search

llvm / llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.
http://llvm.org
Other
28k stars 11.56k forks source link

SME streaming status overloading #100200

Open jroelofs opened 1 month ago

jroelofs commented 1 month ago

We should have an overloading mechanism that dispatches to normal vs streaming implementations of a function at runtime based on the current streaming mode. As an optimization, we could set up dispatch to always call the matching implementation in a normal/streaming caller. In streaming-compatible callers, we'd have to inspect the streaming mode via the compiler-rt builtin.

Use case for this could be to provide a memcpy implementation that keeps the streaming mode active or inactive, and uses the corresponding instructions without a streaming mode change.

jroelofs commented 1 month ago

cc @sdesmalen-arm, @rsandifo-arm, @aemerson for SME bits cc @labrinea, as this is kind of FMV-adjacent

jroelofs commented 1 month ago

breadcrumb for myself: rdar://116990507

efriedma-quic commented 1 month ago

Currently, you can write something like the following. This dispatches as appropriate. It currently doesn't optimize well, but that's something we could fix. Given that, what benefit does a dedicated "overloading" feature provide?

void f_streaming() __arm_streaming;
void f_nonstreaming();
inline void f() __arm_streaming_compatible {
  if (__arm_in_streaming_mode())
    f_streaming();
  else
    f_nonstreaming();
}

(Also, I'm skeptical you'd want to use this for memcpy(); small memcpys only use streaming-compatible instructions anyway, so it's cheaper to check the size before checking whether streaming mode is enabled.)