openblas 0.3.13 fails to compile on macOS

[fxcoudert]

openblas 0.3.13 fails to compile in Homebrew (Homebrew/homebrew-core#66821) with:

gcc-ranlib ../../libopenblasp-r0.3.13.a
 make[2]: gcc-ranlib: No such file or directory
 make[2]: *** [../../libopenblasp-r0.3.13.a] Error 1
 make[1]: *** [lapacklib] Error 2
 make: *** [netlib] Error 2

We run:

system make CC=gcc-10 FC=gfortran libs netlib shared

I have no idea why it's trying to call gcc-ranlib, but we don't have it installed. The system ranlib works perfectly fine, and should be called. We can (and probably will) override this with RANLIB=ranlib but that still seems like a bug that should be fixed here.

[martin-frbg]

I do not think this is due to a change in 0.3.13, possibly the hyphenation in gcc-10 tricks c_check into assuming that you are cross-compiling with a CROSS_SUFFIX (actually prefix) of "gcc-" to add to RANLIB in Makefile.system. Do you see CROSS=1 in your Makefile.conf ?

[fxcoudert]

$ grep CROSS Makefile.conf 
CROSS_SUFFIX=gcc-
CROSS=1

[fxcoudert]

That's because this code is broken since da6d5d6

# Checking cross compile
 $hostos   = `uname -s | sed -e s/\-.*//`;    chop($hostos);
 $hostarch = `uname -m | sed -e s/i.86/x86/`;chop($hostarch);
 $hostarch = `uname -p` if ($hostos eq "AIX");
 $hostarch = `uname -p` if ($hostos eq "AIX" || $hostos eq "SunOS");
 chop($hostarch);

uname -m returns x86_64\n, the first chop (on the same line) removes the newline. But then there is an extra chop on line 10, don't know why, which removes the 4. Later, we end up with architecture= x86_64 and hostarch= x86_6, and therefore identified as cross-compiling.

[fxcoudert]

I'm not sure why chop is being used at all throughout this code. The only thing that should be removed is the trailing newline, so chomp would a lot safer than chop.

[martin-frbg]

Ah, thanks, that is a bit insidious indeed. The Solaris case probably needs its own line anyway IF the AIX case used to work without a chop/chomp. (And the whole thing could go unnoticed on Linux, and apparently all other systems including OSX in our CI as well, as they have gcc-ranlib installed)

[fxcoudert]

I'd still recommend using chomp (instead of chop) for safety. Is there a case in this script where you want to remove something other than a newline? That'd be weird

[martin-frbg]

Probably not - unless some platform returns a trailing blank or has a misbehaving chomp. The "chop" has been inherited from GotoBLAS I believe...