Move emscripten_wasm_worker_self_id to native code. NFC (#26468)

This will speed up this function significantly which is going to be
important for hybrid pthreads + Wasm Workers builds where we need to be
able to distinguish between Wasm Workers and non-Wasm Workers sometimes
at the very local level.

Note that this function returns 0 for all non-Wasm Workers, which
includes both the main thread and pthreads.

Also, move emscripten_current_thread_is_wasm_worker since that is
trivially implemented on top of `_self_id()`

Author: sbc100

src/lib/libsigs.js

@@ -619,7 +619,6 @@ sigs = {
   emscripten_create_wasm_audio_worklet_processor_async__sig: 'vippp',
   emscripten_create_worker__sig: 'ip',
   emscripten_current_thread_is_audio_worklet__sig: 'i',
-  emscripten_current_thread_is_wasm_worker__sig: 'i',
   emscripten_date_now__sig: 'd',
   emscripten_dbg__sig: 'vp',
   emscripten_dbg_backtrace__sig: 'vp',
@@ -853,7 +852,6 @@ sigs = {
   emscripten_wasm_worker_post_function_vi__sig: 'vipi',
   emscripten_wasm_worker_post_function_vii__sig: 'vipii',
   emscripten_wasm_worker_post_function_viii__sig: 'vipiii',
-  emscripten_wasm_worker_self_id__sig: 'i',
   emscripten_webgl_commit_frame__sig: 'i',
   emscripten_webgl_create_context__sig: 'ppp',
   emscripten_webgl_destroy_context__sig: 'ip',

src/lib/libwasm_worker.js

@@ -126,7 +126,7 @@ addToLibrary({
     ___set_stack_limits(wwParams.stackLowestAddress + wwParams.stackSize, wwParams.stackLowestAddress);
 #endif
     // Run the C side Worker initialization for stack and TLS.
-    __emscripten_wasm_worker_initialize(wwParams.stackLowestAddress, wwParams.stackSize);
+    __emscripten_wasm_worker_initialize(wwParams.wwID, wwParams.stackLowestAddress, wwParams.stackSize);
 #if PTHREADS
     // Record the pthread configuration, and whether this Wasm Worker supports synchronous blocking in emscripten_futex_wait().
     // (regular Wasm Workers do, AudioWorklets don't)
@@ -244,16 +244,6 @@ if (ENVIRONMENT_IS_WASM_WORKER
     _wasmWorkers = {};
   },
 
-  emscripten_current_thread_is_wasm_worker: () => {
-#if WASM_WORKERS
-    return ENVIRONMENT_IS_WASM_WORKER;
-#else
-    // implicit return 0;
-#endif
-  },
-
-  emscripten_wasm_worker_self_id: () => wwParams?.wwID,
-
   emscripten_wasm_worker_post_function_v: (id, funcPtr) => {
     _wasmWorkers[id].postMessage({'_wsc': funcPtr, 'x': [] }); // "WaSm Call"
   },

system/include/emscripten/wasm_worker.h

@@ -67,13 +67,15 @@ void emscripten_terminate_wasm_worker(emscripten_wasm_worker_t id);
 void emscripten_terminate_all_wasm_workers(void);
 
 // Returns true if the current thread is executing a Wasm Worker, false
-// otherwise.  Note that calling this function can be relatively slow as it
-// incurs a Wasm->JS transition, so avoid calling it in hot paths.
+// otherwise.
 bool emscripten_current_thread_is_wasm_worker(void);
 
 // Returns a unique ID that identifies the calling Wasm Worker. Similar to
 // pthread_self().  The main browser thread will return 0 as the ID. First Wasm
 // Worker will return 1, and so on.
+// Note: This function also returns 0 when called from other non-Wasm Worker
+// contexts, such as pthreads in a program built with both pthread and Wasm
+// Worker support.
 uint32_t emscripten_wasm_worker_self_id(void);
 
 // emscripten_wasm_worker_post_function_*: Post a pointer to a C/C++ function to

system/lib/wasm_worker/library_wasm_worker.c

@@ -76,3 +76,7 @@ void emscripten_wasm_worker_sleep(int64_t nsecs) {
   int32_t addr = 0;
   emscripten_atomic_wait_u32(&addr, 0, nsecs);
 }
+
+bool emscripten_current_thread_is_wasm_worker() {
+  return emscripten_wasm_worker_self_id() != 0;
+}

system/lib/wasm_worker/wasm_worker_initialize.S

@@ -14,32 +14,47 @@
 .globaltype __tls_size, PTR, immutable
 .functype emscripten_stack_set_limits (PTR /*base*/, PTR /*end*/) -> ()
 
+.section .__self_id,"",@
+.globaltype __self_id, i32
+__self_id:
+
+.section .text,"",@
+
+.globl emscripten_wasm_worker_self_id
+emscripten_wasm_worker_self_id:
+  .functype emscripten_wasm_worker_self_id () -> (i32)
+  global.get __self_id
+  end_function
+
 .globl _emscripten_wasm_worker_initialize
 _emscripten_wasm_worker_initialize:
-  .functype _emscripten_wasm_worker_initialize (PTR /*stackLowestAddress*/, i32 /*stackSize*/) -> ()
+  .functype _emscripten_wasm_worker_initialize (i32 /*selfID*/, PTR /*stackLowestAddress*/, i32 /*stackSize*/) -> ()
   .local PTR, PTR
 
-  // stack_end = stackLowestAddress + (__builtin_wasm_tls_size() + 15) & -16;
   local.get 0
+  global.set __self_id
+
+  // stack_end = stackLowestAddress + (__builtin_wasm_tls_size() + 15) & -16;
+  local.get 1
   global.get __tls_size
   PTR.add
   PTR.const 0xf
   PTR.add
   PTR.const -0x10
   PTR.and
-  local.set 2
+  local.set 3
 
   // stack_base = stackLowestAddress + stackSize;
-  local.get 0
   local.get 1
+  local.get 2
 #ifdef __wasm64__
   i64.extend_i32_u
 #endif
   PTR.add
-  local.set 3
+  local.set 4
 
+  local.get 4
   local.get 3
-  local.get 2
   // emscripten_stack_set_limits(stack_base, stack_end);
   call emscripten_stack_set_limits
 
@@ -53,7 +68,7 @@ _emscripten_wasm_worker_initialize:
 
 
   // __wasm_init_tls(stackLowestAddress);
-  local.get 0
+  local.get 1
   .functype __wasm_init_tls (PTR) -> ()
   call __wasm_init_tls
 
@@ -62,7 +77,7 @@ _emscripten_wasm_worker_initialize:
   // So we must initialize __stack_pointer only *after* completing __wasm_init_tls:
 
   // __stack_pointer = stack_base;
-  local.get 3
+  local.get 4
   global.set __stack_pointer
 
   end_function

test/codesize/audio_worklet_wasm.expected.js

@@ -134,7 +134,7 @@ var K = [], L = a => {
         a = {
             latencyHint: c,
             sampleRate: G[a + 4 >> 2] || void 0,
-            M: 0 > b ? "hardware" : b || "default"
+            N: 0 > b ? "hardware" : b || "default"
         };
     } else a = void 0;
     a = new AudioContext(a);
@@ -204,7 +204,7 @@ var K = [], L = a => {
         });
         e.port.postMessage({
             _boot: 1,
-            N: ba++,
+            M: ba++,
             G: m.wasm,
             H: w,
             K: c,
@@ -246,7 +246,8 @@ function y() {
         C = a.n;
         Y = a.o;
         A = a.k;
-        t ? (Y(u.K, u.F), p || (removeEventListener("message", M), K = K.forEach(L), addEventListener("message", L))) : a.i();
+        t ? (Y(u.M, u.K, u.F), p || (removeEventListener("message", M), K = K.forEach(L), 
+        addEventListener("message", L))) : a.i();
         t || X();
     }));
 }

test/codesize/hello_wasm_worker_wasm.expected.js

@@ -67,7 +67,7 @@ function h() {
         y = b.g;
         z = b.i;
         l = b.h;
-        d ? (z(e.m, e.o), removeEventListener("message", p), k = k.forEach(n), addEventListener("message", n)) : b.f();
+        d ? (z(e.s, e.m, e.o), removeEventListener("message", p), k = k.forEach(n), addEventListener("message", n)) : b.f();
         d || y();
     }));
 }

test/codesize/test_minimal_runtime_code_size_audio_worklet.json

@@ -1,10 +1,10 @@
 {
   "a.html": 515,
   "a.html.gz": 355,
-  "a.js": 4309,
-  "a.js.gz": 2217,
-  "a.wasm": 1329,
+  "a.js": 4313,
+  "a.js.gz": 2220,
+  "a.wasm": 1324,
   "a.wasm.gz": 892,
-  "total": 6153,
-  "total_gz": 3464
+  "total": 6152,
+  "total_gz": 3467
 }

test/codesize/test_minimal_runtime_code_size_hello_wasm_worker.json

@@ -1,10 +1,10 @@
 {
   "a.html": 515,
   "a.html.gz": 355,
-  "a.js": 855,
-  "a.js.gz": 543,
-  "a.wasm": 1841,
-  "a.wasm.gz": 1067,
-  "total": 3211,
-  "total_gz": 1965
+  "a.js": 859,
+  "a.js.gz": 545,
+  "a.wasm": 1847,
+  "a.wasm.gz": 1071,
+  "total": 3221,
+  "total_gz": 1971
 }

tools/emscripten.py

@@ -1184,7 +1184,7 @@ def create_pointer_conversion_wrappers(metadata):
     '_emscripten_thread_free_data': '_p',
     '_emscripten_dlsync_self_async': '_p',
     '_emscripten_proxy_dlsync_async': '_pp',
-    '_emscripten_wasm_worker_initialize': '_p_',
+    '_emscripten_wasm_worker_initialize': '__p_',
     '_emscripten_proxy_poll_finish': '_pp_',
     '_wasmfs_rename': '_pp',
     '_wasmfs_readlink': '_pp',