OpenBLAS 0.3.9 build segfault on ARM64 + 64-bit integer support

[tylerjereddy]

We've been observing this segfault for a while now, so it is not specific to 0.3.9.

Invocation line:
make DYNAMIC_ARCH=1 USE_OPENMP=0 NUM_THREADS=64 BINARY=64 INTERFACE64=1 SYMBOLSUFFIX=64_ OBJCONV=/io/objconv/objconv

Link to build/failure log: https://travis-ci.org/MacPython/openblas-libs/jobs/659863204?utm_medium=notification&utm_source=github_status

Summary of failure:

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

Backtrace for this error:
#0  0xffff9b1509ab
#1  0xffff9b14fa1f
#2  0xffff9b2c65b7
#3  0xbe80d4
make[1]: *** [level1] Segmentation fault (core dumped)
make[1]: *** Waiting for unfinished jobs....
make: *** [tests] Error 2

Example PR exhibiting the segfault: MacPython/openblas-libs#26

If you'd like me to try some things let me know. Thanks!

[tylerjereddy]

To be clear, I have no evidence that this build ever worked, since the scientific Python ecosystem is preparing to support a much wider array of platforms & this is one of the new ones. So just looking for suggestions :)

[isuruf]

Can you run a verbose build?

[brada4]

make should emit crashing commands. At the point it crashes it is not any of openblas code, it is at the end of lapack build, it could be ar or gcc

[martin-frbg]

Crash is in the sdot section of the sblat2 test.

[martin-frbg]

Turns out INTERFACE64=1 was never translated into the appropriate fortran compiler option. Should be fixed now for arm64, but the same problem is likely to exist on other non-x86 architectures as well. ISTR this used to be handled in a common Makefile section where it interfered with some compilers so got thrown out - I still need to look into this.

[tylerjereddy]

Cool, thanks for the work here!