It's too easy to get multi-target target lists in the wrong order

[abadams]

E.g. if you say x86-64,x86-64-sse41,x86-64-avx2, then it uses the first thing in the list that matches, which is not what you want, and it uses the last thing in the list for the runtime, which might not run.

[alexreinking]

Why does it use the last thing in the list for the runtime? If you generate a standalone runtime with -r, it should take gcd(x86-64,x86-64-sse41,x86-64-avx2) = x86-64. That behavior should be mirrored for when you aren't doing that.

[steven-johnson]

Why does it use the last thing in the list for the runtime? If you generate a standalone runtime with -r, it should take gcd(x86-64,x86-64-sse41,x86-64-avx2) = x86-64.

Yes, we should do that too, but that wouldn't change the fact that the multitarget was likely specified that way by mistake.

[abadams]

Oh, then I'm not sure why the above multitarget string produces illegal instructions on machines without avx2.

[steven-johnson]

Oh, then I'm not sure why the above multitarget string produces illegal instructions on machines without avx2.

I interpreted his comment to mean "we should be doing that, but aren't currently"

[abadams]

There's code in Generator.cpp that attempts to compute a gcd target. It uses Target::get_runtime_compatible_target

[abadams]

We should probably add an error when a later target is a strict superset of an earlier target, because that's never useful.