diff --git a/tests/collection.rs b/tests/collection.rs
new file mode 100644
index 0000000..68b0322
--- /dev/null
+++ b/tests/collection.rs
@@ -0,0 +1,461 @@
+use std::collections::BTreeMap;
+use std::env;
+use std::fmt::{self, Write};
+use std::thread;
+
+use regex;
+use regex_automata::{DenseDFA, ErrorKind, Regex, RegexBuilder, StateID, DFA};
+use serde_bytes;
+use toml;
+
+macro_rules! load {
+    ($col:ident, $path:expr) => {
+        $col.extend(RegexTests::load(
+            concat!("../data/tests/", $path),
+            include_bytes!(concat!("../data/tests/", $path)),
+        ));
+    };
+}
+
+lazy_static! {
+    pub static ref SUITE: RegexTestCollection = {
+        let mut col = RegexTestCollection::new();
+        load!(col, "fowler/basic.toml");
+        load!(col, "fowler/nullsubexpr.toml");
+        load!(col, "fowler/repetition.toml");
+        load!(col, "fowler/repetition-long.toml");
+        load!(col, "crazy.toml");
+        load!(col, "flags.toml");
+        load!(col, "iter.toml");
+        load!(col, "no-unicode.toml");
+        load!(col, "unicode.toml");
+        col
+    };
+}
+
+#[derive(Clone, Debug)]
+pub struct RegexTestCollection {
+    pub by_name: BTreeMap<String, RegexTest>,
+}
+
+#[derive(Clone, Debug, Deserialize)]
+pub struct RegexTests {
+    pub tests: Vec<RegexTest>,
+}
+
+#[derive(Clone, Debug, Deserialize)]
+pub struct RegexTest {
+    pub name: String,
+    #[serde(default)]
+    pub options: Vec<RegexTestOption>,
+    pub pattern: String,
+    #[serde(with = "serde_bytes")]
+    pub input: Vec<u8>,
+    #[serde(rename = "matches")]
+    pub matches: Vec<Match>,
+    #[serde(default)]
+    pub captures: Vec<Option<Match>>,
+    #[serde(default)]
+    pub fowler_line_number: Option<u64>,
+}
+
+#[derive(Clone, Copy, Debug, Deserialize, Eq, PartialEq)]
+#[serde(rename_all = "kebab-case")]
+pub enum RegexTestOption {
+    Anchored,
+    CaseInsensitive,
+    NoUnicode,
+    Escaped,
+    #[serde(rename = "invalid-utf8")]
+    InvalidUTF8,
+}
+
+#[derive(Clone, Copy, Deserialize, Eq, PartialEq)]
+pub struct Match {
+    pub start: usize,
+    pub end: usize,
+}
+
+impl RegexTestCollection {
+    fn new() -> RegexTestCollection {
+        RegexTestCollection { by_name: BTreeMap::new() }
+    }
+
+    fn extend(&mut self, tests: RegexTests) {
+        for test in tests.tests {
+            let name = test.name.clone();
+            if self.by_name.contains_key(&name) {
+                panic!("found duplicate test {}", name);
+            }
+            self.by_name.insert(name, test);
+        }
+    }
+
+    pub fn tests(&self) -> Vec<&RegexTest> {
+        self.by_name.values().collect()
+    }
+}
+
+impl RegexTests {
+    fn load(path: &str, slice: &[u8]) -> RegexTests {
+        let mut data: RegexTests = toml::from_slice(slice)
+            .expect(&format!("failed to load {}", path));
+        for test in &mut data.tests {
+            if test.options.contains(&RegexTestOption::Escaped) {
+                test.input = unescape_bytes(&test.input);
+            }
+        }
+        data
+    }
+}
+
+#[derive(Debug)]
+pub struct RegexTester {
+    asserted: bool,
+    results: RegexTestResults,
+    skip_expensive: bool,
+    whitelist: Vec<regex::Regex>,
+    blacklist: Vec<regex::Regex>,
+}
+
+impl Drop for RegexTester {
+    fn drop(&mut self) {
+        // If we haven't asserted yet, then the test is probably buggy, so
+        // fail it. But if we're already panicking (e.g., a bug in the regex
+        // engine), then don't double-panic, which causes an immediate abort.
+        if !thread::panicking() && !self.asserted {
+            panic!("must call RegexTester::assert at end of test");
+        }
+    }
+}
+
+impl RegexTester {
+    pub fn new() -> RegexTester {
+        let mut tester = RegexTester {
+            asserted: false,
+            results: RegexTestResults::default(),
+            skip_expensive: false,
+            whitelist: vec![],
+            blacklist: vec![],
+        };
+        for x in env::var("REGEX_TEST").unwrap_or("".to_string()).split(",") {
+            let x = x.trim();
+            if x.is_empty() {
+                continue;
+            }
+            if x.starts_with("-") {
+                tester = tester.blacklist(&x[1..]);
+            } else {
+                tester = tester.whitelist(x);
+            }
+        }
+        tester
+    }
+
+    pub fn skip_expensive(mut self) -> RegexTester {
+        self.skip_expensive = true;
+        self
+    }
+
+    pub fn whitelist(mut self, name: &str) -> RegexTester {
+        self.whitelist.push(regex::Regex::new(name).unwrap());
+        self
+    }
+
+    pub fn blacklist(mut self, name: &str) -> RegexTester {
+        self.blacklist.push(regex::Regex::new(name).unwrap());
+        self
+    }
+
+    pub fn assert(&mut self) {
+        self.asserted = true;
+        self.results.assert();
+    }
+
+    pub fn build_regex<S: StateID>(
+        &self,
+        mut builder: RegexBuilder,
+        test: &RegexTest,
+    ) -> Option<Regex<DenseDFA<Vec<S>, S>>> {
+        if self.skip(test) {
+            return None;
+        }
+        self.apply_options(test, &mut builder);
+
+        match builder.build_with_size::<S>(&test.pattern) {
+            Ok(re) => Some(re),
+            Err(err) => {
+                if let ErrorKind::Unsupported(_) = *err.kind() {
+                    None
+                } else {
+                    panic!(
+                        "failed to build {:?} with pattern '{:?}': {}",
+                        test.name, test.pattern, err
+                    );
+                }
+            }
+        }
+    }
+
+    pub fn test_all<'a, I, T>(&mut self, builder: RegexBuilder, tests: I)
+    where
+        I: IntoIterator<IntoIter = T, Item = &'a RegexTest>,
+        T: Iterator<Item = &'a RegexTest>,
+    {
+        for test in tests {
+            let builder = builder.clone();
+            let re: Regex = match self.build_regex(builder, test) {
+                None => continue,
+                Some(re) => re,
+            };
+            self.test(test, &re);
+        }
+    }
+
+    pub fn test<'a, D: DFA>(&mut self, test: &RegexTest, re: &Regex<D>) {
+        self.test_is_match(test, re);
+        self.test_find(test, re);
+        // Some tests (namely, fowler) are designed only to detect the
+        // first match even if there are more subsequent matches. To that
+        // end, we only test match iteration when the number of matches
+        // expected is not 1, or if the test name has 'iter' in it.
+        if test.name.contains("iter") || test.matches.len() != 1 {
+            self.test_find_iter(test, re);
+        }
+    }
+
+    pub fn test_is_match<'a, D: DFA>(
+        &mut self,
+        test: &RegexTest,
+        re: &Regex<D>,
+    ) {
+        self.asserted = false;
+
+        let got = re.is_match(&test.input);
+        let expected = test.matches.len() >= 1;
+        if got == expected {
+            self.results.succeeded.push(test.clone());
+            return;
+        }
+        self.results.failed.push(RegexTestFailure {
+            test: test.clone(),
+            kind: RegexTestFailureKind::IsMatch,
+        });
+    }
+
+    pub fn test_find<'a, D: DFA>(&mut self, test: &RegexTest, re: &Regex<D>) {
+        self.asserted = false;
+
+        let got =
+            re.find(&test.input).map(|(start, end)| Match { start, end });
+        if got == test.matches.get(0).map(|&m| m) {
+            self.results.succeeded.push(test.clone());
+            return;
+        }
+        self.results.failed.push(RegexTestFailure {
+            test: test.clone(),
+            kind: RegexTestFailureKind::Find { got },
+        });
+    }
+
+    pub fn test_find_iter<'a, D: DFA>(
+        &mut self,
+        test: &RegexTest,
+        re: &Regex<D>,
+    ) {
+        self.asserted = false;
+
+        let got: Vec<Match> = re
+            .find_iter(&test.input)
+            .map(|(start, end)| Match { start, end })
+            .collect();
+        if got == test.matches {
+            self.results.succeeded.push(test.clone());
+            return;
+        }
+        self.results.failed.push(RegexTestFailure {
+            test: test.clone(),
+            kind: RegexTestFailureKind::FindIter { got },
+        });
+    }
+
+    fn skip(&self, test: &RegexTest) -> bool {
+        if self.skip_expensive {
+            if test.name.starts_with("repetition-long") {
+                return true;
+            }
+        }
+        if !self.blacklist.is_empty() {
+            if self.blacklist.iter().any(|re| re.is_match(&test.name)) {
+                return true;
+            }
+        }
+        if !self.whitelist.is_empty() {
+            if !self.whitelist.iter().any(|re| re.is_match(&test.name)) {
+                return true;
+            }
+        }
+        false
+    }
+
+    fn apply_options(&self, test: &RegexTest, builder: &mut RegexBuilder) {
+        for opt in &test.options {
+            match *opt {
+                RegexTestOption::Anchored => {
+                    builder.anchored(true);
+                }
+                RegexTestOption::CaseInsensitive => {
+                    builder.case_insensitive(true);
+                }
+                RegexTestOption::NoUnicode => {
+                    builder.unicode(false);
+                }
+                RegexTestOption::Escaped => {}
+                RegexTestOption::InvalidUTF8 => {
+                    builder.allow_invalid_utf8(true);
+                }
+            }
+        }
+    }
+}
+
+#[derive(Clone, Debug, Default)]
+pub struct RegexTestResults {
+    /// Tests that succeeded.
+    pub succeeded: Vec<RegexTest>,
+    /// Failed tests, indexed by group name.
+    pub failed: Vec<RegexTestFailure>,
+}
+
+#[derive(Clone, Debug)]
+pub struct RegexTestFailure {
+    test: RegexTest,
+    kind: RegexTestFailureKind,
+}
+
+#[derive(Clone, Debug)]
+pub enum RegexTestFailureKind {
+    IsMatch,
+    Find { got: Option<Match> },
+    FindIter { got: Vec<Match> },
+}
+
+impl RegexTestResults {
+    pub fn assert(&self) {
+        if self.failed.is_empty() {
+            return;
+        }
+        let failures = self
+            .failed
+            .iter()
+            .map(|f| f.to_string())
+            .collect::<Vec<String>>()
+            .join("\n\n");
+        panic!(
+            "found {} failures:\n{}\n{}\n{}\n\n\
+             Set the REGEX_TEST environment variable to filter tests, \n\
+             e.g., REGEX_TEST=crazy-misc,-crazy-misc2 runs every test \n\
+             whose name contains crazy-misc but not crazy-misc2\n\n",
+            self.failed.len(),
+            "~".repeat(79),
+            failures.trim(),
+            "~".repeat(79)
+        )
+    }
+}
+
+impl fmt::Display for RegexTestFailure {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(
+            f,
+            "{}: {}\n    \
+             options: {:?}\n    \
+             pattern: {}\n    \
+             pattern (escape): {}\n    \
+             input: {}\n    \
+             input (escape): {}\n    \
+             input (hex): {}",
+            self.test.name,
+            self.kind.fmt(&self.test)?,
+            self.test.options,
+            self.test.pattern,
+            escape_default(&self.test.pattern),
+            nice_raw_bytes(&self.test.input),
+            escape_bytes(&self.test.input),
+            hex_bytes(&self.test.input)
+        )
+    }
+}
+
+impl RegexTestFailureKind {
+    fn fmt(&self, test: &RegexTest) -> Result<String, fmt::Error> {
+        let mut buf = String::new();
+        match *self {
+            RegexTestFailureKind::IsMatch => {
+                if let Some(&m) = test.matches.get(0) {
+                    write!(buf, "expected match (at {}), but none found", m)?
+                } else {
+                    write!(buf, "expected no match, but found a match")?
+                }
+            }
+            RegexTestFailureKind::Find { got } => write!(
+                buf,
+                "expected {:?}, but found {:?}",
+                test.matches.get(0),
+                got
+            )?,
+            RegexTestFailureKind::FindIter { ref got } => write!(
+                buf,
+                "expected {:?}, but found {:?}",
+                test.matches, got
+            )?,
+        }
+        Ok(buf)
+    }
+}
+
+impl fmt::Display for Match {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "({}, {})", self.start, self.end)
+    }
+}
+
+impl fmt::Debug for Match {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "({}, {})", self.start, self.end)
+    }
+}
+
+fn nice_raw_bytes(bytes: &[u8]) -> String {
+    use std::str;
+
+    match str::from_utf8(bytes) {
+        Ok(s) => s.to_string(),
+        Err(_) => escape_bytes(bytes),
+    }
+}
+
+fn escape_bytes(bytes: &[u8]) -> String {
+    use std::ascii;
+
+    let escaped = bytes
+        .iter()
+        .flat_map(|&b| ascii::escape_default(b))
+        .collect::<Vec<u8>>();
+    String::from_utf8(escaped).unwrap()
+}
+
+fn hex_bytes(bytes: &[u8]) -> String {
+    bytes.iter().map(|&b| format!(r"\x{:02X}", b)).collect()
+}
+
+fn escape_default(s: &str) -> String {
+    s.chars().flat_map(|c| c.escape_default()).collect()
+}
+
+fn unescape_bytes(bytes: &[u8]) -> Vec<u8> {
+    use std::str;
+    use unescape::unescape;
+
+    unescape(&str::from_utf8(bytes).expect("all input must be valid UTF-8"))
+}
diff --git a/tests/regression.rs b/tests/regression.rs
new file mode 100644
index 0000000..c2d2c12
--- /dev/null
+++ b/tests/regression.rs
@@ -0,0 +1,42 @@
+use regex_automata::{dense, DFA};
+
+// A regression test for checking that minimization correctly translates
+// whether a state is a match state or not. Previously, it was possible for
+// minimization to mark a non-matching state as matching.
+#[test]
+fn minimize_sets_correct_match_states() {
+    let pattern =
+        // This is a subset of the grapheme matching regex. I couldn't seem
+        // to get a repro any smaller than this unfortunately.
+        r"(?x)
+            (?:
+                \p{gcb=Prepend}*
+                (?:
+                    (?:
+                        (?:
+                            \p{gcb=L}*
+                            (?:\p{gcb=V}+|\p{gcb=LV}\p{gcb=V}*|\p{gcb=LVT})
+                            \p{gcb=T}*
+                        )
+                        |
+                        \p{gcb=L}+
+                        |
+                        \p{gcb=T}+
+                    )
+                    |
+                    \p{Extended_Pictographic}
+                    (?:\p{gcb=Extend}*\p{gcb=ZWJ}\p{Extended_Pictographic})*
+                    |
+                    [^\p{gcb=Control}\p{gcb=CR}\p{gcb=LF}]
+                )
+                [\p{gcb=Extend}\p{gcb=ZWJ}\p{gcb=SpacingMark}]*
+            )
+        ";
+
+    let dfa = dense::Builder::new()
+        .minimize(true)
+        .anchored(true)
+        .build(pattern)
+        .unwrap();
+    assert_eq!(None, dfa.find(b"\xE2"));
+}
diff --git a/tests/suite.rs b/tests/suite.rs
new file mode 100644
index 0000000..8397194
--- /dev/null
+++ b/tests/suite.rs
@@ -0,0 +1,250 @@
+use regex_automata::{DenseDFA, Regex, RegexBuilder, SparseDFA};
+
+use collection::{RegexTester, SUITE};
+
+#[test]
+fn unminimized_standard() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(false).premultiply(false).byte_classes(false);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+#[test]
+fn unminimized_premultiply() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(false).premultiply(true).byte_classes(false);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+#[test]
+fn unminimized_byte_class() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(false).premultiply(false).byte_classes(true);
+
+    let mut tester = RegexTester::new();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+#[test]
+fn unminimized_premultiply_byte_class() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(false).premultiply(true).byte_classes(true);
+
+    let mut tester = RegexTester::new();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+#[test]
+fn unminimized_standard_no_nfa_shrink() {
+    let mut builder = RegexBuilder::new();
+    builder
+        .minimize(false)
+        .premultiply(false)
+        .byte_classes(false)
+        .shrink(false);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+#[test]
+fn minimized_standard() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(true).premultiply(false).byte_classes(false);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+#[test]
+fn minimized_premultiply() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(true).premultiply(true).byte_classes(false);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+#[test]
+fn minimized_byte_class() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(true).premultiply(false).byte_classes(true);
+
+    let mut tester = RegexTester::new();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+#[test]
+fn minimized_premultiply_byte_class() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(true).premultiply(true).byte_classes(true);
+
+    let mut tester = RegexTester::new();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+#[test]
+fn minimized_standard_no_nfa_shrink() {
+    let mut builder = RegexBuilder::new();
+    builder
+        .minimize(true)
+        .premultiply(false)
+        .byte_classes(false)
+        .shrink(false);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    tester.test_all(builder, SUITE.tests());
+    tester.assert();
+}
+
+// A basic sanity test that checks we can convert a regex to a smaller
+// representation and that the resulting regex still passes our tests.
+//
+// If tests grow minimal regexes that cannot be represented in 16 bits, then
+// we'll either want to skip those or increase the size to test to u32.
+#[test]
+fn u16() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(true).premultiply(false).byte_classes(true);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    for test in SUITE.tests() {
+        let builder = builder.clone();
+        let re: Regex = match tester.build_regex(builder, test) {
+            None => continue,
+            Some(re) => re,
+        };
+        let small_re = Regex::from_dfas(
+            re.forward().to_u16().unwrap(),
+            re.reverse().to_u16().unwrap(),
+        );
+
+        tester.test(test, &small_re);
+    }
+    tester.assert();
+}
+
+// Test that sparse DFAs work using the standard configuration.
+#[test]
+fn sparse_unminimized_standard() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(false).premultiply(false).byte_classes(false);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    for test in SUITE.tests() {
+        let builder = builder.clone();
+        let re: Regex = match tester.build_regex(builder, test) {
+            None => continue,
+            Some(re) => re,
+        };
+        let fwd = re.forward().to_sparse().unwrap();
+        let rev = re.reverse().to_sparse().unwrap();
+        let sparse_re = Regex::from_dfas(fwd, rev);
+
+        tester.test(test, &sparse_re);
+    }
+    tester.assert();
+}
+
+// Test that sparse DFAs work after converting them to a different state ID
+// representation.
+#[test]
+fn sparse_u16() {
+    let mut builder = RegexBuilder::new();
+    builder.minimize(true).premultiply(false).byte_classes(false);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    for test in SUITE.tests() {
+        let builder = builder.clone();
+        let re: Regex = match tester.build_regex(builder, test) {
+            None => continue,
+            Some(re) => re,
+        };
+        let fwd = re.forward().to_sparse().unwrap().to_u16().unwrap();
+        let rev = re.reverse().to_sparse().unwrap().to_u16().unwrap();
+        let sparse_re = Regex::from_dfas(fwd, rev);
+
+        tester.test(test, &sparse_re);
+    }
+    tester.assert();
+}
+
+// Another basic sanity test that checks we can serialize and then deserialize
+// a regex, and that the resulting regex can be used for searching correctly.
+#[test]
+fn serialization_roundtrip() {
+    let mut builder = RegexBuilder::new();
+    builder.premultiply(false).byte_classes(true);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    for test in SUITE.tests() {
+        let builder = builder.clone();
+        let re: Regex = match tester.build_regex(builder, test) {
+            None => continue,
+            Some(re) => re,
+        };
+
+        let fwd_bytes = re.forward().to_bytes_native_endian().unwrap();
+        let rev_bytes = re.reverse().to_bytes_native_endian().unwrap();
+        let fwd: DenseDFA<&[usize], usize> =
+            unsafe { DenseDFA::from_bytes(&fwd_bytes) };
+        let rev: DenseDFA<&[usize], usize> =
+            unsafe { DenseDFA::from_bytes(&rev_bytes) };
+        let re = Regex::from_dfas(fwd, rev);
+
+        tester.test(test, &re);
+    }
+    tester.assert();
+}
+
+// A basic sanity test that checks we can serialize and then deserialize a
+// regex using sparse DFAs, and that the resulting regex can be used for
+// searching correctly.
+#[test]
+fn sparse_serialization_roundtrip() {
+    let mut builder = RegexBuilder::new();
+    builder.byte_classes(true);
+
+    let mut tester = RegexTester::new().skip_expensive();
+    for test in SUITE.tests() {
+        let builder = builder.clone();
+        let re: Regex = match tester.build_regex(builder, test) {
+            None => continue,
+            Some(re) => re,
+        };
+
+        let fwd_bytes = re
+            .forward()
+            .to_sparse()
+            .unwrap()
+            .to_bytes_native_endian()
+            .unwrap();
+        let rev_bytes = re
+            .reverse()
+            .to_sparse()
+            .unwrap()
+            .to_bytes_native_endian()
+            .unwrap();
+        let fwd: SparseDFA<&[u8], usize> =
+            unsafe { SparseDFA::from_bytes(&fwd_bytes) };
+        let rev: SparseDFA<&[u8], usize> =
+            unsafe { SparseDFA::from_bytes(&rev_bytes) };
+        let re = Regex::from_dfas(fwd, rev);
+
+        tester.test(test, &re);
+    }
+    tester.assert();
+}
diff --git a/tests/tests.rs b/tests/tests.rs
new file mode 100644
index 0000000..fb4cd77
--- /dev/null
+++ b/tests/tests.rs
@@ -0,0 +1,25 @@
+#[cfg(feature = "std")]
+#[macro_use]
+extern crate lazy_static;
+#[cfg(feature = "std")]
+extern crate regex;
+#[cfg(feature = "std")]
+extern crate regex_automata;
+#[cfg(feature = "std")]
+extern crate serde;
+#[cfg(feature = "std")]
+extern crate serde_bytes;
+#[cfg(feature = "std")]
+#[macro_use]
+extern crate serde_derive;
+#[cfg(feature = "std")]
+extern crate toml;
+
+#[cfg(feature = "std")]
+mod collection;
+#[cfg(feature = "std")]
+mod regression;
+#[cfg(feature = "std")]
+mod suite;
+#[cfg(feature = "std")]
+mod unescape;
diff --git a/tests/unescape.rs b/tests/unescape.rs
new file mode 100644
index 0000000..43fe04e
--- /dev/null
+++ b/tests/unescape.rs
@@ -0,0 +1,84 @@
+#[derive(Clone, Copy, Eq, PartialEq)]
+enum State {
+    /// The state after seeing a `\`.
+    Escape,
+    /// The state after seeing a `\x`.
+    HexFirst,
+    /// The state after seeing a `\x[0-9A-Fa-f]`.
+    HexSecond(char),
+    /// Default state.
+    Literal,
+}
+
+pub fn unescape(s: &str) -> Vec<u8> {
+    use self::State::*;
+
+    let mut bytes = vec![];
+    let mut state = Literal;
+    for c in s.chars() {
+        match state {
+            Escape => match c {
+                '\\' => {
+                    bytes.push(b'\\');
+                    state = Literal;
+                }
+                'n' => {
+                    bytes.push(b'\n');
+                    state = Literal;
+                }
+                'r' => {
+                    bytes.push(b'\r');
+                    state = Literal;
+                }
+                't' => {
+                    bytes.push(b'\t');
+                    state = Literal;
+                }
+                'x' => {
+                    state = HexFirst;
+                }
+                c => {
+                    bytes.extend(format!(r"\{}", c).into_bytes());
+                    state = Literal;
+                }
+            },
+            HexFirst => match c {
+                '0'..='9' | 'A'..='F' | 'a'..='f' => {
+                    state = HexSecond(c);
+                }
+                c => {
+                    bytes.extend(format!(r"\x{}", c).into_bytes());
+                    state = Literal;
+                }
+            },
+            HexSecond(first) => match c {
+                '0'..='9' | 'A'..='F' | 'a'..='f' => {
+                    let ordinal = format!("{}{}", first, c);
+                    let byte = u8::from_str_radix(&ordinal, 16).unwrap();
+                    bytes.push(byte);
+                    state = Literal;
+                }
+                c => {
+                    let original = format!(r"\x{}{}", first, c);
+                    bytes.extend(original.into_bytes());
+                    state = Literal;
+                }
+            },
+            Literal => match c {
+                '\\' => {
+                    state = Escape;
+                }
+                c => {
+                    bytes.extend(c.to_string().as_bytes());
+                }
+            },
+        }
+    }
+    match state {
+        Escape => bytes.push(b'\\'),
+        HexFirst => bytes.extend(b"\\x"),
+        HexSecond(c) => bytes.extend(format!("\\x{}", c).into_bytes()),
+        Literal => {}
+    }
+    bytes
+}