fork emscript_fast

1 files changed, 651 insertions, 0 deletions

diff --git a/emscripten.py b/emscripten.py
index 75e6711a..8758df52 100755
--- a/emscripten.py
+++ b/emscripten.py
@@ -706,6 +706,657 @@ Runtime.stackRestore = function(top) { asm['stackRestore'](top) };
 
   outfile.close()
 
+# emscript_fast: emscript'en code using the 'fast' compilation path, using
+#                an LLVM backend
+# FIXME: this is just a copy-paste of normal emscript(), and we trample it
+#        if the proper env var is set (see below). we should refactor to
+#        share code between the two, once emscript_fast stabilizes (or,
+#        leaving it separate like it is will make it trivial to rip out
+#        if the experiment fails)
+
+def emscript_fast(infile, settings, outfile, libraries=[], compiler_engine=None,
+             jcache=None, temp_files=None, DEBUG=None, DEBUG_CACHE=None):
+  """Runs the emscripten LLVM-to-JS compiler. We parallelize as much as possible
+
+  Args:
+    infile: The path to the input LLVM assembly file.
+    settings: JSON-formatted settings that override the values
+      defined in src/settings.js.
+    outfile: The file where the output is written.
+  """
+
+  compiler = path_from_root('src', 'compiler.js')
+
+  # Parallelization: We run 3 phases:
+  #   1 aka 'pre'  : Process types and metadata and so forth, and generate the preamble.
+  #   2 aka 'funcs': Process functions. We can parallelize this, working on each function independently.
+  #   3 aka 'post' : Process globals, generate postamble and finishing touches.
+
+  if DEBUG: logging.debug('emscript: ll=>js')
+
+  if jcache: jcache.ensure()
+
+  # Pre-scan ll and alter settings as necessary
+  if DEBUG: t = time.time()
+  ll = open(infile).read()
+  scan(ll, settings)
+  total_ll_size = len(ll)
+  if DEBUG: logging.debug('  emscript: scan took %s seconds' % (time.time() - t))
+
+  # Split input into the relevant parts for each phase
+
+  if DEBUG: t = time.time()
+
+  pre = []
+  funcs = [] # split up functions here, for parallelism later
+
+  meta_start = ll.find('\n!')
+  if meta_start > 0:
+    meta = ll[meta_start:]
+  else:
+    meta = ''
+    meta_start = -1
+
+  start = ll.find('\n') if ll[0] == ';' else 0 # ignore first line, which contains ; ModuleID = '/dir name'
+
+  func_start = start
+  last = func_start
+  while 1:
+    last = func_start
+    func_start = ll.find('\ndefine ', func_start)
+    if func_start > last:
+      pre.append(ll[last:min(func_start+1, meta_start) if meta_start > 0 else func_start+1] + '\n')
+    if func_start < 0:
+      pre.append(ll[last:meta_start] + '\n')
+      break
+    header = ll[func_start+1:ll.find('\n', func_start+1)+1]
+    end = ll.find('\n}', func_start)
+    last = end+3
+    funcs.append((header, ll[func_start+1:last]))
+    pre.append(header + '}\n')
+    func_start = last
+  ll = None
+
+  if DEBUG and len(meta) > 1024*1024: logging.debug('emscript warning: large amounts of metadata, will slow things down')
+  if DEBUG: logging.debug('  emscript: split took %s seconds' % (time.time() - t))
+
+  if len(funcs) == 0:
+    logging.error('No functions to process. Make sure you prevented LLVM from eliminating them as dead (use EXPORTED_FUNCTIONS if necessary, see the FAQ)')
+
+  #if DEBUG:
+  #  logging.debug('========= pre ================\n')
+  #  logging.debug(''.join(pre))
+  #  logging.debug('========== funcs ===============\n')
+  #  for func in funcs:
+  #    logging.debug('\n// ===\n\n', ''.join(func))
+  #  logging.debug('=========================\n')
+
+  # Save settings to a file to work around v8 issue 1579
+  settings_file = temp_files.get('.txt').name
+  def save_settings():
+    global settings_text
+    settings_text = json.dumps(settings, sort_keys=True)
+    s = open(settings_file, 'w')
+    s.write(settings_text)
+    s.close()
+  save_settings()
+
+  # Phase 1 - pre
+  if DEBUG: t = time.time()
+  pre_file = temp_files.get('.pre.ll').name
+  pre_input = ''.join(pre) + '\n' + meta
+  out = None
+  if jcache:
+    keys = [pre_input, settings_text, ','.join(libraries)]
+    shortkey = jcache.get_shortkey(keys)
+    if DEBUG_CACHE: logging.debug('shortkey', shortkey)
+
+    out = jcache.get(shortkey, keys)
+
+    if DEBUG_CACHE and not out:
+      dfpath = os.path.join(get_configuration().TEMP_DIR, "ems_" + shortkey)
+      dfp = open(dfpath, 'w')
+      dfp.write(pre_input)
+      dfp.write("\n\n========================== settings_text\n\n")
+      dfp.write(settings_text)
+      dfp.write("\n\n========================== libraries\n\n")
+      dfp.write("\n".join(libraries))
+      dfp.close()
+      logging.debug('  cache miss, key data dumped to %s' % dfpath)
+
+    if out and DEBUG: logging.debug('  loading pre from jcache')
+  if not out:
+    open(pre_file, 'w').write(pre_input)
+    #print >> sys.stderr, 'running', str([settings_file, pre_file, 'pre'] + libraries).replace("'/", "'") # see funcs
+    out = jsrun.run_js(compiler, compiler_engine, [settings_file, pre_file, 'pre'] + libraries, stdout=subprocess.PIPE, stderr=STDERR_FILE,
+                       cwd=path_from_root('src'))
+    assert '//FORWARDED_DATA:' in out, 'Did not receive forwarded data in pre output - process failed?'
+    if jcache:
+      if DEBUG: logging.debug('  saving pre to jcache')
+      jcache.set(shortkey, keys, out)
+  pre, forwarded_data = out.split('//FORWARDED_DATA:')
+  forwarded_file = temp_files.get('.json').name
+  pre_input = None
+  open(forwarded_file, 'w').write(forwarded_data)
+  if DEBUG: logging.debug('  emscript: phase 1 took %s seconds' % (time.time() - t))
+
+  indexed_functions = set()
+  forwarded_json = json.loads(forwarded_data)
+  for key in forwarded_json['Functions']['indexedFunctions'].iterkeys():
+    indexed_functions.add(key)
+
+  # Phase 2 - func
+
+  cores = int(os.environ.get('EMCC_CORES') or multiprocessing.cpu_count())
+  assert cores >= 1
+  if cores > 1:
+    intended_num_chunks = int(round(cores * NUM_CHUNKS_PER_CORE))
+    chunk_size = max(MIN_CHUNK_SIZE, total_ll_size / intended_num_chunks)
+    chunk_size += 3*len(meta) # keep ratio of lots of function code to meta (expensive to process, and done in each parallel task)
+    chunk_size = min(MAX_CHUNK_SIZE, chunk_size)
+  else:
+    chunk_size = MAX_CHUNK_SIZE # if 1 core, just use the max chunk size
+
+  if DEBUG: t = time.time()
+  if settings.get('ASM_JS'):
+    settings['EXPORTED_FUNCTIONS'] = forwarded_json['EXPORTED_FUNCTIONS']
+    save_settings()
+
+  chunks = cache_module.chunkify(
+    funcs, chunk_size,
+    jcache.get_cachename('emscript_files') if jcache else None)
+
+  #sys.exit(1)
+  #chunks = [chunks[0]] # pick specific chunks for debugging/profiling
+
+  funcs = None
+
+  if jcache:
+    # load chunks from cache where we can # TODO: ignore small chunks
+    cached_outputs = []
+    def load_from_cache(chunk):
+      keys = [settings_text, forwarded_data, chunk]
+      shortkey = jcache.get_shortkey(keys) # TODO: share shortkeys with later code
+      out = jcache.get(shortkey, keys) # this is relatively expensive (pickling?)
+      if out:
+        cached_outputs.append(out)
+        return False
+      return True
+    chunks = filter(load_from_cache, chunks)
+    if len(cached_outputs) > 0:
+      if out and DEBUG: logging.debug('  loading %d funcchunks from jcache' % len(cached_outputs))
+    else:
+      cached_outputs = []
+
+  # TODO: minimize size of forwarded data from funcs to what we actually need
+
+  if len(chunks) > 0:
+    if cores == 1 and total_ll_size < MAX_CHUNK_SIZE:
+      assert len(chunks) == 1, 'no point in splitting up without multiple cores'
+
+    if DEBUG: logging.debug('  emscript: phase 2 working on %d chunks %s (intended chunk size: %.2f MB, meta: %.2f MB, forwarded: %.2f MB, total: %.2f MB)' % (len(chunks), ('using %d cores' % cores) if len(chunks) > 1 else '', chunk_size/(1024*1024.), len(meta)/(1024*1024.), len(forwarded_data)/(1024*1024.), total_ll_size/(1024*1024.)))
+
+    commands = []
+    for i in range(len(chunks)):
+      funcs_file = temp_files.get('.func_%d.ll' % i).name
+      f = open(funcs_file, 'w')
+      f.write(chunks[i])
+      if not jcache:
+        chunks[i] = None # leave chunks array alive (need its length later)
+      f.write('\n')
+      f.write(meta)
+      f.close()
+      commands.append(
+        (i, funcs_file, meta, settings_file, compiler, forwarded_file, libraries, compiler_engine,# + ['--prof'],
+         DEBUG)
+      )
+
+    if len(chunks) > 1:
+      pool = multiprocessing.Pool(processes=cores)
+      outputs = pool.map(process_funcs, commands, chunksize=1)
+    elif len(chunks) == 1:
+      outputs = [process_funcs(commands[0])]
+
+    commands = None
+
+  else:
+    outputs = []
+
+  if jcache:
+    # save chunks to cache
+    for i in range(len(chunks)):
+      chunk = chunks[i]
+      keys = [settings_text, forwarded_data, chunk]
+      shortkey = jcache.get_shortkey(keys)
+      jcache.set(shortkey, keys, outputs[i])
+    if out and DEBUG and len(chunks) > 0: logging.debug('  saving %d funcchunks to jcache' % len(chunks))
+
+  chunks = None
+
+  if jcache: outputs += cached_outputs # TODO: preserve order
+
+  outputs = [output.split('//FORWARDED_DATA:') for output in outputs]
+  for output in outputs:
+    assert len(output) == 2, 'Did not receive forwarded data in an output - process failed? We only got: ' + output[0][-3000:]
+
+  if DEBUG: logging.debug('  emscript: phase 2 took %s seconds' % (time.time() - t))
+  if DEBUG: t = time.time()
+
+  # merge forwarded data
+  if settings.get('ASM_JS'):
+    all_exported_functions = set(settings['EXPORTED_FUNCTIONS']) # both asm.js and otherwise
+    for additional_export in settings['DEFAULT_LIBRARY_FUNCS_TO_INCLUDE']: # additional functions to export from asm, if they are implemented
+      all_exported_functions.add('_' + additional_export)
+    exported_implemented_functions = set()
+  for func_js, curr_forwarded_data in outputs:
+    curr_forwarded_json = json.loads(curr_forwarded_data)
+    forwarded_json['Types']['hasInlineJS'] = forwarded_json['Types']['hasInlineJS'] or curr_forwarded_json['Types']['hasInlineJS']
+    forwarded_json['Types']['usesSIMD'] = forwarded_json['Types']['usesSIMD'] or curr_forwarded_json['Types']['usesSIMD']
+    forwarded_json['Types']['preciseI64MathUsed'] = forwarded_json['Types']['preciseI64MathUsed'] or curr_forwarded_json['Types']['preciseI64MathUsed']
+    for key, value in curr_forwarded_json['Functions']['blockAddresses'].iteritems():
+      forwarded_json['Functions']['blockAddresses'][key] = value
+    for key in curr_forwarded_json['Functions']['indexedFunctions'].iterkeys():
+      indexed_functions.add(key)
+    if settings.get('ASM_JS'):
+      export_bindings = settings['EXPORT_BINDINGS']
+      export_all = settings['EXPORT_ALL']
+      for key in curr_forwarded_json['Functions']['implementedFunctions'].iterkeys():
+        if key in all_exported_functions or export_all or (export_bindings and key.startswith('_emscripten_bind')):
+          exported_implemented_functions.add(key)
+    for key, value in curr_forwarded_json['Functions']['unimplementedFunctions'].iteritems():
+      forwarded_json['Functions']['unimplementedFunctions'][key] = value
+    for key, value in curr_forwarded_json['Functions']['neededTables'].iteritems():
+      forwarded_json['Functions']['neededTables'][key] = value
+
+  if settings.get('ASM_JS'):
+    parts = pre.split('// ASM_LIBRARY FUNCTIONS\n')
+    if len(parts) > 1:
+      pre = parts[0]
+      outputs.append([parts[1]])
+  funcs_js = [output[0] for output in outputs]
+
+  outputs = None
+  if DEBUG: logging.debug('  emscript: phase 2b took %s seconds' % (time.time() - t))
+  if DEBUG: t = time.time()
+
+  # calculations on merged forwarded data
+  forwarded_json['Functions']['indexedFunctions'] = {}
+  i = settings['FUNCTION_POINTER_ALIGNMENT'] # universal counter
+  if settings['ASM_JS']: i += settings['RESERVED_FUNCTION_POINTERS']*settings['FUNCTION_POINTER_ALIGNMENT']
+  base_fp = i
+  table_counters = {} # table-specific counters
+  alias = settings['ASM_JS'] and settings['ALIASING_FUNCTION_POINTERS']
+  sig = None
+  for indexed in indexed_functions:
+    if alias:
+      sig = forwarded_json['Functions']['implementedFunctions'].get(indexed) or forwarded_json['Functions']['unimplementedFunctions'].get(indexed)
+      assert sig, indexed
+      if sig not in table_counters:
+        table_counters[sig] = base_fp
+      curr = table_counters[sig]
+      table_counters[sig] += settings['FUNCTION_POINTER_ALIGNMENT']
+    else:
+      curr = i
+      i += settings['FUNCTION_POINTER_ALIGNMENT']
+    #logging.debug('function indexing ' + str([indexed, curr, sig]))
+    forwarded_json['Functions']['indexedFunctions'][indexed] = curr # make sure not to modify this python object later - we use it in indexize
+
+  def split_32(x):
+    x = int(x)
+    return '%d,%d,%d,%d' % (x&255, (x >> 8)&255, (x >> 16)&255, (x >> 24)&255)
+
+  indexing = forwarded_json['Functions']['indexedFunctions']
+  def indexize_mem(js):
+    return re.sub(r"\"?'?{{ FI_([\w\d_$]+) }}'?\"?,0,0,0", lambda m: split_32(indexing.get(m.groups(0)[0]) or 0), js)
+  def indexize(js):
+    return re.sub(r"'{{ FI_([\w\d_$]+) }}'", lambda m: str(indexing.get(m.groups(0)[0]) or 0), js)
+
+  blockaddrs = forwarded_json['Functions']['blockAddresses']
+  def blockaddrsize_mem(js):
+    return re.sub(r'"?{{{ BA_([\w\d_$]+)\|([\w\d_$]+) }}}"?,0,0,0', lambda m: split_32(blockaddrs[m.groups(0)[0]][m.groups(0)[1]]), js)
+  def blockaddrsize(js):
+    return re.sub(r'"?{{{ BA_([\w\d_$]+)\|([\w\d_$]+) }}}"?', lambda m: str(blockaddrs[m.groups(0)[0]][m.groups(0)[1]]), js)
+
+  pre = blockaddrsize(blockaddrsize_mem(indexize(indexize_mem(pre))))
+
+  if settings.get('ASM_JS'):
+    # move postsets into the asm module
+    class PostSets: js = ''
+    def handle_post_sets(m):
+      PostSets.js = m.group(0)
+      return '\n'
+    pre = re.sub(r'function runPostSets[^}]+}', handle_post_sets, pre)
+
+  #if DEBUG: outfile.write('// pre\n')
+  outfile.write(pre)
+  pre = None
+
+  #if DEBUG: outfile.write('// funcs\n')
+
+  # forward
+  forwarded_data = json.dumps(forwarded_json)
+  forwarded_file = temp_files.get('.2.json').name
+  open(forwarded_file, 'w').write(indexize(forwarded_data))
+  if DEBUG: logging.debug('  emscript: phase 2c took %s seconds' % (time.time() - t))
+
+  # Phase 3 - post
+  if DEBUG: t = time.time()
+  post_file = temp_files.get('.post.ll').name
+  open(post_file, 'w').write('\n') # no input, just processing of forwarded data
+  out = jsrun.run_js(compiler, compiler_engine, [settings_file, post_file, 'post', forwarded_file] + libraries, stdout=subprocess.PIPE, stderr=STDERR_FILE,
+                     cwd=path_from_root('src'))
+  post, last_forwarded_data = out.split('//FORWARDED_DATA:') # if this fails, perhaps the process failed prior to printing forwarded data?
+  last_forwarded_json = json.loads(last_forwarded_data)
+
+  if settings.get('ASM_JS'):
+    post_funcs, post_rest = post.split('// EMSCRIPTEN_END_FUNCS\n')
+    post = post_rest
+
+    # Move preAsms to their right place
+    def move_preasm(m):
+      contents = m.groups(0)[0]
+      outfile.write(contents + '\n')
+      return ''
+    post_funcs = re.sub(r'/\* PRE_ASM \*/(.*)\n', lambda m: move_preasm(m), post_funcs)
+
+    funcs_js += ['\n' + post_funcs + '// EMSCRIPTEN_END_FUNCS\n']
+
+    simple = os.environ.get('EMCC_SIMPLE_ASM')
+    class Counter:
+      i = 0
+      j = 0
+    pre_tables = last_forwarded_json['Functions']['tables']['pre']
+    del last_forwarded_json['Functions']['tables']['pre']
+
+    def make_table(sig, raw):
+      i = Counter.i
+      Counter.i += 1
+      bad = 'b' + str(i)
+      params = ','.join(['p%d' % p for p in range(len(sig)-1)])
+      coercions = ';'.join(['p%d = %s' % (p, shared.JS.make_coercion('p%d' % p, sig[p+1], settings)) for p in range(len(sig)-1)]) + ';'
+      ret = '' if sig[0] == 'v' else ('return %s' % shared.JS.make_initializer(sig[0], settings))
+      start = raw.index('[')
+      end = raw.rindex(']')
+      body = raw[start+1:end].split(',')
+      for j in range(settings['RESERVED_FUNCTION_POINTERS']):
+        body[settings['FUNCTION_POINTER_ALIGNMENT'] * (1 + j)] = 'jsCall_%s_%s' % (sig, j)
+      Counter.j = 0
+      def fix_item(item):
+        Counter.j += 1
+        newline = Counter.j % 30 == 29
+        if item == '0': return bad if not newline else (bad + '\n')
+        return item if not newline else (item + '\n')
+      body = ','.join(map(fix_item, body))
+      return ('function %s(%s) { %s %s(%d); %s }' % (bad, params, coercions, 'abort' if not settings['ASSERTIONS'] else 'nullFunc', i, ret), ''.join([raw[:start+1], body, raw[end:]]))
+
+    infos = [make_table(sig, raw) for sig, raw in last_forwarded_json['Functions']['tables'].iteritems()]
+
+    function_tables_defs = '\n'.join([info[0] for info in infos]) + '\n// EMSCRIPTEN_END_FUNCS\n' + '\n'.join([info[1] for info in infos])
+
+    asm_setup = ''
+    maths = ['Math.' + func for func in ['floor', 'abs', 'sqrt', 'pow', 'cos', 'sin', 'tan', 'acos', 'asin', 'atan', 'atan2', 'exp', 'log', 'ceil', 'imul']]
+    fundamentals = ['Math', 'Int8Array', 'Int16Array', 'Int32Array', 'Uint8Array', 'Uint16Array', 'Uint32Array', 'Float32Array', 'Float64Array']
+    math_envs = ['Math.min'] # TODO: move min to maths
+    asm_setup += '\n'.join(['var %s = %s;' % (f.replace('.', '_'), f) for f in math_envs])
+
+    if settings['PRECISE_F32']: maths += ['Math.fround']
+
+    basic_funcs = ['abort', 'assert', 'asmPrintInt', 'asmPrintFloat'] + [m.replace('.', '_') for m in math_envs]
+    if settings['RESERVED_FUNCTION_POINTERS'] > 0: basic_funcs.append('jsCall')
+    if settings['SAFE_HEAP']: basic_funcs += ['SAFE_HEAP_LOAD', 'SAFE_HEAP_STORE', 'SAFE_HEAP_CLEAR']
+    if settings['CHECK_HEAP_ALIGN']: basic_funcs += ['CHECK_ALIGN_2', 'CHECK_ALIGN_4', 'CHECK_ALIGN_8']
+    if settings['ASSERTIONS']:
+      basic_funcs += ['nullFunc']
+      asm_setup += 'function nullFunc(x) { Module["printErr"]("Invalid function pointer called. Perhaps a miscast function pointer (check compilation warnings) or bad vtable lookup (maybe due to derefing a bad pointer, like NULL)?"); abort(x) }\n'
+
+    basic_vars = ['STACKTOP', 'STACK_MAX', 'tempDoublePtr', 'ABORT']
+    basic_float_vars = ['NaN', 'Infinity']
+
+    if forwarded_json['Types']['preciseI64MathUsed'] or \
+       forwarded_json['Functions']['libraryFunctions'].get('llvm_cttz_i32') or \
+       forwarded_json['Functions']['libraryFunctions'].get('llvm_ctlz_i32'):
+      basic_vars += ['cttz_i8', 'ctlz_i8']
+
+    if settings.get('DLOPEN_SUPPORT'):
+      for sig in last_forwarded_json['Functions']['tables'].iterkeys():
+        basic_vars.append('F_BASE_%s' % sig)
+        asm_setup += '  var F_BASE_%s = %s;\n' % (sig, 'FUNCTION_TABLE_OFFSET' if settings.get('SIDE_MODULE') else '0') + '\n'
+
+    asm_runtime_funcs = ['stackAlloc', 'stackSave', 'stackRestore', 'setThrew'] + ['setTempRet%d' % i for i in range(10)]
+    # function tables
+    function_tables = ['dynCall_' + table for table in last_forwarded_json['Functions']['tables']]
+    function_tables_impls = []
+
+    for sig in last_forwarded_json['Functions']['tables'].iterkeys():
+      args = ','.join(['a' + str(i) for i in range(1, len(sig))])
+      arg_coercions = ' '.join(['a' + str(i) + '=' + shared.JS.make_coercion('a' + str(i), sig[i], settings) + ';' for i in range(1, len(sig))])
+      coerced_args = ','.join([shared.JS.make_coercion('a' + str(i), sig[i], settings) for i in range(1, len(sig))])
+      ret = ('return ' if sig[0] != 'v' else '') + shared.JS.make_coercion('FUNCTION_TABLE_%s[index&{{{ FTM_%s }}}](%s)' % (sig, sig, coerced_args), sig[0], settings)
+      function_tables_impls.append('''
+  function dynCall_%s(index%s%s) {
+    index = index|0;
+    %s
+    %s;
+  }
+''' % (sig, ',' if len(sig) > 1 else '', args, arg_coercions, ret))
+
+      for i in range(settings['RESERVED_FUNCTION_POINTERS']):
+        jsret = ('return ' if sig[0] != 'v' else '') + shared.JS.make_coercion('jsCall(%d%s%s)' % (i, ',' if coerced_args else '', coerced_args), sig[0], settings)
+        function_tables_impls.append('''
+  function jsCall_%s_%s(%s) {
+    %s
+    %s;
+  }
+
+''' % (sig, i, args, arg_coercions, jsret))
+      shared.Settings.copy(settings)
+      asm_setup += '\n' + shared.JS.make_invoke(sig) + '\n'
+      basic_funcs.append('invoke_%s' % sig)
+      if settings.get('DLOPEN_SUPPORT'):
+        asm_setup += '\n' + shared.JS.make_extcall(sig) + '\n'
+        basic_funcs.append('extCall_%s' % sig)
+
+    # calculate exports
+    exported_implemented_functions = list(exported_implemented_functions)
+    exported_implemented_functions.append('runPostSets')
+    exports = []
+    if not simple:
+      for export in exported_implemented_functions + asm_runtime_funcs + function_tables:
+        exports.append("%s: %s" % (export, export))
+      exports = '{ ' + ', '.join(exports) + ' }'
+    else:
+      exports = '_main'
+    # calculate globals
+    try:
+      del forwarded_json['Variables']['globals']['_llvm_global_ctors'] # not a true variable
+    except:
+      pass
+    # If no named globals, only need externals
+    global_vars = map(lambda g: g['name'], filter(lambda g: settings['NAMED_GLOBALS'] or g.get('external') or g.get('unIndexable'), forwarded_json['Variables']['globals'].values()))
+    global_funcs = ['_' + key for key, value in forwarded_json['Functions']['libraryFunctions'].iteritems() if value != 2]
+    def math_fix(g):
+      return g if not g.startswith('Math_') else g.split('_')[1]
+    asm_global_funcs = ''.join(['  var ' + g.replace('.', '_') + '=global.' + g + ';\n' for g in maths]) + \
+                       ''.join(['  var ' + g + '=env.' + math_fix(g) + ';\n' for g in basic_funcs + global_funcs])
+    asm_global_vars = ''.join(['  var ' + g + '=env.' + g + '|0;\n' for g in basic_vars + global_vars]) + \
+                      ''.join(['  var ' + g + '=+env.' + g + ';\n' for g in basic_float_vars])
+    # In linkable modules, we need to add some explicit globals for global variables that can be linked and used across modules
+    if settings.get('MAIN_MODULE') or settings.get('SIDE_MODULE'):
+      assert settings.get('TARGET_LE32'), 'TODO: support x86 target when linking modules (needs offset of 4 and not 8 here)'
+      for key, value in forwarded_json['Variables']['globals'].iteritems():
+        if value.get('linkable'):
+          init = forwarded_json['Variables']['indexedGlobals'][key] + 8 # 8 is Runtime.GLOBAL_BASE / STATIC_BASE
+          if settings.get('SIDE_MODULE'): init = '(H_BASE+' + str(init) + ')|0'
+          asm_global_vars += '  var %s=%s;\n' % (key, str(init))
+
+    # sent data
+    the_global = '{ ' + ', '.join(['"' + math_fix(s) + '": ' + s for s in fundamentals]) + ' }'
+    sending = '{ ' + ', '.join(['"' + math_fix(s) + '": ' + s for s in basic_funcs + global_funcs + basic_vars + basic_float_vars + global_vars]) + ' }'
+    # received
+    if not simple:
+      receiving = ';\n'.join(['var ' + s + ' = Module["' + s + '"] = asm["' + s + '"]' for s in exported_implemented_functions + function_tables])
+    else:
+      receiving = 'var _main = Module["_main"] = asm;'
+
+    # finalize
+
+    if DEBUG: logging.debug('asm text sizes' + str([map(len, funcs_js), len(asm_setup), len(asm_global_vars), len(asm_global_funcs), len(pre_tables), len('\n'.join(function_tables_impls)), len(function_tables_defs.replace('\n', '\n  ')), len(exports), len(the_global), len(sending), len(receiving)]))
+
+    funcs_js = ['''
+%s
+function asmPrintInt(x, y) {
+  Module.print('int ' + x + ',' + y);// + ' ' + new Error().stack);
+}
+function asmPrintFloat(x, y) {
+  Module.print('float ' + x + ',' + y);// + ' ' + new Error().stack);
+}
+// EMSCRIPTEN_START_ASM
+var asm = (function(global, env, buffer) {
+  %s
+  var HEAP8 = new global.Int8Array(buffer);
+  var HEAP16 = new global.Int16Array(buffer);
+  var HEAP32 = new global.Int32Array(buffer);
+  var HEAPU8 = new global.Uint8Array(buffer);
+  var HEAPU16 = new global.Uint16Array(buffer);
+  var HEAPU32 = new global.Uint32Array(buffer);
+  var HEAPF32 = new global.Float32Array(buffer);
+  var HEAPF64 = new global.Float64Array(buffer);
+''' % (asm_setup, "'use asm';" if not forwarded_json['Types']['hasInlineJS'] and not settings['SIDE_MODULE'] and settings['ASM_JS'] == 1 else "'almost asm';") + '\n' + asm_global_vars + '''
+  var __THREW__ = 0;
+  var threwValue = 0;
+  var setjmpId = 0;
+  var undef = 0;
+  var tempInt = 0, tempBigInt = 0, tempBigIntP = 0, tempBigIntS = 0, tempBigIntR = 0.0, tempBigIntI = 0, tempBigIntD = 0, tempValue = 0, tempDouble = 0.0;
+''' + ''.join(['''
+  var tempRet%d = 0;''' % i for i in range(10)]) + '\n' + asm_global_funcs] + ['  var tempFloat = %s;\n' % ('Math_fround(0)' if settings.get('PRECISE_F32') else '0.0')] + ['''
+// EMSCRIPTEN_START_FUNCS
+function stackAlloc(size) {
+  size = size|0;
+  var ret = 0;
+  ret = STACKTOP;
+  STACKTOP = (STACKTOP + size)|0;
+''' + ('STACKTOP = (STACKTOP + 3)&-4;' if settings['TARGET_X86'] else 'STACKTOP = (STACKTOP + 7)&-8;') + '''
+  return ret|0;
+}
+function stackSave() {
+  return STACKTOP|0;
+}
+function stackRestore(top) {
+  top = top|0;
+  STACKTOP = top;
+}
+function setThrew(threw, value) {
+  threw = threw|0;
+  value = value|0;
+  if ((__THREW__|0) == 0) {
+    __THREW__ = threw;
+    threwValue = value;
+  }
+}
+function copyTempFloat(ptr) {
+  ptr = ptr|0;
+  HEAP8[tempDoublePtr] = HEAP8[ptr];
+  HEAP8[tempDoublePtr+1|0] = HEAP8[ptr+1|0];
+  HEAP8[tempDoublePtr+2|0] = HEAP8[ptr+2|0];
+  HEAP8[tempDoublePtr+3|0] = HEAP8[ptr+3|0];
+}
+function copyTempDouble(ptr) {
+  ptr = ptr|0;
+  HEAP8[tempDoublePtr] = HEAP8[ptr];
+  HEAP8[tempDoublePtr+1|0] = HEAP8[ptr+1|0];
+  HEAP8[tempDoublePtr+2|0] = HEAP8[ptr+2|0];
+  HEAP8[tempDoublePtr+3|0] = HEAP8[ptr+3|0];
+  HEAP8[tempDoublePtr+4|0] = HEAP8[ptr+4|0];
+  HEAP8[tempDoublePtr+5|0] = HEAP8[ptr+5|0];
+  HEAP8[tempDoublePtr+6|0] = HEAP8[ptr+6|0];
+  HEAP8[tempDoublePtr+7|0] = HEAP8[ptr+7|0];
+}
+''' + ''.join(['''
+function setTempRet%d(value) {
+  value = value|0;
+  tempRet%d = value;
+}
+''' % (i, i) for i in range(10)])] + [PostSets.js + '\n'] + funcs_js + ['''
+  %s
+
+  return %s;
+})
+// EMSCRIPTEN_END_ASM
+(%s, %s, buffer);
+%s;
+''' % (pre_tables + '\n'.join(function_tables_impls) + '\n' + function_tables_defs.replace('\n', '\n  '), exports, the_global, sending, receiving)]
+
+    if not settings.get('SIDE_MODULE'):
+      funcs_js.append('''
+Runtime.stackAlloc = function(size) { return asm['stackAlloc'](size) };
+Runtime.stackSave = function() { return asm['stackSave']() };
+Runtime.stackRestore = function(top) { asm['stackRestore'](top) };
+''')
+
+    # Set function table masks
+    masks = {}
+    max_mask = 0
+    for sig, table in last_forwarded_json['Functions']['tables'].iteritems():
+      mask = table.count(',')
+      masks[sig] = str(mask)
+      max_mask = max(mask, max_mask)
+    def function_table_maskize(js, masks):
+      def fix(m):
+        sig = m.groups(0)[0]
+        return masks[sig]
+      return re.sub(r'{{{ FTM_([\w\d_$]+) }}}', lambda m: fix(m), js) # masks[m.groups(0)[0]]
+    funcs_js = map(lambda js: function_table_maskize(js, masks), funcs_js)
+
+    if settings.get('DLOPEN_SUPPORT'):
+      funcs_js.append('''
+  asm.maxFunctionIndex = %(max_mask)d;
+  DLFCN.registerFunctions(asm, %(max_mask)d+1, %(sigs)s, Module);
+  Module.SYMBOL_TABLE = SYMBOL_TABLE;
+''' % { 'max_mask': max_mask, 'sigs': str(map(str, last_forwarded_json['Functions']['tables'].keys())) })
+
+  else:
+    function_tables_defs = '\n'.join([table for table in last_forwarded_json['Functions']['tables'].itervalues()])
+    outfile.write(function_tables_defs)
+    funcs_js = ['''
+// EMSCRIPTEN_START_FUNCS
+'''] + funcs_js + ['''
+// EMSCRIPTEN_END_FUNCS
+''']
+
+  # Create symbol table for self-dlopen
+  if settings.get('DLOPEN_SUPPORT'):
+    symbol_table = {}
+    for k, v in forwarded_json['Variables']['indexedGlobals'].iteritems():
+       if forwarded_json['Variables']['globals'][k]['named']:
+         symbol_table[k] = str(v + forwarded_json['Runtime']['GLOBAL_BASE'])
+    for raw in last_forwarded_json['Functions']['tables'].itervalues():
+      if raw == '': continue
+      table = map(string.strip, raw[raw.find('[')+1:raw.find(']')].split(","))
+      for i in range(len(table)):
+        value = table[i]
+        if value != '0':
+          if settings.get('SIDE_MODULE'):
+            symbol_table[value] = 'FUNCTION_TABLE_OFFSET+' + str(i)
+          else:
+            symbol_table[value] = str(i)
+    outfile.write("var SYMBOL_TABLE = %s;" % json.dumps(symbol_table).replace('"', ''))
+
+  for i in range(len(funcs_js)): # do this loop carefully to save memory
+    funcs_js_item = funcs_js[i]
+    funcs_js[i] = None
+    funcs_js_item = indexize(funcs_js_item)
+    funcs_js_item = blockaddrsize(funcs_js_item)
+    outfile.write(funcs_js_item)
+  funcs_js = None
+
+  outfile.write(indexize(post))
+  if DEBUG: logging.debug('  emscript: phase 3 took %s seconds' % (time.time() - t))
+
+  outfile.close()
+
+if os.environ.get('EMCC_FAST_COMPILER'):
+  emscript = emscript_fast
+
 def main(args, compiler_engine, cache, jcache, relooper, temp_files, DEBUG, DEBUG_CACHE):
   # Prepare settings for serialization to JSON.
   settings = {}