-sMAIN_MODULE=2 with -pthread misses exports __emscripten_thread_crashed and __embind_initialize_bindings

[8051Enthusiast]

I'm using (a custom build of) Qt 6.7 with emscripten 3.1.58, with both -pthread and -sMAIN_MODULE=2.
When trying to run the build in a webbrowser, I'm getting an exception because __emscripten_thread_crashed is undefined, which is caused by an exception because __embind_initialize_bindings is undefined.
Setting -sEXPORTED_FUNCTIONS=_main,__emscripten_thread_crashed,__embind_initialize_bindings works around this (but this wasn't needed on 3.1.55 I think).
I'm guessing DCE deletes these because they're only called from javascript?

emcc --version:

emcc (Emscripten gcc/clang-like replacement + linker emulating GNU ld) 3.1.58 (a41843e0860e52c948c1fce20307933c6631c800)
Copyright (C) 2014 the Emscripten authors (see AUTHORS.txt)
This is free and open source software under the MIT license.
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Linker command:

/opt/emsdk/upstream/emscripten/em++ -O3 -DNDEBUG -sMAIN_MODULE=2 -sSTACK_SIZE=5MB -pthread -s PTHREAD_POOL_SIZE=4 -s INITIAL_MEMORY=50MB -s MAXIMUM_MEMORY=4GB -s EXPORTED_RUNTIME_METHODS=UTF16ToString,stringToUTF16,JSEvents,specialHTMLTargets,FS,callMain,ENV -s EXPORT_NAME=yphbt_entry -s MAX_WEBGL_VERSION=2 -s FETCH=1 -s WASM_BIGINT=1 -s STACK_SIZE=5MB -s MODULARIZE=1 -s DISABLE_EXCEPTION_CATCHING=1 -pthread -s ALLOW_MEMORY_GROWTH -sASYNCIFY_IMPORTS=qt_asyncify_suspend_js,qt_asyncify_resume_js -s ERROR_ON_UNDEFINED_SYMBOLS=1 @CMakeFiles/yphbt.dir/objects1 -o yphbt.js @CMakeFiles/yphbt.dir/linkLibs.rsp

[sbc100]

Looks like a real bug which I imagine was introduced in #21701. I can take a look at this tomorrow.

[sbc100]

Actually I can't see how this is possible since __emscripten_thread_crashed and __embind_initialize_bindings should be included via

emscripten/tools/link.py

Line 488 in 0e4c599

'_emscripten_thread_crashed',

and

emscripten/tools/link.py

Line 496 in 0e4c599

settings.REQUIRED_EXPORTS.append('_embind_initialize_bindings')

.

I'm having trouble reproducing as when I build with the above options I see this in the generated JS file:

    var __emscripten_thread_crashed = () =>                                      
        (__emscripten_thread_crashed = wasmExports["_emscripten_thread_crashed"])();

And the wasm file contains that _emscripten_thread_crashed export.

[8051Enthusiast]

Hmm, for me the js file contained the var __emscripten_thread_crashed = ... line, but the exports don't contain it:

Wasm exports

Export[34]:
 - func[21607] <__wasm_call_ctors> -> "__wasm_call_ctors"
 - func[15916] <__wasm_apply_data_relocs> -> "__wasm_apply_data_relocs"
 - func[213]  -> "free"
 - func[13252] <__main_argc_argv> -> "__main_argc_argv"
 - func[250]  -> "malloc"
 - func[15546]  -> "pthread_self"
 - func[15582]  -> "htonl"
 - func[8231]  -> "htons"
 - func[8231]  -> "ntohs"
 - func[15614] <__getTypeName> -> "__getTypeName"
 - func[15613] <_emscripten_tls_init> -> "_emscripten_tls_init"
 - func[5573]  -> "emscripten_builtin_memalign"
 - func[15612]  -> "setThrew"
 - func[15611] <_emscripten_tempret_set> -> "_emscripten_tempret_set"
 - func[15610]  -> "emscripten_stack_set_limits"
 - func[15609] <_emscripten_stack_restore> -> "_emscripten_stack_restore"
 - func[15608] <_emscripten_stack_alloc> -> "_emscripten_stack_alloc"
 - func[15607]  -> "emscripten_stack_get_current"
 - func[8262] <__dl_seterr> -> "__dl_seterr"
 - func[15600] <_emscripten_dlsync_self_async> -> "_emscripten_dlsync_self_async"
 - func[5644] <_emscripten_dlsync_self> -> "_emscripten_dlsync_self"
 - func[15598] <_emscripten_proxy_dlsync_async> -> "_emscripten_proxy_dlsync_async"
 - func[15594] <_emscripten_proxy_dlsync> -> "_emscripten_proxy_dlsync"
 - func[15587] <_emscripten_thread_init> -> "_emscripten_thread_init"
 - func[15564] <_emscripten_run_on_main_thread_js> -> "_emscripten_run_on_main_thread_js"
 - func[15549] <_emscripten_thread_free_data> -> "_emscripten_thread_free_data"
 - func[8209] <_emscripten_thread_exit> -> "_emscripten_thread_exit"
 - func[8194] <_emscripten_timeout> -> "_emscripten_timeout"
 - func[15534] <_emscripten_check_mailbox> -> "_emscripten_check_mailbox"
 - func[8171] <__cxa_decrement_exception_refcount> -> "__cxa_increment_exception_refcount"
 - func[8171] <__cxa_decrement_exception_refcount> -> "__cxa_decrement_exception_refcount"
 - func[15507] <__cxa_can_catch> -> "__cxa_can_catch"
 - func[15506] <__cxa_is_pointer_type> -> "__cxa_is_pointer_type"
 - func[15285] <_emscripten_run_callback_on_thread> -> "_emscripten_run_callback_on_thread"

Those lines you highlighted in link.py also most likely executed, since I'm getting the warnings right above them.
Should I try to make a more minimal reproducer? For reference, this is the Dockerfile that causes the exception, but it takes around 1-2 hours to build without cache so you probably wouldn't want to build that one.

Edit: Also I forgot to mention, this error does not occur in debug mode

[sbc100]

Anything you can do to reduce the side the reproducer would be great. For example, can you try other opt levels (e.g. -O2 or -O1).

You could also try to bisect to figure out if #21701 was really the change that introduced this issue. See https://emscripten.org/docs/contributing/developers_guide.html?highlight=developer#bisecting

[8051Enthusiast]

Ok, so the -sMAIN_MODULE=2 flag was a red herring, the same problem occurs when disabling it. With -O2 it works properly. The following program (still depending on Qt) has the same error:

#include <QThreadPool>
#include <QFuture>
#include <QtConcurrent>

int main() {
	QThreadPool pool;
	QFuture<void> future = QtConcurrent::run(&pool, [](){});
}

As does this one:

#include <QNetworkAccessManager>
#include <QNetworkReply>
int main() {
    auto manager = new QNetworkAccessManager();
    auto request = QNetworkRequest(QUrl("https://example.com"));
    request.setHeader(QNetworkRequest::ContentTypeHeader, "text/plain");
    manager->post(request, "msg");
}

Removing the calls resolves the issue, even if the same Qt libraries are linked with the same build options.
Bisecting didn't work as the commit for #21701 didn't include the statx symbol, which was included in the commit right after it. The commit with statx still has the problem though.

[8051Enthusiast]

Actually, scrap that about -sMAIN_MODULE=2, I don't know if removing it does anything or not because removing it causes the symbols to be minified and I can't grep for __emscripten_thread_crashed anymore.
I added a stub statx and #21701 does indeed introduce the problem.

[dsamo]

Are there any news here? I am facing the same issue without -sMAIN_MODULE=2. From my tests currently using emsdk 3.1.59 compiling with optimizations -Os, -O3 with pthreads is broken, rest work fine.
Compile flags:
-O3 -Wall -Wextra -Wundef -Werror -Wno-error=pthreads-mem-growth -sDISABLE_EXCEPTION_CATCHING=0 -pthread
Link flags:
--bind -sSTACK_SIZE=1MB -sALLOW_MEMORY_GROWTH=1 -sDISABLE_EXCEPTION_CATCHING=0 -sEXCEPTION_STACK_TRACES=1 -sENVIRONMENT=web,worker -sERROR_ON_UNDEFINED_SYMBOLS=1 -sEXPORTED_RUNTIME_METHODS=ccall -sNO_EXIT_RUNTIME=1 -sMAX_WEBGL_VERSION=2 -sMIN_WEBGL_VERSION=1 -sWASM=1 -sTEXTDECODER=0 -lidbfs.js -sWEBSOCKET_URL='wss://' -Wl,--shared-memory,--no-check-features

It seems while the functions are exported they are not visible in the worker context causing them to crash. This is a screenshot from MS Edge console showing one call when the Javascript context is set to top (main thread) and second call inside context em-pthread:

[andreamancuso]

+1

Normally (on 3.1.56) I'd use the following linker flags:

-O3 -pthread -lembind --embind-emit-tsd interface.d.ts -s ENVIRONMENT='web,worker' -s MODULARIZE=1 -s ALLOW_MEMORY_GROWTH=1 -s WASM=1 -s USE_GLFW=3 -s USE_WEBGPU=1 -s NO_FILESYSTEM=1 -s NO_EXIT_RUNTIME=0 -s STANDALONE_WASM=0 -s EXIT_RUNTIME=1 -s ASSERTIONS=1 -s STACK_OVERFLOW_CHECK=2 -s MIN_WEBGL_VERSION=2 -s MAX_WEBGL_VERSION=2 -s DISABLE_EXCEPTION_CATCHING=0 -s PTHREAD_POOL_SIZE=12

I just upgraded to 3.1.59 and added -s STRICT because of #20580

I got:

wasm-ld: error: C:\Users\[user]\AppData\Local\Temp\tmpm1_qwmwllibemscripten_js_symbols.so: undefined symbol: _emscripten_run_callback_on_thread. Required by emscripten_set_fullscreenchange_callback_on_thread
wasm-ld: error: C:\Users\[user]\AppData\Local\Temp\tmpm1_qwmwllibemscripten_js_symbols.so: undefined symbol: _emscripten_run_callback_on_thread. Required by emscripten_set_wheel_callback_on_thread
wasm-ld: error: C:\Users\[user]\AppData\Local\Temp\tmpm1_qwmwllibemscripten_js_symbols.so: undefined symbol: _emscripten_run_callback_on_thread. Required by emscripten_get_element_css_size
wasm-ld: error: C:\Users\[user]\AppData\Local\Temp\tmpm1_qwmwllibemscripten_js_symbols.so: undefined symbol: _emscripten_run_callback_on_thread. Required by emscripten_set_resize_callback_on_thread

Switching to -O2 appears to solve the issue, however I then get:

Any help with this would be much appreciated.

[sbc100]

The _emscripten_run_callback_on_thread undefined symbol can be fixed by adding -lhtml5.js. This is needed because you added -sSTRICT which disables AUTO_JS_LIBRARIES by default.

[dsamo]

I ended up solving my problem by adding __emscripten_thread_crashed and __embind_initialize_bindings to the EXPORTED_FUNCTIONS. But this further points out that there is a problem here because this should not be needed. A guess of mine is that DCE still eliminates these functions in higher optimizations because wasmExports['_emscripten_thread_crashed'] and wasmExports['_embind_initialize_bindings'] appear as undefined.

Are there any news here? I am facing the same issue without -sMAIN_MODULE=2. From my tests currently using emsdk 3.1.59 compiling with optimizations -Os, -O3 with pthreads is broken, rest work fine. Compile flags: -O3 -Wall -Wextra -Wundef -Werror -Wno-error=pthreads-mem-growth -sDISABLE_EXCEPTION_CATCHING=0 -pthread Link flags: --bind -sSTACK_SIZE=1MB -sALLOW_MEMORY_GROWTH=1 -sDISABLE_EXCEPTION_CATCHING=0 -sEXCEPTION_STACK_TRACES=1 -sENVIRONMENT=web,worker -sERROR_ON_UNDEFINED_SYMBOLS=1 -sEXPORTED_RUNTIME_METHODS=ccall -sNO_EXIT_RUNTIME=1 -sMAX_WEBGL_VERSION=2 -sMIN_WEBGL_VERSION=1 -sWASM=1 -sTEXTDECODER=0 -lidbfs.js -sWEBSOCKET_URL='wss://' -Wl,--shared-memory,--no-check-features

It seems while the functions are exported they are not visible in the worker context causing them to crash. This is a screenshot from MS Edge console showing one call when the Javascript context is set to top (main thread) and second call inside context em-pthread:

[andreamancuso]

Thank you.

I changed the linker flags to

-O3 -pthread -lembind --embind-emit-tsd interface.d.ts -lhtml5.js -lhtml5_webgl.js -s STRICT -s ENVIRONMENT='web,worker' -s MODULARIZE=1 -s ALLOW_MEMORY_GROWTH=1 -s WASM=1 -s USE_GLFW=3 -s USE_WEBGPU=1 -s NO_FILESYSTEM=1 -s NO_EXIT_RUNTIME=0 -s STANDALONE_WASM=0 -s EXIT_RUNTIME=1 -s ASSERTIONS=1 -s STACK_OVERFLOW_CHECK=2 -s MIN_WEBGL_VERSION=2 -s MAX_WEBGL_VERSION=2 -s DISABLE_EXCEPTION_CATCHING=0 -sPTHREAD_POOL_SIZE=12

Sadly I am still getting:

I also tried adding -s EXPORTED_FUNCTIONS=__emscripten_run_callback_on_thread but it didn't make a difference.