The general theme of this is to make parser a better independent
library.
The specific thing we do here is replacing callback based TreeSink with
a data structure. That is, rather than calling user-provided tree
construction methods, the parser now spits out a very bare-bones tree,
effectively a log of a DFS traversal.
This makes the parser usable without any *specifc* tree sink, and allows
us to, eg, move tests into this crate.
Now, it's also true that this is a distinction without a difference, as
the old and the new interface are equivalent in expressiveness. Still,
this new thing seems somewhat simpler. But yeah, I admit I don't have a
suuper strong motivation here, just a hunch that this is better.
The general problem we are dealing with here is this:
```
macro_rules! thrice {
($e:expr) => { $e * 3}
}
fn main() {
let x = thrice!(1 + 2);
}
```
we really want this to print 9 rather than 7.
The way rustc solves this is rather ad-hoc. In rustc, token trees are
allowed to include whole AST fragments, so 1+2 is passed through macro
expansion as a single unit. This is a significant violation of token
tree model.
In rust-analyzer, we intended to handle this in a more elegant way,
using token trees with "invisible" delimiters. The idea was is that we
introduce a new kind of parenthesis, "left $"/"right $", and let the
parser intelligently handle this.
The idea was inspired by the relevant comment in the proc_macro crate:
https://doc.rust-lang.org/stable/proc_macro/enum.Delimiter.html#variant.None
> An implicit delimiter, that may, for example, appear around tokens
> coming from a “macro variable” $var. It is important to preserve
> operator priorities in cases like $var * 3 where $var is 1 + 2.
> Implicit delimiters might not survive roundtrip of a token stream
> through a string.
Now that we are older and wiser, we conclude that the idea doesn't work.
_First_, the comment in the proc-macro crate is wishful thinking. Rustc
currently completely ignores none delimiters. It solves the (1 + 2) * 3
problem by having magical token trees which can't be duplicated:
* https://rust-lang.zulipchat.com/#narrow/stream/185405-t-compiler.2Frust-analyzer/topic/TIL.20that.20token.20streams.20are.20magic
* https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/Handling.20of.20Delimiter.3A.3ANone.20by.20the.20parser
_Second_, it's not like our implementation in rust-analyzer works. We
special-case expressions (as opposed to treating all kinds of $var
captures the same) and we don't know how parser error recovery should
work with these dollar-parenthesis.
So, in this PR we simplify the whole thing away by not pretending that
we are doing something proper and instead just explicitly special-casing
expressions by wrapping them into real `()`.
In the future, to maintain bug-parity with `rustc` what we are going to
do is probably adding an explicit `CAPTURED_EXPR` *token* which we can
explicitly account for in the parser.
If/when rustc starts handling delimiter=none properly, we'll port that
logic as well, in addition to special handling.
I don't like our macro tests -- they are brittle and don't inspire
confidence. I think the reason for that is that we try to unit-test
them, but that is at odds with reality, where macro expansion
fundamentally depends on name resolution.
Consider these expples
{ 92 }
async { 92 }
'a: { 92 }
#[a] { 92 }
Previously the tree for them were
BLOCK_EXPR
{ ... }
EFFECT_EXPR
async
BLOCK_EXPR
{ ... }
EFFECT_EXPR
'a:
BLOCK_EXPR
{ ... }
BLOCK_EXPR
#[a]
{ ... }
As you see, it gets progressively worse :) The last two items are
especially odd. The last one even violates the balanced curleys
invariant we have (#10357) The new approach is to say that the stuff in
`{}` is stmt_list, and the block is stmt_list + optional modifiers
BLOCK_EXPR
STMT_LIST
{ ... }
BLOCK_EXPR
async
STMT_LIST
{ ... }
BLOCK_EXPR
'a:
STMT_LIST
{ ... }
BLOCK_EXPR
#[a]
STMT_LIST
{ ... }
FragmentKind played two roles:
* entry point to the parser
* syntactic category of a macro call
These are different use-cases, and warrant different types. For example,
macro can't expand to visibility, but we have such fragment today.
This PR introduces `ExpandsTo` enum to separate this two use-cases.
I suspect we might further split `FragmentKind` into `$x:specifier` enum
specific to MBE, and a general parser entry point, but that's for
another PR!
