[llvm] [ARM] Allow FP reg conversion when copying Sx to Dx (PR #147559)

[via llvm-commits]

eleviant wrote:

> Why do you think the compiler should do this? It's not what gcc does. And in general implicitly performing floating-point conversions seems like a bad idea.

Well, the inline assembly in summary actually instructs to put float value into 64-bit FP reg and that's exactly what I did. GCC seems to put value into s1 instead of d1, which seems confusing. Compiling the same with clang results in undefined behavior in release build w/o asserts, (for me it consumes all available memory and crashes). What's the sane thing to do here in your opinion? Emit an error and exit? Doing what GCC does? 

https://github.com/llvm/llvm-project/pull/147559