Do not inherit LC_CTYPE locale from environment

[nikic]

Apparently, the way setlocale inheritance currently works in PHP is as follows:

• LC_ALL is defaulted to "C".
• LC_CTYPE is inherited from environment.
• Some code (like PCRE) does not inherit LC_CTYPE from environment, it will assume "C" by default instead.

This PR removes locale inheritance completely. On startup we will stay at the default of LC_ALL="C". Changing the locale requires an explicit setlocale() call.

[cmb69]

net-snmp's init_snmp() may call setlocale(LC_CTYPE, "");. We could counteract:

 ext/snmp/snmp.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/ext/snmp/snmp.c b/ext/snmp/snmp.c
index 8464e74aad..47a47f559b 100644
--- a/ext/snmp/snmp.c
+++ b/ext/snmp/snmp.c
@@ -49,6 +49,7 @@
 #ifdef HAVE_UNISTD_H
 #include <unistd.h>
 #endif
+#include <locale.h>
 
 #ifndef __P
 #ifdef __GNUC__
@@ -1971,6 +1972,7 @@ PHP_MINIT_FUNCTION(snmp)
 	le_snmp_session = zend_register_list_destructors_ex(php_snmp_session_destructor, NULL, PHP_SNMP_SESSION_RES_NAME, module_number);
 
 	init_snmp("snmpapp");
+	setlocale(LC_ALL, "C");
 
 #ifdef NETSNMP_DS_LIB_DONT_PERSIST_STATE
 	/* Prevent update of the snmpapp.conf file */

[cmb69]

Seems to work as desired. We would have to fix 2 Windows tests, or apply PR #5481.

[nikic]

For some reason the ext/standard/tests/strings/basename_invalid_path.phpt test fails on MacOS (even after retriggering the build, doesn't seem spurious). There's no diff shown though, just fails.

[nikic]

php_basename uses php_mblen so maybe basename() is locale-sensitive. I don't get why there is no diff though.

[kocsismate]

@nikic I ran this test locally on MacOS, and it shows the following diff:

001+ string(1) "�"
001- string(0) ""

[cmb69]

Maybe the diff can't be rendered due to invalid UTF-8?

Anyhow, from the docs:

basename() is locale aware, so for it to see the correct basename with multibyte character paths, the matching locale must be set using the setlocale() function.

Isn't that lovely. ;)

[nikic]

The basename() issue should be fixed by 90705d4. That was an implementation bug that got uncovered by this change.

[cmb69]

Thanks for fixing this! :)

[KalleZ]

Big +1 for this, it is a good step forward