Box Block, fn_decl, variant and Ty in the AST.

[eddyb]

Note: I only tested on top of my #10670 PR, size reductions come from both change sets.

With this, more enums are shrinked, the most significant one being ast_node, from 104 bytes (master) to 96 (#10670) and now to 32 bytes.

My own testcase requires 200MB less when compiling (not including the other 200MB gained in #10670), and rustc-stage2 is down by about 130MB.

I believe there is more to gain by fiddling with the enums' layouts.

[eddyb]

I removed this clone, but not others (the variant one just below and some in the parser), any suggestions?

[pnkfelix]

oddly #2543 was closed by PR #9088 even though this comment remained in the source. I assume @nikomatsakis did mean to close it, since that PR removed a large number of the copies, but clearly not all of them...

[sanxiyn]

Should this really use @ box instead of ~ box?

[eddyb]

~T doesn't allow sharing - however, I'm preparing a new patch where I use P<T> instead of @T, and P(value) instead of @value to allocate.
This P<T> trick works with type P<T> = &'static T (coming from an arena), and with the upcoming smart pointer additions (the custom deref and borrow traits), type P<T> = Rc<T> could work just as well.

I will try to implement a proper arena, instead of leaking, with safe-ish (always fail on ptr deref) post-drop behavior (maybe by mmap-ing with MAP_NORESERVE and no permissions, after munmap-ing each chunk).

[eddyb]

Nevermind my previous idea, I'm using @T again, but I've kept the P<T> wrappers to help with future refactoring. I still can't use type P<T> = Rc<T> because the smart pointer improvements discussed by @nikomatsakis are required.

[pcwalton]

What improvements are those in particular? I would like to remove @ from the language ASAP…

[emberian]

[emberian]

(Note that this is without -O)

[thestinger]

@pcwalton: AFAIK he just means the dereference operator overload

[eddyb]

@pcwalton custom (auto-)dereference, (auto-)borrow and (optionally) use-after-move handler - so explicit cloning to split ownership isn't required.
But this last one is rare enough, so I can live with a P::split_ownership(&P<T>) wrapper that takes a reference and returns a new P<T> pointing to the same object (ref-count bump for Rc<T>).
Oh, and using Rc<T> without replacing every @T in the AST won't work right now, because the ~RcBox allocation ends up being a @RcBox and then it segfaults (because of the different layout).

[pnkfelix]

For me, sha 468657c fails make during an attempt to build librustdoc atop the stage1 compiler.

[eddyb]

Rebased and updated rustdoc - my bad, I always forget to fix the tests.

[emberian]

Squash and r=me