diff --git a/configure.in b/configure.in
index e0980d36..1748e1dc 100644
--- a/configure.in
+++ b/configure.in
@@ -168,7 +168,7 @@ AC_HEADER_STDC
 AC_HEADER_SYS_WAIT
 AC_HEADER_TIME
 
-AC_CHECK_HEADERS(wctype.h wchar.h)
+AC_CHECK_HEADERS(wctype.h)
 AC_CHECK_HEADERS(fcntl.h limits.h time.h unistd.h)
 AC_CHECK_HEADERS(sigaction.h)
 AC_CHECK_HEADERS(arpa/inet.h)
@@ -279,7 +279,7 @@ AC_FUNC_MMAP
 AC_FUNC_STRFTIME
 AC_CHECK_FUNCS(cfmakeraw gethostbyaddr herror strerror)
 AC_CHECK_FUNCS(popen uname access chmod alarm timegm mremap)
-AC_CHECK_FUNCS(strcasecmp strncasecmp strcasestr strstr strchr strrchr wcwidth)
+AC_CHECK_FUNCS(strcasecmp strncasecmp strcasestr strstr strchr strrchr)
 AC_CHECK_FUNCS(memmove bcopy stpcpy strdup index isdigit mempcpy memrchr)
 AC_CHECK_FUNCS(snprintf vsnprintf asprintf vasprintf)
 AC_CHECK_FUNCS(getifaddrs getpwnam inet_pton inet_ntop)
diff --git a/src/bfu/inpfield.c b/src/bfu/inpfield.c
index 297b84b5..7edd838c 100644
--- a/src/bfu/inpfield.c
+++ b/src/bfu/inpfield.c
@@ -680,12 +680,32 @@ kbd_field(struct dialog_data *dlg_data, struct widget_data *widget_data)
 			if (check_kbd_textinput_key(ev)) {
 				unsigned char *text = widget_data->cdata;
 				int textlen = strlen(text);
-#ifdef CONFIG_UTF8
-				const unsigned char *ins = encode_utf8(get_kbd_key(ev));
-				int inslen = utf8charlen(ins);
-#else  /* !CONFIG_UTF8 */
+#ifndef CONFIG_UTF8
+				/* Both get_kbd_key(ev) and @text
+				 * are in the terminal's charset.  */
 				const int inslen = 1;
-#endif /* !CONFIG_UTF8 */
+#else  /* CONFIG_UTF8 */
+				const unsigned char *ins;
+				int inslen;
+
+				if (term->utf8) {
+					/* get_kbd_key(ev) is in UCS-4,
+					 * and @text is in UTF-8.  */
+					ins = encode_utf8(get_kbd_key(ev));
+					/* get_kbd_key(ev) cannot be L'\0'
+					 * because @check_kbd_textinput_key
+					 * would have rejected it.  So it
+					 * is OK to use @strlen below.  */
+				} else {
+					/* get_kbd_key(ev) is UCS-4, and @text
+					 * is in the terminal's charset.  */
+					int cp = get_opt_codepage_tree(term->spec,
+								       "charset");
+
+					ins = u2cp_no_nbsp(get_kbd_key(ev), cp);
+				}
+				inslen = strlen(ins);
+#endif /* CONFIG_UTF8 */
 
 				if (textlen >= widget_data->widget->datalen - inslen)
 					goto display_field;
diff --git a/src/intl/charsets.c b/src/intl/charsets.c
index 6c296a38..9f788215 100644
--- a/src/intl/charsets.c
+++ b/src/intl/charsets.c
@@ -1,8 +1,5 @@
 /* Charsets convertor */
 
-#ifndef _XOPEN_SOURCE
-#define _XOPEN_SOURCE /* wcwidth() */
-#endif
 #ifndef _GNU_SOURCE
 #define _GNU_SOURCE /* strcasecmp() */
 #endif
@@ -17,9 +14,6 @@
 
 #include <ctype.h>
 #include <stdlib.h>
-#if HAVE_WCHAR_H
-#include <wchar.h>
-#endif
 #if HAVE_WCTYPE_H
 #include <wctype.h>
 #endif
@@ -558,6 +552,30 @@ invalid_arg:
  * Find out number of standard terminal collumns needed for displaying symbol
  * (glyph) which represents Unicode character c.
  *
+ * TODO: Use wcwidth when it is available. This seems to require:
+ * - Make the configure script check whether <wchar.h> and wcwidth exist.
+ * - Define _XOPEN_SOURCE and include <wchar.h>.
+ * - Test that __STDC_ISO_10646__ is defined.  (This macro means wchar_t
+ *   matches ISO 10646 in all locales.)
+ * However, these do not suffice, because wcwidth depends on LC_CTYPE
+ * in glibc-2.3.6.  For instance, wcwidth(0xff20) is -1 when LC_CTYPE
+ * is "fi_FI.ISO-8859-1" or "C", but 2 when LC_CTYPE is "fi_FI.UTF-8".
+ * <features.h> defines __STDC_ISO_10646__ as 200009L, so 0xff20 means
+ * U+FF20 FULLWIDTH COMMERCIAL AT regardless of LC_CTYPE; but this
+ * character is apparently not supported in all locales.  Why is that?
+ * - Perhaps there is standardese that requires supported characters
+ *   to be convertable to multibyte form.  Then ELinks could just pick
+ *   some UTF-8 locale for its wcwidth purposes.
+ * - Perhaps wcwidth can even return different nonnegative values for
+ *   the same ISO 10646 character in different locales.  Then ELinks
+ *   would have to set LC_CTYPE to match at least the terminal's
+ *   charset (which may differ from the LC_CTYPE environment variable,
+ *   especially when the master process is serving a slave terminal).
+ *   But there is no guarantee that the libc supports all the same
+ *   charsets as ELinks does.
+ * For now, it seems safest to avoid the potentially locale-dependent
+ * libc version of wcwidth, and instead use a hardcoded mapping.
+ *
  * @return	2 for double-width glyph, 1 for others.
  * 		TODO: May be extended to return 0 for zero-width glyphs
  * 		(like composing, maybe unprintable too).
@@ -565,10 +583,6 @@ invalid_arg:
 inline int
 unicode_to_cell(unicode_val_T c)
 {
-#if __STDC_ISO_10646__ && HAVE_WCWIDTH
-	if (wcwidth(c) >= 2)
-		return 2;
-#else  /* !__STDC_ISO_10646__ || !HAVE_WCWIDTH */
 	if (c >= 0x1100
 		&& (c <= 0x115f			/* Hangul Jamo */
 		|| c == 0x2329
@@ -584,7 +598,6 @@ unicode_to_cell(unicode_val_T c)
 		|| (c >= 0x20000 && c <= 0x2fffd)
 		|| (c >= 0x30000 && c <= 0x3fffd)))
 		return 2;
-#endif /* !__STDC_ISO_10646__ || !HAVE_WCWIDTH */
 
 	return 1;
 }