Expose LLVM atomic memcpy in intrinsics?

joshlf · 2019-02-17T22:31:23.653Z

Would it be possible to add LLVM’s atomic memcpy operations to core and std's intrinsics modules? In particular, the llvm.memcpy.element.unordered.atomic, llvm.memmove.element.unordered.atomic, and ‘llvm.memset.element.unordered.atomic intrinsics.

What is the process for a) getting this approved and b) going through the process of adding them?

sanxiyn · 2019-02-18T03:53:31.713Z

I think they can be just added, since they will never be stable?

In the long term, I am worried about effect of LLVM intrinsics usage to alternative non-LLVM backends, although so far it seems manageable.

joshlf · 2019-02-18T03:55:41.113Z

sanxiyn:

I think they can be just added, since they will never be stable?

Good point.

sanxiyn:

In the long term, I am worried about effect of LLVM intrinsics usage to alternative non-LLVM backends, although so far it seems manageable.

I agree in the general case, although in this particular case, even if we expose them through a stable facade, it’d be trivial to implement as a simple loop of atomic operations on for non-LLVM backends.

bascule · 2019-02-18T19:04:54.594Z

sanxiyn:

In the long term, I am worried about effect of LLVM intrinsics usage to alternative non-LLVM backends, although so far it seems manageable.

To avoid that these things need stable wrapper APIs, which as it were I happen to be interested in.

I’m not sure exactly how llvm.memset.element.unordered.atomic relates to, say, core::intrinsics::volatile_store (which has a stable wrapper std::ptr::write_volatile), but I’d really like an analogue for that for doing multiple writes of the same value.

I’m thinking something with a name along the lines of one of the following:

core::ptr::write_volatile_many
core::ptr::fill_volatile
core::ptr::set_volatile

joshlf · 2019-02-18T19:20:35.187Z

bascule:

I’m not sure exactly how llvm.memset.element.unordered.atomic relates to, say, core::intrinsics::volatile_store (which has a stable wrapper std::ptr::write_volatile ), but I’d really like an analogue for that for doing multiple writes of the same value.

The volatile intrinsics are not concurrency-safe; it’s UB to use them to operate on memory which is being concurrently modified.

joshlf · 2019-02-20T16:37:10.373Z

OK posted this issue: https://github.com/rust-lang/rust/issues/58599