`parse_tt` currently traverses a `&[TokenTree]` to do matching. But this
is a bad representation for the traversal.
- `TokenTree` is nested, and there's a bunch of expensive and fiddly
state required to handle entering and exiting nested submatchers.
- There are three positions (sequence separators, sequence Kleene ops,
and end of the matcher) that are represented by an index that exceeds
the end of the `&[TokenTree]`, which is clumsy and error-prone.
This commit introduces a new representation called `MatcherLoc` that is
designed specifically for matching. It fixes all the above problems,
making the code much easier to read. A `&[TokenTree]` is converted to a
`&[MatcherLoc]` before matching begins. Despite the cost of the
conversion, it's still a net performance win, because various pieces of
traversal state are computed once up-front, rather than having to be
recomputed repeatedly during the macro matching.
Some improvements worth noting.
- `parse_tt_inner` is *much* easier to read. No more having to compare
`idx` against `len` and read comments to understand what the result
means.
- The handling of `Delimited` in `parse_tt_inner` is now trivial.
- The three end-of-sequence cases in `parse_tt_inner` are now handled in
three separate match arms, and the control flow is much simpler.
- `nameize` is no longer recursive.
- There were two places that issued "missing fragment specifier" errors:
one in `parse_tt_inner()`, and one in `nameize()`. Presumably the
latter was never executed. There's now a single place issuing these
errors, in `compute_locs()`.
- The number of heap allocations done for a `check full` build of
`async-std-1.10.0` (an extreme example of heavy macro use) drops from
11.8M to 2.6M, and most of these occur outside of macro matching.
- The size of `MatcherPos` drops from 64 bytes to 16 bytes. Small enough
that it no longer needs boxing, which partly accounts for the
reduction in allocations.
- The rest of the drop in allocations is due to the removal of
`MatcherKind`, because we no longer need to record anything for the
parent matcher when entering a submatcher.
- Overall it reduces code size by 45 lines.
It's only used in one place, and there we clone and then make a bunch of
modifications. It's clearer if we duplicate more explicitly, and there's
a symmetry now between `sequence()` and `empty_sequence()`.
`parse_tt` needs a way to get from within submatchers make to the
enclosing submatchers. Currently it has two distinct mechanisms for
this:
- `Delimited` submatchers use `MatcherPos::stack` to record stuff about
the parent (and further back ancestors).
- `Sequence` submatchers use `MatcherPosSequence::parent` to point to
the parent matcher position.
Having two mechanisms is really confusing, and it took me a long time to
understand all this.
This commit eliminates `MatcherPos::stack`, and changes `Delimited`
submatchers to use the same mechanism as sequence submatchers. That
mechanism is also changed a bit: instead of storing the entire parent
`MatcherPos`, we now only store the necessary parts from the parent
`MatcherPos`.
Overall this is a small performance win, with the positives outweighing
the negatives, but it's mostly for clarity.
Currently, we detect an exit from a `Delimited` submatcher when `idx`
exceeds the bounds of the current submatcher *and* there is a `stack`
entry.
This commit changes it to something simpler: just look for a
`CloseDelim` token.
Currently, matches within a sequence are recorded in a new empty
`matches` vector. Then when the sequence finishes the matches are merged
into the `matches` vector of the parent.
This commit changes things so that a sequence mp inherits the matches
made so far. This means that additional matches from the sequence don't
need to be merged into the parent. `push_match` becomes more
complicated, and the current sequence depth needs to be tracked. But
it's a sizeable performance win because it avoids one or more
`push_match` calls on every iteration of a sequence.
The commit also removes `match_hi`, which is no longer necessary.
Ignore doc comments in a declarative macro matcher.
Fixes#95267. Reverts to the old behaviour before #95159 introduced a
regression.
r? `@petrochenkov`
It's only needed for macro expansion, not as a general element in the
AST. This commit removes it, adds `NtOrTt` for the parser and macro
expansion cases, and renames the variants in `NamedMatch` to better
match the new type.
Currently it copies a `KleeneOp` and a `Token` out of a
`SequenceRepetition`. It's better to store a reference to the
`SequenceRepetition`, which is now possible due to #95159 having changed
the lifetimes.
The `Lrc` is only relevant within `transcribe()`. There, the `Lrc` is
helpful for the non-`NtTT` cases, because the entire nonterminal is
cloned. But for the `NtTT` cases the inner token tree is cloned (a full
clone) and so the `Lrc` is of no help.
This commit splits the `NtTT` and non-`NtTT` cases, avoiding the useless
`Lrc` in the former case, for the following effect on macro-heavy
crates.
- It reduces the total number of allocations a lot.
- It increases the size of some of the remaining allocations.
- It doesn't affect *peak* memory usage, because the larger allocations
are short-lived.
This overall gives a speed win.
As its name suggests, `TokenTreeOrTokenTreeSlice` is either a single
`TokenTree` or a slice of them. It has methods `len` and `get_tt` that
let it be treated much like an ordinary slice. The reason it isn't an
ordinary slice is that for `TokenTree::Delimited` the open and close
delimiters are represented implicitly, and when they are needed they are
constructed on the fly with `Delimited::{open,close}_tt`, rather than
being present in memory.
This commit changes `Delimited` so the open and close delimiters are
represented explicitly. As a result, `TokenTreeOrTokenTreeSlice` is no
longer needed and `MatcherPos` and `MatcherTtFrame` can just use an
ordinary slice. `TokenTree::{len,get_tt}` are also removed, because they
were only needed to support `TokenTreeOrTokenTreeSlice`.
The change makes the code shorter and a little bit faster on benchmarks
that use macro expansion heavily, partly because `MatcherPos` is a lot
smaller (less data to `memcpy`) and partly because ordinary slice
operations are faster than `TokenTreeOrTokenTreeSlice::{len,get_tt}`.
By putting them in `TtParser`, we can reuse them for every rule in a
macro. With that done, they can be `SmallVec` instead of `Vec`, and this
is a performance win because these vectors are hot and `SmallVec`
operations are a bit slower due to always needing an "inline or heap?"
check.
This type was a small performance win for `html5ever`, which uses a
macro with hundreds of very simple rules that don't contain any
metavariables. But this type is complicated (extra lifetimes) and
perf-neutral for macros that do have metavariables.
This commit removes `MatcherPosHandle`, simplifying things a lot. This
increases the allocation rate for `html5ever` and similar cases a bit,
but makes things easier for follow-up changes that will improve
performance more than what we lost here.
It currently has no state, just the three methods `parse_tt`,
`parse_tt_inner`, and `bb_items_ambiguity_error`.
This commit is large but trivial, and mostly consists of changes to the
indentation of those methods. Subsequent commits will do more.
The current structure makes it hard to tell that there are just four
distinct code paths, depending on how many items there are in `bb_items`
and `next_items`. This commit introduces a `match` that clarifies
things.
There are three `Option` fields in `MatcherPos` that are only used in
tandem. This commit combines them, making the code slightly easier to
read. (It also makes clear that the `sep` field arguably should have
been `Option<Option<Token>>`!)
To avoid the strange style where comments force `else` onto its own
line.
The commit also removes several else-after-return constructs, which can
be hard to read.
