In some rare cases, it was possible for ripgrep's inner literal detector
to extract a set of literals that could produce a false negative. #2884
gives an example: `(?i:e.x|ex)`. In this case, the set extracted can be
discovered by running `rg '(?i:e.x|ex) --trace`:
Seq[E("EX"), E("Ex"), E("eX"), E("ex")]
This extraction leads to building a multi-substring matcher for `EX`,
`Ex`, `eX` and `ex`. Searching the haystack `e-x` produces no match,
and thus, ripgrep shows no matches. But the regex `(?i:e.x|ex)` matches
`e-x`.
The issue at play here was that when two extracted literal sequences
were unioned, we were correctly unioning their "prefix" attribute.
And this in turn leads to those literal sequences being combined
incorrectly via cross product. This case in particular triggers it
because two different optimizations combine to produce an incorrect
result. Firslty, the regex has a common prefix extracted and is
rewritten as `(?i:e(?:.x|x))`. Secondly, the `x` in the first branch of
the alternation has its `prefix` attribute set to `false` (correctly),
which means it can't be cross producted with another concatenation. But
in this case, it is unioned with the `x` from the second branch, and
this results in the union result having `prefix` set to `true`. This
in turn pops up and lets it get cross producted with the `e` prefix,
producing an incorrect literal sequence.
We fix this by changing the implementation of `union` to return
`prefix` set to `true` only when *both* literal sequences being unioned
have `prefix` set to `true`.
Doing this exposed a second bug that was present, but was purely
cosmetic: the extracted literals in this case, after the fix, are
`X` and `x`. They were considered "exact" (i.e., lead to a match),
but of course they are not. Observing an `X` or an `x` does not mean
there is a match. This was fixed by making `choose` always return
an inexact literal sequence. This is perhaps too conservative in
aggregate in some cases, but always correct. The idea here is that if
one is choosing between two concatenations, then it is likely the case
that the sequence returned should be considered inexact. The issue
is that this can lead to avoiding cross products in some cases that
would otherwise be correct. This is bad because it means extracting
shorter literals in some cases. (In general, the longer the literal the
better.) But we prioritize correctness for now and fix it. You can see
a few tests where this shortens some extracted literals.
Fixes#2884
Previously, we had logic to skip our own inner literal optimization if
the regex itself was already (likely) accelerated. It turns out that the
presence of a Unicode word boundary can defeat acceleration to a point.
It's likely enough that even if the underlying regex is accelerated, it
would be prudent to do our own inner literal optimization if the pattern
has a Unicode word boundary.
Normally a Unicode word boundary doesn't defeat literal optimizations,
since even the slower engines can make use of *prefix* literal
optimizations. But a regex can be accelerated via its own inner or
suffix literal optimizations, and those require the use of a DFA (or
lazy DFA). Since DFAs crap out on haystacks that contain a non-ASCII
Unicode scalar value when the regex contains a Unicode word boundary, it
follows that an "accelerated" can still wind up being quite slow.
(An "accelerated" regex can also slow down because of restrictions on
avoiding quadratic behavior, but I believe this happens less frequently
and is not as severe as the slow down as a result of Unicode word
boundaries. Namely, avoiding quadratic behavior just means giving up on
the inner literal optimization for a single search. In which case, the
regex engine can still fall back to a normal forward DFA. That will
definitely be slower than an inner literal optimization done by ripgrep,
but not quite as dramatic as it would be when DFAs can't be used at
all.)
This is mostly a copy of the prefix literal extractor in regex-syntax,
but with a tweaked notion of Seq that keeps track of whether it's a
prefix of an expression or not. If it isn't, then we can't cross it as a
suffix to another Seq.
This new extractor should be a lot more robust than the old one. We
actually will keep going through the regex to try and find the "best"
literals to search for (according to some heuristic).
This does a little bit of refactoring so that we can pass both a
ConfiguredHIR and a Regex to the inner literal extraction routine.
One downside of this approach is that a regex object hangs on to a
ConfiguredHIR. But the extra memory usage is probably negligible. A
benefit though is that converting the HIR to its concrete syntax is now
lazy and only happens when logging is enabled.
This leaves the grep-regex crate in tatters. Pretty much the entire
thing needs to be re-worked. The upshot is that it should result in some
big simplifications. I hope.
The idea here is to drop down and actually use regex-automata 0.3
instead of the regex crate itself.
In a prior commit, we fixed a performance problem with the -w flag by
doing a little extra work to extract literals. It turns out that using
literals in this case when the -w flag is NOT used results in a
performance regression. The reasoning is that we end up using a "fast"
regex as a prefilter when the regex engine itself uses its own
equivalent prefilter, so ripgrep ends up redoing a fair amount of work.
Instead, we only do this extra work when we know the -w flag is enabled.
If a literal is entirely whitespace, then it's quite likely that it is
very common. So when that case occurs, just don't do (inner) literal
optimizations at all.
The regex engine may still make sub-optimal decisions here, but that's a
problem for another day.
Fixes#1087
The top-level listing was just getting a bit too long for my taste. So
put all of the code in one directory and shrink the large top-level mess
to a small top-level mess.
NOTE: This commit only contains renames. The subsequent commit will
actually make ripgrep build again. We do it this way with the naive hope
that this will make it easier for git history to track the renames.
Sigh.
