diff --git a/build.rs b/build.rs
index 51ad927673..1b715931cf 100644
--- a/build.rs
+++ b/build.rs
@@ -11,6 +11,10 @@ fn main() {
         None => return,
     };
 
+    if minor >= 19 {
+        println!("cargo:rustc-cfg=syn_can_use_thread_id");
+    }
+
     // Macro modularization allows re-exporting the `quote!` macro in 1.30+.
     if minor >= 30 {
         println!("cargo:rustc-cfg=syn_can_call_macro_by_path");
diff --git a/src/error.rs b/src/error.rs
index 712b9b7d82..32adb4455c 100644
--- a/src/error.rs
+++ b/src/error.rs
@@ -15,6 +15,7 @@ use proc_macro2::{
 };
 
 use buffer::Cursor;
+use thread::ThreadBound;
 
 /// The result of a Syn parser.
 pub type Result<T> = std::result::Result<T, Error>;
@@ -26,12 +27,20 @@ pub type Result<T> = std::result::Result<T, Error>;
 /// [module documentation]: index.html
 ///
 /// *This type is available if Syn is built with the `"parsing"` feature.*
-#[derive(Debug, Clone)]
+#[derive(Debug)]
 pub struct Error {
-    span: Span,
+    // Span is implemented as an index into a thread-local interner to keep the
+    // size small. It is not safe to access from a different thread. We want
+    // errors to be Send and Sync to play nicely with the Failure crate, so pin
+    // the span we're given to its original thread and assume it is
+    // Span::call_site if accessed from any other thread.
+    span: ThreadBound<Span>,
     message: String,
 }
 
+#[cfg(test)]
+struct _Test where Error: Send + Sync;
+
 impl Error {
     /// Usually the [`ParseStream::error`] method will be used instead, which
     /// automatically uses the correct span from the current position of the
@@ -70,13 +79,21 @@ impl Error {
     /// ```
     pub fn new<T: Display>(span: Span, message: T) -> Self {
         Error {
-            span: span,
+            span: ThreadBound::new(span),
             message: message.to_string(),
         }
     }
 
+    /// The source location of the error.
+    ///
+    /// Spans are not thread-safe so this function returns `Span::call_site()`
+    /// if called from a different thread than the one on which the `Error` was
+    /// originally created.
     pub fn span(&self) -> Span {
-        self.span
+        match self.span.get() {
+            Some(span) => *span,
+            None => Span::call_site(),
+        }
     }
 
     /// Render the error as an invocation of [`compile_error!`].
@@ -87,23 +104,25 @@ impl Error {
     /// [`compile_error!`]: https://doc.rust-lang.org/std/macro.compile_error.html
     /// [`parse_macro_input!`]: ../macro.parse_macro_input.html
     pub fn to_compile_error(&self) -> TokenStream {
+        let span = self.span();
+
         // compile_error!($message)
         TokenStream::from_iter(vec![
-            TokenTree::Ident(Ident::new("compile_error", self.span)),
+            TokenTree::Ident(Ident::new("compile_error", span)),
             TokenTree::Punct({
                 let mut punct = Punct::new('!', Spacing::Alone);
-                punct.set_span(self.span);
+                punct.set_span(span);
                 punct
             }),
             TokenTree::Group({
                 let mut group = Group::new(Delimiter::Brace, {
                     TokenStream::from_iter(vec![TokenTree::Literal({
                         let mut string = Literal::string(&self.message);
-                        string.set_span(self.span);
+                        string.set_span(span);
                         string
                     })])
                 });
-                group.set_span(self.span);
+                group.set_span(span);
                 group
             }),
         ])
@@ -124,6 +143,15 @@ impl Display for Error {
     }
 }
 
+impl Clone for Error {
+    fn clone(&self) -> Self {
+        Error {
+            span: ThreadBound::new(self.span()),
+            message: self.message.clone(),
+        }
+    }
+}
+
 impl std::error::Error for Error {
     fn description(&self) -> &str {
         "parse error"
diff --git a/src/lib.rs b/src/lib.rs
index a5348543a2..b61a2ad6b0 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -569,6 +569,9 @@ mod span;
 #[cfg(all(any(feature = "full", feature = "derive"), feature = "printing"))]
 mod print;
 
+#[cfg(feature = "parsing")]
+mod thread;
+
 ////////////////////////////////////////////////////////////////////////////////
 
 #[cfg(any(feature = "parsing", feature = "full", feature = "derive"))]
diff --git a/src/thread.rs b/src/thread.rs
new file mode 100644
index 0000000000..8a79e951b0
--- /dev/null
+++ b/src/thread.rs
@@ -0,0 +1,81 @@
+use std::fmt::{self, Debug};
+
+use self::thread_id::ThreadId;
+
+/// ThreadBound is a Sync-maker and Send-maker that allows accessing a value
+/// of type T only from the original thread on which the ThreadBound was
+/// constructed.
+pub struct ThreadBound<T> {
+    value: T,
+    thread_id: ThreadId,
+}
+
+unsafe impl<T> Sync for ThreadBound<T> {}
+
+// Send bound requires Copy, as otherwise Drop could run in the wrong place.
+unsafe impl<T: Copy> Send for ThreadBound<T> {}
+
+impl<T> ThreadBound<T> {
+    pub fn new(value: T) -> Self {
+        ThreadBound {
+            value: value,
+            thread_id: thread_id::current(),
+        }
+    }
+
+    pub fn get(&self) -> Option<&T> {
+        if thread_id::current() == self.thread_id {
+            Some(&self.value)
+        } else {
+            None
+        }
+    }
+}
+
+impl<T: Debug> Debug for ThreadBound<T> {
+    fn fmt(&self, formatter: &mut fmt::Formatter) -> fmt::Result {
+        match self.get() {
+            Some(value) => Debug::fmt(value, formatter),
+            None => formatter.write_str("unknown"),
+        }
+    }
+}
+
+#[cfg(syn_can_use_thread_id)]
+mod thread_id {
+    use std::thread;
+
+    pub use std::thread::ThreadId;
+
+    pub fn current() -> ThreadId {
+        thread::current().id()
+    }
+}
+
+#[cfg(not(syn_can_use_thread_id))]
+mod thread_id {
+    use std::sync::atomic::{AtomicUsize, Ordering, ATOMIC_USIZE_INIT};
+
+    thread_local! {
+        static THREAD_ID: usize = {
+            static NEXT_THREAD_ID: AtomicUsize = ATOMIC_USIZE_INIT;
+
+            // Ordering::Relaxed because our only requirement for the ids is
+            // that they are unique. It is okay for the compiler to rearrange
+            // other memory reads around this fetch. It's still an atomic
+            // fetch_add, so no two threads will be able to read the same value
+            // from it.
+            //
+            // The main thing which these orderings affect is other memory reads
+            // around the atomic read, which for our case are irrelevant as this
+            // atomic guards nothing.
+            NEXT_THREAD_ID.fetch_add(1, Ordering::Relaxed)
+        };
+    }
+
+    pub type ThreadId = usize;
+
+    pub fn current() -> ThreadId {
+        THREAD_ID.with(|id| *id)
+    }
+}