gmp-5.0.1 tests fail on AIX 5.2

[Patrick Begou]

I'm sorry, Torbjörn, my last post is a very bad post! I provide
a list of options and tests whithout any explanation, just
saying it fails. I apologise.

So let me start from the begining. I worked on this install for
several days before posting and I do not detail my build process.


I start all the tests  from scratch with:
gzcat gmp-5.0.1.tar.gz |tar xvf -
cd  gmp-5.0.1

1) using default values,
CC=cc ./configure --prefix=/HA/sources/begou/AIX32

configure is unable to detect the right hardware:
....
checking host system type... powerpc630-ibm-aix5.2.0.0
    no, it is a 43P-260 (I have also a P630 and it is not the same architecture)
checking ABI=aix64
    I understand I'm building a 64bit version of the lib
checking compiler cc -O2 -q64 -qtune=pwr3 -qmaxmem=20000
    ./configure select "-q64" for building 64 bits lib
    but if qtune=pwr3 is OK for a P630 it's wrong for a 43P-260
    which need -qtune=pwr2 for a power2 cpu.
.....

make is OK (I can build power3 binaries on a power2 system)
But make check fails for several tests on the power2 (of course,
the binaries  are for a power3) but also on a power3 system
where I have only the runtime C/C++ libraries installed.
It freeze finaly on the 2 hosts with the message:
FAIL: t-invert


2) So I try to tell configure which is the right architecture.
Based on http://gmplib.org/manual/Build-Options.html#Build-Options
I select the power2 architecture.
CC=cc ./configure --prefix=/HA/sources/begou/AIX32 --build=power2
...
checking build system type... power2-ibm-aix
checking host system type... power2-ibm-aix
checking ABI=standard
checking whether cc is xlc... yes
checking compiler cc -O2 -qarch=pwr -qmaxmem=20000 ... yes
...

All seams OK, I do not know what ABI=standard means (32 or 64 bits ?)
but compiler options do not use the -q64 option (so 32 bits is the
default for xlc)
make is successfull.
but with make check several tests fails
FAIL: t-toom8h
FAIL: t-mullo
FAIL: t-invert
FAIL: t-div
FAIL: t-bdiv
====================================
5 of 30 tests failed
Please report to gmp-bugs at gmplib.org
====================================


3) So in a next step I've tried to replace default compiler options
by more restrictive (in optimisation) and more secure (but generating
slower code) options:
-q32				just to be sure to select 32 bits
-g 				include debug information
-qarch=pwr2  -qtune=pwr2 	more precise architecture than -pwr
-qstrict 			limit optimisation agressivity
-qalign=power			force memory alignment
-qinitauto=00			initialise all automatic variables to 00
-O2 -qmaxmem=20000              and I keep these default options.

and launch a new build (no C++ nor fortran options in this test)
CC=cc CFLAGS="-q32 -g -O2 -qarch=pwr2 -qtune=pwr2 -qstrict -qalign=power
-qinitauto=00 -qmaxmem=20000" \
./configure --prefix=/HA/sources/begou/AIX32 --build=power2
checking build system type... power2-ibm-aix
checking host system type... power2-ibm-aix
checking ABI=standard
checking whether cc is xlc... yes
checking compiler cc -q32 -g -O2 -qarch=pwr2 -qtune=pwr2 -qstrict -qalign=power
-qinitauto=00 -qmaxmem=20000 ... yes

make is successfull.
but with make check the same tests fails. So it may not be a problem related to
a too agressive optimisation.
FAIL: t-toom8h
FAIL: t-mullo
FAIL: t-invert
FAIL: t-div
FAIL: t-bdiv

4) Then I try a generic C build as proposed in
http://gmplib.org/manual/Build-Options.html#Build-Options
(I had missed this posibility when reading the doc as I focus on available
architectures) with default options.
CC=cc ./configure --prefix=/HA/sources/begou/AIX32 --build=none

checking build system type... none-none-none
checking host system type... none-none-none
checking ABI=long
checking whether cc is xlc... yes

make is successfull but more tests fail in make check:
FAIL: t-toom8h
FAIL: t-hgcd
FAIL: t-mullo
FAIL: t-invert
FAIL: t-div
FAIL: t-bdiv
====================================
6 of 30 tests failed
Please report to gmp-bugs at gmplib.org
====================================


5) For the two last builds (3 and 4) I've tried to increase the available
memory with the linker option -bmaxdata:2000000000 (program can use up
to 2Gbytes of memory) to be sure it was not a system error (this IBM
has 3Gbytes RAM).
The same tests fail.

May be these details clarify my attempts to build gmp on AIX 5.2.

Patrick
-- 
===============================================================
|  Equipe M.O.S.T.         | http://most.hmg.inpg.fr          |
|  Patrick BEGOU           |       ------------               |
|  LEGI                    | mailto:Patrick.Begou at hmg.inpg.fr |
|  BP 53 X                 | Tel 04 76 82 51 35               |
|  38041 GRENOBLE CEDEX    | Fax 04 76 82 52 71               |
===============================================================