On Tue, Sep 5, 2017 at 10:43 PM, Grant <[email protected]> wrote: >>>>>> Now I'm running into "trap invalid opcode" errors on the older >>>>>> systems. Can I disable some of the newer CPU instruction sets on the >>>>>> master laptop when compiling to hopefully generate binaries that will >>>>>> work on the older systems? >>>>> >>>>> Yes. You need to find out CPU_FLAGS_X86 and "-march=" values the >>>>> machines have, and use that in make.conf. Run the commands... >>>>> >>>>> cpuinfo2cpuflags-x86 >>>>> gcc -c -Q -march=native --help=target | grep march= >>>>> >>>>> ...on the target machines. This will tell you what "native" is and >>>>> what CPU_FLAGS_X86 values to use. >>>>> >>>>> https://gcc.gnu.org/onlinedocs/gcc-6.3.0/gcc/x86-Options.html#x86-Options >>>>> lists available "march=" options, and what instruction sets they support. >>>>> E.g. my old core2 desktop shows... >>>>> >>>>> [d531][waltdnes][~] cpuinfo2cpuflags-x86 >>>>> CPU_FLAGS_X86="mmx mmxext sse sse2 sse3 ssse3" >>>>> >>>>> [d531][waltdnes][~] gcc -c -Q -march=native --help=target | grep march= >>>>> -march= core2 >>>>> >>>>> >>>>> Unless the laptops are really old, you can probably get away with... >>>>> CFLAGS="-O2 march=core2 -mfpmath=sse -fopenmp -fomit-frame-pointer -pipe >>>>> -fno-unwind-tables -fno-asynchronous-unwind-tables" >>>>> >>>>> booby trap 1) Unless all machines are Intel "Atom" family, do *NOT* use >>>>> a "march=" that implements the "movbe" instruction. >>>>> >>>>> booby trap 2) If you throw in any AMD-based machines proceed with care. >>>>> >>>>> Can you post the output of... >>>>> gcc -c -Q -march=native --help=target | grep march= >>>>> ...for all the target machines? >>>> >>>> >>>> Let's see how -mtune=native goes and resort to the above if necessary. >>>> It doesn't look too bad though. >>> >>> >>> emerge -e world has finished and pushed and -mtune=native seems to >>> have solved the issue for now. >>> >> >> You might be interested in "-march=x86-64 -mtune=generic" though this >> will mean you might miss out on some optimizations. > > > If I could miss out on optimizations, what is the advantage compared > to -mtune=native? Better compatibility across CPUs? >
Well, having looked it up (again), -march will break backwards compatibility while -mtune=native will not if you stay within a family of processors. So, yes. One of the finer points of -mtune is that is specifies processor cache sizes. If this is set to an incorrect value you could get very poor performance relative to leaving it unset (with -mtune=generic). For the exact differences, try: gcc -march=native -mtune=native -E -v - </dev/null 2>&1 | grep cc1 echo | gcc -dM -E - -march=native Which will show you the expanded compilation flags and macro definitions, respectively.
