[PATCH] D69212: [RISCV] Sign-extend 32-bit integer inline assembly operands on RV64I

[James Clarke via Phabricator via llvm-commits]

jrtc27 added a comment.

So, I'm not sure whether this is actually something we want to make work or not. You can already end up with both inconsistencies between compilers and inconsistencies within a compiler for how 16-bit values get represented when passed as register operands, and the fact that we're going against the grain and making i32 an illegal type just makes that also apply to i32; it's nothing new, just more widespread. I wrote a bunch of stuff exploring how this is all already broken for i16 over on the corresponding FreeBSD revision that found this issue (https://reviews.freebsd.org/D22084#482766), but trying to support this sounds like opening a can of worms, and it seems likely that we will still end up with subtle differences, ones which might rear their heads only once there is sufficient inlining (and perhaps due to LTO), removing the sanitisation applied by the calling convention. Moreover, only guaranteeing GCC compatibility (and self-consistency) for i32 and iPTR (whatever that may be) across architectures seems a bit arbitrary. Yes, i32 is a bit special due to C's integer promotion rules, but it feels like a slippery slope towards standardising on i16 and i8 too. Yes, I went and wrote this patch, but having further explored the issue and thought about it, my inclination is that we should just tell people to not do this, and put in the explicit sign-extending or zero-extending casts they want in order to get an XLEN-sized value. Especially since there is not likely to be much RISC-V inline assembly out there at the moment that relies on this. What do others think? Perhaps Clang could gain more than the current `warn_asm_mismatched_size_modifier` (which is currently about whether the register you're getting due to your modifiers matches what your input is, e.g. if the input gets promoted to an i32 on AArch64 you should be using a W register with `%w0` rather than `%0`; it's not warning about the promotion itself).


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69212/new/

https://reviews.llvm.org/D69212