Compiler assertion failure when building Bullet (ammo.js) with SSE support enabled.

[juj]

Set up to build with the Emscripten branch https://github.com/juj/emscripten/tree/sse2 , then

git clone https://github.com/juj/ammo.js.git
cd ammo.js
git checkout sse
python make.py

Observed:

The build fails with

Linking CXX static library libBulletCollision.a
[100%] Built target BulletCollision
LLVM ERROR: I->getType() == I->getOperand(0)->getType()
Stack dump:
0.  Program arguments: C:/code/emsdk/clang/fastcomp/build_incoming_vs2010_64/bin/RelWithDebInfo\llc c:\users\jukka\appdata\local\temp\tmpcly5lc\temp.bc -march=js -filetype=asm -o c:\users\jukka\appdata\local\temp\tmp2mrtlm.4.js -O3 -emscripten-max-setjmps=20 
1.  Running pass 'Expand and lower illegal >i32 operations into 32-bit chunks' on module 'c:\users\jukka\appdata\local\temp\tmpcly5lc\temp.bc'.
Traceback (most recent call last):
  File "C:\code\emsdk\emscripten\incoming\emscripten.py", line 1655, in <module>
    _main(environ=os.environ)
  File "C:\code\emsdk\emscripten\incoming\emscripten.py", line 1643, in _main
    temp_files.run_and_clean(lambda: main(
  File "C:\code\emsdk\emscripten\incoming\tools\tempfiles.py", line 39, in run_and_clean
    return func()
  File "C:\code\emsdk\emscripten\incoming\emscripten.py", line 1651, in <lambda>
    DEBUG_CACHE=DEBUG_CACHE,
  File "C:\code\emsdk\emscripten\incoming\emscripten.py", line 1538, in main
    jcache=jcache, temp_files=temp_files, DEBUG=DEBUG, DEBUG_CACHE=DEBUG_CACHE)
  File "C:\code\emsdk\emscripten\incoming\emscripten.py", line 805, in emscript_fast
    start_funcs = backend_output.index(start_funcs_marker)
ValueError: substring not found
Traceback (most recent call last):
  File "C:\code\emsdk\emscripten\incoming\emcc", line 1281, in <module>
    final = shared.Building.emscripten(final, append_ext=False, extra_args=extra_args)
  File "C:\code\emsdk\emscripten\incoming\tools\shared.py", line 1501, in emscripten
    assert os.path.exists(filename + '.o.js'), 'Emscripten failed to generate .js'
AssertionError: Emscripten failed to generate .js

[sunfishcode]

Following the steps above, I get this error:

ImportError: No module named tools.shared

Is there a python package I should install for this?

[kripken]

tools/shared.py is in emscripten. Which make.py should find using EMSCRIPTEN_ROOT in ~/.emscripten. Is that perhaps not set up right?

[sunfishcode]

That was it. I was able to reproduce the failure, and I've now checked in several fixes to emscripten-fastcomp. Locally, I also made a few minor edits to the bullet source code, to eliminate _mm_move_sd in 3 places, like this:

•    v2 = btCastdTo128f(_mm_move_sd(btCastfTo128d(v2), btCastfTo128d(vT))); // z0 z1 z2 0
  

•    v2 = _mm_shuffle_ps(vT, v2, BT_SHUFFLE(0, 1, 2, 3)); // z0 z1 z2 0
  

(though note that I'm not yet completely confident that that's the right shuffle mask in the end). since _mm_move_sd uses float64x2 which we don't yet support.

With these fixes, the build succeeds. It doesn't validate as asm.js because it's not yet built with -s PRECISE_F32. See [0] for backgroud info; I'll look into hacking the build scripts to add that. And, it doesn't yet run successfully yet; there are errors at runtime that I haven't investigated yet.

[0] https://bugzilla.mozilla.org/show_bug.cgi?id=1077175

[sunfishcode]

Ok, with several more fixes checked in, plus one hack, I can now build, validate, and run bullet with SIMD. It still needs the changes to avoid _mm_move_sd mentioned above, and it needs this one hack in builds/temp.js to work around the fact that OdinMonkey doesn't support storeXYZ yet [0]:

-SIMD_float32x4_storeXYZ(HEAPU8, $scevgep, $tmpA$sroa$0$0$vec$extract);
+HEAPF32[$scevgep>>2] = $tmpA$sroa$0$0$vec$extract.x;
+HEAPF32[$scevgep + 4>>2] = $tmpA$sroa$0$0$vec$extract.y;
+HEAPF32[$scevgep + 8>>2] = $tmpA$sroa$0$0$vec$extract.z;

[0] https://bugzilla.mozilla.org/show_bug.cgi?id=1113338

[juj]

Works in current incoming.