Use C11 _Thread_local for current locale. NFC (#26486)

We already do a similar thing for `__errno_location`.

This means that locale usage no longer depends on pthread_self, which
means it works correctly in Wasm Workers.

I guess the codesize savings here come from the fact that the TLS data
doesn't need to initialized programmatically anymore.

Author: sbc100

src/struct_info_generated.json

@@ -1036,11 +1036,11 @@
             "p_proto": 8
         },
         "pthread": {
-            "__size__": 132,
-            "profilerBlock": 112,
-            "stack": 52,
-            "stack_size": 56,
-            "waiting_async": 128
+            "__size__": 124,
+            "profilerBlock": 104,
+            "stack": 48,
+            "stack_size": 52,
+            "waiting_async": 120
         },
         "pthread_attr_t": {
             "__size__": 44,

src/struct_info_generated_wasm64.json

@@ -1036,11 +1036,11 @@
             "p_proto": 16
         },
         "pthread": {
-            "__size__": 232,
-            "profilerBlock": 200,
-            "stack": 88,
-            "stack_size": 96,
-            "waiting_async": 228
+            "__size__": 216,
+            "profilerBlock": 184,
+            "stack": 80,
+            "stack_size": 88,
+            "waiting_async": 212
         },
         "pthread_attr_t": {
             "__size__": 88,

system/lib/libc/musl/src/internal/locale_impl.h

@@ -45,9 +45,17 @@ hidden char *__gettextdomain(void);
 #define C_LOCALE ((locale_t)&__c_locale)
 #define UTF8_LOCALE ((locale_t)&__c_dot_utf8_locale)
 
+#ifdef __EMSCRIPTEN__
+extern _Thread_local locale_t __tls_locale;
+
+#define CURRENT_LOCALE (__tls_locale)
+
+#define CURRENT_UTF8 (!!__tls_locale->cat[LC_CTYPE])
+#else
 #define CURRENT_LOCALE (__pthread_self()->locale)
 
 #define CURRENT_UTF8 (!!__pthread_self()->locale->cat[LC_CTYPE])
+#endif
 
 #undef MB_CUR_MAX
 #define MB_CUR_MAX (CURRENT_UTF8 ? 4 : 1)

system/lib/libc/musl/src/internal/pthread_impl.h

@@ -39,7 +39,10 @@ struct pthread {
 
 	/* Part 2 -- implementation details, non-ABI. */
 	int tid;
+#ifndef __EMSCRIPTEN__
+	// Emscripten uses C11 _Thread_local instead for errno
 	int errno_val;
+#endif
 	volatile int detach_state;
 	volatile int cancel;
 	volatile unsigned char canceldisable, cancelasync;
@@ -60,7 +63,10 @@ struct pthread {
 	} robust_list;
 	int h_errno_val;
 	volatile int timer_id;
+#ifndef __EMSCRIPTEN__
+	// Emscripten uses C11 _Thread_local instead for locale
 	locale_t locale;
+#endif
 	volatile int killlock[1];
 	char *dlerror_buf;
 	void *stdio_locks;

system/lib/libc/musl/src/locale/uselocale.c

@@ -2,13 +2,24 @@
 #include "pthread_impl.h"
 #include "libc.h"
 
+#ifdef __EMSCRIPTEN__
+_Thread_local locale_t __tls_locale = &libc.global_locale;
+#endif
+
 locale_t __uselocale(locale_t new)
 {
+#ifdef __EMSCRIPTEN__
+	locale_t old = __tls_locale;
+	locale_t global = &libc.global_locale;
+
+	if (new) __tls_locale = new == LC_GLOBAL_LOCALE ? global : new;
+#else
 	pthread_t self = __pthread_self();
 	locale_t old = self->locale;
 	locale_t global = &libc.global_locale;
 
 	if (new) self->locale = new == LC_GLOBAL_LOCALE ? global : new;
+#endif
 
 	return old == global ? LC_GLOBAL_LOCALE : old;
 }

system/lib/pthread/library_pthread.c

@@ -147,7 +147,6 @@ void _emscripten_init_main_thread(void) {
   // Main thread ID is always 1.  It can't be 0 because musl assumes
   // tid is always non-zero.
   __main_pthread.tid = getpid();
-  __main_pthread.locale = &libc.global_locale;
   // pthread struct prev and next should initially point to itself (see __init_tp),
   // this is used by pthread_key_delete for deleting thread-specific data.
   __main_pthread.next = __main_pthread.prev = &__main_pthread;

system/lib/pthread/pthread_create.c

@@ -182,7 +182,6 @@ int __pthread_create(pthread_t* restrict res,
   // pthread struct robust_list head should point to itself.
   new->robust_list.head = &new->robust_list.head;
 
-  new->locale = &libc.global_locale;
   if (attr._a_detach) {
     new->detach_state = DT_DETACHED;
   } else {

system/lib/pthread/pthread_self_stub.c

@@ -30,7 +30,6 @@ extern int __stack_low;
 
 __attribute__((constructor))
 static void init_pthread_self(void) {
-  __main_pthread.locale = &libc.global_locale;
   __main_pthread.tid = getpid();
   __main_pthread.stack = &__stack_high;
   __main_pthread.stack_size = ((size_t)&__stack_high) - ((size_t)&__stack_low);

test/codesize/test_codesize_cxx_ctors1.json

@@ -1,10 +1,10 @@
 {
   "a.out.js": 19214,
   "a.out.js.gz": 7981,
-  "a.out.nodebug.wasm": 132658,
-  "a.out.nodebug.wasm.gz": 49938,
-  "total": 151872,
-  "total_gz": 57919,
+  "a.out.nodebug.wasm": 132638,
+  "a.out.nodebug.wasm.gz": 49927,
+  "total": 151852,
+  "total_gz": 57908,
   "sent": [
     "__cxa_throw",
     "_abort_js",

test/codesize/test_codesize_cxx_ctors2.json

@@ -1,10 +1,10 @@
 {
   "a.out.js": 19191,
   "a.out.js.gz": 7966,
-  "a.out.nodebug.wasm": 132082,
-  "a.out.nodebug.wasm.gz": 49598,
-  "total": 151273,
-  "total_gz": 57564,
+  "a.out.nodebug.wasm": 132064,
+  "a.out.nodebug.wasm.gz": 49586,
+  "total": 151255,
+  "total_gz": 57552,
   "sent": [
     "__cxa_throw",
     "_abort_js",