it was a fun idea to get easy perf wins. it's clear it's a failed experiment
Conversation
I like the performance gains but I also like the concept that pointers point to the type of objects they say they point to. I almost always see -fno-strict-overflow as an indication of a failing codebase.
4
3
I hear you but the language as she is used in practice encourages type punning in sorta critical stuff like the famous socket example
3
8
What are the performance consequences in practice of disabling strict aliasing in C, anyway? Rust doesn't have it and it didn't actually tell LLVM about noalias until relatively recently, and I think enabling noalias only got ~ 5% avg so far (though that'll probably improve).
2
Rust can provide much better aliasing information than C based on the properties of & and &mut which is equivalent to putting restrict on nearly all pointers in C code. It hasn't been able to leverage that as much as it should be able to on top of LLVM due to optimization bugs.
2
5
&T for any type T without interior mutability (which is tracked by the type system) and any &mut U type can be marked with the same noalias markers used by Clang for restrict since noalias means no memory dependencies between them and is valid even if the address is the same.
2
4
Rust is actually doing that now, though, isn't it? I think it was enabled several months ago. Performance definitely improved, I just think ~5% isn't enough to justify something as intrusive as strict aliasing is for C (and it should be able to do *better* than strict aliasing).
1
1
Not saying that can't improve with time, and like I said maybe C code has characteristics that make it benefit more from TBAA than I'd expect, I'm mostly just curious whether strict aliasing is even worth it for C.
2
1
Rust is currently only adding noalias markers to function parameters when the parameter is Box<T>, &mut T or &T with the &T case limited to when the type T has no interior mutability (i.e. it does not contain any UnsafeCell such as Cell, RefCell, Mutex, etc. use internally).
2
2
Rust isn't using scoped noalias metadata yet due to a mix of LLVM bugs and the need to clearly define the semantics for unsafe code. The LLVM scoped noalias metadata also has limitations in how it's designed. It can only mark the outermost pointer used in a function parameter.
2
2
LLVM converts function parameter metadata into scoped noalias metadata so you can manually add no-op functions passing through Box<T>, &T or &mut T in order to get a scoped noalias metadata marker added to your code. Means you can at least work around it when micro-optimizing.
Probably, if they haven't added any scoped noalias metadata by that point they'll be stripping out tons of the scoped noalias metadata that's generated by LLVM when it inlines functions from the noalias function parameters.
1
1
Show replies