Apply data relocation before calling static constructors. (#17308)

Until now we have applied data reloctions at the beginning of
`__wasm_call_ctors`.  However, in #17295, it was observed that this
is not valid, since relocation need to be applied to all loaded
libraries before any static constructor can be run.

To deal with this issue we now export `__wasm_apply_data_relocs`
separately and call this function on all libraries before calling any
`__wasm_call_ctors` function.

This change must land before the corresponding llvm-side change:
https://reviews.llvm.org/D128515.  The tests for this change are
temporarily disabled until the llvm change gets rolled in.

Author: sbc100

emcc.py

@@ -2094,6 +2094,11 @@ def phase_linker_setup(options, state, newargs, user_settings):
     # in standalone mode, crt1 will call the constructors from inside the wasm
     settings.REQUIRED_EXPORTS.append('__wasm_call_ctors')
 
+  if settings.RELOCATABLE:
+    # TODO(https://reviews.llvm.org/D128515): Make this mandatory once
+    # llvm change lands
+    settings.EXPORT_IF_DEFINED.append('__wasm_apply_data_relocs')
+
   if settings.RELOCATABLE and not settings.DYNAMIC_EXECUTION:
     exit_with_error('cannot have both DYNAMIC_EXECUTION=0 and RELOCATABLE enabled at the same time, since RELOCATABLE needs to eval()')
 

src/library_dylink.js

@@ -654,6 +654,14 @@ var LibraryDylink = {
               __ATINIT__.push(init);
             }
           }
+          var applyRelocs = moduleExports['__wasm_apply_data_relocs'];
+          if (applyRelocs) {
+            if (runtimeInitialized) {
+              applyRelocs();
+            } else {
+              __RELOC_FUNCS__.push(applyRelocs);
+            }
+          }
 #if USE_PTHREADS
         }
 #endif

src/preamble.js

@@ -338,6 +338,10 @@ var __ATMAIN__    = []; // functions called when main() is to be run
 var __ATEXIT__    = []; // functions called during shutdown
 var __ATPOSTRUN__ = []; // functions called after the main() is called
 
+#if RELOCATABLE
+var __RELOC_FUNCS__ = [];
+#endif
+
 var runtimeInitialized = false;
 
 #if EXIT_RUNTIME
@@ -393,6 +397,9 @@ function initRuntime() {
   err('__set_stack_limits: ' + _emscripten_stack_get_base() + ', ' + _emscripten_stack_get_end());
 #endif
   ___set_stack_limits(_emscripten_stack_get_base(), _emscripten_stack_get_end());
+#endif
+#if RELOCATABLE
+  callRuntimeCallbacks(__RELOC_FUNCS__);
 #endif
   <<< ATINITS >>>
   callRuntimeCallbacks(__ATINIT__);
@@ -1051,6 +1058,10 @@ function createWasm() {
     addOnInit(Module['asm']['__wasm_call_ctors']);
 #endif
 
+#if hasExportedFunction('___wasm_apply_data_relocs')
+    __RELOC_FUNCS__.push(Module['asm']['__wasm_apply_data_relocs']);
+#endif
+
 #if ABORT_ON_WASM_EXCEPTIONS
     instrumentWasmTableWithAbort();
 #endif

tests/test_other.py

@@ -1808,11 +1808,18 @@ def test_dylink_pthread_bigint_em_js(self):
     self.do_runf(test_file('core/test_em_js.cpp'))
 
   @node_pthreads
-  def test_dylink_pthread_comdat(self):
+  @parameterized({
+    '': (False,),
+    'flipped': (True,),
+  })
+  def test_dylink_pthread_comdat(self, flipped):
     # Test that the comdat info for `Foo`, which is defined in the side module,
     # is visible to the main module.
     create_file('foo.h', r'''
     struct Foo {
+      Foo() {
+        method();
+      }
       // Making this method virtual causes the comdat group for the
       // class to only be defined in the side module.
       virtual void method() const;
@@ -1823,6 +1830,12 @@ def test_dylink_pthread_comdat(self):
       #include <typeinfo>
       #include <emscripten/console.h>
 
+      // Foo constructor calls a virtual function, with the vtable defined
+      // in the side module. This verifies that the side module's data
+      // reloctions are applied before calling static constructors in the
+      // main module.
+      Foo g_foo;
+
       int main() {
         _emscripten_outf("main: Foo typeid: %s", typeid(Foo).name());
 
@@ -1839,14 +1852,21 @@ def test_dylink_pthread_comdat(self):
         _emscripten_outf("side: Foo typeid: %s", typeid(Foo).name());
       }
       ''')
+    if flipped:
+      side = 'main.cpp'
+      main = 'side.cpp'
+    else:
+      self.skipTest('https://reviews.llvm.org/D128515')
+      side = 'side.cpp'
+      main = 'main.cpp'
     self.run_process([
       EMCC,
       '-o', 'libside.wasm',
-      'side.cpp',
+      side,
       '-pthread', '-Wno-experimental',
       '-sSIDE_MODULE=1'])
     self.do_runf(
-      'main.cpp',
+      main,
       'main: Foo typeid: 3Foo\nside: Foo typeid: 3Foo\n',
       emcc_args=[
         '-pthread', '-Wno-experimental',