That’s not a guarantee on the Linux world either, but at least you do have the option of recompiling your distro to not use those options.
There’s talks from some distros to start dropping support for such old CPUs because it’s holding back newer CPUs that could run even faster by using those instructions.
Is it really that hard to include a fallback though? Obviously there’s a way to collect the information without that flag. I suppose if you didn’t want to take a performance hitting checking the flag all the time it could become a compile option (I would think anyone running that old of hardware would be willing to learn how to compile the kernel anyway), but there should be options available to keep the support available some how?
That’s pretty much exactly how it works already. You compile with -march=x86-64-v4 and it’ll use SSE and AVX all over the place.
glibc does the runtime thing, but only once on application startup where the dynamic linker will link the version of the function optimized for your CPU. But it’s a manual process on glibc’s part, the variants are written by hand.
Not every project cares enough to do it dynamically like that and it would be a nightmare that way.
The fallback is, recompile with -march=x86-64 which will only use the base set of instructions. Or -march=i486 if you want to run on absolutely ancient hardware.
That’s not a guarantee on the Linux world either, but at least you do have the option of recompiling your distro to not use those options.
There’s talks from some distros to start dropping support for such old CPUs because it’s holding back newer CPUs that could run even faster by using those instructions.
Is it really that hard to include a fallback though? Obviously there’s a way to collect the information without that flag. I suppose if you didn’t want to take a performance hitting checking the flag all the time it could become a compile option (I would think anyone running that old of hardware would be willing to learn how to compile the kernel anyway), but there should be options available to keep the support available some how?
That’s pretty much exactly how it works already. You compile with
-march=x86-64-v4and it’ll use SSE and AVX all over the place.glibc does the runtime thing, but only once on application startup where the dynamic linker will link the version of the function optimized for your CPU. But it’s a manual process on glibc’s part, the variants are written by hand.
Not every project cares enough to do it dynamically like that and it would be a nightmare that way.
The fallback is, recompile with
-march=x86-64which will only use the base set of instructions. Or-march=i486if you want to run on absolutely ancient hardware.