[libc unicode] Make glob sort order match bash

Call setlocale() with LC_ALL instead of LC_CTYPE

GNU readline calls it with LC_TYPE

Author: Andy C

bin/oils_for_unix.py

@@ -83,14 +83,13 @@ def CaperDispatch():
 
 def InitLocale(environ):
     # type: (Dict[str, str]) -> None
-    """Set GLOBAL libc locale from environment, and CHECK that it's valid.
-
-    Note: LC_COLLATE/LC_ALL might be necessary for glob
-    LANG= is the default, LC_ALL= sets all of them
-    https://unix.stackexchange.com/questions/576701/what-is-the-difference-between-lang-c-and-lc-all-c
-    """
+    """Set the GLOBAL libc locale from the env, and CHECK that it's valid."""
     try:
-        locale_name = pylocale.setlocale(pylocale.LC_CTYPE, '')
+        # Note: LC_ALL (rather than LC_CTYPE) makes glob order match bash
+        #
+        # https://unix.stackexchange.com/questions/576701/what-is-the-difference-between-lang-c-and-lc-all-c
+        # LANG= is the default, LC_ALL= sets all of them
+        locale_name = pylocale.setlocale(pylocale.LC_ALL, '')
 
         # passing None queries it
         #lo = locale.setlocale(locale.LC_CTYPE, None)
@@ -109,8 +108,9 @@ def InitLocale(environ):
         #log('codeset %s', codeset)
 
         if not match.IsUtf8Codeset(codeset):
-            print_stderr("oils warning: codeset %r doesn't look like UTF-8" % codeset)
-            print_stderr('              Set OILS_LOCALE_OK=1 to remove this message')
+            print_stderr("oils warning: codeset %r doesn't look like UTF-8" %
+                         codeset)
+            print_stderr('              OILS_LOCALE_OK=1 removes this message')
 
 
 # TODO: Hook up valid applets (including these) to completion

cpp/libc.h

@@ -40,9 +40,14 @@ int sleep_until_error(double seconds);
 
 // pylib/locale_.py
 namespace pylocale {
+
 const int CODESET = ::CODESET;
+
+constexpr int lc_all = LC_ALL;
 constexpr int lc_ctype = LC_CTYPE;
+#undef LC_ALL
 #undef LC_CTYPE
+const int LC_ALL = lc_all;
 const int LC_CTYPE = lc_ctype;
 
 class Error {
@@ -53,6 +58,7 @@ class Error {
 };
 BigStr* setlocale(int category, BigStr* locale);
 BigStr* nl_langinfo(int item);
+
 }  // namespace pylocale
 
 #endif  // LIBC_H

pylib/pylocale.py

@@ -9,6 +9,7 @@
 import _locale  # type: ignore
 
 CODESET = _locale.CODESET  # type: int
+LC_ALL = _locale.LC_ALL  # type: int
 LC_CTYPE = _locale.LC_CTYPE  # type: int
 
 

spec/glob.test.sh

@@ -259,26 +259,24 @@ __a__
 
 #### Glob ordering respects LC_COLLATE (zsh respects this too)
 
-# bug from test/gold
-# mksh ksh osh - hello-test.sh comes first
+# test/spec-common.sh sets LC_ALL=C.UTF_8
+unset LC_ALL
+
+touch hello hello.py hello_preamble.sh hello-test.sh
+echo h*
+
 # bash - hello_preamble.h comes first
-#
 # But ord('_') == 95 
 #     ord('-') == 45
 
 # https://serverfault.com/questions/122737/in-bash-are-wildcard-expansions-guaranteed-to-be-in-order
 
-touch hello hello.py hello_preamble.sh hello-test.sh
+#LC_COLLATE=C.UTF-8
+LC_COLLATE=en_US.UTF-8  # en_US is necessary
 echo h*
 
-# Spec tests set LC_ALL=C.UTF_8
-unset LC_ALL
-
-# the en_US does it
-LC_COLLATE=en_US.UTF-8
-#LC_COLLATE=C.UTF-8
+LC_COLLATE=en_US.UTF-8 $SH -c 'echo h*'
 
-echo h*
 
 # Doesn't work, probably because
 #LC_COLLATE=en_US.UTF-8
@@ -287,11 +285,13 @@ echo h*
 ## STDOUT:
 hello hello-test.sh hello.py hello_preamble.sh
 hello hello_preamble.sh hello.py hello-test.sh
+hello hello_preamble.sh hello.py hello-test.sh
 ## END
 
 ## N-I dash/mksh/ash STDOUT:
 hello hello-test.sh hello.py hello_preamble.sh
 hello hello-test.sh hello.py hello_preamble.sh
+hello hello-test.sh hello.py hello_preamble.sh
 ## END
 
 

test/gold.sh

@@ -111,7 +111,7 @@ test-dollar-sq() { _compare $GOLD_DIR/dollar-sq.sh; }
 # 2025-06: This is a sort order issue that doesn't show up in CI (related to
 #          LC_COLLATE)
 #          I reproduced this issue in spec/glob case #39
-TODO-test-word-eval() { _compare $GOLD_DIR/word-eval.sh; }
+test-word-eval() { _compare $GOLD_DIR/word-eval.sh; }
 
 test-abuild() {
   _compare $GOLD_DIR/abuild.sh is_function is_function