grep -F could erroneously fail to match in non-UTF8 multibyte locales

This fixes a bug that can strike only when using a non-UTF8 multibyte
locale like ja_JP.SHIFT_JIS.

Consider this example: it would mistakenly fail to match before
this patch:

  printf '\203AA\n'|LC_ALL=ja_JP.SHIFT_JIS src/grep -F A

When searching for a single byte that happens to be the latter
byte of a multibyte character, and the target byte also follows
that multibyte character, grep -F would advance an internal pointer
by one byte too many, thus missing the target byte.  A test case
for this bug is already included in tests/sjis-mb.

* src/kwsearch.c (Fexecute): Skip one byte less, after matched middle of a
multi-byte character.  Introduced by commit v2.18-119-gfb7d538.

2 files changed, 21 insertions, 3 deletions

diff --git a/NEWS b/NEWS
index 15975288..c5202dbf 100644
--- a/NEWS
+++ b/NEWS
@@ -48,6 +48,13 @@ GNU grep NEWS                                    -*- outline -*-
   of a multibyte character when using a '^'-anchored alternate in a pattern,
   leading it to print non-matching lines.  [bug present since "the beginning"]
 
+  grep -F Y no longer fails to match in non-UTF8 multibyte locales like
+  Shift-JIS, when the input contains a 2-byte character, XY, followed by
+  the single-byte search pattern, Y.  grep would find the first, middle-
+  of-multibyte matching "Y", and then mistakenly advance an internal
+  pointer one byte too far, skipping over the target "Y" just after that.
+  [bug introduced in grep-2.19]
+
   grep -E rejected unmatched ')', instead of treating it like '\)'.
   [bug present since "the beginning"]
 
diff --git a/src/kwsearch.c b/src/kwsearch.c
index aa965f62..1335a269 100644
--- a/src/kwsearch.c
+++ b/src/kwsearch.c
@@ -133,9 +133,20 @@ Fexecute (char const *buf, size_t size, size_t *match_size,
       if (!match_lines && MB_CUR_MAX > 1 && !using_utf8 ()
           && mb_goback (&mb_start, beg + offset, buf + size) != 0)
         {
-          /* The match was a part of multibyte character, advance at least
-             one byte to ensure no infinite loop happens.  */
-          beg = mb_start;
+          /* We have matched a single byte that is not at the beginning of a
+             multibyte character.  mb_goback has advanced MB_START past that
+             multibyte character.  Now, we want to position BEG so that the
+             next kwsexec search starts there.  Thus, to compensate for the
+             for-loop's BEG++, above, subtract one here.  This code is
+             unusually hard to reach, and exceptionally, let's show how to
+             trigger it here:
+
+               printf '\203AA\n'|LC_ALL=ja_JP.SHIFT_JIS src/grep -F A
+
+             That assumes the named locale is installed.
+             Note that your system's shift-JIS locale may have a different
+             name, possibly including "sjis".  */
+          beg = mb_start - 1;
           continue;
         }
       beg += offset;