While the feature has a great value, it is right now breaking the build of
lang/gcc. Given the importance of lang/gcc it is better to revert now and
reapply the patch once it has been fixed and passes an exp-run on all supported
version
With hat: portmgr
Upstream gcc 4.8 doesn't have support for this - it'll create threads,
but it won't do any of the thread affinity stuff for FreeBSD.
This allows for OMP_PROC_BIND=true to bind threads to their initial
CPUs, leading to some pretty drastic improvements in performance
for certain NUMA workloads.
Approved by: gerald
This was causing the gcc packages to be generated with a
/usr/local/libdata/ldconfig/gcc file. All were conflicting. Bump
PORTREVISION to fix packages built during this time.
With hat: portmgr
Reported by: sunpoet
The larger part of the patch is a backport from gcc trunk which is sent
upstream for approval.
Thanks to Sean Bruno for testing, Andrew Turner for explaining me fine details
and Gerald for approving.
Approved by: gerald (maintainer)
script to assume the BUILD_CONFIG is set to bootstrap-debug, instead of
letting it auto-detect.
With clang 3.5.0 this auto-detection can fail, due to a discrepancy [1]
[2] in its debug information, when objects are produced with and without
-g. When the auto-detection fails, gcc will compare objects with full
debug information during the stage comparisons, and this sometimes
causes those stage comparisons to fail unexpectedly.
[1] http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20141222/250134.html
[2] http://llvm.org/PR22046
Approved by: gerald (maintainer)
to GCC 4.8.3.
This entails updating the lang/gcc port as well as changing the default
in Mk/bsd.default-versions.mk, and it replaces the CONFLICT between the
lang/gcc and lang/gcc47 ports by lang/gcc48.
GCC now uses C++ as its implementation language and performs more
aggressive loop analysis which can be disabled via the
-fno-aggressive-loop-optimizations command-line option.
Compilation of extremely large functions has been signficantly improved,
as have interprocedural optimizations.
A new optimization level -Og has been introduced. It addresses the need
for fast compilation and a superior debugging experience while providing
a reasonable level of run-time performance. This should be better
suitable for development than the default -O0.
A new local register allocator (LRA) has been implemented, which replaces
the 26 year old reload pass and improves generated code quality. For now
it is active on the x86 and x86-64 targets.
AddressSanitizer, a fast memory error detector, has been added and can be
enabled via -fsanitize=address.
Each diagnostic emitted now includes the original source line and a caret
indicating the column.
The new option -Wpedantic is an alias for -pedantic, which is now deprecated.
The C++ frontend and associated run-time library libstdc++ have gained
support for many additional C++11 features. As with previous releases
the Fortrand frontend has seen many improvements as well.
Support for the AArch64 has been added, and there are many improvements
to the x86/x86-64 backend and others.
See http://gcc.gnu.org/gcc-4.8/changes.html for an extense list of changes;
http://gcc.gnu.org/gcc-4.8/porting_to.html for information on how to port
to that new version.
PR: 192025
Tested by: antoine (-exp runs)
series that closes this branch.
Extend full-regression-test by running contrib/test_summary.
Also clean *.la files in LIBEXEC, and recursively so, there and for
TARGLIB.
Since FreeBSD 8.4 and FreeBSD 9.1 make(1) do support :tu and :tl as a
replacement for :U and :L (which has been marked as deprecated)
bmake which is the default on FreeBSD 10+ only support by default
:tu/:tl a hack has been added at the time to support :U and :L to ease
migration. This hack is now not necessary anymore
Note that this makes the ports tree incompatible with make(1) from
FreeBSD 8.3 or earlier
With hat: portmgr
- Add pkg-message that references the need to use -Wl,-rpath=... . [1]
- Replace USE_BZIP2 by USES=tar:bzip2.
- No longer install rebuild-gcj-db47 (which requires bash among others)
and its man page.
Bump PORTREVISION.
PR: 185902 [1]
GCC 4.6.4 to GCC 4.7.3. This entails updating the lang/gcc port as
well as changing the default in Mk/bsd.default-versions.mk.
This adds powerpc64 as a supported architecture (and removes ia64,
though it can be supported by manually installing lang/gcc48).
New binaries %%GNU_HOST%%-gcc-ar47, %%GNU_HOST%%-gcc-nm47, and
%%GNU_HOST%%-gcc-ranlib47 are provided to support link-time
optimization (LTO) which scales significantly better.
And it adds support for indirect functions (IFUNCS), experimental
support for transactional memory in the compiler as well as a supporting
run-time library called libitm, a new string length optimization pass,
and support for atomic operations specifying the C++11/C11 memory model.
Version 3.1 of the OpenMP specification is now supported for the C,
C++, and Fortran compilers.
GCC accepts the options -std=c11 and -std=gnu11 for the C11 revision
of the ISO C standard which inlcude support for unicode strings,
nonreturning functions (_Noreturn and <stdnoreturn.h>), alignment
support (_Alignas, _Alignof, max_align_t, <stdalign.h>), and a
__builtin_complex built-in function.
The C++ frontend now accepts the -std=c++11, -std=gnu++11, and
-Wc++11-compat options and implements many C++11 features of the
language including extended friends syntax, explicit override
control, non-static data member initializers, user-defined literals,
alias declarations, delegating constructors, atomic classes, and more.
The C++ standard library and Fortran frontend have received many
improvements. See http://gcc.gnu.org/gcc-4.7/changes.html for an
extense list of changes; http://gcc.gnu.org/gcc-4.7/porting_to.html
for information on how to port to that new version.
PR: 182136
Supported by: Christoph Moench-Tegeder <cmt@burggraben.net> (fixing many ports)
Tested by: bdrewery (two -exp runs)
area to silence some warnings some are concerned about. [1]
No longer run ccache-update-links as part of post-install which, in
the world of staging, no longer is what it used to be. Rely on the
existing @exec and @unexec in pkg-plist instead. [2]
Submitted by: miwi [1]
Discussed with: antoine [1][2]
Since we now configure with --with-gmp=${LOCALBASE} this is no longer
necessary, and due to bugs in binutils (which should not install ansidecl.h
into ${PREFIX}/include, fixed with revision 336642 [1]) and GCC (which
should search its own include directories with higher priority) could
lead to build failures.
PR: 184327 [1]
the staging infrastructure and had us remove info/gcc46 ourselves. [1]
This has now been addressed in the general infrastructure and actually
causes warnings in some cases. [2]
PR: 184178 [1]
Reported by: amdmi3 [2]
name of this port and avoid a package name collision with other GCC
ports. This also allows us to remove LATEST_LINK.
And it finally allows for a simple and proper CONFLICTS between
lang/gcc and lang/gcc46.
with new versions of texinfo. [1]
Adopt the new LIB_DEPENDS standard and replace USE_GMAKE by USES=gmake.
Remove an obsolete dependency on bison. [2]
STAGEify.
PR: 183342 [1]
Reported by: Christoph Moench-Tegeder <cmt@burggraben.net> [1], marino [2]
- Switched to automake 1.11.6, see CVE-2012-3386.
- #14669: Fixed extraction of CC from gmp.h.
- Fixed case of intermediate zero real or imaginary part in mpc_fma,
found by hydra with GMP_CHECK_RANDOMIZE=1346362345.
This is on top of the following changes from version 1.0
- Licence change towards LGPLv3+ for the code and GFDLv1.3+ (with no
invariant sections) for the documentation.
- 100% of all lines are covered by tests
- Renamed functions
. mpc_mul_2exp to mpc_mul_2ui
. mpc_div_2exp to mpc_div_2ui
- 0^0, which returned (NaN,NaN) previously, now returns (1,+0).
- Removed compatibility with K&R compilers, which was untestable due
to lack of such compilers.
- New functions
. mpc_log10
. mpc_mul_2si, mpc_div_2si
- Speed-ups
. mpc_fma
- Bug fixes
. mpc_div and mpc_norm now return a value indicating the effective
rounding direction, as the other functions.
. mpc_mul, mpc_sqr and mpc_norm now return correct results even if
there are over- or underflows during the computation.
. mpc_asin, mpc_proj, mpc_sqr: Wrong result when input variable has
infinite part and equals output variable is corrected.
. mpc_fr_sub: Wrong return value for imaginary part is corrected.
Convert to the new LIB_DEPENDS standard and remove hard-coded
.so versions from a couple of dependent ports.
Bump PORTREVISIONS of all dependent ports.
PR: 183141
Approved by: portmgr (bdrewery)
on FreeBSD 10, and amd64 on earlier versions.
SSP_UNSAFE is added to disable in a port if it fails to build, but
this should only be used in rare circumstances such as kernel modules.
Otherwise, the port may just be failing due to lack of respecting
LDFLAGS.
On FreeBSD 10, this uses an ldscript in /usr/lib/libc.so to pull in
libssp_nonshared.a to address issues linking on i386 [1].
On earlier FreeBSD versions the WITH_SSP knob will add -lssp_nonshared
to LDFLAGS on i386. This is not needed on amd64. However, several hundred
ports do not currently respect LDFLAGS, so this support is disabled currently
as it causes build failures if a dependency is looking for the stack_chk
symbols.
Many thanks to jlh@ for this as he had many years of patience in getting
all of the necessary pieces [1][2] in.
[1] http://svnweb.freebsd.org/base/head/lib/libc/libc.ldscript?revision=251668&view=markup
PR: ports/138228 [2]
Submitted by: jlh (bsd.ssp.mk based on)
Reviewed by: bapt
With hat: portmgr
exp-runs done: 37 over a month on 91i386,91amd64,10i386,10amd64
Instead of hardcoding the compiler target as FreeBSD, use the OPSYS
variable to it. This makes no practical difference for FreeBSD, but
it helps DragonFly get properly configured.
Approved by: gerald
if present instead of relying on the port we actually depend on.
The issue is that /usr/include/iconv.h has #include <stdbool.h> which
in turn, since both are included very late in the game, conflicts with
similar definintions by libcpp itself.
libstdc++ does not seem to require adjustments, so PR 161417 may not
be relevant in full.
PR: 161417
It's value is "--with-libiconv-prefix=/usr/local" for systems
before 100043 with ports libiconv and to use at systems post
100043 with base iconv it's value is "" (NULL).
Co-authors: bapt, madpilot and bsam (me)
bootstrap of GCC, as opposed to just running a simple build.
Bootstrapping is actually the default upstream, we disable it
by default for the stable flavors of GCC since that is a huge
win in terms of build time of the port. No change in default
behavior for this port.
(When bootstrapping, use bootstrap-lean instead of default/vanilla
bootstrap to save on disk space.)
Remove redundant setting of DISTNAME.
It brings bison as a build dependency in case it is set the following way:
USES= bison or USES= bison:build
it brings bison as a run dependency in case it is set the following way:
USES= bison:run
it brings bison both as a run and build dependency in case it the set the following way:
USES= bison:both
While here trim some headers
Convert some USE_GNOME= gnomehack to USES= pathfix
instead of including a pre-built version of the Eclipse Java Compiler
(ECJ) ourselves. [1]
Replace the use of DISTFILES by DISTNAME, since we are now down to
one in all cases.
Make binutils a build dependency as well (not just a run-time dependency).
PR: 175072 [1]
architectures where the library is not supported. [1]
Revamp the handling of different languages (frontends, run-times)
by GCC to allow for orthogonal setting them in the future. [2]
Print the list of languages being built as part of pre-everything
and shorten the output of operating system and version there.
Reported by: linimon [1]
Inspired by: jkim [2]
Feature safe: yes
in our ports tree. This speeds up a full port/package test cycle
by 31% on a 4-core system; a simple build/install will benefit even
more.
(This may impact compile time by this compiler a bit, depending on
how well the system compiler optimizes this codebase. And we only
should do this for stable, established versions of GCC.)
Remove ABI version numbers for all library dependencies.
