so: when building compile args, meson is deduplicating flags. When a
compiler argument is appended, a later appearance of a dedup'ed is going
to remove a earlier one. If the argument is prepended, the element
*before* the new one is going to be removed. And that is where the
problem reported in https://github.com/mesonbuild/meson/pull/7119 is
coming in. In the revision linked there, the order of replacement in the
prepend case was revesered.
With this patch, we restore this behaviour again.
When meson is currently being run with a python that seems to have been
installed from the Windows Store, replace the general WindowsApps
directory in search paths with dirname(sys.executable), and also handle
failures with pathlib.resolve on WindowsApps exe files.
Use the IDE's OpenMP flag instead of adding /openmp to additional
arguments. The IDE appears to override /openmp in additional arguments
with the IDE setting, which defaults to false, leading to binaries built
without OpenMP.
The linker that comes with MSVC does not understand the /openmp flag.
This results in a string of
LINK : warning LNK4044: unrecognized option '/openmp'; ignored
warnings, one for each static_library linked with an executable.
Avoid this by only setting the linker openmp flag when the compiler is
not MSVC.
It's assumed that where we use DEPFILE in command or rspfile_content, it
can be quoted by quoting the ninja variable (e.g. $DEPFILE ->
'$DEPFILE')
This is nearly always true, but not for gcc response files, where
backslash is always an escape, even inside single quotes.
So this fails if the value of DEPFILE contains backslashes (e.g. a
Windows path)
Do some special casing, adding DEPFILE_UNQUOTED, so that the value of
depfile is not shell quoted (so ninja can use it to locate the depfile
to read), but the value of DEPFILE used in command or rspfile_content is
shell/response file quoted)
(It would seem this also exists as a more general problem with built-in
ninja variables: '$out' appearing in command is fine, unless one of the
output filenames contains a single quote. Although forbidding shell
metacharacters in filenames seems a reasonable way to solve that.)
(How does this even work, currently? Backslashes in the value of all
ninja variables, including DEPFILE were escaped, which protected them
against being treated as escapes in the gcc response file. And
fortunately, the empty path elements indicated by a double backslash in
the value of depfile are ignored when ninja opens that file to read it.)
Now that all command-line escaping for ninja is dealt with in the ninja
backend, escape_extra_args() shouldn't need to do anything.
But tests of existing behaviour rely on all backslashes in defines being
C escaped: This means that Windows-style paths including backslashes can
be safely used, but makes it impossible to have a define containing a C
escape.
We avoided having to get this right previously, as we'd always use a
response file if possible.
But this is so insane, I can't imagine it's right.
See also: subprocess.list2cmdline() internal method
In certain exotic configurations, the style of quoting expected in the
response file may not match that expected by the shell.
e.g. under MSYS2, ninja invokes commands via CreateProcess (which
results in cmd-style quoting processed by parse_cmdline or
CommandLineToArgvW), but gcc will use sh-style quoting in any response
file it reads.
Future work: The rspfile quoting style should be a method of the
compiler or linker object, rather than hardcoded in ninjabackend.
(In fact, can_linker_accept_rsp() should be extended to do this, since
if we can accept rsp, we should know the quoting style)
Rather than ad-hoc avoiding quoting where harmful, identify arguments
which contain shell constructs and ninja variables, and don't apply
quoting to those arguments.
This is made more complex by some arguments which might contain ninja
variables anywhere, not just at start, e.g. '/Fo$out'
(This implementation would fall down if there was an argument which
contained both a literal $ or shell metacharacter and a ninja variable,
but there are no instances of such a thing and it seems unlikely)
$DEPFILE needs special treatment. It's used in the special variable
depfile, so it's value can't be shell quoted (as it used as a filename
to read by ninja). So instead that variable needs to be shell quoted
when it appears in a command.
(Test common/129, which uses a depfile with a space in it's name,
exercises that)
If 'targetdep' is not in raw_names, test cases/rust all fail.
We need to count rsp and non-rsp references separately, which we need to
do after build statement variables have been set so we can tell the
difference, which introduces a bit of complexity.
Writing rsp files on Windows is moderately expensive, so only use them
when the command line is long enough to need them.
This also makes the output of 'ninja -v' useful more often (something
like 'cl @exec@exe/main.c.obj.rsp' is not very useful if you don't have
the response file to look at)
For a rule where using a rspfile is possible, write rspfile and
non-rspfile versions of that rule. Choose which one to use for each
build statement, depending on the anticpated length of the command line.
This is done with the new cmake subprojects options object
that is similar to the already exisiting configuration data
object. It is consumed by the new `options` kwarg of the
cmake.subproject function.
the previous optimizations from 4524088d38
were not relaly good, and not really scaleable, since only the lookup
was improved. However, the really heavy calls to remove have not been
improved.
With this commit we are refactoring CompilerArgs into a data structure
which does not use remove at all. This works that we are building a pre
and post list, which gets flushed into __container at some point.
However, we build pre and post by deduplicating forward. Later on, when
we are flushing pre and post into __container, we are deduplicating
backwards the list, so we are not changing behaviour here.
This overall cuts off 10s of the efl configuration time. Further more
this improves configure times on arm devices a lot more, since remove
does seem to be a lot slower there. In general this results in the fact
that __iadd__ is not within the top 5 of costly functions in
generate_single_complie.
Nirbheek Chauhan
Marcel Hollerbach
Eric Dodd
D Scott Phillips
Dylan Baker
Michael Hirsch
TheQwertiest
Seungha Yang
Kyrylo Polezhaiev
Jussi Pakkanen
Peter Harris
Dan Kegel
Jon Turney
Daniel Mensinger
Richard Brown
Joshua Gawley
Lisa White
Mike Gilbert
jonathanmist