Export mutable wasm globals as JS objects. NFC (#25530)

For immutable wasm globals we currently export them as simple JS numbers
at build time. This is because data symbol addresses are exported as
immutable globals.

However, its also useful to export real wasm globals. For now we treat
mutable globals and immutable global differently here but we should look
into perhaps changing that in the future.

Author: sbc100

ChangeLog.md

@@ -20,6 +20,10 @@ See docs/process.md for more on how version tagging works.
 
 4.0.17 (in development)
 -----------------------
+- Mutable Wasm globals can now be exported from native code.  Currently these
+  cannot be declared in C/C++ but can be defined and exported in assembly code.
+  This currently only works for mutable globals since immutables are already
+  (and continue to be) exported as plain JS numbers. (#25530)
 - Minimum Firefox version was bumped up to Firefox 68 ESR, since older Firefox
   versions are not able to run the parallel browser harness: (#25493)
   - Firefox: v65 -> v68

src/lib/libdylink.js

@@ -230,24 +230,33 @@ var LibraryDylink = {
       }
 #endif
 
-      GOT[symName] ||= new WebAssembly.Global({'value': '{{{ POINTER_WASM_TYPE }}}', 'mutable': true});
-      if (replace || GOT[symName].value == 0) {
+
+      var existingEntry = GOT[symName] && GOT[symName].value != 0;
+      if (replace || !existingEntry) {
 #if DYLINK_DEBUG == 2
         dbg(`updateGOT: before: ${symName} : ${GOT[symName].value}`);
 #endif
+        var newValue;
         if (typeof value == 'function') {
-          GOT[symName].value = {{{ to64('addFunction(value)') }}};
+          newValue = {{{ to64('addFunction(value)') }}};
 #if DYLINK_DEBUG == 2
           dbg(`updateGOT: FUNC: ${symName} : ${GOT[symName].value}`);
 #endif
         } else if (typeof value == {{{ POINTER_JS_TYPE }}}) {
-          GOT[symName].value = value;
+          newValue = value;
         } else {
-          err(`unhandled export type for '${symName}': ${typeof value}`);
+          // The GOT can only contain addresses (i.e data addresses or function
+          // addresses so we currently ignore other types export here.
+#if DYLINK_DEBUG
+          dbg(`updateGOT: ignoring ${symName} due to its type: ${typeof value}`);
+#endif
+          continue;
         }
 #if DYLINK_DEBUG == 2
-        dbg(`updateGOT:  after: ${symName} : ${GOT[symName].value} (${value})`);
+        dbg(`updateGOT:  after: ${symName} : ${newValue} (${value})`);
 #endif
+        GOT[symName] ||= new WebAssembly.Global({'value': '{{{ POINTER_WASM_TYPE }}}', 'mutable': true});
+        GOT[symName].value = newValue;
       }
 #if DYLINK_DEBUG
       else if (GOT[symName].value != value) {
@@ -260,29 +269,43 @@ var LibraryDylink = {
 #endif
   },
 
+  $isImmutableGlobal__internal: true,
+  $isImmutableGlobal: (val) => {
+    if (val instanceof WebAssembly.Global) {
+      try {
+        val.value = val.value;
+      } catch {
+        return true;
+      }
+    }
+    return false;
+  },
+
   // Applies relocations to exported things.
   $relocateExports__internal: true,
-  $relocateExports__deps: ['$updateGOT'],
+  $relocateExports__deps: ['$updateGOT', '$isImmutableGlobal'],
   $relocateExports__docs: '/** @param {boolean=} replace */',
   $relocateExports: (exports, memoryBase, replace) => {
-    var relocated = {};
-
-    for (var e in exports) {
-      var value = exports[e];
+    function relocateExport(name, value) {
 #if SPLIT_MODULE
       // Do not modify exports synthesized by wasm-split
-      if (e.startsWith('%')) {
-        relocated[e] = value
-        continue;
+      if (name.startsWith('%')) {
+        return value;
       }
 #endif
-      // Detect wasm global exports. These represent data addresses
+      // Detect immuable wasm global exports. These represent data addresses
       // which are relative to `memoryBase`
-      if (value instanceof WebAssembly.Global) {
-        value = value.value;
-        value += {{{ to64('memoryBase') }}};
+      if (isImmutableGlobal(value)) {
+        return value.value + {{{ to64('memoryBase') }}};
       }
-      relocated[e] = value;
+
+      // Return unmodified value (no relocation required).
+      return value;
+    }
+
+    var relocated = {};
+    for (var e in exports) {
+      relocated[e] = relocateExport(e, exports[e])
     }
     updateGOT(relocated, replace);
     return relocated;

src/settings_internal.js

@@ -16,8 +16,8 @@
 // underscore.
 var WASM_EXPORTS = [];
 
-// Similar to above but only includes the global/data symbols.
-var WASM_GLOBAL_EXPORTS = [];
+// Similar to above but only includes the data symbols (address exports).
+var DATA_EXPORTS = [];
 
 // An array of all symbols exported from all the side modules specified on the
 // command line.

test/codesize/test_codesize_hello_dylink.json

@@ -1,10 +1,10 @@
 {
-  "a.out.js": 26912,
-  "a.out.js.gz": 11470,
+  "a.out.js": 26919,
+  "a.out.js.gz": 11469,
   "a.out.nodebug.wasm": 18567,
   "a.out.nodebug.wasm.gz": 9199,
-  "total": 45479,
-  "total_gz": 20669,
+  "total": 45486,
+  "total_gz": 20668,
   "sent": [
     "__heap_base",
     "__indirect_function_table",

test/codesize/test_codesize_hello_dylink_all.json

@@ -1,7 +1,7 @@
 {
-  "a.out.js": 245822,
+  "a.out.js": 245829,
   "a.out.nodebug.wasm": 597746,
-  "total": 843568,
+  "total": 843575,
   "sent": [
     "IMG_Init",
     "IMG_Load",

test/core/test_wasm_global.c

@@ -0,0 +1,43 @@
+#include <stdio.h>
+#include <emscripten/em_asm.h>
+#include <emscripten.h>
+
+__asm__(
+".section .data.my_global,\"\",@\n"
+".globl my_global\n"
+".globaltype my_global, i32\n"
+"my_global:\n"
+);
+
+int get_global() {
+  int val;
+  // Without volatile here this test fails in O1 and above.
+  __asm__ volatile ("global.get my_global\n"
+                    "local.set %0\n" : "=r" (val));
+  return val;
+}
+
+void set_global(int val) {
+  __asm__("local.get %0\n"
+          "global.set my_global\n" : : "r" (val));
+}
+
+int main() {
+  printf("in main: %d\n", get_global());
+  set_global(42);
+  printf("new value: %d\n", get_global());
+  EM_ASM({
+      // With the ESM integration, the Wasm global be exported as a regular
+      // number.  Otherwise it will be a WebAssembly.Global object.
+#ifdef ESM_INTEGRATION
+      assert(typeof _my_global == 'number', typeof _my_global);
+      out('from js:', _my_global);
+      _my_global += 1
+#else
+      assert(typeof _my_global == 'object', typeof _my_global);
+      out('from js:', _my_global.value);
+      _my_global.value += 1
+#endif
+  });
+  printf("done: %d\n", get_global());
+}

test/core/test_wasm_global.out

@@ -0,0 +1,4 @@
+in main: 0
+new value: 42
+from js: 42
+done: 43

test/test_core.py

@@ -9677,6 +9677,19 @@ def test_externref_emjs(self, dynlink):
       self.set_setting('MAIN_MODULE', 2)
     self.do_core_test('test_externref_emjs.c')
 
+  @parameterized({
+    '': [False],
+    'dylink': [True],
+  })
+  @no_esm_integration('https://github.com/emscripten-core/emscripten/issues/25543')
+  def test_wasm_global(self, dynlink):
+    if dynlink:
+      self.check_dylink()
+      self.set_setting('MAIN_MODULE', 2)
+    if self.get_setting('WASM_ESM_INTEGRATION'):
+      self.cflags.append('-DESM_INTEGRATION')
+    self.do_core_test('test_wasm_global.c', cflags=['-sEXPORTED_FUNCTIONS=_main,_my_global'])
+
   def test_syscall_intercept(self):
     self.do_core_test('test_syscall_intercept.c')
 

tools/building.py

@@ -810,7 +810,7 @@ def metadce(js_file, wasm_file, debug_info, last):
     exports = settings.WASM_EXPORTS
   else:
     # Ignore exported wasm globals.  Those get inlined directly into the JS code.
-    exports = sorted(set(settings.WASM_EXPORTS) - set(settings.WASM_GLOBAL_EXPORTS))
+    exports = sorted(set(settings.WASM_EXPORTS) - set(settings.DATA_EXPORTS))
 
   extra_info = '{ "exports": [' + ','.join(f'["{asmjs_mangle(x)}", "{x}"]' for x in exports) + ']}'
 

tools/emscripten.py

@@ -111,7 +111,7 @@ def update_settings_glue(wasm_file, metadata, base_metadata):
     settings.WASM_EXPORTS = base_metadata.all_exports
   else:
     settings.WASM_EXPORTS = metadata.all_exports
-  settings.WASM_GLOBAL_EXPORTS = list(metadata.global_exports.keys())
+  settings.DATA_EXPORTS = list(metadata.data_exports.keys())
   settings.HAVE_EM_ASM = bool(settings.MAIN_MODULE or len(metadata.em_asm_consts) != 0)
 
   # start with the MVP features, and add any detected features.
@@ -269,9 +269,9 @@ def trim_asm_const_body(body):
   return body
 
 
-def create_global_exports(global_exports):
+def create_data_exports(data_exports):
   lines = []
-  for k, v in global_exports.items():
+  for k, v in data_exports.items():
     if shared.is_internal_global(k):
       continue
 
@@ -408,11 +408,11 @@ def emscript(in_wasm, out_wasm, outfile_js, js_syms, finalize=True, base_metadat
     other_exports = base_metadata.other_exports
     # We want the real values from the final metadata but we only want to
     # include names from the base_metadata.  See phase_link() in link.py.
-    global_exports = {k: v for k, v in metadata.global_exports.items() if k in base_metadata.global_exports}
+    data_exports = {k: v for k, v in metadata.data_exports.items() if k in base_metadata.data_exports}
   else:
     function_exports = metadata.function_exports
     other_exports = metadata.other_exports
-    global_exports = metadata.global_exports
+    data_exports = metadata.data_exports
 
   if settings.ASYNCIFY == 1:
     function_exports['asyncify_start_unwind'] = webassembly.FuncType([webassembly.Type.I32], [])
@@ -421,7 +421,7 @@ def emscript(in_wasm, out_wasm, outfile_js, js_syms, finalize=True, base_metadat
     function_exports['asyncify_stop_rewind'] = webassembly.FuncType([], [])
 
   parts = [pre]
-  parts += create_module(metadata, function_exports, global_exports, other_exports,
+  parts += create_module(metadata, function_exports, data_exports, other_exports,
                          forwarded_json['librarySymbols'], forwarded_json['nativeAliases'])
   parts.append(post)
   settings.ALIASES = list(forwarded_json['nativeAliases'].keys())
@@ -1008,10 +1008,10 @@ def create_receiving(function_exports, other_exports, library_symbols, aliases):
   return '\n'.join(receiving)
 
 
-def create_module(metadata, function_exports, global_exports, other_exports, library_symbols, aliases):
+def create_module(metadata, function_exports, data_exports, other_exports, library_symbols, aliases):
   module = []
   module.append(create_receiving(function_exports, other_exports, library_symbols, aliases))
-  module.append(create_global_exports(global_exports))
+  module.append(create_data_exports(data_exports))
 
   sending = create_sending(metadata, library_symbols)
   if settings.WASM_ESM_INTEGRATION: