See https://github.com/kripken/emscripten/pull/7512 for details.

Also don't pass js as the second arg since its the default.

This means the EMTEST_SAVE_DIR works for these directories too

It also means we don't need to do the cleanup ourselves but can
rely on the common cleanup code in runner.py.

Also, remove a couple of cases where an extra temp directory
was used for no reason.

This makes the EMCC_DEBUG-logged commands, like wasm-ld, use inputs that are not cleaned up when emcc.py exits, so that when debugging you can literally copy and paste those commands from the EMCC_DEBUG logging to re-run them.

show an error when trying to do OUTLINING_LIMIT with the LLVM wasm backend. this gives us a clear error, and also prevents some wasted work on CI (#7711)

This makes memcpy around 572 bytes, as opposed to 1,400 which is the size of musl's. This is noticeable on small programs where this can save 2-3% or so.

The memcpy impl here is a conversion from the asm.js one which we benchmarked extensively in the past.

Note that this only affects the wasm backend, as we build this into compiler-rt.

Also fix the compilation of compiler-rt, which did not include -D__EMSCRIPTEN__ before.

Fixes #7639

We can now use emsdk [install|activate] latest-upstream to get the wasm backend builds, without hacks to get the lkgr etc. The emsdk also fetches a proper node for us, so we can remove that hack as well.

* no-exit-runtime for wasm backend

* update

* Enable usage of binaryen port with wasm backend

* remove unneeded test and make metadce test not care much about llvm backend code sizes

* fix

* [ci skip]

* run no-exit-runtime first, so that opts we run later can remove the import if no longer used

See #7692

The prefix of an import may differ from the prefix in js, in that case (because of how prefixing is different between the wasm backend and asm2wasm, extra _ for the latter).

This gets the wasm backend to parity with asm2wasm in metadce removing unnecessary JS library code.

openposixtestsuite: pthread_detach() test 2-1.c does not want to see pthread_detach() function return ESRCH return value, but always just EINVAL.

The second (`sig`) argument to `addFunction` is only used under certain
specific circumstance. Only assert in this code path, but preserve the
warning in all cases.

This takes a wasm function and adds it directly to the table.

This is useful for `dlsym()` which knows that it has a native wasm
function but doesn't know its signature so cannot supply the second
argument to `addFunctio()`.

It takes 273s on my machine.

Normalize access to EM_CACHE variable, and avoid tripping up when comparing paths in cache.get(), in case they are the same modulo back vs forward slashes or letter case

Importing the table from JS matches the behaviour of ams2wasm, allows
for table manipulation from JS (e.g in addFunction) and is essential for
dynamic linking where the table must be shared between modules.

This change also sets wasmTableSize in the wasm backend path based on
the metadata provided by wasm-emscripten-finalize.

Clean up our handling of feature flags, and pass the proper ones to asm2wasm and wasm-opt. This should fix current pthreads test breakage.

(in the benchmark script we use wasm-opt in a very specific way, and can just enable all the flags there)

This fixes current SIMD breakage, as we just added binaryen SIMD support, and need to tell emscripten not to use it.

- Remove single argument use of `self.in_dir()`.
  All tests run in their own directory, so `self.in_dir('foo')` is
  normally equivialent to just `'foo'`
- Same for os.path.join(self.get_dir(), ...)
- Use new utility function `create_test_file` in place of open('w')
  for consistency.

This avoids having to set settings back to their default value
which results in longer command lines than necessary.

We had a place in emscripten.py where we manually set up stackAlloc etc. This nonstandard location confused metadce, which saw that those functions had extra definitions and uses and hence did not eliminate them.

Removing that extra location reveals the possible reason for it, that it handled the _ prefix difference. Adding those special functions to the list of other special functions (that don't need prefixing) appears makes things work again.

This tests how well metadce does, counting exports and imports and code size etc.

I increased the code size slack to 10%, as there may be more churn on the wasm backend side due to LLVM codegen changes.