automata: rejigger DFA start state computation

It turns out that requiring callers to provide an `Input` (and thus a
`&[u8]` haystack) is a bit onerous for all cases. Namely, part of the
point of `regex-automata` was to expose enough guts to make it tractable
to write a streaming regex engine. A streaming regex engine, especially
one that does a byte-at-a-time loop, is somewhat antithetical to having
a haystack in a single `&[u8]` slice. This made computing start states
possible but very awkward and quite unclear in terms of what the
implementation would actually do with the haystack.

This commit fixes that by exposing a lower level `start_state` method on
both of the DFAs that can be called without materializing an `Input`.
Instead, callers must create a new `start::Config` value which provides
all of the information necessary for the DFA to compute the correct
start state. This in turn also exposes the `crate::util::start` module.

This is ultimately a breaking change because it adds a new required
method to the `Automaton` trait. It also makes `start_state_forward` and
`start_state_reverse` optional. It isn't really expected for callers to
implement the `Automaton` trait themselves (and perhaps I will seal it
so we can do such changes in the future without it being breaking), but
still, this is technically breaking.

Callers using `start_state_forward` or `start_state_reverse` with either
DFA remain unchanged and unaffected.

Closes #1031

CHANGELOG.md

@@ -1,6 +1,13 @@
 TBD
 ===
 
+New features:
+
+* [FEATURE(regex-automata) #1031](https://github.com/rust-lang/regex/pull/1031):
+DFAs now have a `start_state` method that doesn't use an `Input`.
+
+Bug fixes:
+
 * [BUG #1046](https://github.com/rust-lang/regex/issues/1046):
 Fix a bug that could result in incorrect match spans when using a Unicode word
 boundary and searching non-ASCII strings.

regex-automata/src/dfa/automaton.rs

@@ -7,6 +7,7 @@ use crate::{
         prefilter::Prefilter,
         primitives::{PatternID, StateID},
         search::{Anchored, HalfMatch, Input, MatchError},
+        start,
     },
 };
 
@@ -226,21 +227,50 @@ pub unsafe trait Automaton {
     /// ```
     fn next_eoi_state(&self, current: StateID) -> StateID;
 
-    /// Return the ID of the start state for this lazy DFA when executing a
-    /// forward search.
+    /// Return the ID of the start state for this DFA for the given starting
+    /// configuration.
     ///
     /// Unlike typical DFA implementations, the start state for DFAs in this
     /// crate is dependent on a few different factors:
     ///
     /// * The [`Anchored`] mode of the search. Unanchored, anchored and
     /// anchored searches for a specific [`PatternID`] all use different start
     /// states.
-    /// * The position at which the search begins, via [`Input::start`]. This
-    /// and the byte immediately preceding the start of the search (if one
-    /// exists) influence which look-behind assertions are true at the start
-    /// of the search. This in turn influences which start state is selected.
-    /// * Whether the search is a forward or reverse search. This routine can
-    /// only be used for forward searches.
+    /// * Whether a "look-behind" byte exists. For example, the `^` anchor
+    /// matches if and only if there is no look-behind byte.
+    /// * The specific value of that look-behind byte. For example, a `(?m:^)`
+    /// assertion only matches when there is either no look-behind byte, or
+    /// when the look-behind byte is a line terminator.
+    ///
+    /// The [starting configuration](start::Config) provides the above
+    /// information.
+    ///
+    /// This routine can be used for either forward or reverse searches.
+    /// Although, as a convenience, if you have an [`Input`], then it may
+    /// be more succinct to use [`Automaton::start_state_forward`] or
+    /// [`Automaton::start_state_reverse`]. Note, for example, that the
+    /// convenience routines return a [`MatchError`] on failure where as this
+    /// routine returns a [`StartError`].
+    ///
+    /// # Errors
+    ///
+    /// This may return a [`StartError`] if the search needs to give up when
+    /// determining the start state (for example, if it sees a "quit" byte).
+    /// This can also return an error if the given configuration contains an
+    /// unsupported [`Anchored`] configuration.
+    fn start_state(
+        &self,
+        config: &start::Config,
+    ) -> Result<StateID, StartError>;
+
+    /// Return the ID of the start state for this DFA when executing a forward
+    /// search.
+    ///
+    /// This is a convenience routine for calling [`Automaton::start_state`]
+    /// that converts the given [`Input`] to a [start
+    /// configuration](start::Config). Additionally, if an error occurs, it is
+    /// converted from a [`StartError`] to a [`MatchError`] using the offset
+    /// information in the given [`Input`].
     ///
     /// # Errors
     ///
@@ -251,23 +281,30 @@ pub unsafe trait Automaton {
     fn start_state_forward(
         &self,
         input: &Input<'_>,
-    ) -> Result<StateID, MatchError>;
+    ) -> Result<StateID, MatchError> {
+        let config = start::Config::from_input_forward(input);
+        self.start_state(&config).map_err(|err| match err {
+            StartError::Quit { byte } => {
+                let offset = input
+                    .start()
+                    .checked_sub(1)
+                    .expect("no quit in start without look-behind");
+                MatchError::quit(byte, offset)
+            }
+            StartError::UnsupportedAnchored { mode } => {
+                MatchError::unsupported_anchored(mode)
+            }
+        })
+    }
 
-    /// Return the ID of the start state for this lazy DFA when executing a
-    /// reverse search.
+    /// Return the ID of the start state for this DFA when executing a reverse
+    /// search.
     ///
-    /// Unlike typical DFA implementations, the start state for DFAs in this
-    /// crate is dependent on a few different factors:
-    ///
-    /// * The [`Anchored`] mode of the search. Unanchored, anchored and
-    /// anchored searches for a specific [`PatternID`] all use different start
-    /// states.
-    /// * The position at which the search begins, via [`Input::start`]. This
-    /// and the byte immediately preceding the start of the search (if one
-    /// exists) influence which look-behind assertions are true at the start
-    /// of the search. This in turn influences which start state is selected.
-    /// * Whether the search is a forward or reverse search. This routine can
-    /// only be used for reverse searches.
+    /// This is a convenience routine for calling [`Automaton::start_state`]
+    /// that converts the given [`Input`] to a [start
+    /// configuration](start::Config). Additionally, if an error occurs, it is
+    /// converted from a [`StartError`] to a [`MatchError`] using the offset
+    /// information in the given [`Input`].
     ///
     /// # Errors
     ///
@@ -278,7 +315,18 @@ pub unsafe trait Automaton {
     fn start_state_reverse(
         &self,
         input: &Input<'_>,
-    ) -> Result<StateID, MatchError>;
+    ) -> Result<StateID, MatchError> {
+        let config = start::Config::from_input_reverse(input);
+        self.start_state(&config).map_err(|err| match err {
+            StartError::Quit { byte } => {
+                let offset = input.end();
+                MatchError::quit(byte, offset)
+            }
+            StartError::UnsupportedAnchored { mode } => {
+                MatchError::unsupported_anchored(mode)
+            }
+        })
+    }
 
     /// If this DFA has a universal starting state for the given anchor mode
     /// and the DFA supports universal starting states, then this returns that
@@ -1798,6 +1846,14 @@ unsafe impl<'a, A: Automaton + ?Sized> Automaton for &'a A {
         (**self).next_eoi_state(current)
     }
 
+    #[inline]
+    fn start_state(
+        &self,
+        config: &start::Config,
+    ) -> Result<StateID, StartError> {
+        (**self).start_state(config)
+    }
+
     #[inline]
     fn start_state_forward(
         &self,
@@ -2015,6 +2071,90 @@ impl OverlappingState {
     }
 }
 
+/// An error that can occur when computing the start state for a search.
+///
+/// Computing a start state can fail for a few reasons, either based on
+/// incorrect configuration or even based on whether the look-behind byte
+/// triggers a quit state. Typically one does not need to handle this error
+/// if you're using [`Automaton::start_state_forward`] (or its reverse
+/// counterpart), as that routine automatically converts `StartError` to a
+/// [`MatchError`] for you.
+///
+/// This error may be returned by the [`Automaton::start_state`] routine.
+///
+/// This error implements the `std::error::Error` trait when the `std` feature
+/// is enabled.
+///
+/// This error is marked as non-exhaustive. New variants may be added in a
+/// semver compatible release.
+#[non_exhaustive]
+#[derive(Clone, Debug)]
+pub enum StartError {
+    /// An error that occurs when a starting configuration's look-behind byte
+    /// is in this DFA's quit set.
+    Quit {
+        /// The quit byte that was found.
+        byte: u8,
+    },
+    /// An error that occurs when the caller requests an anchored mode that
+    /// isn't supported by the DFA.
+    UnsupportedAnchored {
+        /// The anchored mode given that is unsupported.
+        mode: Anchored,
+    },
+}
+
+impl StartError {
+    pub(crate) fn quit(byte: u8) -> StartError {
+        StartError::Quit { byte }
+    }
+
+    pub(crate) fn unsupported_anchored(mode: Anchored) -> StartError {
+        StartError::UnsupportedAnchored { mode }
+    }
+}
+
+#[cfg(feature = "std")]
+impl std::error::Error for StartError {}
+
+impl core::fmt::Display for StartError {
+    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
+        match *self {
+            StartError::Quit { byte } => write!(
+                f,
+                "error computing start state because the look-behind byte \
+                 {:?} triggered a quit state",
+                crate::util::escape::DebugByte(byte),
+            ),
+            StartError::UnsupportedAnchored { mode: Anchored::Yes } => {
+                write!(
+                    f,
+                    "error computing start state because \
+                     anchored searches are not supported or enabled"
+                )
+            }
+            StartError::UnsupportedAnchored { mode: Anchored::No } => {
+                write!(
+                    f,
+                    "error computing start state because \
+                     unanchored searches are not supported or enabled"
+                )
+            }
+            StartError::UnsupportedAnchored {
+                mode: Anchored::Pattern(pid),
+            } => {
+                write!(
+                    f,
+                    "error computing start state because \
+                     anchored searches for a specific pattern ({}) \
+                     are not supported or enabled",
+                    pid.as_usize(),
+                )
+            }
+        }
+    }
+}
+
 /// Runs the given overlapping `search` function (forwards or backwards) until
 /// a match is found whose offset does not split a codepoint.
 ///