GCC 4.8/OS X: need -flax-vector-conversions

[Tim Smith]

Compiling plink2 against Accelerate on OS X with gcc 4.8 with the provided Makefile fails, generating errors like:

/System/Library/Frameworks/vecLib.framework/Headers/vBasicOps.h: In function 'vUInt16 vU16HalfMultiply(vUInt16, vUInt16)':

/System/Library/Frameworks/vecLib.framework/Headers/vBasicOps.h:352:84: note: use -flax-vector-conversions to permit conversions between vectors with differing element types or numbers of subparts

   vUInt16   __vbasicops_vB) { return _mm_mullo_epi16(__vbasicops_vA, __vbasicops_vB); }

                                                                                    ^

/System/Library/Frameworks/vecLib.framework/Headers/vBasicOps.h:352:84: error: cannot convert 'vUInt16 {aka __vector(8) short unsigned int}' to '__m128i {aka __vector(2) long long int}' for argument '1' to '__m128i _mm_mullo_epi16(__m128i, __m128i)'

/System/Library/Frameworks/vecLib.framework/Headers/vBasicOps.h: In function 'vSInt16 vS16HalfMultiply(vSInt16, vSInt16)':

/System/Library/Frameworks/vecLib.framework/Headers/vBasicOps.h:377:84: error: cannot convert 'vSInt16 {aka __vector(8) short int}' to '__m128i {aka __vector(2) long long int}' for argument '1' to '__m128i _mm_mullo_epi16(__m128i, __m128i)'

   vSInt16   __vbasicops_vB) { return _mm_mullo_epi16(__vbasicops_vA, __vbasicops_vB); }

There's full output at https://gist.github.com/tdsmith/142117fcab430f022962.

Taking gcc's advice and adding -flax-vector-conversions to CFLAGS in the Makefile allows the build to finish.

Thanks,

Tim

[Christopher Chang]

Try adding "#define __ACCELERATE__" immediately under "#include <Accelerate/Accelerate.h>" on line 10 of plink_matrix.h .  (http://permalink.gmane.org/gmane.comp.lang.r.mac/8269 implies that something like this might work.  If it does fix the problem, I'll add it to the main codebase.)

[Christopher Chang]

Hmm.  Is this a 64-bit Intel Mac?

On Monday, June 16, 2014 2:19:42 AM UTC+8, Tim Smith wrote:

I saw the same thread and tried that both by adding -D__ACCELERATE__ to CFLAGS and just now by adding the #define in plink_matrix.h as you suggested; neither had any effect for me.

[Tim Smith]

Yes, it is. I'm using Homebrew's gcc package.

[Christopher Chang]

If there's some Homebrew + OS X 10.8+ detection logic you can suggest for the Makefile, and your testing suggests the -flax-vector-conversions plink2 build doesn't have any problems, I can update the Makefile accordingly.  I can't justify adding that flag for all Mac builds, though.

[Tim Smith]

On Sunday, June 15, 2014 7:11:20 PM UTC-7, Christopher Chang wrote:

If there's some Homebrew + OS X 10.8+ detection logic you can suggest for the Makefile, and your testing suggests the -flax-vector-conversions plink2 build doesn't have any problems, I can update the Makefile accordingly.  I can't justify adding that flag for all Mac builds, though.

Roger that. What compiler are you using to generate the Mac builds now? The flag appeared in gcc 4.3 [1] and enables the default behavior from 4.2. I was kind of hoping this hadn't been seen yet because you were building with 4.2. If that's not the case, there's probably deeper magic involved and I'm much less sanguine about this solution.

Thanks,

Tim

[1] https://gcc.gnu.org/gcc-4.3/changes.html 

[Christopher Chang]

I have a fairly standard Xcode setup, thus I'm building with 4.2 due to well-known Apple licensing issues.  So, okay, I'll go ahead and add Mac gcc version detection logic along the lines of http://stackoverflow.com/questions/5188267/checking-the-gcc-version-in-a-makefile .