- Declare a PATCH_DEPENDS on emulators/linux_base-8 only if actually using
RTPpatch to apply an Intel provided binary patch.
- Turn the GCC-compatibility of ICC on by default for FreeBSD >= 502108;
except for one bug which is worked around by this port and will be fixed
in src later FreeBSD gained support for using the GCC-compatibility along
with the patch to compile the kernel with ICC (but the ICC 8.0 series
wasn't configurable/hackable enough to actually use it on FreeBSD, which
resulted into the aforementioned bug).
- On FreeBSD >= 502108 default to using libstdc++ from the base as STL
instead of STLport unless "-cxxlib-icc" is passed to icpc (made possible
by turning on the GCC-compatibility and the compatibility to GCC 3.3 and
3.4 which was added to ICC 8.1). On FreeBSD < 502108 STLport i.e.
devel/stlport-icc is and will continue to be the only STL available.
Update the instructions displayed by the post-install target accordingly.
- Put the wrappers for glibc specific symbols and other GNU/Linux compat
hacks into their own library "libiccfbsd" and teach the ld-wrapper to
injected this lib instead of adding these things to the Intel libcxa and
libcxaguard. Beginning with ICC 8.1 non of the Intel libs is "guaranteed"
to be linked into resulting executable (this is actually a fix in ICC
as libcxa and libcxaguard are C++ only). This fixes linking against libm
with icc amongst other things [1].
- Clean the ld-wrapper up a bit. Stop trying to create a perfect world for
the real ld(1) regarding superfluous linkage options, ICC natively passes
far to many of them to the linker that we easily could remove them all.
- Change the ld-wrapper to allow for bootstrapping STLport in a bit
different way that we used to do it, required to make devel/stlport-icc
build correctly again.
- Use fmt(1) to print the infos displayed by the post-install target so
the text is formated properly after the included variables are expanded [2].
Todo: - Rework the freaking thread library selection via the PTHREAD_LIBS
environment variable by the ld-wrapper, this causes really annoying
problems when compiling ports with ICC. Some functionality analogous
to the GCC "-pthread" option (which is also known by ICC but is not
documented and doesn't do the right thing for FreeBSD) would be great.
- Make devel/stlport-icc build again with ICC 8.1 after devel/stlport
has been updated to 4.6.2 (PR 73604). Patch for 4.5.3 already done.
Reported by: Dan Nelson <dnelson@allantgroup.com> [1]
Courtesy of: netchild [2]
Approved by: netchild
- For changes since the 8.0 series see the installed C++ReleaseNotes.htm
but note that information given there doesn't necessarily apply to ICC
on FreeBSD, e.g. -cxxlib-gcc isn't the default on FreeBSD yet and this
port also doesn't install the Eclipse and CDT IDEs.
- ICC now unfortunately requires emulators/linux_base-8.
- Works fine for compiling C source.
- A 6.0-current GENERIC kernel compiles and boots.
- The devel/stlport-icc port currently can't link the exception handling
testsuite with this ICC version (due to relying on a missbehaviour of
the old ICC versions) and has to be changed in a way that doesn't break
lang/icc7.
- Support for using the GCC-compatibility of ICC on FreeBSD and using
the GNU libstdc++ as the STL with ICC is in the works.
o Like with the system GCC, default to libpthread for the threads library
on FreeBSD >= 502102.
Approved by: netchild
In joint forces with: netchild
- add intel-patch target to easy porting effort of future versions [1]
- remove intel debugger rpm, as long as we don't have a libthread_db
we can't use it [2]
Note: The stlport-icc exception handling test will still fail with this
version.
Suggested by (sort of): Marius Strobl <marius@alchemy.franken.de> [1]
Noticed by: Marius Strobl <marius@alchemy.franken.de> [2]
As Intel uses it's own directory for ifc and icc, we don't conflict with
ifc anymore.
Because of ABI changes, you have to recompile C++ programs (don't forget
stlport-icc).
Note that this port is a _work in progress_:
- Icc allows to use an already installed libstdc++ from gcc, this doesn't
work yet on FreeBSD. Libstdc++ on 4.x is too old, so it's unlikely we
can add support for it. The headers of libstdc++ shipping with FreeBSD
5.2-CURRENT use GCCisms not (yet) supported by icc, the hardcoded search
path for them also doesn't fit for FreeBSD 5.2-CURRENT.
- We've incorporated parts (cxa) of the FreeBSD >= 502101 libc on < 502101
systems. It's tested on 4.x, but not on FreeBSD < 502101.
- Not all (new) options (including GCC compatibility) are thoroughly
tested.
When encountering problems please report to me first instead of directly
contacting Intel.
Ackknowledgements:
- Bradley T Hughes <bhughes@trolltech.com> for PR 59552, it resulted in
a modification of our libc (C++ DSO Object Destruction API) we
incorporate in the port on < 502101 systems.
- Marius Strobl <marius@alchemy.franken.de> for his help with the port
(e.g. ld.c, cxa).
- Transform some warnings into errors as suggested by some included
docs (some kind of MSVC compatibility which isn't reverted in icc
for linux).
ld.c:
- add possibility to use a different threads lib via PTHREAD_LIBS
variable (e.g. PTHREAD_LIBS=-lthr) [1]
this may be subject to change when gcc learns how to handle our
different threads libs
- refactor some code [1][2]
- remove mailwrapper license, there's no code from mailwrapper
anymore [2]
- correct the order of libc and libc_r [1][2]
Submitted by: mi [1]
Submitted by: Marius Strobl <marius@alchemy.franken.de> [2]
Reviewed by: Marius Strobl <marius@alchemy.franken.de> [1]
- fix [dfi]vec.h with stlport-iostreams
- do not install a Windows header (mathf.h)
- do not install libompstub (depends on pthread_atfork(), see PR 17437)
Submitted by: marius@alchemy.franken.de
- point to the icc errata after make install
Note:
Object files built with version 6.0 of Intel(R) Compilers may not link
properly with objects built with version 7.0 of Intel(R) Compilers.
Rebuilding of the object files is suggested.
Changes between the submitted version and the committed version:
- update to 078 instead of 075
- treat excessive parameters in function calls as an error instead
of a warning (removed compatibility with a not named MS product, as
suggested by Intel)
- allow to override the INTEL_FLEXLM_LICENSE variable
From the submitter:
- Unbreak usage of signal.h on -current and RELENG_5_0.
(committers note: see below)
- Fix ownership of installed files when built as non-root.
- Fix the patched versions of the icc- and the icpc-script to work with options
containing spaces (e.g. '-DFOOBAR="foo bar"').
- Install 'clicense' and 'csupport' in the docs-dir even if NOPORTDOCS is
defined, the later is not optional but required by 'icid'.
- Move the ld-wrapper to a sub-dir and thus out of what is normally set in
PATH to prevent footshooting (this isn't perfect but simple until someone
reworks this port to provide FreeBSD-compatibility in another way... or
Intel releases a native version).
- Patch the headers to better fit for FreeBSD (hopefully...).
Notes: - Icc7 is more picky about unknow options than the previous versions,
i.e. the gcc-options normally supplied in CFLAGS unless altered in
/etc/make.conf. Therefore compilation of ports with icc likely fail
if the standard CFLAGS are not unset (e.g. `make CC=icc CFLAGS=""`)
or replaced by options valid for icc.
- When compiling multi-threaded C-code make sure to link with icc and
option '-mt' (for ports e.g. via PTHREAD_CFLAGS) otherwise libc_r
won't be linked in correctly.
Submitted by: marius@alchemy.franken.de
Additional information:
- rev 1.20 of src/sys/i386/include/signal.h introduced __aligned which
isn't handled in sys/cdefs.h for non GCC compilers (at least not in
publically available sources)
- the FreeBSD stdarg.h is missing a macro for va_copy() in the non GCC case,
we handle it in the port, but be prepared to get problems when the base
system gets fixed
- we don't use the icc-buildin of alignof()
- Intel provides a float.h which has different floatingpoint values
(e.g. MAX_FLOAT) than we have in the base system, in the port we use the
FreeBSD header (Warner knows about the issue... at least he got a mail
from me and Marius)
- we replace __wchar_t with __ct_rune_t because icc has a build-in type
for it, this affects "typedef __ct_rune_t __wchar_t;" in sys/_types.h.
It isn't known if this may be evil...
- icc doesn't need the option "-mt" anymore to generate thread safe code,
but our FreeBSD version still needs it to be able to choose libc_r
instead of libc
If a src-committer wants to address any of the above mentioned issues he
should first contact me, I may have already some fixes in my local tree.
- Re-enable patching of iccvars.(c)sh (solely called by icid and therefore
overlooked), accidentally broken by last commit.
- Fix generation of multi-threaded code, triggered by the (badly documented)
compiler option "-mt".
- Disable compiler options "-openmp*" and "-parallel", they depend on
libguide which itself depends on pthread_atfork() (not provided by libc_r),
therefore don't install libguide.
Note: All compiler options are now believed to either work on FreeBSD or
be disabled (Maintainer note: '-prof_gen' is under investigation).
- Add a message about installing stlport-icc in order to get full C++ support.
- Remove workaround for generation of shared objects with binutils/ld 2.12.x,
appears to be fixed (unfortunately no confirmation by Intel Support, "That
is good news <...> then we should just close this issue." *sigh*).
Submitted by: marius@alchemy.franken.de
- Add note about a problem on P4 and how to solve it.
Suggested by: Martin Kahlert <martin.kahlert@infineon.com>
- Remove outdated information in pkg-descr.
- use an empty MASTER_SITES to not get mail from fenner's distfile survey
Suggested by: fenner
- correct the message in IGNORE after obrien's change to linux-base
- fill in the product name in docs/csupport
