Allow SHARED_MEMORY flag to be used independently. (emscripten-core#22683)

Currently, there are a couple of restrictions on the `SHARED_MEMORY`
flag:
* Attempting to use the `SHARED_MEMORY` flag without the `PTHREADS` or
`WASM_WORKERS` flags simply fails, because we don't have variations of
the system libraries that have this flag standalone. Instead, emcc
attempts to use the single-threaded system libraries, which are not
compatible with the user's build because they don't have the correct
wasm feature flags (i.e. -matomics and -mbulk-memory). This PR adds
`SHARED_MEMORY`-only variations of the system libraries, which allows
users to build with a standalone `SHARED_MEMORY` flag without error.
* Currently, `SHARED_MEMORY` also implies `IMPORTED_MEMORY`. However,
`IMPORTED_MEMORY` is only important for builds that use the `PTHREADS`
or `WASM_WORKERS` flags, since they actually do multi-threading and need
to read and write to the same memory across workers. This PR changes
this so that if a standalone `SHARED_MEMORY` flag is specified, it can
still avoid importing the memory.

The use-case for enabling the `SHARED_MEMORY` flag without actually
having an actual threading implementation is to be compatible with other
modules which may import that memory whose import declaration specifies
the memory as shared. This is an important use-case for Flutter, as we
want to create two variations of the emscripten-generated module (one
single-threaded and one multi-threaded, depending on whether the
environment supports it) while keeping the module that imports the
emscripten-generated module the same (which means keeping the memory
import declaration as shared).

Author: eyebrowsoffire

system/lib/compiler-rt/stack_limits.S

@@ -15,14 +15,17 @@
 #endif
 
 .globaltype __stack_pointer, PTR
+.globl __stack_pointer
 
 .section .globals,"",@
 
 # TODO(sbc): It would be nice if these we initialized directly
 # using PTR.const rather than using the `emscripten_stack_init`
 .globaltype __stack_end, PTR
 __stack_end:
+.globl __stack_end
 .globaltype __stack_base, PTR
+.globl __stack_base
 __stack_base:
 
 .section .text,"",@
@@ -82,57 +85,6 @@ emscripten_stack_get_free:
   PTR.sub
   end_function
 
-#ifdef __EMSCRIPTEN_WASM_WORKERS__
-# TODO: Relocate the following to its own file wasm_worker.S, but need to figure out how to reference
-# __stack_base and __stack_end globals from a separate file as externs in order for that to work.
-.globl _emscripten_wasm_worker_initialize
-_emscripten_wasm_worker_initialize:
-  .functype _emscripten_wasm_worker_initialize (PTR /*stackLowestAddress*/, i32 /*stackSize*/) -> ()
-
-  // __stack_end = stackLowestAddress + (__builtin_wasm_tls_size() + 15) & -16;
-  local.get 0
-  .globaltype __tls_size, PTR, immutable
-  global.get __tls_size
-  PTR.add
-  PTR.const 0xf
-  PTR.add
-  PTR.const -0x10
-  PTR.and
-  global.set __stack_end
-
-  // __stack_base = stackLowestAddress + stackSize;
-  local.get 0
-  local.get 1
-#ifdef __wasm64__
-  i64.extend_i32_u
-#endif
-  PTR.add
-  global.set __stack_base
-
-// TODO: We'd like to do this here to avoid JS side calls to __set_stack_limits.
-//       (or even better, we'd like to avoid duplicate versions of the stack variables)
-// See https://github.com/emscripten-core/emscripten/issues/16496
-//  global.get __stack_base
-//  global.get __stack_end
-//  .functype __set_stack_limits (PTR, PTR) -> ()
-//  call __set_stack_limits
-
-  // __wasm_init_tls(stackLowestAddress);
-  local.get 0
-  .functype __wasm_init_tls (PTR) -> ()
-  call __wasm_init_tls
-
-  // N.b. The function __wasm_init_tls above does not need
-  // __stack_pointer initialized, and it destroys the value it was set to.
-  // So we must initialize __stack_pointer only *after* completing __wasm_init_tls:
-
-  // __stack_pointer = __stack_base;
-  global.get __stack_base
-  global.set __stack_pointer
-
-  end_function
-#endif
-
 # Add emscripten_stack_init to static ctors
 .section .init_array.1,"",@
 .p2align ALIGN

system/lib/dlmalloc.c

@@ -21,13 +21,12 @@
 /* XXX Emscripten Tracing API. This defines away the code if tracing is disabled. */
 #include <emscripten/trace.h>
 
-#ifdef __EMSCRIPTEN_WASM_WORKERS__
+#ifdef __EMSCRIPTEN_SHARED_MEMORY__
 #define USE_LOCKS 1
 #endif
 
 /* Make malloc() and free() threadsafe by securing the memory allocations with pthread mutexes. */
 #if __EMSCRIPTEN_PTHREADS__
-#define USE_LOCKS 1
 #define USE_SPIN_LOCKS 0 // Ensure we use pthread_mutex_t.
 #endif
 

system/lib/wasm_worker/wasm_worker_initialize.S

@@ -0,0 +1,64 @@
+.extern __stack_pointer
+.extern __stack_base
+.extern __stack_end
+
+#ifdef __wasm64__
+#define PTR i64
+#define ALIGN 3
+#define PTRSTORE .int64
+#else
+#define PTR i32
+#define ALIGN 2
+#define PTRSTORE .int32
+#endif
+
+.globaltype __stack_pointer, PTR
+.globaltype __stack_end, PTR
+.globaltype __stack_base, PTR
+
+.globl _emscripten_wasm_worker_initialize
+_emscripten_wasm_worker_initialize:
+  .functype _emscripten_wasm_worker_initialize (PTR /*stackLowestAddress*/, i32 /*stackSize*/) -> ()
+
+  // __stack_end = stackLowestAddress + (__builtin_wasm_tls_size() + 15) & -16;
+  local.get 0
+  .globaltype __tls_size, PTR, immutable
+  global.get __tls_size
+  PTR.add
+  PTR.const 0xf
+  PTR.add
+  PTR.const -0x10
+  PTR.and
+  global.set __stack_end
+
+  // __stack_base = stackLowestAddress + stackSize;
+  local.get 0
+  local.get 1
+#ifdef __wasm64__
+  i64.extend_i32_u
+#endif
+  PTR.add
+  global.set __stack_base
+
+// TODO: We'd like to do this here to avoid JS side calls to __set_stack_limits.
+//       (or even better, we'd like to avoid duplicate versions of the stack variables)
+// See https://github.com/emscripten-core/emscripten/issues/16496
+//  global.get __stack_base
+//  global.get __stack_end
+//  .functype __set_stack_limits (PTR, PTR) -> ()
+//  call __set_stack_limits
+
+  // __wasm_init_tls(stackLowestAddress);
+  local.get 0
+  .functype __wasm_init_tls (PTR) -> ()
+  call __wasm_init_tls
+
+  // N.b. The function __wasm_init_tls above does not need
+  // __stack_pointer initialized, and it destroys the value it was set to.
+  // So we must initialize __stack_pointer only *after* completing __wasm_init_tls:
+
+  // __stack_pointer = __stack_base;
+  global.get __stack_base
+  global.set __stack_pointer
+
+  end_function

test/core/test_dlfcn_self.exports

@@ -9,7 +9,9 @@ __progname_full
 __sig_actions
 __sig_pending
 __signgam
+__stack_base
 __stack_chk_guard
+__stack_end
 __threwValue
 __timezone
 __tzname

test/core/test_em_asm_unicode.out

@@ -1 +1 @@
-hello world…
+hello world…

test/core/test_longjmp.out

@@ -1,4 +1,4 @@
 first
 result: 1 1
 second
-result: 2 -1
+result: 2 -1

tools/link.py

@@ -1413,7 +1413,7 @@ def phase_linker_setup(options, state, newargs):
       'removeRunDependency',
     ]
 
-  if settings.SHARED_MEMORY or settings.RELOCATABLE or settings.ASYNCIFY_LAZY_LOAD_CODE:
+  if settings.PTHREADS or settings.WASM_WORKERS or settings.RELOCATABLE or settings.ASYNCIFY_LAZY_LOAD_CODE:
     settings.IMPORTED_MEMORY = 1
 
   set_initial_memory()

tools/system_libs.py

@@ -742,12 +742,17 @@ def vary_on(cls):
 
   @classmethod
   def get_default_variation(cls, **kwargs):
-    return super().get_default_variation(is_mt=settings.PTHREADS, is_ww=settings.WASM_WORKERS and not settings.PTHREADS, **kwargs)
+    return super().get_default_variation(
+      is_mt=settings.PTHREADS,
+      is_ww=settings.SHARED_MEMORY and not settings.PTHREADS,
+      **kwargs
+    )
 
   @classmethod
   def variations(cls):
     combos = super(MTLibrary, cls).variations()
-    # To save on # of variations, pthreads and Wasm workers when used together, just use pthreads variation.
+
+    # These are mutually exclusive, only one flag will be set at any give time.
     return [combo for combo in combos if not combo['is_mt'] or not combo['is_ww']]
 
 
@@ -1452,9 +1457,19 @@ def get_default_variation(cls, **kwargs):
     return super().get_default_variation(stub=not settings.WASM_WORKERS, **kwargs)
 
   def get_files(self):
+    files = []
+    if self.is_stub:
+      files = [
+        'library_wasm_worker_stub.c'
+      ]
+    else:
+      files = [
+        'library_wasm_worker.c',
+        'wasm_worker_initialize.S',
+      ]
     return files_in_path(
         path='system/lib/wasm_worker',
-        filenames=['library_wasm_worker_stub.c' if self.is_stub else 'library_wasm_worker.c'])
+        filenames=files)
 
   def can_use(self):
     # see src/library_wasm_worker.js