Remove expired port:

2019-10-29 lang/gcc5: Unsupported by upstream. Use GCC 9 or newer instead.

Now that GCC 9 is the canonical version in the Ports Collection,
recommend that (instead of GCC 9) in the deprecation notes of ports
for old versions of GCC.

After some 19 months of deprecation, mark this port for expiration in
two months (end of October 2019).

GCC 5 went end-of-life upstream in October 2017, so this port is now
scheduled for removal more than two years after that.  Nothing in
the Ports Collection depends on it any longer (and has not for a
while), and the default version of GCC in the ports tree is GCC 9,
four major releases later.

While here stll create a proper USES block in the Makefile. [1]

Reported by:	portlint [1]

lang/gcc*: Hide pkg-message during upgrades

PR:		239419
Approved by:	gerald (maintainer)

Recommend the use of GCC 8 (instead of GCC 7 until now) to replace
these older versions of GCC - GCC 4.8, GCC 5, and GCC 6 - which have
been end-of-lifed upstream many moons ago.  GCC 8 has been the default
version of GCC in the Ports Collection for a while and as such proven
itself, plus of all versions it is most likely to be present/used.

No functional change, just updated advice to our users.

On powerpcspe configure GCC --with-cpu=8548 instead of --with-cpu=8540,
which caused the build to fail after the update to binutils 2.31 and was
factually incorrect anyways (since the oldest we support is 8548).

PR:		235393
Reported by:	jhibbits

Install texinfo files (GNU info) into ${PREFIX}/share/info

After a discussion on the mailing list on moving manpages to
${PREFIX}/share/man for consistency with base where it is
installed in usr/share/man, it appeared the same should happen
to GNU info files which were installed under share in base and
not in ports.

Now texinfo is not in base on any of the supported version of FreeBSD
it is possible to proceed to this move and it is easier to do than
the manpage change.

Other benefit than consistency are less patching: all build tools but
cmake are expecting info files to be under share/info and cmake (patched here)
was having an exception for BSD so the patch makes FreeBSD case less
specific for them

Bump revision of all impacted ports

PR:		232907
exp-run by:	antoine
Differential Revision:	https://reviews.freebsd.org/D17816

Recommend the use of GCC 7 or later over this port, since that is now
the default version of GCC for the Ports Collection.

And backport r478722 | gerald | 2018-09-01 from lang/gcc7: [1]

Disable the build/use of libssp/gets-chk since FreeBSD 12 and later
do not feature gets() any longer.

PR:		222796, 231066 [1]
Differential Revision:	https://reviews.freebsd.org/D12298 [1]

After an include, PLIST_SUB must be appended to, not overwritten.

If it is overwritten, many values will be clobbered, and
pain will ensue.

PR:		230864
Submitted by:	mat
exp-runs by:	antoine

Run "/usr/sbin/service ldconfig restart" for USE_LDCONFIG instead of
"ldconfig -m <path>" so the order of ldconfig search directories after
package installation is the same as after a reboot.  The original command
simply appends the path to the list of directories while the ldconfig rc.d
script uses "sort -u".

Bump lang/gcc* which are known to install libraries with exactly the same
name so the library loaded at runtime depends on the order of the search
directories.

PR:		228046
Approved by:	portmgr (antoine)

Simplify MASTER_SITES in all GCC-ports related to end-of-lifed releases
(gcc47, gcc48, gcc49, and gcc5), taking advantage of the default for GCC
releases provided by MASTER_SITES_ABBREVS.

Reported by:	portlint

Update mpfr to 4.0.1 release and bump PORTVERSION of dependent ports.

Allow gcc compilers to be built on new arch 'powerpcspe'.

Tested for no-harm on amd64.

Submitted by:		jhibbits
Approved by:		portmgr (tier-2 blanket)
Differential Revision:	D13126

Bump PORTREVISIONs of all users of math/mpc that we just updated to
version 1.1.0 (via revision 464079).

GCC 5 reached end-of-life last year, so add a deprecation note like
lang/gcc49 and earlier already carry and recommend GCC 6 or later
instead of GCC 5.

PR:		225629

Fix builds of lang/gcc{48,49,5} with clang 6.0.0

Since clang++ 6.0.0 now defaults to -std=gnu++14 (similar to g++ 6 and
higher), building gcc48, gcc49 or gcc5 produce quite a number of
"invalid suffix on literal; C++11 requires a space between literal and
identifier" errors.  This is because in many places, double quotes are
directly followed by printf helper macros like HOST_WIDE_INT_PRINT.

In gcc 6 and later, spaces were added between the double quotes and the
macros, to fix this issue, but for earlier versions, use a combination
of find, grep and sed to add them mechanically throughout the respective
source trees.

In addition, gcc5 needs a regular patch to fix an incorrect call to
error(), which should have been error_at().  (This was a mismerge by
upstream.)

Approved by:	gerald (maintainer)
PR:		224927
MFH:		2018Q1

Add support for armv7. [1]

Rework the architecture-specific special settings, sorting snippets and
using .elif instead of distinct .if sequences.

Tested by:	andreast [1]

Remove second instance of install-strip patch that was applied twice.
No functional change.

Pointy hat to:	self

Backport two fixes from lang/gcc6:

[1] Make sure what we install is stripped (i.e., debug info is removed).

    (For more background see revisions 454177 and 454422.)

[2] Add a patch that we pulled into gcc6-devel via upstream a week ago
    that addresses a real-world issue around threading and unwinding as
    files/patch-freebsd-unwind.h .

Bump PORTREVISION since [2] is a functional change and [1] changes the
package.

Reported by:	Ports QA Framework, miwi, sobomax [1]
Discussed with:	tijl, miwi [1]
Tested by:	sobomax [1]
Differential Revision:	https://reviews.freebsd.org/D10357 [1]

Remove conflict with lang/gcc5-devel which does not exist any more
since revision 452563.

Backport revision 454177 from lang/gcc7:

Make sure what we install is stripped (i.e., debug info is removed).

The straightforward way is setting INSTALL_TARGET to install-strip,
which is supported by the upstream GCC build machinery.

Unfortunately this fails when running as regular user (non-root)
since strip requires write permission to the files in question,
and we install binaries as r-xr-xr-x by default.  To work around
that we need to set BINMODE to allow for write access by the user,
something that's common on GNU/Linux (which is why this probably
has not been noticed there).  This is not necessary when running
as root.

Merge revisions 437281

  This adds a man page for gcov-dump5 (introduced recently) and also
  one for gcov-tool5 (which we have had for a while).

and 436904

  This brings a new little utility program gcov-dump6 to dump code
  coverage data (unfortunately without a man page or documentation).

from lang/gcc5-devel into pkg-plist which I had missed in the update
to GCC 5.5 a few minutes ago.

Update to the GCC 5.5 release.  This brings more than 250 final bug
fixes compared to GCC 5.4 and is the last release planned on the GCC 5
branch which is now closed.

files/patch-aarch64-support, files/patch-disable-armvhf-config.gcc,
files/patch-libgcc-config-arm-unwind-arm, and files/patch-x86-64-fix-m16
all have been merged upstream in between GCC 5.4 and 5.5 and can thus
be removed; the same is the case for most of files/patch-libc++.

Finally, the tarball is now compressed using xz instead of bzip2.

PR:		216266

Graphite loop optimizations are an experimental feature of GCC, disabled
by default (and guarded by the GRAPHITE option).

Now that we have both GCC 6 and GCC 7 in the tree and GCC 5 is going end
of live upstream soon, remove Graphite support from the GCC 5-related
ports.  Anyone using Graphite is better served by the newer versions of
GCC.

Now that the default version of GCC in the tree is GCC 6 (the lang/gcc6
port), remove Java support (incl. the JAVA option) from lang/gcc5.  Only
one other port actually relies on this and this change speeds up the build
and reduces the size of this port/package quite a bit.

Do not quote constant strings compared with ${ARCH}. [1]  Remove
some commented (and thus disabled) logic around this on the way.

This brings the active lang/gcc* release-based ports in sync with
their respective lang/gcc*-devel twins.

PR:		221905 [1]
Submitted by:	linimon [1]

Move ONLY_FOR_ARCHS and BROKEN_sparc64 together and up into a block
of their own (per guidance from portlint and the manual).

Remove workaround for PR 219274 that bapt@ had put in place in 2014
from pkg-plist for lang/gcc5 and lang/gcc6 (which won't be updated
too soon otherwise).

PR:		219274

Remove headers being created by GCC's fixincludes machinery from the
installation / packaging to avoid breakage when FreeBSD's headers are
changing afterwards.  Several fellow committers have strongly indicated
that our headers do not need the kind of adjustments that GCC performs.

Alphabetize ARCH definitions.

No need to bump PORTREVISION because package does not change.

Reviewed by:	gerald
Approved by:	portmgr (tier-2 blanket)

Adjust WWW URL to refer to https and avoid a trailing slash.

Remove CONFLICT with lang/gcc (which now pulls in lang/gcc5 by default).

Instead add mutual CONFLICTS between lang/gcc5 and lang/gcc5-devel which
we actually missed.

Make _Unwind_GetIP, _Unwind_GetIPInfo and _Unwind_SetIP available as functions
for arm*-*-freebsd*.
This patch is already pushed upstream to all active gcc branches.
GCC-5, GCC-6, GCC-7 and trunk. The gcc?-devel ports will catch up these bits
with the next update.
Once a new release for gcc6 or gcc5 is done, this patch will be obsolete.

Approved by: gerald@ (maintainer)

- Fix shebang

Approved by:	gerald (maintainer via mail)

Define WCHAR_T for aarch64 on all active gcc (gcc/gcc5 and gcc6) releases.
This define is already in upstream.
The gcc*-devel ports will pickup the commit from upstream.

Submitted by:	kan@
Approved by:	gerald@ (maintainer)

Fix armv6 bootstrap. This piece of config is already in the upstream tree and
will go away once gcc-5.5 is released.

Approved by:	gerald (maintainer)

Sync files/java-patch-hier with the lang/gcc port (the only difference
being a marker to avoid a portlint warning re patch format).

Pet portlint wrt. patch file format.

Add files/patch-x86-64-fix-m16 which already has been accepted in
the upstream GCC 5 branch (and hence is in lang/gcc5-devel) and
which we carry in lang/gcc.

It should become obsolete when/if GCC 5.5 releases.

Provide the JAVA option also for powerpc64, in alignment with the
lang/gcc5-devel and lang/gcc ports.

Update lang/gcc and hence the default version of GCC in the Ports
Collection (requested by USE_GCC=yes and various USES=compiler
invocations) from GCC 4.9.4 to GCC 5.4.

files/patch-arm-support and files/patch-gcc_system.h have become
obsolete.  New patches files/patch-arm-unwind-cxx-support and
files/patch-libc++ help support arm targets and new libc++ in base.

ONLY_FOR_ARCHS now also includes arm.

A new option GRAPHITE_DESC, off by default for now, adds support for
Graphite loop optimizations.

Finally, conflicts with other lang/gcc* ports are adjusted suitably.

Backport the bits needed for aarch64-*-freebsd* support from the gcc5-devel
port.
These bits will go away once GCC releases the next gcc5.

Approved by:	gerald (maintainer)

Remove traces of armv6hf which no longer exists as an arch. [1]

Remove files/patch-unwind-ia64.h: we have not been supporting ia64 with
this release series, i.e., ONLY_FOR_ARCHS does not include ia64.

No PORTREVISION bump since nothing should actually change for
existing/supported platforms.

Reported by:	andreast [1]

No longer define CPE_VERSION for ports tracking releases of GCC since
there the default of PORTREVISION is just fine.

Sort ONLY_FOR_ARCHS.

Similar to bug 212465, lang/gcc5 doesn't compile with recent versions of
libc++, because it attempts to redefine abort():

In file included from
/wrkdirs/usr/ports/lang/gcc5/work/gcc-5.4.0/gcc/auto-profile.c:25:
In file included from /usr/include/c++/v1/map:446:
/usr/include/c++/v1/functional:1398:2: error: no member named 'fancy_abort' in
namespace 'std::__1'; did you mean simply 'fancy_abort'?
        _VSTD::abort();
        ^~~~~~~
/usr/include/c++/v1/__config:383:15: note: expanded from macro '_VSTD'
#define _VSTD std::_LIBCPP_NAMESPACE
              ^

Patch this in the same way as the other gcc ports, by including <new> in
gcc/system.h, and moving a few includes to before "system.h".

Approved by:	gerald (maintainer)
PR:		216266
MFH:		2017Q1

The output of tools like awk, date, sort, tr,... depends on the current
locale set by the user.  Add LANG=C and LC_ALL=C at the beginning of
bsd.port.mk and export them so all commands are executed with the C locale.
LC_ALL=C overrides all other LC_* variables.  LANG is used by setlocale(3)
as default value for LC_* variables, so normally it isn't used when LC_ALL
is set, but there's code out there that looks at LANG directly so it's safer
to set it as well.  The only commands not captured by this are !=
assignments before any inclusion of bsd.port.*mk.

Introduce USE_LOCALE=<locale> that adds LANG=<locale> and LC_ALL=<locale> to
CONFIGURE_ENV and MAKE_ENV so upstream build systems can be executed with a
different locale (e.g. USE_LOCALE=en_US.UTF-8).

PR:		215882
Exp-run by:	antoine
Approved by:	portmgr (antoine)

Omit the definition of DISTVERSION which is only necessary for ports
tracking snapshots of GCC and simplify the definition of GCC_VERSION
(but keep the latter as a variable to align the gcc5-devel and gcc5 ports).

${RM} already has -f.

PR:		213570
Submitted by:	mat
Exp-run by:	antoine
Sponsored by:	Absolight

Mark as broken on sparc64.

Approved by:	portmgr (tier-2 blanket)

Default powerpc64 to bootstrapping (option BOOTSTRAP) since otherwise
GCC can be mis-built, leading to an internal compiler error building
libgcc/libgcov.c, at least on FreeBSD 11.

Adjust OPTIONS_DEFINE_powerpc64 and OPTIONS_DEFAULT_powerpc64
incrementally (with +=) to avoid overwriting settings defined
at the top of the Makefile (or child ports). [1]

Submitted by:	swills [1]
Reported by:	swills

Backport the following from lang/gcc5-devel:

Only override CONFIGURE_TARGET for amd64 which is x86-64/x86_64 for the
rest of the world including GNU and GCC.  For all other architectures
it already defaults to the value we were setting.

Backport the following from lang/gcc5-devel:

GCC uses an AWK script to generate source code that helps process
command-line options.  According to POSIX, string comparisons (and
hence sorting) are to be performed based on the locale's collating
order.  Alas GNU AWK only does so in POSIX mode, whereas starting
with FreeBSD 11 we do so by default, running into a bug (or false
assumption) with that script used by GCC.

Setting MAKE_ARGS such that AWK is always invoked in the C locale
works around this bug.

PR:		210122, 211742
Submitted by:	jkim

Update to the GCC 5.4 release.

files/patch-build-without-bootstrap and files/patch-gcc-freebsd-powerpc64
(ELFv2 support for FreeBSD PowerPC64) are now upstream, so drop them.

Due to timeing of the release freeze files/patch-armv6-hf-support has not
been integrated in this upstream release yet.

Skip armv6hf support and move it into armv6.

Discussed with: gerald@

lang/gcc*: convert to CONFIGURE_OUTSOURCE

PR:		208294, 208309
Exp-run by:	antoine
Approved by:	gerald (maintainer)
Differential Revision:	https://reviews.freebsd.org/D4157

Remove ${PORTSDIR}/ from dependencies, categories h, i, j, k, and l.

With hat:	portmgr
Sponsored by:	Absolight

Bring the fixes mentioned below from upstream to our gcc-5.3 release port.

[1]: https://gcc.gnu.org/ml/gcc-patches/2015-12/msg02057.html
[2]: https://gcc.gnu.org/ml/gcc-patches/2015-12/msg00685.html

PR:		205394, 205440
Approved by:	gerald (maintainer)

Add two fixes which have missed the gcc-5.3 release date:
- Fix --disable-bootstrap build.
- Add ELFv2 support for FreeBSD PowerPC64.

Approved by:	gerald (maintainer)

Update to the GCC 5.3 release.

Reestablish the GRAPHITE option.  Still off by default, as originally.

Pet overeager portlint (when it comes to patch format).

Disable GRAPHITE option for the time being, to be restored later this
month with the update to GCC 5.3.

PR:		204925
Reported by:	amdmi3

Add a MULTILIB option to gcc{,48,49,5} for powerpc64

This change is the same as r400632, which updated gcc[56]-devel, but now
for gcc{,48,49,5}.  This change is the second attempt at doing this: the
first attempt went in r401072 and was reverted in r401074 because the diff
was bogus and enabled the new MULTILIB option under all platforms instead
of just powerpc64.

This fixes the build of gcc{,48,49,5} under powerpc64 when the system
is built without the lib32 libraries.

More in detail:

If the system is built with lib32 support (WITH_LIB32, which is the default),
building gcc from ports results in a compiler that can target both 64-bit and

Backport support for Graphite loop optimizations via a new option
GRAPHITE from gcc5-devel.

This is off by default for the time being and pulls in devel/isl
as an additional dependency.

Revert r401072.

I'm not sure what happened exactly but I think I committed the change from
the wrong client.  The applied change enabled the MULTILIB option for all
architectures and not only powerpc64.  Let's just revert the commit and do
it properly from scratch; other things might be wrong so I wanna take a
closer look, and it's best to just revert quickly.

Add a MULTILIB option to gcc{,48,49,5} for powerpc64

This change is the same as r400632, which updated gcc[56]-devel, but now
for gcc{,48,49,5}.  Waited a week to ensure the change caused nothing to go
horribly wrong but this change is very low risk because it only affects
powerpc64.

This fixes the build of gcc{,48,49,5} under powerpc64 when the system
is built without the lib32 libraries.

More in detail:

If the system is built with lib32 support (WITH_LIB32, which is the default),
building gcc from ports results in a compiler that can target both 64-bit and
32-bit binaries on powerpc64.  However, when lib32 support is disabled
(WITHOUT_LIB32), gcc should only be built with 64-bit support or otherwise
the build fails.

To fix this, explicitly disable 32-bit support when /usr/lib32 is not present
and add a MULTILIB option (which is only defined for powerpc64 when 32-bit
support is possible and defaults to yes to preserve the current behavior) to
allow the user to explicitly control this feature.

Approved by:    gerald (maintainer), bdrewery (mentor), andreast
Differential Revision:  https://reviews.freebsd.org/D3952

Add missing USES=compiler, needed for ${COMPILER_TYPE} checks

PR:		203540

Remove deprecated @exec/@unexec from ports using ccache-update-links

Remove UNIQUENAME and LATEST_LINK.

UNIQUENAME was never unique, it was only used by USE_LDCONFIG and now,
we won't have conflicts there.

Use PKGBASE instead of LATEST_LINK in PKGLATESTFILE, the *only* consumer
is pkg-devel, and it works just fine without LATEST_LINK as pkg-devel
has the correct PKGNAME anyway.

Now that UNIQUENAME is gone, OPTIONSFILE is too. (it's been called
OPTIONS_FILE now.)

Reviewed by:	antoine, bapt
Exp-run by:	antoine
Sponsored by:	Absolight
Differential Revision:	https://reviews.freebsd.org/D3336

By default libtool replaces -export-symbols <file> with -retain-symbols-file
<file> on ELF systems, but this doesn't really do what -export-symbols is
meant to do.  On GNU ELF systems it converts <file> to a simple version
script first and then uses -version-script instead of -retain-symbols-file.
Let USES=libtool patch libtool scripts to do this on all systems with GNU
ld(1).

Bump PORTREVISION on all ports where the build log contains -export-symbols.

audio/calf: This port builds a module that now exports only one function,
but it also builds a number of executables that link to this module and
expect to see other functions.  Because it's already a bit dodgy to link to
a module (libtool warns about this) let the module continue to export only
one function and instead build an ordinary library from the same source that
the executables can link to.  Fix a number of other issues in the same

Update from GCC 5.1 (the first release of GCC 5) to GCC 5.2.

This mostly brings bug fixes across teh board plus, thanks to
andreast@, support for position independent code directly upstream,
so we can remove files/patch-pie-support.

Backport PIE support for FreeBSD from GCC trunk.

Approved by:	gerald

Merge MASTER_SITES and MASTER_SITE_SUBDIR into just the former.

Suggested by:	mat

Fix MASTER_SITE_SUBDIR.

Reported by:	jbeich
Pointy Hat to:	whatever injects distcache.FreeBSD.org into my testing

lang/gcc5: Remove Dragonfly's Makefile; it's no longer needed

Now that the JAVA exclusion is part of the main Makefile, the only thing
Makefile.DragonFly was doing was supporting pre-gcc5 dev branch.  Those
users have a static package available to them, and in general they are
expected to upgrade to post-gcc5 in base anyway.  Thus the DF makefile
is no longer necessary.

Update to the 5.1 release, the first release of the GCC 5 series.

The default mode for C is now -std=gnu11 instead of -std=gnu89.
New warning options -Wc90-c99-compat and -Wc99-c11-compat may
prove useful on that front.

The C++ front end now has full C++14 language support including
C++14 variable templates, C++14 aggregates with non-static data
member initializers, C++14 extended constexpr, and more.
The Standard C++ Library (libstdc++) has full C++11 support and
experimental full C++14 support.  It uses a new ABI by default.

There have been significant improvements to inter-procedural optimizations
and link-time optimization such as One Definition Rule based merging of C++
types as well as register allocation.

Update to the 20150421 snapshot of GCC 5.0.1 -- nearly the 5.1 release.

Update the comment that describes the relationship of PORTVERSION and
GCC_VERSION.

Update to the 20150414 snapshot of GCC 5.0.1 and adjust to this
particularity of the new GCC versioning scheme where the third
component of the version number is bumped when the release branches.
Use three component version numbers again.

Update to the 20150405 snapshot of GCC 5.0.

Update to the 20150329 snapshot of GCC 5.0.

lang/gcc(46,47,48,49,5): Use OPTIONS_EXCLUDE_DragonFly to block JAVA

The JAVA frontend doesn't build on DragonFly on any release.  The new
OPTIONS_EXCLUDE_${OPSYS} feature is a nice way to avoid the use of
Makefile.DragonFly (most are in dports, but one is in lang/gcc5).

The recent addition of CXXFLAGS to lang/gcc5 prevents Makefile.DragonFly
on lang/gcc5 from being removed outright.  There are a couple of options
available to allow its removal, but I'll need to discuss with Gerald.

Approved by:	DragonFly blanket

Fix UNIQUENAME not being unique after recent PORTNAME shuffle.

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

Update to the 20150322 snapshot of GCC 5.0.

Add CPE information. [1]

Use PKGNAMESUFFIX so that PORTNAME falls back to plain gcc and we
can avoid setting DISTNAME and CPE_PRODUCT. [2]

PR:		198260 [1]
Submitted by:	shun.fbsd.pr@dropcut.net [1]
Suggested by:	mat [2]

lang/gcc5: Support DragonFly's gcc50 base compiler in specific scenario

Update to the 20150315 snapshot of GCC 5.0.

Update to the 20150308 snapshot of GCC 5.0.

Update to the 20150301 snapshot of GCC 5.0.

Update to the 20150222 snapshot of GCC 5.0.

Update to the 20150215 snapshot of GCC 5.0.

Update to the 20150208 snapshot of GCC 5.0.

Update to the 20150201 snapshot of GCC 5.0.

This now successfully builds libffi again, so we can enable JAVA by
default on i386 and amd64.

PR:		197171

Also disable Java and libffi on amd64 until the next snapshot of
GCC 5.0.

PR:		197171

Allow to build gcc for armv6(hf).

Approved by:	gerald	(maintainer)

Update to the 20150125 snapshot of GCC 5.0.  Sadly libffi is still
broken thus Java remains disabled on i386.

Update to the 20150118 snapshot of GCC 5.0.

libgcj-5.0.pc is now properly called libgcj-5.pc; adjust pkg-plist.

libffi is broken on FreeBSD i386 since 2015-01-11, cf.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64779, so disable it
for the time being.

Update to the 20150111 snapshot of GCC 5.0.

Welcome FreeBSD/ARM support to upstream GCC! [1]

Thanks to:	andreast [1]

Update to the 20150104 snapshot of GCC 5.0.

Update to the 20141228 snapshot of GCC 5.0.

When building the gcc ports using a full bootstrap, tell the configure
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)