Read exports from side modules at link time

This means we can reason about the symbols which are or are not defined
across the entire program which allows for ERROR_ON_UNDEFINED_SYMBOLS to
be enabled for any build that includes all its side modules on the
command line.

Author: sbc100

emcc.py

@@ -631,13 +631,17 @@ def process_dynamic_libs(dylibs):
     imports = webassembly.get_imports(dylib)
     new_exports = []
     for imp in imports:
-      if imp.type not in (webassembly.ExternType.FUNC, webassembly.ExternType.GLOBAL):
+      if imp.kind not in (webassembly.ExternType.FUNC, webassembly.ExternType.GLOBAL):
         continue
       new_exports.append(imp.field)
     logger.debug('Adding exports based on `%s`: %s', dylib, new_exports)
     shared.Settings.EXPORTED_FUNCTIONS.extend(shared.asmjs_mangle(e) for e in new_exports)
     shared.Settings.DEFAULT_LIBRARY_FUNCS_TO_INCLUDE.extend(new_exports)
 
+    exports = webassembly.get_exports(dylib)
+    for export in exports:
+      shared.Settings.SIDE_MODULE_SYMBOLS.append(shared.asmjs_mangle(export.name))
+
 
 def unmangle_symbols_from_cmdline(symbols):
   def unmangle(x):
@@ -1447,9 +1451,13 @@ def default_setting(name, new_default):
           '__heap_base',
           '__stack_pointer',
       ]
-      # This needs to be exported on the Module object too so it's visible
-      # to side modules too.
-      shared.Settings.EXPORTED_FUNCTIONS += ['___heap_base']
+      shared.Settings.EXPORTED_FUNCTIONS += [
+          # This needs to be exported on the Module object too so it's visible
+          # to side modules too.
+          '___heap_base',
+          # Unconditional dependency in library_dylink.js
+          '_setThrew',
+      ]
       if shared.Settings.MINIMAL_RUNTIME:
         exit_with_error('MINIMAL_RUNTIME is not compatible with relocatable output')
       if shared.Settings.WASM2JS:

src/compiler.js

@@ -67,6 +67,7 @@ if (settingsFile) {
 EXPORTED_FUNCTIONS = set(EXPORTED_FUNCTIONS);
 EXCEPTION_CATCHING_ALLOWED = set(EXCEPTION_CATCHING_ALLOWED);
 IMPLEMENTED_FUNCTIONS = set(IMPLEMENTED_FUNCTIONS);
+SIDE_MODULE_SYMBOLS = set(SIDE_MODULE_SYMBOLS);
 INCOMING_MODULE_JS_API = set(INCOMING_MODULE_JS_API);
 
 RUNTIME_DEBUG = LIBRARY_DEBUG || GL_DEBUG;

src/jsifier.js

@@ -49,6 +49,18 @@ function stringifyWithFunctions(obj) {
   }
 }
 
+function isDefined(symName) {
+  if (symName in IMPLEMENTED_FUNCTIONS || symName in SIDE_MODULE_SYMBOLS) {
+    return true;
+  }
+  // 'invoke_' symbols are created at runtime in libary_dylink.py so can
+  // always be considered as defined.
+  if (RELOCATABLE && symName.startsWith('_invoke_')) {
+    return true;
+  }
+  return false;
+}
+
 // JSifier
 function JSify(functionsOnly) {
   var mainPass = !functionsOnly;
@@ -147,7 +159,7 @@ function JSify(functionsOnly) {
       var noExport = false;
 
       if (!LibraryManager.library.hasOwnProperty(ident)) {
-        if (!(finalName in IMPLEMENTED_FUNCTIONS) && !LINKABLE) {
+        if (!isDefined(finalName) && !LINKABLE) {
           var msg = 'undefined symbol: ' + ident;
           if (dependent) msg += ' (referenced by ' + dependent + ')';
           if (ERROR_ON_UNDEFINED_SYMBOLS) {

src/library.js

@@ -35,6 +35,14 @@ LibraryManager.library = {
     {{{ makeSetTempRet0('$i') }}};
   },
 
+#if SAFE_HEAP
+  // Trivial wrappers around runtime functions that make these symbols available
+  // to native code.
+  segfault: function() { segfault(); },
+  alignfault: function() { alignfault(); },
+  ftfault: function() { ftfault(); },
+#endif
+
   // ==========================================================================
   // JavaScript <-> C string interop
   // ==========================================================================

src/modules.js

@@ -443,6 +443,14 @@ function exportRuntime() {
     'abort',
   ];
 
+  if (SAFE_HEAP) {
+    runtimeElements = runtimeElements.concat([
+      'segfault',
+      'ftfault',
+      'alignfault',
+    ])
+  }
+
   if (USE_OFFSET_CONVERTER) {
     runtimeElements.push('WasmOffsetConverter');
   }

src/settings_internal.js

@@ -193,3 +193,5 @@ var HEAP_BASE = 0;
 // Also set for STANDALONE_WASM since the _start function is needed to call
 // static ctors, even if there is no user main.
 var HAS_MAIN = 0;
+
+var SIDE_MODULE_SYMBOLS = [];

tests/other/metadce/hello_world_O3_MAIN_MODULE_2.exports

@@ -2,6 +2,7 @@ __wasm_call_ctors
 dynCall_jiji
 main
 malloc
+setThrew
 stackAlloc
 stackRestore
 stackSave

tests/other/metadce/hello_world_O3_MAIN_MODULE_2.funcs

@@ -9,6 +9,7 @@ $dlmalloc
 $legalstub$dynCall_jiji
 $main
 $sbrk
+$setThrew
 $stackAlloc
 $stackRestore
 $stackSave

tests/other/test_split_main_module.out

@@ -1,3 +1,3 @@
-wrote profile of 236 bytes (allocated 236 bytes)
+wrote profile of 240 bytes (allocated 240 bytes)
 Hello from main!
 Hello from lib!

tests/test_core.py

@@ -3594,6 +3594,7 @@ def dylink_testf(self, main, side, expected=None, force_c=False, main_emcc_args=
 
     # side settings
     self.clear_setting('MAIN_MODULE')
+    self.clear_setting('ERROR_ON_UNDEFINED_SYMBOLS')
     self.set_setting('SIDE_MODULE')
     side_suffix = 'wasm' if self.is_wasm() else 'js'
     if isinstance(side, list):
@@ -3608,10 +3609,14 @@ def dylink_testf(self, main, side, expected=None, force_c=False, main_emcc_args=
     self.set_setting('MAIN_MODULE', main_module)
     self.clear_setting('SIDE_MODULE')
     if auto_load:
+      # Normally we don't report undefined symbols when linking main modules but
+      # in this case we know all the side modules are specified on the command line.
+      # TODO(sbc): Make this the default one day
+      self.set_setting('ERROR_ON_UNDEFINED_SYMBOLS')
       self.emcc_args += main_emcc_args
       self.emcc_args.append('liblib.so')
-      if force_c:
-        self.emcc_args.append('-nostdlib++')
+    if force_c:
+      self.emcc_args.append('-nostdlib++')
 
     if isinstance(main, list):
       # main is just a library