Instead of a `kind` field containting a `TokenNode` variant, a
TokenTree is now an enum with variants of different types (Literal, Op,
Term, etc). Note that a TokenTree could be a sequence of TokenTrees if it is a
Group variant.
Other notes:
I'm unsure about the set_span call in Builder::emit_if, but I did not want to
throw away the passed in Span.
Parsing relies on destructuring references to the values associated with
TokenTree enum variants
It doesn't seem as easy to compose/chain TokenStreams as it is to
collect a Vec<TokenTree> into a TokenStream. There is probably some
iterator API I could use instead. See `match_arms` and build.rs
Refs #121
This also disallows `@let` in non-block contexts, e.g.
img src=@let foo = bar; ;
This yields a 20 - 50 ns speedup in benchmarks, as well as reducing code
size by a slight amount.
When Rust and HTML syntax differ, Maud tends to side with Rust syntax.
This can be seen with string literals, for example, where we use
backslash escapes instead of HTML entities. Using `;` to terminate void
elements is consistent with this idea.
Moreover, this reduces confusion around the generated code. Maud does
not insert an extra slash into the HTML output (as per the spec) but
the syntax may imply otherwise.
This confusion may have been the cause of a [bug I found in the
wild][1], where the code omitted the trailing slash on an `input`
element. I suspect that this is because the author thought that a
trailing slash in the Maud template would lead to one in the HTML
output. Switching to semicolons would prevent this misconception.
[1]: https://github.com/anowell/quasar/pull/3
