GMP 6.0.0a on AMD A8-6600K

[Kevin Buckley]

> I'll try again tonight (errors on a box a at home) and send you the full
> details
>
> Bare with me: I am not normally this poor!

and my English is normally better than the mis-spelling of "Bear" would
suggest too !

On the Ubuntu host

tar xf /srv/domains/linuxfromscratch.org/LFS-7.7/gmp-6.0.0a.tar.xz
cd gmp-6.0.0/
./configure \
  --prefix=/usr/local \
  --enable-cxx \
  --docdir=/usr/local/share/doc/gmp-6.0.0a

Notes:

1) This link gets created.

gmp-mparam.h -> mpn/x86_64/bd2/gmp-mparam.h

Cf the "bd2" with the "bd1" in the compilation's -mtune and
-march flags: two different views of the same processor ?

2) Went with a "make -k check" so you can see all the failures.

3) I noticed a few possibly related messages in the dmesg output
    on the host that also appear inside the VM's dmesg - this might
    be the show-stopper.

Attached:

ub1204-config.status
ub1204-config.log

make          2>&1 | tee ub1204-make.out
make html     2>&1 | tee ub1204-make_html.out
make -k check 2>&1 | tee ub1204-make-k_check.out

Attached

ub1204-make.out
ub1204-make_html.out
ub1024-make-k_check.out

ub1204-dmesg

Repeated same inside the Xen VM (CentOS 6.6)

Notes:

gmp-mparam.h -> mpn/x86_64/bd2/gmp-mparam.h

Attached:

cent66-config.status
cent66-config.log
cent66-make.out
cent66-make_html.out
cent66-make-k_check.out

cent66-dmesg

Repeated same inside the LFS chroot

Notes:

LFS chroot build uses --prefix=/usr not /usr/local

gmp-mparam.h -> mpn/x86_64/bd2/gmp-mparam.h

Attached:

lfs_77-config.status
lfs_77-config.log
lfs_77-make.out
lfs_77-make_html.out
lfs_77-make-k_check.out
lfs_77-dmesg


I believe that the "trap invalid opcode" messages seen in the dmesg
are probably the root cause, or the symptom of the root cause, of the
issue but I attach all the other stuff to avoid having to read any minds !


Also downloaded the latest GMP snapshot, gmp-6.0.0-20150409.tar.lz,
as per Marc's suggestion and built it in the Ubuntu 12.04 box.

Notes:

1) Configure's summary stanza ends with a slightly Host type to 6.0.0a

  Version:           GNU MP 6.0.0
  Host type:         piledrivernoavx-unknown-linux-gnu
  ABI:               64
  Install prefix:    /usr/local
  Compiler:          gcc -std=gnu99
  Static libraries:  yes
  Shared libraries:  yes

2) The gmp-mparam.h link is, as before on the Ubuntu 12.04 box,
linking to a "bd2" file:

gmp-mparam.h -> mpn/x86_64/bd2/gmp-mparam.h

3) The compilations still use the "bd1" annotations

4) I still see "Illegal Instruction" failures in the "make check"

# grep " tests " ub1204-20150409-make-k_check.out
All 8 tests passed
3 of 45 tests failed
16 of 63 tests failed
1 of 14 tests failed
3 of 27 tests failed
All 7 tests passed
2 of 3 tests failed
13 of 20 tests failed

although there seems to be one less test in the snapshot, compared to
the 6.0.0a greps, where all three showed:

All 9 tests passed
3 of 45 tests failed
16 of 63 tests failed
1 of 14 tests failed
3 of 27 tests failed
All 7 tests passed
2 of 3 tests failed
13 of 20 tests failed


Attached:

ub1204-20150409-config.status
ub1204-20150409-config.log
ub1204-20150409-make.out
ub1204-20150409-make_html.out
ub1024-20150409-make-k_check.out

=======

When I saw the same failures on the "bare metal" Ubuntu box, I didn't
bother with building on the VM or the LFS chroot.

What I will also do, once I reboot out of the Xen environment, is to try
the  6.0.0a and snapshot sources against a "pure" kernel: just for
completeness.

Still hoping this is something really obvious that I have missed,
Kevin
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 20150409.tar.xz
Type: application/octet-stream
Size: 128612 bytes
Desc: not available
URL: <https://gmplib.org/list-archives/gmp-bugs/attachments/20150409/ceab4dc6/attachment-0001.obj>