search

rust-lang / spec

The Rust specification
Apache License 2.0
103 stars 7 forks source link

Policy: Implementation-specific rules? #29

Open ehuss opened 8 months ago

ehuss commented 8 months ago

Should we ever allow implementation-specific rules? If so, how should that be represented?

In the past, there has been a sentiment from some in the project to never allow that. I largely agree with that. However, I think there will be gray areas that will be difficult to navigate with such a prohibition.

chorman0773 commented 8 months ago

Unspecified/nondeterministic is fairly well used within the operational semantics, and certain things outside of there basically have to:

Additionally, I would generally regard rustc --target x86_64-unknown-linux-gnu and rustc --target aarch64-pc-windows-msvc as being separate "implementations", insofar as they have different input parameters to the emulated abstract machine. Target-defined is, IMO, a subtype of implementation-defined.

digama0 commented 8 months ago

Additionally, I would generally regard rustc --target x86_64-unknown-linux-gnu and rustc --target aarch64-pc-windows-msvc as being separate "implementations", insofar as they have different input parameters to the emulated abstract machine. Target-defined is, IMO, a subtype of implementation-defined.

I don't really agree with this. I would expect a rust spec to say that size_of::<usize>() == 8 when compiling for the target x86_64-unknown-linux-gnu and this is not an implementation choice. This is just as much defined behavior as x + y performing a panicking add on debug and wrapping on release. I don't think we should copy the C/C++ practice of lumping together target-specific and implementation-defined behavior.

(This does not address the practical aspects of actually documenting all 100+ targets in the spec. This is just an aspirational statement about how I think the spec should be in the absence of other constraints.)

ehuss commented 8 months ago

Target-defined is, IMO, a subtype of implementation-defined.

Thanks for bringing that up. I think it is a distinct enough question that I forgot to include, and opened #44 for that.