Fix a bug in memrnchr

The bytes check was incorrect, leading to bugs in some cases.

Author: glebm

src/memrnchr.rs

@@ -3,13 +3,6 @@
 // We use this mainly to skip repeated `/`. If there is only one slash, `memrnchr` performs the same
 // as a naive version (e.g. `rposition`). However, it is much faster in pathological cases.
 
-const LO_U64: u64 = 0x0101010101010101;
-const HI_U64: u64 = 0x8080808080808080;
-
-// use truncation
-const LO_USIZE: usize = LO_U64 as usize;
-const HI_USIZE: usize = HI_U64 as usize;
-
 #[cfg(target_pointer_width = "32")]
 const USIZE_BYTES: usize = 4;
 #[cfg(target_pointer_width = "64")]
@@ -49,10 +42,8 @@ pub fn memrnchr(x: u8, text: &[u8]) -> Option<usize> {
       let u = *(ptr.offset(offset as isize - 2 * USIZE_BYTES as isize) as *const usize);
       let v = *(ptr.offset(offset as isize - USIZE_BYTES as isize) as *const usize);
 
-      // break if there is a matching byte
-      let zu = contains_zero_byte(u ^ repeated_x);
-      let zv = contains_zero_byte(v ^ repeated_x);
-      if !zu || !zv {
+      // break if there is no matching byte
+      if u & repeated_x != usize::max_value() || v & repeated_x != usize::max_value() {
         break;
       }
     }
@@ -63,18 +54,6 @@ pub fn memrnchr(x: u8, text: &[u8]) -> Option<usize> {
   text[..offset].iter().rposition(|elt| *elt != x)
 }
 
-/// Return `true` if `x` contains any zero byte.
-///
-/// From *Matters Computational*, J. Arndt
-///
-/// "The idea is to subtract one from each of the bytes and then look for
-/// bytes where the borrow propagated all the way to the most significant
-/// bit."
-#[inline]
-fn contains_zero_byte(x: usize) -> bool {
-  x.wrapping_sub(LO_USIZE) & !x & HI_USIZE != 0
-}
-
 #[cfg(target_pointer_width = "32")]
 #[inline]
 fn repeat_byte(b: u8) -> usize {