Compile errors on mac OSX: "No such instruction"

[Taco Cohen]

Hi there,

Thanks for writing what looks like some really awesome spherical harmonics transform code.

I'm trying to compile the library on mac osx 10.9.5. I run ./configure, and get this: https://gist.github.com/anonymous/b67445218ca911cf05ba

Looks fine so far. Next I run 'make', but I get a long list of errors of the following kind:

gcc -march=native -O2 -I/usr/local/include -L/usr/local/lib -ffast-math -fomit-frame-pointer -std=gnu99 -D_GNU_SOURCE -O2 -c sht_init.c -o sht_init.o

/var/folders/s8/c5hbsqw97yn88qhh5khyjp8h0000gn/T//ccGy3z71.s:318:no such instruction: `vcomisd LC5(%rip), %xmm0'

/var/folders/s8/c5hbsqw97yn88qhh5khyjp8h0000gn/T//ccGy3z71.s:326:no such instruction: `vmovsd LC6(%rip), %xmm2'

/var/folders/s8/c5hbsqw97yn88qhh5khyjp8h0000gn/T//ccGy3z71.s:352:no such instruction: `vmovsd (%r11,%rdx,8), %xmm1'

/var/folders/s8/c5hbsqw97yn88qhh5khyjp8h0000gn/T//ccGy3z71.s:353:no such instruction: `vandpd %xmm2, %xmm1,%xmm1'

Any idea what is wrong?

[Nathanaël Schaeffer]

Hello,

I'm not an OSX user, but this looks like a problem with your compiler not really supporting avx. Could you tell me the result of these commands: "gcc --version" and "sysctl -n machdep.cpu.brand_string"

Anyway, as a quick workaround at the expense of a significant performance drop, you can try

./configure --disable-simd

Another solution with probably less performance drop is to try to use clang instead of gcc, which should also support avx:

./configure CC=clang

[Taco Cohen]

Hi Nathanael,

Thanks for the swift response. Here is the output for gcc --version: gcc --version gcc (GCC) 5.0.0 20140824 (experimental) Copyright (C) 2014 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

sysctl -n machdep.cpu.brand_string Intel(R) Core(TM) i7-4850HQ CPU @ 2.30GHz

---

When I use --disable-simd, I still get the "no such instruction" errors. When I use CC=clang, compilation seems to work, but when I try to run the SHT_example.py I get:

ImportError: dlopen(/Users/user/anaconda/lib/python2.7/site-packages/_shtns.so, 2): Symbol not found: __an12_l Referenced from: /Users/user/anaconda/lib/python2.7/site-packages/_shtns.so Expected in: dynamic lookup

(as you can see, I use the Anaconda Python distribution; I set the PYTHON variable to /Users/user/anaconda/bin/python before compiling)

[Nathanaël Schaeffer]

Hello,

So there are two different problems now !

1) No such instruction

So your cpu does indeed support these instructions. It seems to be a problem in your compiling tool chain. I've found a similar problem reported on stackoverflow, with a possible workaround : "Adding -Wa,-q to the compiler flags" (second answer). Could you try it by running :

./configure --enable-python CFLAGS="-Wa,-q"

2) non-working python module when compiled with clang

I can reproduce this on my system too, I have no idea why clang fails at building a proper library here. I will look into it...

[Nathanaël Schaeffer]

• problem 1 (see above comment) is a problem with the compile chain on OSX.
• problem 2 is now fixed in v2.6.1