intel-gmmlib package requires more CPU features than Void baselineand is compiled with -march=corei7

[ericonr]

Reported by @sgn on IRC.

Per https://github.com/intel/gmmlib/blob/2072b0d1e8ba2cba2f94bc2c1fda89d6e457a50b/Source/GmmLib/Linux.cmake :

    # General optimization options
    -march=${GMMLIB_MARCH}
    -mpopcnt
    -msse
    -msse2
    -msse3
    -mssse3
    -msse4
    -msse4.1
    -msse4.2
    -mfpmath=sse
    -finline-functions
    -fno-short-enums
    -Wa,--noexecstack
    -fno-strict-aliasing
    # Common defines
    -DUSE_MMX
    -DUSE_SSE
    -DUSE_SSE2
    -DUSE_SSE3
    -DUSE_SSSE3

GMMLIB_MARCH is set to corei7 by default, but we can fix that with configure_args. What I also worry about are all the -msse* switches, since afaik they can affect generated code that isn't directly using compiler intrinsics for them.

[tornaria]

It seems, according to https://github.com/intel/gmmlib#supported-platforms, that the supported graphic devices already imply the cpu supports all that.

[ericonr]

It seems corei7 is mostly equivalent to nehalem:

  {"corei7", PROCESSOR_NEHALEM, CPU_NEHALEM, PTA_NEHALEM},

I don't know if it makes sense to force a different target, then - corei7 probably generates slightly better code here than x86_64 would. Maybe we it might just be enough to document these constraints in the template itself. Thanks for the hint!