Tony Reix

Hi, Thanks for the fix! I've tested it in my environment, and that works fine! I have seen another place where I'm sure the -O2 fmadd instructions generates an issue....

Yes. Sure. I'm now looking at issues without -O2 still with the "geometry" module . I'll warn you if I encounter other issues like this one, or others. I already...

More info. On AIX, we have a GCC option -ffp-contract=off which enables to NOT use fmadd . When using it while building test "legendre_stieltjes_test" with -O2, I can see that...

We have a recent Intel machine with AVX. And we can see that the issues are there too. However, please note that the Intel code which is generated is generic...

I have checked on a Intel hardware machine with AVX & AVX2 : the above test does fail too on such Intel HW when compiled with -O2 and -march=native ....

Since we worked to work-around this for being able to generate RPM files, I had to rebuild it all without the work-around, and then original above issue has disappeared and...

> ... Actually endianness could be the problem here - Hummm. Since that works fine for Power8/64bit, I'm not sure it is Endianess. > but perhaps gdb can be coaxed...

> also worth using assembly-free GENERIC target, because POWER8 assembly was unlikely on PPC604e or whatever was the latest 32bit PPC running AIX Hummm So, maybe this 32bit assembler code...

> I see nothing wrong with the format line labeled 9994, probably something is scribbling over parts of the stack Thanks for providing your expertise. Hummmm I gonna relaunch the...

The problem may be not stable, since it now shows the first issue, but with no core: ``` cd OpenBLAS-0.3.10/32bit/power8/test OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 ./sblat3 < ./sblat3.dat ******* FATAL ERROR - PARAMETER...