Currently, calls or other operations (such as MUL_LONG on 32-bit platforms) that return multiple registers must be stored (either to the stack or elsewhere). This means that promoted structs (whose fields have been promoted to allow them to be separately tracked, optimized and enregistered) will be forced to the stack.
This is the underlying issue in #5112 and #8571, but support for this would also enable MUL_LONG to be supported without forcing the destination to be in memory, as well as enabling hardware intrinsics such as Bmi2.MultiplyNoFlags to return a ValueTuple and allow the result to go directly to a pair of registers.

category:cq
theme:structs
skill-level:expert
cost:large