Consider using pair mode to return scalar pair bools as i1

[cuviper]

Starting in #51583, we're representing scalar pair bools as i8 in LLVM aggregates to match their memory storage, whereas they are i1 as immediate values.

When a pair is the argument to a function, we use PassMode::Pair and pass each part like independent immediate values. We don't use that mode for return values though, so a paired bool will be extended to i8 for return, then truncated back to i1 when the caller unpacks it.

Quoting @eddyb in #51583 (comment):

I wonder if they should be using the pair "passing mode" that arguments do, and create a LLVM aggregate type on the fly, using the immediate types for the pair components. That way we'd get {i1, i1} for returns, but everything else would see {i8, i8}.

Not sure it's worth the complexity though. When inlining, LLVM should collapse the zext followed by trunc, just like it gets rid of packing into a pair and unpacking it.

[eddyb]

I think that for compatibility with backends that support "true multiple return" (i.e. Cranelift), we should be using PassMode::Pair for returns as well, with more explicit packing. cc @sunfishcode

[sunfishcode]

That sounds right to me. Even in LLVM, an aggregate return type in LLVM is essentially LLVM's convention for multiple return values, so it sounds desirable to use i1 there too.