LOOKUP_PKG_RULE is called many times and spends a lot of time searching
for rules when the default rule is the most common. This simply
avoids exhaustive searching since we can determine beforehand if there
are multiple rules.
Also drops some of the gmsl functions if favour of normal make
variables.
First commit adds sort/strip so print-deps-for-build-pkg output can be
easily compared.
```
$ git checkout HEAD~1
$ time make print-deps-for-build-pkg \
MXE_TARGETS="`echo {i686-w64-mingw32,x86_64-w64-mingw32}.{static,shared}`"> orig
real 0m17.564s
user 0m16.410s
sys 0m0.783s
$ git checkout speedup
$ time make print-deps-for-build-pkg \
MXE_TARGETS="`echo {i686-w64-mingw32,x86_64-w64-mingw32}.{static,shared}`"> new
real 0m4.036s
user 0m3.624s
sys 0m0.326s
$ diff -u orig new
$
```
There are various cases where a download succeeds but the checksum
is wrong:
- upstream re-releases tarball without new version
- sourceforge returns 200 with a text message about outage
- other strange glitches
In these cases, we currently fail and stop the build. This PR allows
for falling through the next URLs until it will eventually get to
the checksum-suffixed version.
This ensures builds are always (potentially) repeatable as the checksum
associated with the version is reachable. Next step would be to change
all the backups to use suffixed tarballs.
* change libssh2/apache/xiph URLs to canonical ones
* resubmit patch to remove unnecessary
--no-check-certificate option. The URLs
are non-HTTPS anyway
* blas: restore URL and point to versioned archive
It is needed to specify a path to custom wget binary without
removing the argument passed by WGET (--user-agent).
Also can be used to add --no-check-certificate back.
It is used in libotr (toolkit/Makefile.am) to define a list
of symlinks to the main man page. Search for MANLINKS reveals
that it is used in many other projects.
The rationale: bd2c77f4c2 (commitcomment-21229420)
> If you absolutely want to disable secure transport I think it'd be a
> better choice to use the http:// protocol instead, making the
> insecurity unambiguously visible in logs/screen output. (Not sure if
> GitLab supports it, but the first two does I think.)
Regular downloads of packages are verified by checksums, so
--no-check-certificate doesn't compromise the build system,
but the checksums themselves are often updated with update-checksum-%
which in turn calls the regular package download mechanism, so there
is a possibility of downloading and sealing a poisoned file.
On the one hand, old systems may still rely on --no-check-certificate,
so it is not nice to completely disable it for regular downloads.
However keeping this option enabled for backup servers only is enough
to support such systems because of the fallback mechanism.
On the other hand, download from a backup doesn't make sense while
updating a package, because the package is definetely not in the backup yet.
So --no-check-certificate is now enabled only for backup servers
and backup servers are disabled while updating packages.
See https://github.com/mxe/mxe/pull/1694#issuecomment-285324739
- move cmake configuration from mxe-conf to cmake-conf
- replace `echo` with templates for readability and maintenance
- allow packages to set other dep files
- set CMAKE_POLICY_DEFAULT_CMPNNNN in wrapper since
`cmake_minimum_required` or `cmake_policy` can't be set in
toolchain (closes#971)
Minimal implementation to strip the [largest files][lf-gist] by
default, mostly made up of gcc/binutils and test programs.
gdal and geos both produce large libraries, but the libs themselves
aren't worth stripping, it's the 20 odd programs produced by gdal
with those libs statically linked that consume the most space.
I'm leaving these undocumented as the defaults seems reasonable and
the interface may well change when we enable debug/release variants.
closes#985closes#1249
[lf-gist]:https://github.com/mxe/mxe/issues/1249#issuecomment-193392038
Tony Theodore
Viktor Szakats
Boris Nagaev