ffi: pass non-empty slice when haystack is empty

To work around likely bugs in (older versions of) PCRE2. Namely, at one
point, PCRE2 would dereference the haystack pointer even when the length
was zero.

This was reported in #10 and we worked around this in #11 by passing a
pointer to a const `&[]`, with the (erroneous) presumption that this
would be a valid pointer to dereference. In retrospect though, this was
a little silly, because you should never be dereferencing a pointer to
an empty slice. It's not valid. Alas, at that time, Rust did actually
hand you a valid pointer that could be dereferenced. But [this
PR][rust-pull] changed that. And thus, we're back to where we started:
handing buggy versions of PCRE2 a zero length haystack with a dangling
pointer.

So we fix this once and for all by passing a slice of length 1, but with
a haystack length of 0, to the PCRE2 search routine when searching an
empty haystack. This will guarantee the provision of a dereferencable
pointer should PCRE2 decide to dereference it.

Fixes #42

[rust-pull]: rust-lang/rust#123936

src/ffi.rs

@@ -433,20 +433,35 @@ impl MatchData {
         start: usize,
         options: u32,
     ) -> Result<bool, Error> {
-        // When the subject is empty, we use an empty slice
-        // with a known valid pointer. Otherwise, slices derived
-        // from, e.g., an empty `Vec<u8>` may not have a valid
-        // pointer, since creating an empty `Vec` is guaranteed
-        // to not allocate.
-        const EMPTY: &[u8] = &[];
+        // When the subject is empty, we use an NON-empty slice with a known
+        // valid pointer. Otherwise, slices derived from, e.g., an empty
+        // `Vec<u8>` may not have a valid pointer, since creating an empty
+        // `Vec` is guaranteed to not allocate.
+        //
+        // We use a non-empty slice since it is otherwise difficult
+        // to guarantee getting a dereferencable pointer. Which makes
+        // sense, because the slice is empty, the pointer should never be
+        // dereferenced!
+        //
+        // Alas, older versions of PCRE2 did exactly this. While that bug has
+        // been fixed a while ago, it still seems to pop up[1]. So we try
+        // harder.
+        //
+        // Note that even though we pass a non-empty slice in this case, we
+        // still pass a length of zero. This just provides a pointer that won't
+        // explode if you try to dereference it.
+        //
+        // [1]: https://github.com/BurntSushi/rust-pcre2/issues/42
+        static SINGLETON: &[u8] = &[0];
+        let len = subject.len();
         if subject.is_empty() {
-            subject = EMPTY;
+            subject = SINGLETON;
         }
 
         let rc = pcre2_match_8(
             code.as_ptr(),
             subject.as_ptr(),
-            subject.len(),
+            len,
             start,
             options,
             self.as_mut_ptr(),