a list and knowing that you
need to pass it on as a list, and then having to re-allocate as
`list(mylist)` "because the type annotations says it could be a str or
tuple".
Except it cannot be a str, because if it is then the application is
fatally flawed and logic errors occur to disastrous end user effects,
and the type annotations:
- do not enforce their promises of annotating types
- fail to live up to "minimal runtime penalties" due to all the `list()`
Shun this broken concept, by hardening the type annotations. As it turns
out, we do not actually need any of this covariance or protocol-ism for
a list of strings! The whole attempt was a slow, buggy waste of time.
There are two problems here: a typing problem, and an algorithm problem.
We expect it to always be passed to CustomTarget() as a list, but we ran
list() on it, which became horribly mangled if you violated the types
and passed a string instead. This caused weird*er* errors and didn't
even do anything. We want to do all validation in the interpreter,
anyway, and make the build level dumb.
Meanwhile we type it as accepting a T.Sequence, which technically
permits... a string, actually. This isn't intentional; the point of
using T.Sequence is out of a misguided idea that APIs are supposed to be
"technically correct" by allowing "anything that fulfills an interface",
which is a flawed concept because we aren't using interfaces here, and
also because "technically string fulfills the same interface as a list,
if we're talking sequences".
Basically:
- mypy is broken by design, because it typechecks "python", not "what we
wish python to be"
- we do not actually need to graciously permit passing tuples instead of
lists
As far as historic implementations of this logic go, we have formerly:
- originally, typeslistified anything
- switched to accepting list from the interpreter, redundantly ran list()
on the list we got, and mishandling API violations passing a string
(commit 11f9638035)
- switched to accepting anything, stringlistifying it if it was not
`None`, mishandling `[None]`, and invoking list(x) on a brand new list
from stringlistify (commit 157d438835)
- stopped stringlistify, just accept T.List[str | None] and re-cast to
list, violates typing because we use/handle plain None too
(commit a8521fef70)
- break typing by declaring we accept a simple string, which still
results in mishandling by converting 'foo' -> ['f', 'o', 'o']
(commit ac576530c4)
All of this. ALL of it. Is because we tried to be fancy and say we
accept T.Tuple; the only version of this logic that has ever worked
correctly is the original untyped do-all-validation-in-the-build-phase
typeslistified version.
Let's just call it what it is. We want a list | None, and we handle it too.
This is no longer implicitly installed due to libxslt. Actually though,
we don't need to test this in order to ensure that the custom dependency
works -- we have other jobs that test it, and the config-tool handling
itself won't suddenly fail on msys2 specifically.
In particular, CUDA 12 removes support for Kepler (3.x) entirely.
Unusually, however, it does not introduce any new architectures,
or even compute capabilities.
When finding a py.dependency() we try to use pkg-config. We then apply
our own custom base class, which replaces self.name with the informative
comment "override the name from the "real" dependency lookup", to which
I can only say "uhhh why". Why do we want to do that???
It turns out we don't, it was just a really old legacy design because we
had a SystemDependency with a .pkgdep attribute hiding the real
dependency bizarro-land style. We cleaned that up in commit
4d67dd19e5 and as part of that, we
*shifted over* the self.name assignment to preserve the visible effects,
sort of. We didn't have a *reason* to override the name, we just did it
because... we weren't sure whether it mattered.
Unfortunately it very much does matter the other way -- we don't want
it. We can pass this dependency to the pkgconfig module, which uses the
name attribute to fill out the `Requires: ` field. Also, the name should
name what we have. :p
Get rid of this bizarre historic quirk. Since we have proper
dependencies here, we should go all in.
Fixes https://github.com/ufo-kit/ufo-core/pull/185#issuecomment-1328224996
It is only used by Environment.get_exe_wrapper() and every callers were
handling None already. Type annotation was wrong, it already could
return None for the case an exe wrapper is needed but none is provided.
Jussi Pakkanen
Nirbheek Chauhan
Oliver Kraitschy
Paolo Bonzini
Michael Hirsch
Xavier Claessens
Amit D
Eli Schwartz
Tristan Partin
Benjamin Gilbert
Brad Smith
Andreas Deininger
Phil Jones
Ewout ter Hoeven
Olexa Bilaniuk
L. E. Segovia