RFC: Rvalue lifetimes

[nikomatsakis]

We need to clarify our story on rvalues. @graydon, @bblum and I have discussed this. I wrote up my latest thoughts on the situation more here:

http://smallcultfollowing.com/babysteps/blog/2012/09/15/rvalue-lifetimes/

Currently I am tending towards the "DWIM"-style rule (obviously I'd have to spell it out more specifically). I was thinking that a purely inference-based solution, while flexible, seems less specifiable. Or rather, to standardize it, you'd have to specify the inference technique, which seems undesirable.

[nikomatsakis]

Here is the rule I had in mind (based on C++). For any given temporary, you would walk up the AST, examining the parent nodes, to find the first one for which a rule is defined:

• match rvalue {...}: lifetime of rvalue is IES
• foo(..., rvalue, ...): lifetime of rvalue is IES
• ... + ... (etc): lifetime of rvalue is IES
• x = ..., lifetime of rvalue is block where x is declared
• lvalue = ..., where lvalue is some path owned by a local x, same as previous line
• '...;' (any statement), lifetime is that statement
• loop or function body: loop is that block

"IES" == innermost enclosing statement

I think this more-or-less works "as expected" in all cases. But I'm not entirely sure, probably the thing to do is to implement it and see what kind of borrowck errors result.

It might be easier to specify the rule by the cases where the result is not the innermost enclosing statement...

[bblum]

I'm not sure how I feel about this issue. Personally, I might even prefer if the compiler punted on some of the harder cases, saying something like "I couldn't find a lifetime for this rvalue that wouldn't be unexpected - please use a temporary to manually indicate the intended lifetime".

[graydon]

Isn't the region inference algorithm something we'll need to spec anyway? Just for the sake of defining borrow/use errors?

[nikomatsakis]

@graydon yes I realized that later.

[catamorphism]

Just a note for the future that part of resolving this issue means refactoring trans to be more declarative about how it handles cleanups (i.e. relating them to AST nodes rather than calling add_cleanup hither and yon).

[nikomatsakis]

nominating for milestone #1

[graydon]

accepted for well-defined milestone