regex_automata/meta/

strategy.rs

use core::{
    fmt::Debug,
    panic::{RefUnwindSafe, UnwindSafe},
};

use alloc::{borrow::Cow, format, sync::Arc};

use regex_syntax::hir::{literal, Hir};

use crate::{
    meta::{
        error::{BuildError, RetryError, RetryFailError, RetryQuadraticError},
        regex::{Cache, RegexInfo},
        reverse_inner, reverse_suffix, wrappers,
    },
    nfa::thompson::{self, WhichCaptures, NFA},
    util::{
        captures::{Captures, GroupInfo},
        look::LookMatcher,
        prefilter::{self, Prefilter, PrefilterI},
        primitives::{NonMaxUsize, PatternID},
        search::{Anchored, HalfMatch, Input, Match, MatchKind, PatternSet},
    },
};

/// A trait that represents a single meta strategy. Its main utility is in
/// providing a way to do dynamic dispatch over a few choices.
///
/// Why dynamic dispatch? I actually don't have a super compelling reason, and
/// importantly, I have not benchmarked it with the main alternative: an enum.
/// I went with dynamic dispatch initially because the regex engine search code
/// really can't be inlined into caller code in most cases because it's just
/// too big. In other words, it is already expected that every regex search
/// will entail at least the cost of a function call.
///
/// I do wonder whether using enums would result in better codegen overall
/// though. It's a worthwhile experiment to try. Probably the most interesting
/// benchmark to run in such a case would be one with a high match count. That
/// is, a benchmark to test the overall latency of a search call.
pub(super) trait Strategy:
    Debug + Send + Sync + RefUnwindSafe + UnwindSafe + 'static
{
    #[allow(dead_code)]
    fn name(&self) -> Cow<'static, str>;

    fn group_info(&self) -> &GroupInfo;

    fn create_cache(&self) -> Cache;

    fn reset_cache(&self, cache: &mut Cache);

    fn is_accelerated(&self) -> bool;

    fn memory_usage(&self) -> usize;

    fn search(&self, cache: &mut Cache, input: &Input<'_>) -> Option<Match>;

    fn search_half(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Option<HalfMatch>;

    fn is_match(&self, cache: &mut Cache, input: &Input<'_>) -> bool;

    fn search_slots(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        slots: &mut [Option<NonMaxUsize>],
    ) -> Option<PatternID>;

    fn which_overlapping_matches(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        patset: &mut PatternSet,
    );
}

pub(super) fn new(
    info: &RegexInfo,
    hirs: &[&Hir],
) -> Result<Arc<dyn Strategy>, BuildError> {
    // At this point, we're committed to a regex engine of some kind. So pull
    // out a prefilter if we can, which will feed to each of the constituent
    // regex engines.
    let pre = if info.is_always_anchored_start() {
        // PERF: I'm not sure we necessarily want to do this... We may want to
        // run a prefilter for quickly rejecting in some cases. The problem
        // is that anchored searches overlap quite a bit with the use case
        // of "run a regex on every line to extract data." In that case, the
        // regex always matches, so running a prefilter doesn't really help us
        // there. The main place where a prefilter helps in an anchored search
        // is if the anchored search is not expected to match frequently. That
        // is, the prefilter gives us a way to possibly reject a haystack very
        // quickly.
        //
        // Maybe we should do use a prefilter, but only for longer haystacks?
        // Or maybe we should only use a prefilter when we think it's "fast"?
        //
        // Interestingly, I think we currently lack the infrastructure for
        // disabling a prefilter based on haystack length. That would probably
        // need to be a new 'Input' option. (Interestingly, an 'Input' used to
        // carry a 'Prefilter' with it, but I moved away from that.)
        debug!("skipping literal extraction since regex is anchored");
        None
    } else if let Some(pre) = info.config().get_prefilter() {
        debug!(
            "skipping literal extraction since the caller provided a prefilter"
        );
        Some(pre.clone())
    } else if info.config().get_auto_prefilter() {
        let kind = info.config().get_match_kind();
        let prefixes = crate::util::prefilter::prefixes(kind, hirs);
        // If we can build a full `Strategy` from just the extracted prefixes,
        // then we can short-circuit and avoid building a regex engine at all.
        if let Some(pre) = Pre::from_prefixes(info, &prefixes) {
            debug!(
                "found that the regex can be broken down to a literal \
                 search, avoiding the regex engine entirely",
            );
            debug!("using {} strategy", pre.name());
            return Ok(pre);
        }
        // This now attempts another short-circuit of the regex engine: if we
        // have a huge alternation of just plain literals, then we can just use
        // Aho-Corasick for that and avoid the regex engine entirely.
        //
        // You might think this case would just be handled by
        // `Pre::from_prefixes`, but that technique relies on heuristic literal
        // extraction from the corresponding `Hir`. That works, but part of
        // heuristics limit the size and number of literals returned. This case
        // will specifically handle patterns with very large alternations.
        //
        // One wonders if we should just roll this our heuristic literal
        // extraction, and then I think this case could disappear entirely.
        if let Some(pre) = Pre::from_alternation_literals(info, hirs) {
            debug!(
                "found plain alternation of literals, \
                 avoiding regex engine entirely and using Aho-Corasick"
            );
            debug!("using {} strategy", pre.name());
            return Ok(pre);
        }
        prefixes.literals().and_then(|strings| {
            debug!(
                "creating prefilter from {} literals: {:?}",
                strings.len(),
                strings,
            );
            Prefilter::new(kind, strings)
        })
    } else {
        debug!("skipping literal extraction since prefilters were disabled");
        None
    };
    let mut core = Core::new(info.clone(), pre.clone(), hirs)?;
    // Now that we have our core regex engines built, there are a few cases
    // where we can do a little bit better than just a normal "search forward
    // and maybe use a prefilter when in a start state." However, these cases
    // may not always work or otherwise build on top of the Core searcher.
    // For example, the reverse anchored optimization seems like it might
    // always work, but only the DFAs support reverse searching and the DFAs
    // might give up or quit for reasons. If we had, e.g., a PikeVM that
    // supported reverse searching, then we could avoid building a full Core
    // engine for this case.
    core = match ReverseAnchored::new(core) {
        Err(core) => core,
        Ok(ra) => {
            debug!("using {} strategy", ra.name());
            return Ok(Arc::new(ra));
        }
    };
    core = match ReverseSuffix::new(core, hirs) {
        Err(core) => core,
        Ok(rs) => {
            debug!("using {} strategy", rs.name());
            return Ok(Arc::new(rs));
        }
    };
    core = match ReverseInner::new(core, hirs) {
        Err(core) => core,
        Ok(ri) => {
            debug!("using {} strategy", ri.name());
            return Ok(Arc::new(ri));
        }
    };
    debug!("using {} strategy", core.name());
    Ok(Arc::new(core))
}

#[derive(Clone, Debug)]
struct Pre<P> {
    pre: P,
    group_info: GroupInfo,
}

impl<P: PrefilterI> Pre<P> {
    fn new(pre: P) -> Arc<dyn Strategy> {
        // The only thing we support when we use prefilters directly as a
        // strategy is the start and end of the overall match for a single
        // pattern. In other words, exactly one implicit capturing group. Which
        // is exactly what we use here for a GroupInfo.
        let group_info = GroupInfo::new([[None::<&str>]]).unwrap();
        Arc::new(Pre { pre, group_info })
    }
}

// This is a little weird, but we don't actually care about the type parameter
// here because we're selecting which underlying prefilter to use. So we just
// define it on an arbitrary type.
impl Pre<()> {
    /// Given a sequence of prefixes, attempt to return a full `Strategy` using
    /// just the prefixes.
    ///
    /// Basically, this occurs when the prefixes given not just prefixes,
    /// but an enumeration of the entire language matched by the regular
    /// expression.
    ///
    /// A number of other conditions need to be true too. For example, there
    /// can be only one pattern, the number of explicit capture groups is 0, no
    /// look-around assertions and so on.
    ///
    /// Note that this ignores `Config::get_auto_prefilter` because if this
    /// returns something, then it isn't a prefilter but a matcher itself.
    /// Therefore, it shouldn't suffer from the problems typical to prefilters
    /// (such as a high false positive rate).
    fn from_prefixes(
        info: &RegexInfo,
        prefixes: &literal::Seq,
    ) -> Option<Arc<dyn Strategy>> {
        let kind = info.config().get_match_kind();
        // Check to see if our prefixes are exact, which means we might be
        // able to bypass the regex engine entirely and just rely on literal
        // searches.
        if !prefixes.is_exact() {
            return None;
        }
        // We also require that we have a single regex pattern. Namely,
        // we reuse the prefilter infrastructure to implement search and
        // prefilters only report spans. Prefilters don't know about pattern
        // IDs. The multi-regex case isn't a lost cause, we might still use
        // Aho-Corasick and we might still just use a regular prefilter, but
        // that's done below.
        if info.pattern_len() != 1 {
            return None;
        }
        // We can't have any capture groups either. The literal engines don't
        // know how to deal with things like '(foo)(bar)'. In that case, a
        // prefilter will just be used and then the regex engine will resolve
        // the capture groups.
        if info.props()[0].explicit_captures_len() != 0 {
            return None;
        }
        // We also require that it has zero look-around assertions. Namely,
        // literal extraction treats look-around assertions as if they match
        // *every* empty string. But of course, that isn't true. So for
        // example, 'foo\bquux' never matches anything, but 'fooquux' is
        // extracted from that as an exact literal. Such cases should just run
        // the regex engine. 'fooquux' will be used as a normal prefilter, and
        // then the regex engine will try to look for an actual match.
        if !info.props()[0].look_set().is_empty() {
            return None;
        }
        // Finally, currently, our prefilters are all oriented around
        // leftmost-first match semantics, so don't try to use them if the
        // caller asked for anything else.
        if kind != MatchKind::LeftmostFirst {
            return None;
        }
        // The above seems like a lot of requirements to meet, but it applies
        // to a lot of cases. 'foo', '[abc][123]' and 'foo|bar|quux' all meet
        // the above criteria, for example.
        //
        // Note that this is effectively a latency optimization. If we didn't
        // do this, then the extracted literals would still get bundled into
        // a prefilter, and every regex engine capable of running unanchored
        // searches supports prefilters. So this optimization merely sidesteps
        // having to run the regex engine at all to confirm the match. Thus, it
        // decreases the latency of a match.

        // OK because we know the set is exact and thus finite.
        let prefixes = prefixes.literals().unwrap();
        debug!(
            "trying to bypass regex engine by creating \
             prefilter from {} literals: {:?}",
            prefixes.len(),
            prefixes,
        );
        let choice = match prefilter::Choice::new(kind, prefixes) {
            Some(choice) => choice,
            None => {
                debug!(
                    "regex bypass failed because no prefilter could be built"
                );
                return None;
            }
        };
        let strat: Arc<dyn Strategy> = match choice {
            prefilter::Choice::Memchr(pre) => Pre::new(pre),
            prefilter::Choice::Memchr2(pre) => Pre::new(pre),
            prefilter::Choice::Memchr3(pre) => Pre::new(pre),
            prefilter::Choice::Memmem(pre) => Pre::new(pre),
            prefilter::Choice::Teddy(pre) => Pre::new(pre),
            prefilter::Choice::ByteSet(pre) => Pre::new(pre),
            prefilter::Choice::AhoCorasick(pre) => Pre::new(pre),
        };
        Some(strat)
    }

    /// Attempts to extract an alternation of literals, and if it's deemed
    /// worth doing, returns an Aho-Corasick prefilter as a strategy.
    ///
    /// And currently, this only returns something when 'hirs.len() == 1'. This
    /// could in theory do something if there are multiple HIRs where all of
    /// them are alternation of literals, but I haven't had the time to go down
    /// that path yet.
    fn from_alternation_literals(
        info: &RegexInfo,
        hirs: &[&Hir],
    ) -> Option<Arc<dyn Strategy>> {
        use crate::util::prefilter::AhoCorasick;

        let lits = crate::meta::literal::alternation_literals(info, hirs)?;
        let ac = AhoCorasick::new(MatchKind::LeftmostFirst, &lits)?;
        Some(Pre::new(ac))
    }
}

// This implements Strategy for anything that implements PrefilterI.
//
// Note that this must only be used for regexes of length 1. Multi-regexes
// don't work here. The prefilter interface only provides the span of a match
// and not the pattern ID. (I did consider making it more expressive, but I
// couldn't figure out how to tie everything together elegantly.) Thus, so long
// as the regex only contains one pattern, we can simply assume that a match
// corresponds to PatternID::ZERO. And indeed, that's what we do here.
//
// In practice, since this impl is used to report matches directly and thus
// completely bypasses the regex engine, we only wind up using this under the
// following restrictions:
//
// * There must be only one pattern. As explained above.
// * The literal sequence must be finite and only contain exact literals.
// * There must not be any look-around assertions. If there are, the literals
// extracted might be exact, but a match doesn't necessarily imply an overall
// match. As a trivial example, 'foo\bbar' does not match 'foobar'.
// * The pattern must not have any explicit capturing groups. If it does, the
// caller might expect them to be resolved. e.g., 'foo(bar)'.
//
// So when all of those things are true, we use a prefilter directly as a
// strategy.
//
// In the case where the number of patterns is more than 1, we don't use this
// but do use a special Aho-Corasick strategy if all of the regexes are just
// simple literals or alternations of literals. (We also use the Aho-Corasick
// strategy when len(patterns)==1 if the number of literals is large. In that
// case, literal extraction gives up and will return an infinite set.)
impl<P: PrefilterI> Strategy for Pre<P> {
    fn name(&self) -> Cow<'static, str> {
        Cow::Owned(format!("prefilter {}", self.pre.name()))
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn group_info(&self) -> &GroupInfo {
        &self.group_info
    }

    fn create_cache(&self) -> Cache {
        Cache {
            capmatches: Captures::all(self.group_info().clone()),
            pikevm: wrappers::PikeVMCache::none(),
            backtrack: wrappers::BoundedBacktrackerCache::none(),
            onepass: wrappers::OnePassCache::none(),
            hybrid: wrappers::HybridCache::none(),
            revhybrid: wrappers::ReverseHybridCache::none(),
        }
    }

    fn reset_cache(&self, _cache: &mut Cache) {}

    fn is_accelerated(&self) -> bool {
        self.pre.is_fast()
    }

    fn memory_usage(&self) -> usize {
        self.pre.memory_usage()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search(&self, _cache: &mut Cache, input: &Input<'_>) -> Option<Match> {
        if input.is_done() {
            return None;
        }
        if input.get_anchored().is_anchored() {
            return self
                .pre
                .prefix(input.haystack(), input.get_span())
                .map(|sp| Match::new(PatternID::ZERO, sp));
        }
        self.pre
            .find(input.haystack(), input.get_span())
            .map(|sp| Match::new(PatternID::ZERO, sp))
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_half(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Option<HalfMatch> {
        self.search(cache, input).map(|m| HalfMatch::new(m.pattern(), m.end()))
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn is_match(&self, cache: &mut Cache, input: &Input<'_>) -> bool {
        self.search(cache, input).is_some()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_slots(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        slots: &mut [Option<NonMaxUsize>],
    ) -> Option<PatternID> {
        let m = self.search(cache, input)?;
        if let Some(slot) = slots.get_mut(0) {
            *slot = NonMaxUsize::new(m.start());
        }
        if let Some(slot) = slots.get_mut(1) {
            *slot = NonMaxUsize::new(m.end());
        }
        Some(m.pattern())
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn which_overlapping_matches(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        patset: &mut PatternSet,
    ) {
        if self.search(cache, input).is_some() {
            patset.insert(PatternID::ZERO);
        }
    }
}

#[derive(Debug)]
struct Core {
    info: RegexInfo,
    pre: Option<Prefilter>,
    nfa: NFA,
    nfarev: Option<NFA>,
    pikevm: wrappers::PikeVM,
    backtrack: wrappers::BoundedBacktracker,
    onepass: wrappers::OnePass,
    hybrid: wrappers::Hybrid,
    dfa: wrappers::DFA,
}

impl Core {
    fn new(
        info: RegexInfo,
        pre: Option<Prefilter>,
        hirs: &[&Hir],
    ) -> Result<Core, BuildError> {
        let mut lookm = LookMatcher::new();
        lookm.set_line_terminator(info.config().get_line_terminator());
        let thompson_config = info.config().to_thompson_config();
        let nfa = thompson::Compiler::new()
            .configure(thompson_config.clone())
            .build_many_from_hir(hirs)
            .map_err(BuildError::nfa)?;
        // It's possible for the PikeVM or the BB to fail to build, even though
        // at this point, we already have a full NFA in hand. They can fail
        // when a Unicode word boundary is used but where Unicode word boundary
        // support is disabled at compile time, thus making it impossible to
        // match. (Construction can also fail if the NFA was compiled without
        // captures, but we always enable that above.)
        let pikevm = wrappers::PikeVM::new(&info, pre.clone(), &nfa)?;
        let backtrack =
            wrappers::BoundedBacktracker::new(&info, pre.clone(), &nfa)?;
        // The onepass engine can of course fail to build, but we expect it to
        // fail in many cases because it is an optimization that doesn't apply
        // to all regexes. The 'OnePass' wrapper encapsulates this failure (and
        // logs a message if it occurs).
        let onepass = wrappers::OnePass::new(&info, &nfa);
        // We try to encapsulate whether a particular regex engine should be
        // used within each respective wrapper, but the DFAs need a reverse NFA
        // to build itself, and we really do not want to build a reverse NFA if
        // we know we aren't going to use the lazy DFA. So we do a config check
        // up front, which is in practice the only way we won't try to use the
        // DFA.
        let (nfarev, hybrid, dfa) =
            if !info.config().get_hybrid() && !info.config().get_dfa() {
                (None, wrappers::Hybrid::none(), wrappers::DFA::none())
            } else {
                // FIXME: Technically, we don't quite yet KNOW that we need
                // a reverse NFA. It's possible for the DFAs below to both
                // fail to build just based on the forward NFA. In which case,
                // building the reverse NFA was totally wasted work. But...
                // fixing this requires breaking DFA construction apart into
                // two pieces: one for the forward part and another for the
                // reverse part. Quite annoying. Making it worse, when building
                // both DFAs fails, it's quite likely that the NFA is large and
                // that it will take quite some time to build the reverse NFA
                // too. So... it's really probably worth it to do this!
                let nfarev = thompson::Compiler::new()
                    // Currently, reverse NFAs don't support capturing groups,
                    // so we MUST disable them. But even if we didn't have to,
                    // we would, because nothing in this crate does anything
                    // useful with capturing groups in reverse. And of course,
                    // the lazy DFA ignores capturing groups in all cases.
                    .configure(
                        thompson_config
                            .which_captures(WhichCaptures::None)
                            .reverse(true),
                    )
                    .build_many_from_hir(hirs)
                    .map_err(BuildError::nfa)?;
                let dfa = if !info.config().get_dfa() {
                    wrappers::DFA::none()
                } else {
                    wrappers::DFA::new(&info, pre.clone(), &nfa, &nfarev)
                };
                let hybrid = if !info.config().get_hybrid() {
                    wrappers::Hybrid::none()
                } else if dfa.is_some() {
                    debug!("skipping lazy DFA because we have a full DFA");
                    wrappers::Hybrid::none()
                } else {
                    wrappers::Hybrid::new(&info, pre.clone(), &nfa, &nfarev)
                };
                (Some(nfarev), hybrid, dfa)
            };
        Ok(Core {
            info,
            pre,
            nfa,
            nfarev,
            pikevm,
            backtrack,
            onepass,
            hybrid,
            dfa,
        })
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn try_search_mayfail(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Option<Result<Option<Match>, RetryFailError>> {
        if let Some(e) = self.dfa.get(input) {
            trace!("using full DFA for search at {:?}", input.get_span());
            Some(e.try_search(input))
        } else if let Some(e) = self.hybrid.get(input) {
            trace!("using lazy DFA for search at {:?}", input.get_span());
            Some(e.try_search(&mut cache.hybrid, input))
        } else {
            None
        }
    }

    fn search_nofail(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Option<Match> {
        let caps = &mut cache.capmatches;
        caps.set_pattern(None);
        // We manually inline 'try_search_slots_nofail' here because we need to
        // borrow from 'cache.capmatches' in this method, but if we do, then
        // we can't pass 'cache' wholesale to to 'try_slots_no_hybrid'. It's a
        // classic example of how the borrow checker inhibits decomposition.
        // There are of course work-arounds (more types and/or interior
        // mutability), but that's more annoying than this IMO.
        let pid = if let Some(ref e) = self.onepass.get(input) {
            trace!("using OnePass for search at {:?}", input.get_span());
            e.search_slots(&mut cache.onepass, input, caps.slots_mut())
        } else if let Some(ref e) = self.backtrack.get(input) {
            trace!(
                "using BoundedBacktracker for search at {:?}",
                input.get_span()
            );
            e.search_slots(&mut cache.backtrack, input, caps.slots_mut())
        } else {
            trace!("using PikeVM for search at {:?}", input.get_span());
            let e = self.pikevm.get();
            e.search_slots(&mut cache.pikevm, input, caps.slots_mut())
        };
        caps.set_pattern(pid);
        caps.get_match()
    }

    fn search_half_nofail(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Option<HalfMatch> {
        // Only the lazy/full DFA returns half-matches, since the DFA requires
        // a reverse scan to find the start position. These fallback regex
        // engines can find the start and end in a single pass, so we just do
        // that and throw away the start offset to conform to the API.
        let m = self.search_nofail(cache, input)?;
        Some(HalfMatch::new(m.pattern(), m.end()))
    }

    fn search_slots_nofail(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        slots: &mut [Option<NonMaxUsize>],
    ) -> Option<PatternID> {
        if let Some(ref e) = self.onepass.get(input) {
            trace!(
                "using OnePass for capture search at {:?}",
                input.get_span()
            );
            e.search_slots(&mut cache.onepass, input, slots)
        } else if let Some(ref e) = self.backtrack.get(input) {
            trace!(
                "using BoundedBacktracker for capture search at {:?}",
                input.get_span()
            );
            e.search_slots(&mut cache.backtrack, input, slots)
        } else {
            trace!(
                "using PikeVM for capture search at {:?}",
                input.get_span()
            );
            let e = self.pikevm.get();
            e.search_slots(&mut cache.pikevm, input, slots)
        }
    }

    fn is_match_nofail(&self, cache: &mut Cache, input: &Input<'_>) -> bool {
        if let Some(ref e) = self.onepass.get(input) {
            trace!(
                "using OnePass for is-match search at {:?}",
                input.get_span()
            );
            e.search_slots(&mut cache.onepass, input, &mut []).is_some()
        } else if let Some(ref e) = self.backtrack.get(input) {
            trace!(
                "using BoundedBacktracker for is-match search at {:?}",
                input.get_span()
            );
            e.is_match(&mut cache.backtrack, input)
        } else {
            trace!(
                "using PikeVM for is-match search at {:?}",
                input.get_span()
            );
            let e = self.pikevm.get();
            e.is_match(&mut cache.pikevm, input)
        }
    }

    fn is_capture_search_needed(&self, slots_len: usize) -> bool {
        slots_len > self.nfa.group_info().implicit_slot_len()
    }
}

impl Strategy for Core {
    fn name(&self) -> Cow<'static, str> {
        Cow::Borrowed("core")
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn group_info(&self) -> &GroupInfo {
        self.nfa.group_info()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn create_cache(&self) -> Cache {
        Cache {
            capmatches: Captures::all(self.group_info().clone()),
            pikevm: self.pikevm.create_cache(),
            backtrack: self.backtrack.create_cache(),
            onepass: self.onepass.create_cache(),
            hybrid: self.hybrid.create_cache(),
            revhybrid: wrappers::ReverseHybridCache::none(),
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn reset_cache(&self, cache: &mut Cache) {
        cache.pikevm.reset(&self.pikevm);
        cache.backtrack.reset(&self.backtrack);
        cache.onepass.reset(&self.onepass);
        cache.hybrid.reset(&self.hybrid);
    }

    fn is_accelerated(&self) -> bool {
        self.pre.as_ref().map_or(false, |pre| pre.is_fast())
    }

    fn memory_usage(&self) -> usize {
        self.info.memory_usage()
            + self.pre.as_ref().map_or(0, |pre| pre.memory_usage())
            + self.nfa.memory_usage()
            + self.nfarev.as_ref().map_or(0, |nfa| nfa.memory_usage())
            + self.onepass.memory_usage()
            + self.dfa.memory_usage()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search(&self, cache: &mut Cache, input: &Input<'_>) -> Option<Match> {
        // We manually inline try_search_mayfail here because letting the
        // compiler do it seems to produce pretty crappy codegen.
        return if let Some(e) = self.dfa.get(input) {
            trace!("using full DFA for full search at {:?}", input.get_span());
            match e.try_search(input) {
                Ok(x) => x,
                Err(_err) => {
                    trace!("full DFA search failed: {_err}");
                    self.search_nofail(cache, input)
                }
            }
        } else if let Some(e) = self.hybrid.get(input) {
            trace!("using lazy DFA for full search at {:?}", input.get_span());
            match e.try_search(&mut cache.hybrid, input) {
                Ok(x) => x,
                Err(_err) => {
                    trace!("lazy DFA search failed: {_err}");
                    self.search_nofail(cache, input)
                }
            }
        } else {
            self.search_nofail(cache, input)
        };
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_half(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Option<HalfMatch> {
        // The main difference with 'search' is that if we're using a DFA, we
        // can use a single forward scan without needing to run the reverse
        // DFA.
        if let Some(e) = self.dfa.get(input) {
            trace!("using full DFA for half search at {:?}", input.get_span());
            match e.try_search_half_fwd(input) {
                Ok(x) => x,
                Err(_err) => {
                    trace!("full DFA half search failed: {_err}");
                    self.search_half_nofail(cache, input)
                }
            }
        } else if let Some(e) = self.hybrid.get(input) {
            trace!("using lazy DFA for half search at {:?}", input.get_span());
            match e.try_search_half_fwd(&mut cache.hybrid, input) {
                Ok(x) => x,
                Err(_err) => {
                    trace!("lazy DFA half search failed: {_err}");
                    self.search_half_nofail(cache, input)
                }
            }
        } else {
            self.search_half_nofail(cache, input)
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn is_match(&self, cache: &mut Cache, input: &Input<'_>) -> bool {
        if let Some(e) = self.dfa.get(input) {
            trace!(
                "using full DFA for is-match search at {:?}",
                input.get_span()
            );
            match e.try_search_half_fwd(input) {
                Ok(x) => x.is_some(),
                Err(_err) => {
                    trace!("full DFA half search failed: {_err}");
                    self.is_match_nofail(cache, input)
                }
            }
        } else if let Some(e) = self.hybrid.get(input) {
            trace!(
                "using lazy DFA for is-match search at {:?}",
                input.get_span()
            );
            match e.try_search_half_fwd(&mut cache.hybrid, input) {
                Ok(x) => x.is_some(),
                Err(_err) => {
                    trace!("lazy DFA half search failed: {_err}");
                    self.is_match_nofail(cache, input)
                }
            }
        } else {
            self.is_match_nofail(cache, input)
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_slots(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        slots: &mut [Option<NonMaxUsize>],
    ) -> Option<PatternID> {
        // Even if the regex has explicit capture groups, if the caller didn't
        // provide any explicit slots, then it doesn't make sense to try and do
        // extra work to get offsets for those slots. Ideally the caller should
        // realize this and not call this routine in the first place, but alas,
        // we try to save the caller from themselves if they do.
        if !self.is_capture_search_needed(slots.len()) {
            trace!("asked for slots unnecessarily, trying fast path");
            let m = self.search(cache, input)?;
            copy_match_to_slots(m, slots);
            return Some(m.pattern());
        }
        // If the onepass DFA is available for this search (which only happens
        // when it's anchored), then skip running a fallible DFA. The onepass
        // DFA isn't as fast as a full or lazy DFA, but it is typically quite
        // a bit faster than the backtracker or the PikeVM. So it isn't as
        // advantageous to try and do a full/lazy DFA scan first.
        //
        // We still theorize that it's better to do a full/lazy DFA scan, even
        // when it's anchored, because it's usually much faster and permits us
        // to say "no match" much more quickly. This does hurt the case of,
        // say, parsing each line in a log file into capture groups, because
        // in that case, the line always matches. So the lazy DFA scan is
        // usually just wasted work. But, the lazy DFA is usually quite fast
        // and doesn't cost too much here.
        if self.onepass.get(&input).is_some() {
            return self.search_slots_nofail(cache, &input, slots);
        }
        let m = match self.try_search_mayfail(cache, input) {
            Some(Ok(Some(m))) => m,
            Some(Ok(None)) => return None,
            Some(Err(_err)) => {
                trace!("fast capture search failed: {_err}");
                return self.search_slots_nofail(cache, input, slots);
            }
            None => {
                return self.search_slots_nofail(cache, input, slots);
            }
        };
        // At this point, now that we've found the bounds of the
        // match, we need to re-run something that can resolve
        // capturing groups. But we only need to run on it on the
        // match bounds and not the entire haystack.
        trace!(
            "match found at {}..{} in capture search, \
             using another engine to find captures",
            m.start(),
            m.end(),
        );
        let input = input
            .clone()
            .span(m.start()..m.end())
            .anchored(Anchored::Pattern(m.pattern()));
        Some(
            self.search_slots_nofail(cache, &input, slots)
                .expect("should find a match"),
        )
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn which_overlapping_matches(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        patset: &mut PatternSet,
    ) {
        if let Some(e) = self.dfa.get(input) {
            trace!(
                "using full DFA for overlapping search at {:?}",
                input.get_span()
            );
            let _err = match e.try_which_overlapping_matches(input, patset) {
                Ok(()) => return,
                Err(err) => err,
            };
            trace!("fast overlapping search failed: {_err}");
        } else if let Some(e) = self.hybrid.get(input) {
            trace!(
                "using lazy DFA for overlapping search at {:?}",
                input.get_span()
            );
            let _err = match e.try_which_overlapping_matches(
                &mut cache.hybrid,
                input,
                patset,
            ) {
                Ok(()) => {
                    return;
                }
                Err(err) => err,
            };
            trace!("fast overlapping search failed: {_err}");
        }
        trace!(
            "using PikeVM for overlapping search at {:?}",
            input.get_span()
        );
        let e = self.pikevm.get();
        e.which_overlapping_matches(&mut cache.pikevm, input, patset)
    }
}

#[derive(Debug)]
struct ReverseAnchored {
    core: Core,
}

impl ReverseAnchored {
    fn new(core: Core) -> Result<ReverseAnchored, Core> {
        if !core.info.is_always_anchored_end() {
            debug!(
                "skipping reverse anchored optimization because \
                 the regex is not always anchored at the end"
            );
            return Err(core);
        }
        // Note that the caller can still request an anchored search even when
        // the regex isn't anchored at the start. We detect that case in the
        // search routines below and just fallback to the core engine. This
        // is fine because both searches are anchored. It's just a matter of
        // picking one. Falling back to the core engine is a little simpler,
        // since if we used the reverse anchored approach, we'd have to add an
        // extra check to ensure the match reported starts at the place where
        // the caller requested the search to start.
        if core.info.is_always_anchored_start() {
            debug!(
                "skipping reverse anchored optimization because \
                 the regex is also anchored at the start"
            );
            return Err(core);
        }
        // Only DFAs can do reverse searches (currently), so we need one of
        // them in order to do this optimization. It's possible (although
        // pretty unlikely) that we have neither and need to give up.
        if !core.hybrid.is_some() && !core.dfa.is_some() {
            debug!(
                "skipping reverse anchored optimization because \
                 we don't have a lazy DFA or a full DFA"
            );
            return Err(core);
        }
        Ok(ReverseAnchored { core })
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn try_search_half_anchored_rev(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Result<Option<HalfMatch>, RetryFailError> {
        // We of course always want an anchored search. In theory, the
        // underlying regex engines should automatically enable anchored
        // searches since the regex is itself anchored, but this more clearly
        // expresses intent and is always correct.
        let input = input.clone().anchored(Anchored::Yes);
        if let Some(e) = self.core.dfa.get(&input) {
            trace!(
                "using full DFA for reverse anchored search at {:?}",
                input.get_span()
            );
            e.try_search_half_rev(&input)
        } else if let Some(e) = self.core.hybrid.get(&input) {
            trace!(
                "using lazy DFA for reverse anchored search at {:?}",
                input.get_span()
            );
            e.try_search_half_rev(&mut cache.hybrid, &input)
        } else {
            unreachable!("ReverseAnchored always has a DFA")
        }
    }
}

// Note that in this impl, we don't check that 'input.end() ==
// input.haystack().len()'. In particular, when that condition is false, a
// match is always impossible because we know that the regex is always anchored
// at the end (or else 'ReverseAnchored' won't be built). We don't check that
// here because the 'Regex' wrapper actually does that for us in all cases.
// Thus, in this impl, we can actually assume that the end position in 'input'
// is equivalent to the length of the haystack.
impl Strategy for ReverseAnchored {
    fn name(&self) -> Cow<'static, str> {
        Cow::Borrowed("reverse anchored")
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn group_info(&self) -> &GroupInfo {
        self.core.group_info()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn create_cache(&self) -> Cache {
        self.core.create_cache()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn reset_cache(&self, cache: &mut Cache) {
        self.core.reset_cache(cache);
    }

    fn is_accelerated(&self) -> bool {
        // Since this is anchored at the end, a reverse anchored search is
        // almost certainly guaranteed to result in a much faster search than
        // a standard forward search.
        true
    }

    fn memory_usage(&self) -> usize {
        self.core.memory_usage()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search(&self, cache: &mut Cache, input: &Input<'_>) -> Option<Match> {
        if input.get_anchored().is_anchored() {
            return self.core.search(cache, input);
        }
        match self.try_search_half_anchored_rev(cache, input) {
            Err(_err) => {
                trace!("fast reverse anchored search failed: {_err}");
                self.core.search_nofail(cache, input)
            }
            Ok(None) => None,
            Ok(Some(hm)) => {
                Some(Match::new(hm.pattern(), hm.offset()..input.end()))
            }
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_half(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Option<HalfMatch> {
        if input.get_anchored().is_anchored() {
            return self.core.search_half(cache, input);
        }
        match self.try_search_half_anchored_rev(cache, input) {
            Err(_err) => {
                trace!("fast reverse anchored search failed: {_err}");
                self.core.search_half_nofail(cache, input)
            }
            Ok(None) => None,
            Ok(Some(hm)) => {
                // Careful here! 'try_search_half' is a *forward* search that
                // only cares about the *end* position of a match. But
                // 'hm.offset()' is actually the start of the match. So we
                // actually just throw that away here and, since we know we
                // have a match, return the only possible position at which a
                // match can occur: input.end().
                Some(HalfMatch::new(hm.pattern(), input.end()))
            }
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn is_match(&self, cache: &mut Cache, input: &Input<'_>) -> bool {
        if input.get_anchored().is_anchored() {
            return self.core.is_match(cache, input);
        }
        match self.try_search_half_anchored_rev(cache, input) {
            Err(_err) => {
                trace!("fast reverse anchored search failed: {_err}");
                self.core.is_match_nofail(cache, input)
            }
            Ok(None) => false,
            Ok(Some(_)) => true,
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_slots(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        slots: &mut [Option<NonMaxUsize>],
    ) -> Option<PatternID> {
        if input.get_anchored().is_anchored() {
            return self.core.search_slots(cache, input, slots);
        }
        match self.try_search_half_anchored_rev(cache, input) {
            Err(_err) => {
                trace!("fast reverse anchored search failed: {_err}");
                self.core.search_slots_nofail(cache, input, slots)
            }
            Ok(None) => None,
            Ok(Some(hm)) => {
                if !self.core.is_capture_search_needed(slots.len()) {
                    trace!("asked for slots unnecessarily, skipping captures");
                    let m = Match::new(hm.pattern(), hm.offset()..input.end());
                    copy_match_to_slots(m, slots);
                    return Some(m.pattern());
                }
                let start = hm.offset();
                let input = input
                    .clone()
                    .span(start..input.end())
                    .anchored(Anchored::Pattern(hm.pattern()));
                self.core.search_slots_nofail(cache, &input, slots)
            }
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn which_overlapping_matches(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        patset: &mut PatternSet,
    ) {
        // It seems like this could probably benefit from a reverse anchored
        // optimization, perhaps by doing an overlapping reverse search (which
        // the DFAs do support). I haven't given it much thought though, and
        // I'm currently focus more on the single pattern case.
        self.core.which_overlapping_matches(cache, input, patset)
    }
}

#[derive(Debug)]
struct ReverseSuffix {
    core: Core,
    pre: Prefilter,
}

impl ReverseSuffix {
    fn new(core: Core, hirs: &[&Hir]) -> Result<ReverseSuffix, Core> {
        if !core.info.config().get_auto_prefilter() {
            debug!(
                "skipping reverse suffix optimization because \
                 automatic prefilters are disabled"
            );
            return Err(core);
        }
        // Also like the reverse inner optimization, a reverse suffix encodes
        // leftmost-first match semantics.
        if core.info.config().get_match_kind() != MatchKind::LeftmostFirst {
            debug!(
                "skipping reverse suffix optimization because \
                 match kind is {:?} but this only supports leftmost-first",
                core.info.config().get_match_kind(),
            );
            return Err(core);
        }
        // Like the reverse inner optimization, we don't do this for regexes
        // that are always anchored. It could lead to scanning too much, but
        // could say "no match" much more quickly than running the regex
        // engine if the initial literal scan doesn't match. With that said,
        // the reverse suffix optimization has lower overhead, since it only
        // requires a reverse scan after a literal match to confirm or reject
        // the match. (Although, in the case of confirmation, it then needs to
        // do another forward scan to find the end position.)
        //
        // Note that the caller can still request an anchored search even
        // when the regex isn't anchored. We detect that case in the search
        // routines below and just fallback to the core engine. Currently this
        // optimization assumes all searches are unanchored, so if we do want
        // to enable this optimization for anchored searches, it will need a
        // little work to support it.
        if core.info.is_always_anchored_start() {
            debug!(
                "skipping reverse suffix optimization because \
                 the regex is always anchored at the start",
            );
            return Err(core);
        }
        // Only DFAs can do reverse searches (currently), so we need one of
        // them in order to do this optimization. It's possible (although
        // pretty unlikely) that we have neither and need to give up.
        if !core.hybrid.is_some() && !core.dfa.is_some() {
            debug!(
                "skipping reverse suffix optimization because \
                 we don't have a lazy DFA or a full DFA"
            );
            return Err(core);
        }
        if core.pre.as_ref().map_or(false, |p| p.is_fast()) {
            debug!(
                "skipping reverse suffix optimization because \
                 we already have a prefilter that we think is fast"
            );
            return Err(core);
        }
        let kind = core.info.config().get_match_kind();
        let suffixes = crate::util::prefilter::suffixes(kind, hirs);
        let lcs = match suffixes.longest_common_suffix() {
            None => {
                debug!(
                    "skipping reverse suffix optimization because \
                     a longest common suffix could not be found",
                );
                return Err(core);
            }
            Some(lcs) if lcs.is_empty() => {
                debug!(
                    "skipping reverse suffix optimization because \
                     the longest common suffix is the empty string",
                );
                return Err(core);
            }
            Some(lcs) => lcs,
        };
        let pre = match Prefilter::new(kind, &[lcs]) {
            Some(pre) => pre,
            None => {
                debug!(
                    "skipping reverse suffix optimization because \
                     a prefilter could not be constructed from the \
                     longest common suffix",
                );
                return Err(core);
            }
        };
        if !pre.is_fast() {
            debug!(
                "skipping reverse suffix optimization because \
                 while we have a suffix prefilter, it is not \
                 believed to be 'fast'"
            );
            return Err(core);
        }
        if !reverse_suffix::has_no_earlier_match(hirs, &lcs) {
            debug!(
                "skipping reverse suffix optimization because \
                 an earlier suffix match could be a complete match \
                 inside of a larger match"
            );
            return Err(core);
        }
        Ok(ReverseSuffix { core, pre })
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn try_search_half_start(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Result<Option<HalfMatch>, RetryError> {
        let mut span = input.get_span();
        let mut min_start = 0;
        loop {
            let litmatch = match self.pre.find(input.haystack(), span) {
                None => break,
                Some(span) => span,
            };
            trace!("reverse suffix scan found suffix match at {litmatch:?}");
            let revinput = input
                .clone()
                .anchored(Anchored::Yes)
                .span(input.start()..litmatch.end);
            if let Some(hm) =
                self.try_search_half_rev_limited(cache, &revinput, min_start)?
            {
                return Ok(Some(hm));
            }

            if span.start >= span.end {
                break;
            }
            span.start = litmatch.start.checked_add(1).unwrap();
            min_start = litmatch.end;
        }
        Ok(None)
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn try_search_half_fwd(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Result<Option<HalfMatch>, RetryFailError> {
        if let Some(e) = self.core.dfa.get(&input) {
            trace!(
                "using full DFA for forward reverse suffix search at {:?}",
                input.get_span()
            );
            e.try_search_half_fwd(&input)
        } else if let Some(e) = self.core.hybrid.get(&input) {
            trace!(
                "using lazy DFA for forward reverse suffix search at {:?}",
                input.get_span()
            );
            e.try_search_half_fwd(&mut cache.hybrid, &input)
        } else {
            unreachable!("ReverseSuffix always has a DFA")
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn try_search_half_rev_limited(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        min_start: usize,
    ) -> Result<Option<HalfMatch>, RetryError> {
        if let Some(e) = self.core.dfa.get(&input) {
            trace!(
                "using full DFA for reverse suffix search at {:?}, \
                 but will be stopped at {} to avoid quadratic behavior",
                input.get_span(),
                min_start,
            );
            e.try_search_half_rev_limited(&input, min_start)
        } else if let Some(e) = self.core.hybrid.get(&input) {
            trace!(
                "using lazy DFA for reverse suffix search at {:?}, \
                 but will be stopped at {} to avoid quadratic behavior",
                input.get_span(),
                min_start,
            );
            e.try_search_half_rev_limited(&mut cache.hybrid, &input, min_start)
        } else {
            unreachable!("ReverseSuffix always has a DFA")
        }
    }
}

impl Strategy for ReverseSuffix {
    fn name(&self) -> Cow<'static, str> {
        Cow::Borrowed("reverse suffix")
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn group_info(&self) -> &GroupInfo {
        self.core.group_info()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn create_cache(&self) -> Cache {
        self.core.create_cache()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn reset_cache(&self, cache: &mut Cache) {
        self.core.reset_cache(cache);
    }

    fn is_accelerated(&self) -> bool {
        self.pre.is_fast()
    }

    fn memory_usage(&self) -> usize {
        self.core.memory_usage() + self.pre.memory_usage()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search(&self, cache: &mut Cache, input: &Input<'_>) -> Option<Match> {
        if input.get_anchored().is_anchored() {
            return self.core.search(cache, input);
        }
        match self.try_search_half_start(cache, input) {
            Err(RetryError::Quadratic(_err)) => {
                trace!("reverse suffix optimization failed: {_err}");
                self.core.search(cache, input)
            }
            Err(RetryError::Fail(_err)) => {
                trace!("reverse suffix reverse fast search failed: {_err}");
                self.core.search_nofail(cache, input)
            }
            Ok(None) => None,
            Ok(Some(hm_start)) => {
                let fwdinput = input
                    .clone()
                    .anchored(Anchored::Pattern(hm_start.pattern()))
                    .span(hm_start.offset()..input.end());
                match self.try_search_half_fwd(cache, &fwdinput) {
                    Err(_err) => {
                        trace!(
                            "reverse suffix forward fast search failed: {_err}"
                        );
                        self.core.search_nofail(cache, input)
                    }
                    Ok(None) => {
                        unreachable!(
                            "suffix match plus reverse match implies \
                             there must be a match",
                        )
                    }
                    Ok(Some(hm_end)) => Some(Match::new(
                        hm_start.pattern(),
                        hm_start.offset()..hm_end.offset(),
                    )),
                }
            }
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_half(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Option<HalfMatch> {
        if input.get_anchored().is_anchored() {
            return self.core.search_half(cache, input);
        }
        match self.try_search_half_start(cache, input) {
            Err(RetryError::Quadratic(_err)) => {
                trace!("reverse suffix half optimization failed: {_err}");
                self.core.search_half(cache, input)
            }
            Err(RetryError::Fail(_err)) => {
                trace!(
                    "reverse suffix reverse fast half search failed: {_err}"
                );
                self.core.search_half_nofail(cache, input)
            }
            Ok(None) => None,
            Ok(Some(hm_start)) => {
                // This is a bit subtle. It is tempting to just stop searching
                // at this point and return a half-match with an offset
                // corresponding to where the suffix was found. But the suffix
                // match does not necessarily correspond to the end of the
                // proper leftmost-first match. Consider /[a-z]+ing/ against
                // 'tingling'. The first suffix match is the first 'ing', and
                // the /[a-z]+/ matches the 't'. So if we stopped here, then
                // we'd report 'ting' as the match. But 'tingling' is the
                // correct match because of greediness.
                let fwdinput = input
                    .clone()
                    .anchored(Anchored::Pattern(hm_start.pattern()))
                    .span(hm_start.offset()..input.end());
                match self.try_search_half_fwd(cache, &fwdinput) {
                    Err(_err) => {
                        trace!(
                            "reverse suffix forward fast search failed: {_err}"
                        );
                        self.core.search_half_nofail(cache, input)
                    }
                    Ok(None) => {
                        unreachable!(
                            "suffix match plus reverse match implies \
                             there must be a match",
                        )
                    }
                    Ok(Some(hm_end)) => Some(hm_end),
                }
            }
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn is_match(&self, cache: &mut Cache, input: &Input<'_>) -> bool {
        if input.get_anchored().is_anchored() {
            return self.core.is_match(cache, input);
        }
        match self.try_search_half_start(cache, input) {
            Err(RetryError::Quadratic(_err)) => {
                trace!("reverse suffix half optimization failed: {_err}");
                self.core.is_match_nofail(cache, input)
            }
            Err(RetryError::Fail(_err)) => {
                trace!(
                    "reverse suffix reverse fast half search failed: {_err}"
                );
                self.core.is_match_nofail(cache, input)
            }
            Ok(None) => false,
            Ok(Some(_)) => true,
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_slots(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        slots: &mut [Option<NonMaxUsize>],
    ) -> Option<PatternID> {
        if input.get_anchored().is_anchored() {
            return self.core.search_slots(cache, input, slots);
        }
        if !self.core.is_capture_search_needed(slots.len()) {
            trace!("asked for slots unnecessarily, trying fast path");
            let m = self.search(cache, input)?;
            copy_match_to_slots(m, slots);
            return Some(m.pattern());
        }
        let hm_start = match self.try_search_half_start(cache, input) {
            Err(RetryError::Quadratic(_err)) => {
                trace!("reverse suffix captures optimization failed: {_err}");
                return self.core.search_slots(cache, input, slots);
            }
            Err(RetryError::Fail(_err)) => {
                trace!(
                    "reverse suffix reverse fast captures search failed: \
                     {_err}"
                );
                return self.core.search_slots_nofail(cache, input, slots);
            }
            Ok(None) => return None,
            Ok(Some(hm_start)) => hm_start,
        };
        trace!(
            "match found at {}..{} in capture search, \
             using another engine to find captures",
            hm_start.offset(),
            input.end(),
        );
        let start = hm_start.offset();
        let input = input
            .clone()
            .span(start..input.end())
            .anchored(Anchored::Pattern(hm_start.pattern()));
        self.core.search_slots_nofail(cache, &input, slots)
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn which_overlapping_matches(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        patset: &mut PatternSet,
    ) {
        self.core.which_overlapping_matches(cache, input, patset)
    }
}

#[derive(Debug)]
struct ReverseInner {
    core: Core,
    preinner: Prefilter,
    nfarev: NFA,
    hybrid: wrappers::ReverseHybrid,
    dfa: wrappers::ReverseDFA,
}

impl ReverseInner {
    fn new(core: Core, hirs: &[&Hir]) -> Result<ReverseInner, Core> {
        if !core.info.config().get_auto_prefilter() {
            debug!(
                "skipping reverse inner optimization because \
                 automatic prefilters are disabled"
            );
            return Err(core);
        }
        // Currently we hard-code the assumption of leftmost-first match
        // semantics. This isn't a huge deal because 'all' semantics tend to
        // only be used for forward overlapping searches with multiple regexes,
        // and this optimization only supports a single pattern at the moment.
        if core.info.config().get_match_kind() != MatchKind::LeftmostFirst {
            debug!(
                "skipping reverse inner optimization because \
                 match kind is {:?} but this only supports leftmost-first",
                core.info.config().get_match_kind(),
            );
            return Err(core);
        }
        // It's likely that a reverse inner scan has too much overhead for it
        // to be worth it when the regex is anchored at the start. It is
        // possible for it to be quite a bit faster if the initial literal
        // scan fails to detect a match, in which case, we can say "no match"
        // very quickly. But this could be undesirable, e.g., scanning too far
        // or when the literal scan matches. If it matches, then confirming the
        // match requires a reverse scan followed by a forward scan to confirm
        // or reject, which is a fair bit of work.
        //
        // Note that the caller can still request an anchored search even
        // when the regex isn't anchored. We detect that case in the search
        // routines below and just fallback to the core engine. Currently this
        // optimization assumes all searches are unanchored, so if we do want
        // to enable this optimization for anchored searches, it will need a
        // little work to support it.
        if core.info.is_always_anchored_start() {
            debug!(
                "skipping reverse inner optimization because \
                 the regex is always anchored at the start",
            );
            return Err(core);
        }
        // Only DFAs can do reverse searches (currently), so we need one of
        // them in order to do this optimization. It's possible (although
        // pretty unlikely) that we have neither and need to give up.
        if !core.hybrid.is_some() && !core.dfa.is_some() {
            debug!(
                "skipping reverse inner optimization because \
                 we don't have a lazy DFA or a full DFA"
            );
            return Err(core);
        }
        if core.pre.as_ref().map_or(false, |p| p.is_fast()) {
            debug!(
                "skipping reverse inner optimization because \
                 we already have a prefilter that we think is fast"
            );
            return Err(core);
        } else if core.pre.is_some() {
            debug!(
                "core engine has a prefix prefilter, but it is \
                 probably not fast, so continuing with attempt to \
                 use reverse inner prefilter"
            );
        }
        let prefilter = match reverse_inner::InnerPrefilter::new(hirs) {
            Some(prefilter) => prefilter,
            // N.B. the 'new' function emits debug messages explaining
            // why we bailed out here.
            None => return Err(core),
        };
        if !reverse_inner::has_no_earlier_match(
            &prefilter.prefix,
            &prefilter.literals,
        ) {
            debug!(
                "skipping reverse inner optimization because an inner \
                 literal match could be confirmed before an earlier match"
            );
            return Err(core);
        }
        debug!("building reverse NFA for prefix before inner literal");
        let thompson_config = core
            .info
            .config()
            .to_thompson_config()
            .reverse(true)
            .which_captures(WhichCaptures::None);
        let result = thompson::Compiler::new()
            .configure(thompson_config)
            .build_from_hir(&prefilter.prefix);
        let nfarev = match result {
            Ok(nfarev) => nfarev,
            Err(_err) => {
                debug!(
                    "skipping reverse inner optimization because the \
                     reverse NFA failed to build: {}",
                    _err,
                );
                return Err(core);
            }
        };
        debug!("building reverse DFA for prefix before inner literal");
        let dfa = if !core.info.config().get_dfa() {
            wrappers::ReverseDFA::none()
        } else {
            wrappers::ReverseDFA::new(&core.info, &nfarev)
        };
        let hybrid = if !core.info.config().get_hybrid() {
            wrappers::ReverseHybrid::none()
        } else if dfa.is_some() {
            debug!(
                "skipping lazy DFA for reverse inner optimization \
                 because we have a full DFA"
            );
            wrappers::ReverseHybrid::none()
        } else {
            wrappers::ReverseHybrid::new(&core.info, &nfarev)
        };
        Ok(ReverseInner { core, preinner: prefilter.pre, nfarev, hybrid, dfa })
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn try_search_full(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Result<Option<Match>, RetryError> {
        let mut span = input.get_span();
        let mut min_match_start = 0;
        let mut min_pre_start = 0;
        loop {
            let litmatch = match self.preinner.find(input.haystack(), span) {
                None => break,
                Some(span) => span,
            };
            if litmatch.start < min_pre_start {
                trace!(
                    "found inner prefilter match at {litmatch:?}, which starts \
                     before the end of the last forward scan at {min_pre_start}, \
                     quitting to avoid quadratic behavior",
                );
                return Err(RetryError::Quadratic(RetryQuadraticError::new()));
            }
            trace!("reverse inner scan found inner match at {litmatch:?}");
            let revinput = input
                .clone()
                .anchored(Anchored::Yes)
                .span(input.start()..litmatch.start);
            // Note that in addition to the literal search above scanning past
            // our minimum start point, this routine can also return an error
            // as a result of detecting possible quadratic behavior if the
            // reverse scan goes past the minimum start point. That is, the
            // literal search might not, but the reverse regex search for the
            // prefix might!
            if let Some(hm_start) = self.try_search_half_rev_limited(
                cache,
                &revinput,
                min_match_start,
            )? {
                let fwdinput = input
                    .clone()
                    .anchored(Anchored::Pattern(hm_start.pattern()))
                    .span(hm_start.offset()..input.end());
                match self.try_search_half_fwd_stopat(cache, &fwdinput)? {
                    Err(stopat) => {
                        min_pre_start = stopat;
                        span.start = litmatch.start.checked_add(1).unwrap();
                    }
                    Ok(hm_end) => {
                        return Ok(Some(Match::new(
                            hm_start.pattern(),
                            hm_start.offset()..hm_end.offset(),
                        )));
                    }
                }
            }

            if span.start >= span.end {
                break;
            }
            span.start = litmatch.start.checked_add(1).unwrap();
            min_match_start = litmatch.end;
        }
        Ok(None)
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn try_search_half_fwd_stopat(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Result<Result<HalfMatch, usize>, RetryFailError> {
        if let Some(e) = self.core.dfa.get(&input) {
            trace!(
                "using full DFA for forward reverse inner search at {:?}",
                input.get_span()
            );
            e.try_search_half_fwd_stopat(&input)
        } else if let Some(e) = self.core.hybrid.get(&input) {
            trace!(
                "using lazy DFA for forward reverse inner search at {:?}",
                input.get_span()
            );
            e.try_search_half_fwd_stopat(&mut cache.hybrid, &input)
        } else {
            unreachable!("ReverseInner always has a DFA")
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn try_search_half_rev_limited(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        min_start: usize,
    ) -> Result<Option<HalfMatch>, RetryError> {
        if let Some(e) = self.dfa.get(&input) {
            trace!(
                "using full DFA for reverse inner search at {:?}, \
                 but will be stopped at {} to avoid quadratic behavior",
                input.get_span(),
                min_start,
            );
            e.try_search_half_rev_limited(&input, min_start)
        } else if let Some(e) = self.hybrid.get(&input) {
            trace!(
                "using lazy DFA for reverse inner search at {:?}, \
                 but will be stopped at {} to avoid quadratic behavior",
                input.get_span(),
                min_start,
            );
            e.try_search_half_rev_limited(
                &mut cache.revhybrid,
                &input,
                min_start,
            )
        } else {
            unreachable!("ReverseInner always has a DFA")
        }
    }
}

impl Strategy for ReverseInner {
    fn name(&self) -> Cow<'static, str> {
        Cow::Borrowed("reverse inner")
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn group_info(&self) -> &GroupInfo {
        self.core.group_info()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn create_cache(&self) -> Cache {
        let mut cache = self.core.create_cache();
        cache.revhybrid = self.hybrid.create_cache();
        cache
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn reset_cache(&self, cache: &mut Cache) {
        self.core.reset_cache(cache);
        cache.revhybrid.reset(&self.hybrid);
    }

    fn is_accelerated(&self) -> bool {
        self.preinner.is_fast()
    }

    fn memory_usage(&self) -> usize {
        self.core.memory_usage()
            + self.preinner.memory_usage()
            + self.nfarev.memory_usage()
            + self.dfa.memory_usage()
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search(&self, cache: &mut Cache, input: &Input<'_>) -> Option<Match> {
        if input.get_anchored().is_anchored() {
            return self.core.search(cache, input);
        }
        match self.try_search_full(cache, input) {
            Err(RetryError::Quadratic(_err)) => {
                trace!("reverse inner optimization failed: {_err}");
                self.core.search(cache, input)
            }
            Err(RetryError::Fail(_err)) => {
                trace!("reverse inner fast search failed: {_err}");
                self.core.search_nofail(cache, input)
            }
            Ok(matornot) => matornot,
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_half(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
    ) -> Option<HalfMatch> {
        if input.get_anchored().is_anchored() {
            return self.core.search_half(cache, input);
        }
        match self.try_search_full(cache, input) {
            Err(RetryError::Quadratic(_err)) => {
                trace!("reverse inner half optimization failed: {_err}");
                self.core.search_half(cache, input)
            }
            Err(RetryError::Fail(_err)) => {
                trace!("reverse inner fast half search failed: {_err}");
                self.core.search_half_nofail(cache, input)
            }
            Ok(None) => None,
            Ok(Some(m)) => Some(HalfMatch::new(m.pattern(), m.end())),
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn is_match(&self, cache: &mut Cache, input: &Input<'_>) -> bool {
        if input.get_anchored().is_anchored() {
            return self.core.is_match(cache, input);
        }
        match self.try_search_full(cache, input) {
            Err(RetryError::Quadratic(_err)) => {
                trace!("reverse inner half optimization failed: {_err}");
                self.core.is_match_nofail(cache, input)
            }
            Err(RetryError::Fail(_err)) => {
                trace!("reverse inner fast half search failed: {_err}");
                self.core.is_match_nofail(cache, input)
            }
            Ok(None) => false,
            Ok(Some(_)) => true,
        }
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn search_slots(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        slots: &mut [Option<NonMaxUsize>],
    ) -> Option<PatternID> {
        if input.get_anchored().is_anchored() {
            return self.core.search_slots(cache, input, slots);
        }
        if !self.core.is_capture_search_needed(slots.len()) {
            trace!("asked for slots unnecessarily, trying fast path");
            let m = self.search(cache, input)?;
            copy_match_to_slots(m, slots);
            return Some(m.pattern());
        }
        let m = match self.try_search_full(cache, input) {
            Err(RetryError::Quadratic(_err)) => {
                trace!("reverse inner captures optimization failed: {_err}");
                return self.core.search_slots(cache, input, slots);
            }
            Err(RetryError::Fail(_err)) => {
                trace!("reverse inner fast captures search failed: {_err}");
                return self.core.search_slots_nofail(cache, input, slots);
            }
            Ok(None) => return None,
            Ok(Some(m)) => m,
        };
        trace!(
            "match found at {}..{} in capture search, \
             using another engine to find captures",
            m.start(),
            m.end(),
        );
        let input = input
            .clone()
            .span(m.start()..m.end())
            .anchored(Anchored::Pattern(m.pattern()));
        self.core.search_slots_nofail(cache, &input, slots)
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn which_overlapping_matches(
        &self,
        cache: &mut Cache,
        input: &Input<'_>,
        patset: &mut PatternSet,
    ) {
        self.core.which_overlapping_matches(cache, input, patset)
    }
}

/// Copies the offsets in the given match to the corresponding positions in
/// `slots`.
///
/// In effect, this sets the slots corresponding to the implicit group for the
/// pattern in the given match. If the indices for the corresponding slots do
/// not exist, then no slots are set.
///
/// This is useful when the caller provides slots (or captures), but you use a
/// regex engine that doesn't operate on slots (like a lazy DFA). This function
/// lets you map the match you get back to the slots provided by the caller.
#[cfg_attr(feature = "perf-inline", inline(always))]
fn copy_match_to_slots(m: Match, slots: &mut [Option<NonMaxUsize>]) {
    let slot_start = m.pattern().as_usize() * 2;
    let slot_end = slot_start + 1;
    if let Some(slot) = slots.get_mut(slot_start) {
        *slot = NonMaxUsize::new(m.start());
    }
    if let Some(slot) = slots.get_mut(slot_end) {
        *slot = NonMaxUsize::new(m.end());
    }
}

// We only test which strategy we get when all literal features are enabled.
// Other cases are less substantially less interesting.
//
// We also don't test this on miri since it takes forever.
//
// We also require `unicode-perl` since some regexes use `\w`, `\d` and `\s`.
#[cfg(all(
    feature = "perf-literal-substring",
    feature = "perf-literal-multisubstring",
    feature = "unicode-perl",
    not(miri),
))]
#[cfg(test)]
mod tests {
    use alloc::{format, string::String, vec::Vec};

    use crate::{meta::regex::Config, util::syntax};

    use super::*;

    fn teddy_probably_available() -> bool {
        cfg!(any(target_arch = "x86_64", target_arch = "aarch64"))
    }

    #[track_caller]
    fn strategy_with(patterns: &[&str], config: Config) -> Arc<dyn Strategy> {
        let hirs = syntax::parse_many(patterns).unwrap();
        let hirs: Vec<&Hir> = hirs.iter().collect();
        let info = RegexInfo::new(config, &hirs);
        new(&info, &hirs).unwrap()
    }

    #[track_caller]
    fn assert_strategy(name: &'static str, patterns: &[&str]) {
        assert_strategy_with(name, patterns, Config::new());
    }

    #[track_caller]
    fn assert_strategy_with(
        name: &'static str,
        patterns: &[&str],
        config: Config,
    ) {
        let strategy = strategy_with(patterns, config);
        assert_eq!(name, strategy.name().as_ref());
    }

    fn literal_alternation(count: usize) -> String {
        let mut pattern = String::new();
        for i in 0..count {
            if i > 0 {
                pattern.push('|');
            }
            pattern.push_str(&format!("needle{i:05}"));
        }
        pattern
    }

    #[test]
    fn pre_from_prefixes_accepts_exact_literal() {
        assert_strategy("prefilter memchr", &["a"]);
        assert_strategy("prefilter memchr2", &["a|b"]);
        assert_strategy("prefilter memchr3", &["a|b|c"]);
        assert_strategy("prefilter memmem", &["Sherlock"]);
        if teddy_probably_available() {
            assert_strategy(
                "prefilter teddy",
                &["Samwise|Gandalf|Holmes|Watson"],
            );
        }
    }

    #[test]
    fn pre_from_prefixes_rejects_inexact_literal() {
        assert_strategy("core", &["a+"]);
    }

    #[test]
    fn pre_from_prefixes_rejects_empty_literal() {
        assert_strategy("core", &[""]);
    }

    #[test]
    fn pre_from_prefixes_rejects_multiple_patterns() {
        assert_strategy("core", &["a", "b"]);
        assert_strategy("core", &["a|b", "c|d"]);
    }

    #[test]
    fn pre_from_prefixes_rejects_captures() {
        assert_strategy("core", &["(a)"]);
        // ... but not when it's a non-capture.
        assert_strategy("prefilter memchr", &["(?:a)"]);
        // ... or when there is a capture, but it's optimized out.
        assert_strategy("prefilter memchr", &["(){0}a"]);
    }

    #[test]
    fn pre_from_prefixes_rejects_look_around() {
        assert_strategy("core", &[r"a\b"]);
    }

    #[test]
    fn pre_from_prefixes_rejects_non_leftmost_first() {
        assert_strategy_with(
            "core",
            &["a"],
            Config::new().match_kind(MatchKind::All),
        );
    }

    #[test]
    fn pre_from_alternation_literals_accepts_large_alternation() {
        let pattern = literal_alternation(3_000);
        assert_strategy("prefilter aho-corasick", &[&pattern]);
    }

    #[test]
    fn pre_from_alternation_literals_rejects_small_alternation() {
        let pattern = literal_alternation(2_999);
        assert_strategy("core", &[&pattern]);
    }

    #[test]
    fn pre_from_alternation_literals_rejects_captures() {
        let pattern = format!("({})", literal_alternation(3_000));
        assert_strategy("core", &[&pattern]);
    }

    #[test]
    fn pre_from_alternation_literals_rejects_look_around() {
        let pattern = format!(r"(?:{})\b", literal_alternation(3_000));
        assert_strategy("core", &[&pattern]);
    }

    #[test]
    fn pre_from_alternation_literals_rejects_non_leftmost_first() {
        let pattern = literal_alternation(3_000);
        assert_strategy_with(
            "core",
            &[&pattern],
            Config::new().match_kind(MatchKind::All),
        );
    }

    #[test]
    fn pre_from_alternation_literals_rejects_multiple_patterns() {
        let pattern1 = literal_alternation(3_000);
        let pattern2 = literal_alternation(3_000);
        assert_strategy("core", &[&pattern1, &pattern2]);
    }

    #[test]
    fn core_selected_when_no_special_strategy_applies() {
        assert_strategy("core", &["[a-z]+"]);
    }

    #[test]
    fn core_selected_when_auto_prefilter_is_disabled() {
        assert_strategy_with(
            "core",
            &[r"\w+Holmes"],
            Config::new().auto_prefilter(false),
        );
    }

    #[test]
    fn core_selected_when_reverse_engines_are_disabled() {
        assert_strategy_with(
            "core",
            &[r"\w+Holmes"],
            Config::new().dfa(false).hybrid(false),
        );
    }

    #[test]
    fn reverse_anchored_accepts_end_anchored() {
        assert_strategy("reverse anchored", &[r"\w+Holmes$"]);
    }

    #[test]
    fn reverse_anchored_accepts_multiple_end_anchored_patterns() {
        assert_strategy("reverse anchored", &[r"\w+Holmes$", r"\w+Watson$"]);
    }

    #[test]
    fn reverse_anchored_accepts_hybrid_without_full_dfa() {
        assert_strategy_with(
            "reverse anchored",
            &[r"\w+Holmes$"],
            Config::new().dfa(false),
        );
    }

    #[test]
    fn reverse_anchored_rejects_start_and_end_anchored() {
        assert_strategy("core", &[r"^\w+Holmes$"]);
    }

    #[test]
    fn reverse_anchored_rejects_without_reverse_engine() {
        assert_strategy_with(
            "core",
            &[r"\w+Holmes$"],
            Config::new().dfa(false).hybrid(false),
        );
    }

    #[test]
    fn reverse_suffix_accepts_prefix_without_internal_suffix() {
        assert_strategy("reverse suffix", &[r"\d+XYZ"]);
        assert_strategy("reverse suffix", &[r"[a-q][^u-z]{13}x"]);
    }

    #[test]
    fn reverse_suffix_rejects_internal_suffix_with_word_boundary() {
        assert_strategy("core", &[r"\b\w+nn\b"]);
    }

    #[test]
    fn reverse_suffix_accepts_disjoint_class_separator() {
        assert_strategy("reverse suffix", &[r"\w+\s+Holmes"]);
        assert_strategy("reverse suffix", &[r"\w+\d*\s+Holmes"]);
        assert_strategy("reverse suffix", &[r"(?:\w+|[.-]+)+\s+Holmes"]);
        assert_strategy("reverse suffix", &[r"\b\w+\s+Holmes"]);
        assert_strategy("reverse suffix", &[r"\w+\sHolmes"]);
        assert_strategy("reverse suffix", &[r"[a-z]+[0-9]+Holmes"]);
        assert_strategy("reverse suffix", &[r"(\w+)(\s+)Holmes"]);
        assert_strategy("reverse suffix", &[r"(?-u:[A-Za-z]+[0-9]+Holmes)"]);
    }

    #[test]
    fn reverse_suffix_class_separator_is_conservative() {
        assert_strategy("core", &[r"\w+\w+Holmes"]);
        assert_strategy("core", &[r"\w+\s*Holmes"]);
        assert_strategy("core", &[r"[a-z]+[0-9a]+xyz"]);
        assert_strategy("core", &[r"[a-z]+[0-9]+a1"]);
        assert_strategy("core", &[r"(?:.abc)?[a]+[b]+c"]);
    }

    #[test]
    fn reverse_suffix_rejects_safe_nfa_overlap() {
        assert_strategy("reverse inner", &[r"(a|aa)b"]);
    }

    #[test]
    fn reverse_suffix_accepts_fixed_length_prefix() {
        assert_strategy("reverse suffix", &[r"[A-Z][0-9]XYZ"]);
    }

    #[test]
    fn reverse_suffix_rejects_multiple_patterns_with_common_suffix() {
        assert_strategy("core", &[r"\d+XYZ", r"\w+XYZ"]);
    }

    #[test]
    fn reverse_suffix_rejects_auto_prefilter_disabled() {
        assert_strategy_with(
            "core",
            &[r"\d+XYZ"],
            Config::new().auto_prefilter(false),
        );
    }

    #[test]
    fn reverse_suffix_rejects_non_leftmost_first() {
        assert_strategy_with(
            "core",
            &[r"\d+XYZ"],
            Config::new().match_kind(MatchKind::All),
        );
    }

    #[test]
    fn reverse_suffix_rejects_anchored_start() {
        assert_strategy("core", &[r"^\d+XYZ"]);
    }

    #[test]
    fn reverse_suffix_rejects_variable_length_prefix() {
        assert_strategy("core", &[r"(?:[A-Za-z]ab)?b"]);
    }

    #[test]
    fn reverse_suffix_rejects_without_reverse_engine() {
        assert_strategy_with(
            "core",
            &[r"\d+XYZ"],
            Config::new().dfa(false).hybrid(false),
        );
    }

    #[test]
    fn reverse_suffix_rejects_existing_fast_prefilter() {
        assert_strategy("core", &[r"abc\w+XYZ"]);
    }

    #[test]
    fn reverse_suffix_rejects_no_literal_suffix() {
        assert_strategy("core", &[r"[a-z]+"]);
    }

    #[test]
    fn reverse_suffix_rejects_multiple_patterns_without_common_suffix() {
        assert_strategy("core", &[r"\d+XYZ", r"\w+ABC"]);
    }

    #[test]
    fn reverse_suffix_rejects_multiple_patterns_when_first_looks_safe() {
        assert_strategy("core", &[r"\d+b", r".bb|b"]);
        assert_strategy("core", &[r"\d+b", r"ab"]);
    }

    #[test]
    fn reverse_suffix_rejects_unsafe_overlap() {
        assert_strategy("core", &[r".abb|b"]);
        assert_strategy("core", &[r".bb|b"]);
    }

    #[test]
    fn reverse_suffix_rejects_fully_absorbed_literal() {
        assert_strategy("core", &[r"(?:[a-wyz]{3}|[a-wyz]).b"]);
    }

    #[test]
    fn reverse_suffix_rejects_disjoint_trailing_component() {
        assert_strategy("core", &[r"(?:[ac-z]{2}b[ac-z])?[ac-z]b"]);
    }

    #[test]
    fn reverse_suffix_rejects_internal_candidate_before_absorber() {
        assert_strategy("core", &[r"(?:[a-z]cb|c)[a-z]*b"]);
    }

    #[test]
    fn reverse_inner_accepts_disjoint_inner_literal() {
        assert_strategy("reverse inner", &[r"\w+@\w+"]);
        assert_strategy(
            "reverse inner",
            &[r"(?P<email>[.\pL]+@(?P<domain>[.\pL]+))"],
        );
    }

    #[test]
    fn reverse_inner_accepts_fixed_length_prefix() {
        assert_strategy("reverse inner", &[r"[A-Z][0-9]@(foo|bar)"]);
    }

    #[test]
    fn reverse_inner_accepts_disjoint_class_separator() {
        assert_strategy("reverse inner", &[r"\w+\s+Holmes\s+\w+"]);
        assert_strategy("reverse inner", &[r"\w+\d*\s+Holmes\s+\w+"]);
        assert_strategy("reverse inner", &[r"\b\w+\s+Holmes\s+\w+"]);
        if teddy_probably_available() {
            assert_strategy(
                "reverse inner",
                &[r"\w+\s+(?:Holmes|Watson)\s+\w+"],
            );
        }
    }

    #[test]
    fn reverse_inner_accepts_leading_disjoint_class_separator() {
        assert_strategy("reverse inner", &[r"\s[A-Za-z]{0,12}ing\s"]);
        assert_strategy("reverse inner", &[r"\s[A-Za-z]*ing\s"]);
        assert_strategy("reverse inner", &[r"\b\s[A-Za-z]{0,12}ing\s"]);
        assert_strategy("reverse inner", &[r"(\s)([A-Za-z]{0,12})ing\s"]);
    }

    #[test]
    fn reverse_inner_leading_class_separator_is_conservative() {
        assert_strategy("core", &[r"\s[\sA-Za-z]{0,12}ing\s"]);
        assert_strategy("core", &[r"\s*[A-Za-z]{0,12}ing\s"]);
    }

    #[test]
    fn reverse_inner_rejects_auto_prefilter_disabled() {
        assert_strategy_with(
            "core",
            &[r"\w+@\w+"],
            Config::new().auto_prefilter(false),
        );
    }

    #[test]
    fn reverse_inner_rejects_non_leftmost_first() {
        assert_strategy_with(
            "core",
            &[r"\w+@\w+"],
            Config::new().match_kind(MatchKind::All),
        );
    }

    #[test]
    fn reverse_inner_rejects_anchored_start() {
        assert_strategy("core", &[r"^\w+@\w+"]);
    }

    #[test]
    fn reverse_inner_rejects_without_reverse_engine() {
        assert_strategy_with(
            "core",
            &[r"\w+@\w+"],
            Config::new().dfa(false).hybrid(false),
        );
    }

    #[test]
    fn reverse_inner_rejects_existing_fast_prefilter() {
        assert_strategy("core", &[r"abc\w+@\w+"]);
    }

    #[test]
    fn reverse_inner_rejects_multiple_patterns() {
        assert_strategy("core", &[r"\w+@\w+", r"\w+#\w+"]);
    }

    #[test]
    fn reverse_inner_rejects_no_top_level_concat() {
        assert_strategy("core", &[r"\w+"]);
    }

    #[test]
    fn reverse_inner_rejects_unsafe_overlap() {
        assert_strategy("core", &[r"a.*@\w+"]);
    }

    #[test]
    fn reverse_inner_rejects_literal_crossing_prefix_boundary() {
        assert_strategy("core", &[r"(?:[ac-z]{3}|[ac-z])(?:aba|baa)"]);
    }

    #[test]
    fn reverse_inner_rejects_internal_separator_alignment() {
        assert_strategy("core", &[r"(?:.abc)?[a]+[b]+c.*"]);
    }

    #[test]
    fn rebar_benchmarks() {
        assert_strategy("reverse suffix", &[r"[a-z]shing"]);
        assert_strategy("reverse suffix", &[r"[a-q][^u-z]{23}x"]);
    }
}