A recent modification in binutils (see
https://sourceware.org/ml/binutils/2015-04/msg00332.html) introduced a change
in the way the ppc assembly instructions "dcbt" and "dcbtst" are handled.
When compiling on a generic ppc processor, the 3-arguments form of those
instructions is no longer accepted, and one must instead use a 2-arguments
form (ignoring the TH argument).
Incidentally, the binutils change also fixed a bug in the way those
instructions were handled. On a generic ppc processor, they used to be
interpreted in the so-called "embedded" form ("dcbt TH, RA, RB", only used on
some embedded machines), while they should have been interpreted in the
so-called "server" form ("dcbt RA, RB, TH"). The ATLAS assembly directive were
apparently written in "server" form, and were therefore mis-assemblied.
Ideally, this patch should be improved in order to use the TH argument on
machines that support it (only needed in atlas_prefetch.h; other instances use
the default zero value for TH).
Last-Update: 2015-11-23
Gbp-Pq: Name powerpc-dcbt.patch
Some of the existing optimized files/cases for PowerPC
contain assembly instructions with implicit big-endian
behavior - thus incorrect for the little-endian mode -
incurring tests failures during the build (i.e., FTBFS).
This is being worked on; this is the workaround for now.
Note: this is an alternative for Michael Normand's patch
'atlas.3.10.2-ppc64le_do_not_use_files_with_lvx.patch',
in order not to touch file lists/indexes, but only C
source code, so the C preprocessor can avoid code per-arch.
For more details, see:
https://bugzilla.redhat.com/show_bug.cgi?id=1080073#c40
In practice this type will have the same effect than the UNKNOWN machine type,
except that it does not try to autodetect the machine CPU.
This is useful when trying to build a generic package on an arch for which
ATLAS does not define a generic machine type. For example, on a armel buildd
with ARMv7 CPU, ATLAS would detect ARMv7 if provided the UNKNOWN type.
Last-Update: 2017-08-16
Gbp-Pq: Name generic.patch
The "GNU" test is moved at the end, otherwise kfreebsd is recognized as Linux.
This is necessary in order to have correct linker flags.
Last-Update: 2017-08-28
Gbp-Pq: Name kfreebsd.patch
The mips{,el} ports use the O32 ABI.
See https://www.linux-mips.org/wiki/MIPS_ABI_History for more details on ABIs.
Last-Update: 2017-08-25
Gbp-Pq: Name mips.patch
Rename ATLAS incomplete liblapack.a to liblapack_atlas.a
Forwarded: not-needed
Last-Update: 2017-08-08
The liblapack.a created by ATLAS build system is incomplete: it only contains
a small subset of the full LAPACK API (only those optimized by ATLAS).
Hence rename this library to liblapack_atlas.a to avoid a name conflict with
the full liblapack.a.
In particular, note that the "-Ss flapack" option of the configure script does
not work as expected as of ATLAS 3.10.3.
Last-Update: 2017-08-08
Gbp-Pq: Name rename-lapack-atlas.patch
* Fix missing symbols in liblapack.so.3 when libblas.so.3 is not the one from
ATLAS.
The problem was that, in 3.10.3-9, liblapack.so.3 was no longer linked to
libatlas.so.3, despite an explicit -latlas in the link command.
This was due to the combination of two factors:
— ld now defaulting to --as-needed in bullseye
— libatlas.so.3 symbols being erroneously embedded in libblas.so.3
The fix consists in no longer embedding libatlas.so.3 symbols inside
libblas.so.3 (this has always been the intent, for some reason this had not
been correctly done). (Closes: #947827)
* debian/rules: simplify using debhelper 12.8’s execute_{before,after}_* rules
* Bump to debhelper 13
— this allows dropping the B-D on dh-exec
— mention libsatlas.so and libtatlas.so in debian/not-installed
* Separate the {build,binary}-arch rules from {build,binary}-indep
* Drop debian/NEWS.Debian: it has never been installed, since it had an
incorrect name
* Bump S-V to 4.5.0
* gfortran-10.patch: new patch, fixes FTBFS against gfortran 10
(Closes: #957015)
projects / atlas.git / log