Make libsyntax compile with minimal amount of feature flags

[erickt]

This a meta-bug that's tracking getting libsyntax to compile with the smallest possible feature flag set so that I can use it as the foundation for syntex. I've done a bunch of the work in #24487, but here are some more longer term dependencies that need some planning:

Here's what needs to be done:

#[feature(core)]:

• Int::zero() in ast
• ToPrimitive in codemap
• usize::to_u32() in codemap
• ptr::read_and_drop() in fold
• slice::ref_slice() in parser

#[feature(collections)]:

• std::collections::BitSet
• String::escape_default

#[feature(libc)]:

• libc::c_uint in codemap.rs
• libc::isatty(libc::STDERR_FILENO) in diagnostic.rs

#[feature(unicode)]:

• char.width() in diagnostic.rs
• char.is_xid_start() in lexer
• char.is_xid_continue() in lexer

#[feature(path_ext)]:

• Path::exists() - easily replaced with call to metadata

#[feature(str_char)]:

• str::char_at() - replace with s.chars().take()
• str::slice_shift_char()

#[feature(rustc_private)]:

• arena
• fmt_macros
• term - can use https://github.com/rust-lang/term

[erickt]

@alexcrichton: Regarding ptr::read_and_drop, the primary user of this code is syntax::ptr::P::map, which uses it to avoid an alloc when folding over ast. I could copy that functionality into libsyntax at the cost of having to copy some unstable drop code as well, which seems slightly risky. I could also reimplement map to just make a new allocation, at the cost of it being ~40% slower. Do you think this is on a critical-enough path that copying ptr::read_and_drop() is worth it? I'd be happy to compare the compile times of rust with and without the extra alloc to see if there's any real impact.

[erickt]

@alexcrichton: Interestingly enough, I just added another variation that modified P::map into:

    pub fn map3<F>(mut self, f: F) -> P<T> where
        F: FnOnce(&mut T),
    {
        f(&mut *self.ptr);
        self
    }

And it's ~66% faster. So maybe I should just replace all the users with that pattern. It's unfortunately a little less convenient, but it might not be that bad if most of the users are isolated to syntax::fold.

edit: here's the benchmark.

[alexcrichton]

@erickt if we could get by with the &mut T variant I think that would be best. The current implementation is not panic-safe into the future as eventually there will be no drop flag and calling read_and_drop followed by an arbitrary closure will not be sound (e.g. a panic in the closure will drop invalid data).

You can also just replace the map3 function with foo(&mut ptr) due to deref coercions, so if you're able to finagle it by just removing usage of the map function itself that may also be best.

[erickt]

@alexcrichton: Regarding BitSet, it's only used in attr.rs in order to speed up tracking if an attribute has been used or not. We got three options for syntex:

1. Replace it in libsyntax with HashSet with the assumption that we don't actually have that many attributes, that the overhead of using a whole usize might not really impact memory usage.
2. Pull BitSet library into an external crate and use it in syntex.
3. Have a patch in syntex that swaps attr.rs's use of BitSet with HashSet.

What do you think is best?

[alexcrichton]

@erickt in general I'm hesitant to contort libsyntax too much, so I would optimize for the least invasive solution. I'm not sure what the stability track is for BitSet, as that may be a plausible course of action.

[jonas-schievink]

Why is this tagged A-metadata?

[Mark-Simulacrum]

Closing in favor of (newer) #41732.