Roll on 2011, and my lovely Matrix44.Multiply(Vector4[] foo) neither inlines the internal Matrix44.Multiply(Vector4 bar) nor does it SIMD any of it even if I inline it myself. Ditto for Matrix34.Multiply(ArrayOfStructs baz).
The CLR does inline, but it has restrictions that you're probably hitting. But it doesn't do processor specific opts. That does suck.
Then there are the kind of performance guided global optimizations that you could do with a language like C#, using data-driven compilers, or manual-driven attributes.
The CLR does have a profile based optimization system, but according to this article, it is internal-only. http://msdn.microsoft.com/en-us/magazine/dd569747.aspx Oh well.