- USE_CROSS_COMPILATION activates it, CROSS_DESTDIR specifies root of
the target filesystem
- derive _CROSS_DESTDIR from CROSS_DESTDIR or MAKEOBJDIR
- buildlink3.mk prefixes the files to symlink with _CROSS_DESTDIR
- compiler/gcc.mk knows about the target prefix (e.g. i386--netbsdelf)
- PKG_DBDIR is prefixed with _CROSS_DESTDIR
- package-install and bin-install are not called with su
- install and strip are redirected to the tool version
- links for the target specific ar, as, ld, nm, objdump, ranlib and
strip are added
- compiler wrapper detect if linking is requested or not
- special command sinks for CPP and CC/CXX add the cross-compile magic:
- modify include dirs to get the target /usr/include
- modify linker dirs and runpath to use target /usr/lib at link time,
but keep correct rpath entries
Supported-by: Google SoC 2007
Basic tests by he@ on Sparc. Review from jlam@.
line is no longer printed by default. To get that behavior back, you can
set the variable WRAPPER_DEBUG to "yes".
This avoids some noise during the build. The old behavior was mostly
useful on Solaris and IRIX, where a -std=c99 or -c99 option had been
added to the compiler's arguments by the wrapper. This caused
diagnostics that were hard to understand, since the user did not specify
these flags and there was no sign that they had been added.
--version-script works are few enough to warrant fixing each one of them
instead. Silenty dropping the flag makes it impossible to detect whether or
not --version-script is supported. Pointed out by joerg@
option. Instead of just merging them, the wrapper fails. Let's see which
packages get broken by that and then fix them.
I've also heard rumours that a plain -L is used as an alias to -nostdlib
in some compilers. This would have been handled wrong up to now, since
it definitely wasn't intended to be merged with the following argument.
-32 does _not_ belong to ABI 32. Instead, the mapping is:
ABI Options
(undef) -o32 -32
32 -n32
64 -64
While here, allowing -OPT:* to pass through to the compiler, since it
will understand it.
This should fix PR 35729.
wrapper_abi variable. This is used in the MIPSpro wrapper to reject any
package that tries to compile a file for the wrong ABI. This is not just
a warning because such a package most likely has more severe problems.
The gcc -Wall option seems to be similar to the MIPSpro -fullwarn
option, so map them.
-xcode=pic13 and -xcode=pic32, as suggested in PR 35555.
This also fixes and old bug where -fPIC (32 bit) had been translated to
-kpic (13 bit), but apparently that doesn't make a big difference in
practice.
transformations can decide whether they happen in a C or a C++ compiler.
Now the wrappers don't depend on the basename of the executed compiler
wrapper anymore.
Fixes PR 35555.
options. Some others (c99, CC) can. To avoid linker errors about
"illegal option -- 1", these options are converted to -xO0 and -xO1,
which work.
Committed during the freeze to avoid breakage of packages. The code that
passed the -O0 and -O1 options through is relatively new, so there may
not have shown packages that break because of this.
the wrapped command and exit. This will be used by libtool which needs
to figure out the true command name. Yeah, it is ugly behaviour from
libtool, but I'm not going to rewrite libtool.
Change ok-ed by jlam.
may lead to failure for some packages, a note to the user is printed
when a wrapped command fails, including the "real" command line of the
wrapped command.
This change arose from the implicit addition of the -std=c99, -std=gnu99
or -c99 options, which made some packages fail with obscure error
messages. This change helps users get to the real cause more quickly.
When discarding arguments, split_arg is set to yes, so that no empty
arguments are added to the resulting command line. A more informative
debugging message is also printed in that case.
Added a debugging message to transform_to(), for consistency with the
other functions.
Added some code to check for unknown options.
Reindented the file so that the compiler options are better visible.
They had been hidden by the massive block comments before. Also, every
option has got its own line so that one only needs to read the first
column to find the relevant code.
make the wrapper code less frightening. For example, instead of
arg="${arg#-Wl,}"
$debug_log $wrapperlog " (transform-sunpro-cc) to: $arg"
addtocache=yes
one can simply write
transform_to "${arg-Wl,}"
The functions that are currently implemented are: transform_to,
transform_discard, transform_to_nocache, transform_discard,
transform_pass and transform_pass_unknown.
cache together with MAKE_FLAGS=-j. The cache file is now regenerated
into a temporary file and then renamed (atomically) to the "real" cache
file.
Fixed some code that would fail when run in "set -e" mode. If you want to
test the "set -e" mode, just uncomment the line in wrapper.sh.
command line
$debug_log $arguments
tries to execute the empty command. This results in error messages of
the form
cc[50]: : cannot execute (IRIX)
bash: : command not found (Bash)
: permission denied (NetBSD /bin/sh)
Setting IFS to the original value as soon as possible fixes these bugs.
Fixes PR 34135.
http://mail-index.netbsd.org/tech-pkg/2006/07/30/0005.html
2. Too much is logged in the .wrapper.log
The wrappers log quite a lot of detailed information by
default, with the result that my .work.log file is 176236028
bytes {for ghc}.
The wrapper functionality is mature, so this detailed
information is seldom needed, and you can in general determine
what happened by looking at the input/output-lines (i.e.
those lines marked as "[*]" and "<.>" in the .work.log).
I would therefore like to change the default behavior to
only log the input/output-lines.
Flip the switch for WRAPPER_DEBUG from "yes" to "no" by default so
that the detailed and rather verbose "push", "pop", "transform", etc.
messages are not printed. While here, also make WRAPPER_DEBUG a
pkgsrc-private variable by prepending with an underscore.
compiler toolchain until the configure stage, move the "wrapper" stage
back to between "patch" and "configure".
This makes "make extract NO_DEPENDS=" work again for all packages.
when passing through the barrier. This ensures the PATH (passed via
PKGSRC_MAKE_ENV) is correctly set for all phases after the barrier.
This fixes a bug in "interactive" pkgsrc use, where if you have no
work directory and type "make build && make install", then the "install"
step does not have a PATH set to include all the wrapper and tools
directories.
introducing the concept of a "barrier". We separate the user-invokable
targets into ones that must happen before the barrier, and ones that
must happen after the barrier. The ones that happen after the barrier
are run in a sub-make process. In this case, the targets that must
be run after the barrier are from the "wrapper" step and beyond. We
rewrite the various "flow" targets, e.g. wrapper, configure, build,
etc., so that they of the right form to use the barrier target.
This now completely removes the concept of PKG_PHASE from pkgsrc. It
is replaced with the concept of "before" and "after" the barrier, and
this state can be checked by testing for the existence of the barrier
cookie file. Because we've removed most of the recursive makes, there
is now nowhere to hook the PKG_ERROR_HANDLER.* commands, so remove
them for now.
As part of this commit, put back the logic that conditionalized the
sources for the various cookie files. Because the sources are all
"phony" targets, they were always run, regardless of whether or not
the cookie file already existed. Now, if a cookie file exists, then
that entire phase associated with that cookie file is skipped.
Lastly, fix a thinko in configure/bsd.configure.mk where setting
NO_CONFIGURE in a package Makefile would manage to skip the "wrapper"
step altogether. Fix this by correctly noting "wrapper" and not
"patch" as the preceding step to "configure".
of the logic from fetch/fetch.mk into flavor/pkg/check.mk, so that
check-vulnerable can be used as a source target.
Make check-vulnerable a source target for every phase of the build
workflow, which ensures that it is always run if the user starts a
new phase from the command line.
Fix the cookie-generation targets so that they don't append, only
overwrite to the cookie file. This works around potential problems
due to recursive makes.
Move the cookie checks so that they surround the corresponding phase
target. The presence of the cookie should now inform the make process
to avoid doing any processing of phases that occur before the phase
corresponding to the cookie.
and into their own directories. Also do some cleanups with build/_build
and pkginstall -- we get rid of _build and simply run pkginstall as
part of the "build" target.
Introduce a new mechanism to handle varying directory depths under
${WRKSRC} in which we find files to override, e.g. configure, config.*,
libtool, etc. OVERRIDE_DIRDEPTH is a package-settable variable that
specifies how far under ${WRKSRC} the various targets should look,
and it defaults to "2". We preserve the
meaning of the various *_OVERRIDE variables, so if they are defined,
then their values supersede the OVERRIDE_DIRDEPTH mechanism.
devel/tla will need to specially set OVERRIDE_DIRDEPTH to 3 (see log
for revision 1.1857 for bsd.pkg.mk -- to be done in a separate commit.
