[clang] [RISCV] Disable rounding of aggregate return/arguments to iXLen. (PR #184736)

[Craig Topper via cfe-commits]

topperc wrote:

> Please can you add some testcases with e.g. `struct even_tinier { char a, b; }`. Right now we only have tests for 32-bit structs full of integers (`struct tiny`), or smaller structs with FP values.
> 
> Another question I do have about this is whether we should be using `getExtend` rather than `getDirect`, to make some specific assertions about the top bits of the aggregate. Maybe we're not able to, given the specific behaviour of `RISCVABIInfo::extendType`. I'm not entirely sure.

Wouldn't getExtend put back a zero extend we're trying to get rid of?

https://github.com/llvm/llvm-project/pull/184736