Merge branch 'incoming'

5 files changed, 827 insertions, 46 deletions

diff --git a/tools/file_packager.py b/tools/file_packager.py
index 1443d165..cc030f59 100644
--- a/tools/file_packager.py
+++ b/tools/file_packager.py
@@ -39,7 +39,7 @@ TODO:        You can also provide .crn files yourself, pre-crunched. With this o
              to dds files in the browser, exactly the same as if this tool compressed them.
 '''
 
-import os, sys, shutil, random, uuid
+import os, sys, shutil, random, uuid, ctypes
 
 import shared
 from shared import Compression, execute, suffix, unsuffixed
@@ -50,6 +50,8 @@ if len(sys.argv) == 1:
 See the source for more details.'''
   sys.exit(0)
 
+DEBUG = os.environ.get('EMCC_DEBUG')
+
 data_target = sys.argv[1]
 
 IMAGE_SUFFIXES = ('.jpg', '.png', '.bmp')
@@ -150,6 +152,32 @@ function assert(check, msg) {
 }
 '''
 
+# Win32 code to test whether the given file has the hidden property set.
+def has_hidden_attribute(filepath):
+  if sys.platform != 'win32':
+    return False
+    
+  try:
+    attrs = ctypes.windll.kernel32.GetFileAttributesW(unicode(filepath))
+    assert attrs != -1
+    result = bool(attrs & 2)
+  except (AttributeError, AssertionError):
+    result = False
+  return result
+
+# The packager should never preload/embed any directories that have a component starting with '.' in them,
+# or if the file is hidden (Win32). Note that this filter ONLY applies to directories. Explicitly specified single files
+# are always preloaded/embedded, even if they start with a '.'.
+def should_ignore(filename):
+  if has_hidden_attribute(filename):
+    return True
+    
+  components = filename.replace('\\\\', '/').replace('\\', '/').split('/')
+  for c in components:
+    if c.startswith('.') and c != '.' and c != '..':
+      return True
+  return False
+
 # Expand directories into individual files
 def add(arg, dirname, names):
   # rootpathsrc: The path name of the root directory on the local FS we are adding to emscripten virtual FS.
@@ -158,8 +186,12 @@ def add(arg, dirname, names):
   for name in names:
     fullname = os.path.join(dirname, name)
     if not os.path.isdir(fullname):
-      dstpath = os.path.join(rootpathdst, os.path.relpath(fullname, rootpathsrc)) # Convert source filename relative to root directory of target FS.
-      data_files.append({ 'srcpath': fullname, 'dstpath': dstpath, 'mode': mode })
+      if should_ignore(fullname):
+        if DEBUG:
+          print >> sys.stderr, 'Skipping hidden file "' + fullname + '" from inclusion in the emscripten virtual file system.'
+      else:
+        dstpath = os.path.join(rootpathdst, os.path.relpath(fullname, rootpathsrc)) # Convert source filename relative to root directory of target FS.
+        data_files.append({ 'srcpath': fullname, 'dstpath': dstpath, 'mode': mode })
 
 for file_ in data_files:
   if os.path.isdir(file_['srcpath']):
@@ -171,6 +203,8 @@ for file_ in data_files:
   if file_['dstpath'].endswith('/'): # If user has submitted a directory name as the destination but omitted the destination filename, use the filename from source file
     file_['dstpath'] = file_['dstpath'] + os.path.basename(file_['srcpath'])
   if file_['dstpath'].startswith('./'): file_['dstpath'] = file_['dstpath'][2:] # remove redundant ./ prefix
+  if DEBUG:
+    print >> sys.stderr, 'Packaging file "' + file_['srcpath'] + '" to VFS in path "' + file_['dstpath'] + '".'
 
 # Remove duplicates (can occur naively, for example preload dir/, preload dir/subdir/)
 seen = {}
@@ -202,7 +236,7 @@ if crunch:
     function requestDecrunch(filename, data, callback) {
       decrunchWorker.postMessage({
         filename: filename,
-        data: data,
+        data: new Uint8Array(data),
         callbackID: decrunchCallbacks.length
       });
       decrunchCallbacks.push(callback);
diff --git a/tools/js-optimizer.js b/tools/js-optimizer.js
index 07317e0a..b02ae3cc 100644
--- a/tools/js-optimizer.js
+++ b/tools/js-optimizer.js
@@ -429,19 +429,47 @@ function simplifyExpressionsPre(ast) {
   // 'useful' mathops already |0 anyhow.
 
   function simplifyBitops(ast) {
-    var SAFE_BINARY_OPS = set('+', '-', '*'); // division is unsafe as it creates non-ints in JS; mod is unsafe as signs matter so we can't remove |0's
+    var SAFE_BINARY_OPS;
+    if (asm) {
+      SAFE_BINARY_OPS = set('+', '-'); // division is unsafe as it creates non-ints in JS; mod is unsafe as signs matter so we can't remove |0's; mul does not nest with +,- in asm
+    } else {
+      SAFE_BINARY_OPS = set('+', '-', '*');
+    }
+    var COERCION_REQUIRING_OPS = set('sub', 'unary-prefix'); // ops that in asm must be coerced right away
+    var COERCION_REQUIRING_BINARIES = set('*', '/', '%'); // binary ops that in asm must be coerced
     var ZERO = ['num', 0];
-    var rerun = true;
-    while (rerun) {
-      rerun = false;
-      traverse(ast, function process(node, type, stack) {
-        if (type == 'binary' && node[1] == '|') {
-          if (node[2][0] == 'num' && node[3][0] == 'num') {
-            return ['num', node[2][1] | node[3][1]];
-          } else if (jsonCompare(node[2], ZERO) || jsonCompare(node[3], ZERO)) {
+
+    function removeMultipleOrZero() {
+      var rerun = true;
+      while (rerun) {
+        rerun = false;
+        traverse(ast, function process(node, type, stack) {
+          if (type == 'binary' && node[1] == '|') {
+            if (node[2][0] == 'num' && node[3][0] == 'num') {
+              return ['num', node[2][1] | node[3][1]];
+            }
+            var go = false;
+            if (jsonCompare(node[2], ZERO)) {
+              // canonicalize order
+              var temp = node[3];
+              node[3] = node[2];
+              node[2] = temp;
+              go = true;
+            } else if (jsonCompare(node[3], ZERO)) {
+              go = true;
+            }
+            if (!go) {
+              stack.push(1);
+              return;
+            }
             // We might be able to remove this correction
             for (var i = stack.length-1; i >= 0; i--) {
-              if (stack[i] == 1) {
+              if (stack[i] >= 1) {
+                if (asm) {
+                  if (stack[stack.length-1] < 2 && node[2][0] == 'call') break; // we can only remove multiple |0s on these
+                  if (stack[stack.length-1] < 1 && (node[2][0] in COERCION_REQUIRING_OPS ||
+                                                    (node[2][0] == 'binary' && node[2][1] in COERCION_REQUIRING_BINARIES))) break; // we can remove |0 or >>2
+                }
                 // we will replace ourselves with the non-zero side. Recursively process that node.
                 var result = jsonCompare(node[2], ZERO) ? node[3] : node[2], other;
                 // replace node in-place
@@ -453,23 +481,23 @@ function simplifyExpressionsPre(ast) {
                 return process(result, result[0], stack);
               } else if (stack[i] == -1) {
                 break; // Too bad, we can't
-              } else if (asm) {
-                break; // we must keep a coercion right on top of a heap access in asm mode
               }
             }
+            stack.push(2); // From here on up, no need for this kind of correction, it's done at the top
+                           // (Add this at the end, so it is only added if we did not remove it)
+          } else if (type == 'binary' && node[1] in USEFUL_BINARY_OPS) {
+            stack.push(1);
+          } else if ((type == 'binary' && node[1] in SAFE_BINARY_OPS) || type == 'num' || type == 'name') {
+            stack.push(0); // This node is safe in that it does not interfere with this optimization
+          } else {
+            stack.push(-1); // This node is dangerous! Give up if you see this before you see '1'
           }
-          stack.push(1); // From here on up, no need for this kind of correction, it's done at the top
-                         // (Add this at the end, so it is only added if we did not remove it)
-        } else if (type == 'binary' && node[1] in USEFUL_BINARY_OPS) {
-          stack.push(1);
-        } else if ((type == 'binary' && node[1] in SAFE_BINARY_OPS) || type == 'num' || type == 'name') {
-          stack.push(0); // This node is safe in that it does not interfere with this optimization
-        } else {
-          stack.push(-1); // This node is dangerous! Give up if you see this before you see '1'
-        }
-      }, null, []);
+        }, null, []);
+      }
     }
 
+    removeMultipleOrZero();
+
     // & and heap-related optimizations
 
     var heapBits, heapUnsigned;
@@ -480,7 +508,7 @@ function simplifyExpressionsPre(ast) {
       return true;
     }
 
-    var hasTempDoublePtr = false;
+    var hasTempDoublePtr = false, rerunOrZeroPass = false;
 
     traverse(ast, function(node, type) {
       if (type == 'name') {
@@ -515,7 +543,6 @@ function simplifyExpressionsPre(ast) {
                  node[2][0] == 'binary' && node[2][1] == '<<' && node[2][3][0] == 'num' &&
                  node[2][2][0] == 'sub' && node[2][2][1][0] == 'name') {
         // collapse HEAPU?8[..] << 24 >> 24 etc. into HEAP8[..] | 0
-        // TODO: run this before | 0 | 0 removal, because we generate | 0
         var amount = node[3][1];
         var name = node[2][2][1][1];
         if (amount == node[2][3][1] && parseHeap(name)) {
@@ -524,6 +551,7 @@ function simplifyExpressionsPre(ast) {
             node[1] = '|';
             node[2] = node[2][2];
             node[3][1] = 0;
+            rerunOrZeroPass = true;
             return node;
           }
         }
@@ -556,6 +584,8 @@ function simplifyExpressionsPre(ast) {
       }
     });
 
+    if (rerunOrZeroPass) removeMultipleOrZero();
+
     if (asm) {
       if (hasTempDoublePtr) {
         traverse(ast, function(node, type) {
@@ -707,18 +737,11 @@ function simplifyExpressionsPre(ast) {
   }
 
   function asmOpts(fun) {
-    // 1. Add final returns when necessary
-    // 2. Remove unneeded coercions on function calls that have no targets (eliminator removed it)
+    // Add final returns when necessary
     var returnType = null;
     traverse(fun, function(node, type) {
       if (type == 'return' && node[1]) {
         returnType = detectAsmCoercion(node[1]);
-      } else if (type == 'stat') {
-        var inner = node[1];
-        if ((inner[0] == 'binary' && inner[1] in ASSOCIATIVE_BINARIES && inner[2][0] == 'call' && inner[3][0] == 'num') ||
-            (inner[0] == 'unary-prefix' && inner[1] == '+' && inner[2][0] == 'call')) {
-          node[1] = inner[2];
-        }
       }
     });
     // Add a final return if one is missing.
diff --git a/tools/shared.py b/tools/shared.py
index 8a172d9c..c16c9115 100644
--- a/tools/shared.py
+++ b/tools/shared.py
@@ -273,7 +273,7 @@ def check_llvm_version():
   except Exception, e:
     logging.warning('Could not verify LLVM version: %s' % str(e))
 
-EXPECTED_NODE_VERSION = (0,6,8)
+EXPECTED_NODE_VERSION = (0,8,0)
 
 def check_node_version():
   try:
@@ -295,7 +295,7 @@ def check_node_version():
 # we re-check sanity when the settings are changed)
 # We also re-check sanity and clear the cache when the version changes
 
-EMSCRIPTEN_VERSION = '1.4.9'
+EMSCRIPTEN_VERSION = '1.5.0'
 
 def generate_sanity():
   return EMSCRIPTEN_VERSION + '|' + get_llvm_target()
@@ -454,9 +454,12 @@ EMSCRIPTEN_TEMP_DIR = configuration.EMSCRIPTEN_TEMP_DIR
 DEBUG_CACHE = configuration.DEBUG_CACHE
 CANONICAL_TEMP_DIR = configuration.CANONICAL_TEMP_DIR
 
-level = logging.DEBUG if os.environ.get('EMCC_DEBUG') else logging.INFO
-logging.basicConfig(level=level, format='%(levelname)-8s %(name)s: %(message)s')
-  
+logging.basicConfig(format='%(levelname)-8s %(name)s: %(message)s')
+def set_logging():
+  logger = logging.getLogger()
+  logger.setLevel(logging.DEBUG if os.environ.get('EMCC_DEBUG') else logging.INFO)
+set_logging()
+
 if not EMSCRIPTEN_TEMP_DIR:
   EMSCRIPTEN_TEMP_DIR = tempfile.mkdtemp(prefix='emscripten_temp_', dir=configuration.TEMP_DIR)
   def clean_temp():
@@ -1091,7 +1094,7 @@ set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)''' % { 'winfix': '' if not WINDOWS e
 
   @staticmethod
   def can_build_standalone():
-    return not Settings.BUILD_AS_SHARED_LIB and not Settings.LINKABLE
+    return not Settings.BUILD_AS_SHARED_LIB and not Settings.LINKABLE and not Settings.EXPORT_ALL
 
   @staticmethod
   def can_use_unsafe_opts():
@@ -1285,6 +1288,9 @@ set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)''' % { 'winfix': '' if not WINDOWS e
       emcc_debug = os.environ.get('EMCC_DEBUG')
       if emcc_debug: del os.environ['EMCC_DEBUG']
 
+      emcc_optimize_normally = os.environ.get('EMCC_OPTIMIZE_NORMALLY')
+      if emcc_optimize_normally: del os.environ['EMCC_OPTIMIZE_NORMALLY']
+
       def make(opt_level):
         raw = relooper + '.raw.js'
         Building.emcc(os.path.join('relooper', 'Relooper.cpp'), ['-I' + os.path.join('relooper'), '--post-js',
@@ -1315,6 +1321,7 @@ set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)''' % { 'winfix': '' if not WINDOWS e
     finally:
       os.chdir(curr)
       if emcc_debug: os.environ['EMCC_DEBUG'] = emcc_debug
+      if emcc_optimize_normally: os.environ['EMCC_OPTIMIZE_NORMALLY'] = emcc_optimize_normally
       if not ok:
         logging.error('bootstrapping relooper failed. You may need to manually create relooper.js by compiling it, see src/relooper/emscripten')
         1/0
diff --git a/tools/test-js-optimizer-asm-pre-output.js b/tools/test-js-optimizer-asm-pre-output.js
index 25d521ab..0e95580f 100644
--- a/tools/test-js-optimizer-asm-pre-output.js
+++ b/tools/test-js-optimizer-asm-pre-output.js
@@ -6,9 +6,19 @@ function a() {
   f(8);
   HEAP[1024] = 5;
   HEAP[1024] = 5;
-  whee(12, 13);
-  whee(12, 13);
+  whee(12, 13) | 0;
+  +whee(12, 13);
   f((g = t(), g + g | 0) | 0);
+  f() | 0;
+  f((h() | 0) + 5 | 0);
+  f(x + y + z | 0);
+  +f();
+  f(+(+h() + 5));
+  $140 = $p_3_i + (-$mantSize_0_i | 0) | 0;
+  f(g() | 0);
+  f(g() | 0 & -1);
+  f((g() | 0) >> 2);
+  $56 = _fcntl() | 0 | 1;
 }
 function b($this, $__n) {
   $this = $this | 0;
@@ -34,7 +44,7 @@ function b($this, $__n) {
     $23 = HEAP32[($this + 4 & 16777215) >> 2] | 0;
   }
   if (($14 - $23 | 0) >>> 0 < $__n >>> 0) {
-    __ZNSt3__112basic_stringIcNS_11char_traitsIcEENS_9allocatorIcEEE9__grow_byEjjjjjj($this, $14, ($__n - $14 | 0) + $23 | 0, $23, $23);
+    __ZNSt3__112basic_stringIcNS_11char_traitsIcEENS_9allocatorIcEEE9__grow_byEjjjjjj($this, $14, $__n - $14 + $23 | 0, $23, $23);
     $30 = HEAP8[$4 & 16777215] | 0;
   } else {
     $30 = $13;
@@ -161,4 +171,352 @@ function boxx($this, $aabb, $xf, $childIndex) {
   HEAPF32[$51 + 4 >> 2] = $_sroa_0_0_insert_insert$1;
   return;
 }
+function _main($argc, $argv) {
+  $argc = $argc | 0;
+  $argv = $argv | 0;
+  var $def_i21 = 0, $def_i = 0, $world = 0, $bd = 0, $shape = 0, $shape1 = 0, $bd2 = 0, $result = 0, $6 = 0, $WARMUP_0 = 0, $14 = 0, $15 = 0, $17 = 0, $i_09_i_i = 0, $j_08_i_i = 0, $34 = 0, $j_1_i_i = 0, $38 = 0, $46 = 0, $48 = 0, $50 = 0, $54 = 0, $i_05_i_i_i = 0, $56 = 0, $62 = 0, $_lcssa_i_i_i = 0, $87 = 0, $96 = 0, $97 = 0, $98 = 0, $112 = 0, $115 = 0, $116 = 0, $118 = 0, $121 = 0, $126 = 0, $135 = 0, $137 = 0, $174 = 0, $176 = 0, $177 = 0, $178 = 0, $179 = 0, $180 = 0, $181 = 0, $182 = 0, $183 = 0, $185 = 0, $186 = 0, $188 = 0, $189 = 0, $190 = 0, $191 = 0, $192 = 0, $193 = 0, $194 = 0, $195 = 0, $196 = 0, $i_057 = 0, $x_sroa_0_0_load303656 = +0, $x_sroa_1_4_load313755 = +0, $j_052 = 0, $y_sroa_0_0_load283451 = +0, $y_sroa_1_4_load293550 = +0, $y_sroa_0_0_insert_insert$1 = +0, $205 = 0, $208 = 0, $209 = 0, $213 = 0, $223 = 0, $236 = 0, $i3_042 = 0, $241 = 0, $242 = 0, $243 = 0, $i4_038 = 0, $245 = 0, $260 = +0, $_0 = 0, label = 0, __stackBase__ = 0;
+  __stackBase__ = STACKTOP;
+  STACKTOP = STACKTOP + 103416 | 0;
+  $def_i21 = __stackBase__ | 0;
+  $def_i = __stackBase__ + 32 | 0;
+  $world = __stackBase__ + 64 | 0;
+  $bd = __stackBase__ + 103096 | 0;
+  $shape = __stackBase__ + 103152 | 0;
+  $shape1 = __stackBase__ + 103200 | 0;
+  $bd2 = __stackBase__ + 103352 | 0;
+  $result = __stackBase__ + 103408 | 0;
+  do {
+    if (($argc | 0) > 1) {
+      $6 = HEAP8[HEAP32[$argv + 4 >> 2] | 0] | 0;
+      if (($6 | 0) == 49) {
+        HEAP32[2414] = 35;
+        $WARMUP_0 = 5;
+        break;
+      } else if (($6 | 0) == 50) {
+        HEAP32[2414] = 161;
+        $WARMUP_0 = 32;
+        break;
+      } else if (($6 | 0) == 51) {
+        label = 43;
+        break;
+      } else if (($6 | 0) == 52) {
+        HEAP32[2414] = 2331;
+        $WARMUP_0 = 320;
+        break;
+      } else if (($6 | 0) == 53) {
+        HEAP32[2414] = 5661;
+        $WARMUP_0 = 640;
+        break;
+      } else if (($6 | 0) == 48) {
+        $_0 = 0;
+        STACKTOP = __stackBase__;
+        return $_0 | 0;
+      } else {
+        _printf(3512, (tempInt = STACKTOP, STACKTOP = STACKTOP + 8 | 0, HEAP32[tempInt >> 2] = $6 - 48, tempInt) | 0) | 0;
+        $_0 = -1;
+        STACKTOP = __stackBase__;
+        return $_0 | 0;
+      }
+    } else {
+      label = 43;
+    }
+  } while (0);
+  if ((label | 0) == 43) {
+    HEAP32[2414] = 333;
+    $WARMUP_0 = 64;
+  }
+  $14 = $world | 0;
+  $15 = $world + 8 | 0;
+  HEAP32[$15 >> 2] = 128;
+  HEAP32[$world + 4 >> 2] = 0;
+  $17 = _malloc(1024) | 0;
+  HEAP32[$world >> 2] = $17;
+  _memset($17 | 0, 0, HEAP32[$15 >> 2] << 3 | 0);
+  _memset($world + 12 | 0, 0, 56);
+  $j_08_i_i = 0;
+  $i_09_i_i = 1;
+  while (1) {
+    if (!(($j_08_i_i | 0) < 14)) {
+      label = 49;
+      break;
+    }
+    if (($i_09_i_i | 0) > (HEAP32[9600 + ($j_08_i_i << 2) >> 2] | 0)) {
+      $34 = $j_08_i_i + 1 | 0;
+      HEAP8[$i_09_i_i + 8952 | 0] = $34 & 255;
+      $j_1_i_i = $34;
+    } else {
+      HEAP8[$i_09_i_i + 8952 | 0] = $j_08_i_i & 255;
+      $j_1_i_i = $j_08_i_i;
+    }
+    $38 = $i_09_i_i + 1 | 0;
+    if (($38 | 0) < 641) {
+      $j_08_i_i = $j_1_i_i;
+      $i_09_i_i = $38;
+    } else {
+      break;
+    }
+  }
+  if ((label | 0) == 49) {
+    ___assert_func(3248, 73, 6448, 3360);
+    return 0;
+  }
+  HEAP32[$world + 102468 >> 2] = 0;
+  HEAP32[$world + 102472 >> 2] = 0;
+  HEAP32[$world + 102476 >> 2] = 0;
+  HEAP32[$world + 102864 >> 2] = 0;
+  HEAP32[$world + 102872 >> 2] = -1;
+  $46 = $world + 102884 | 0;
+  HEAP32[$46 >> 2] = 16;
+  HEAP32[$world + 102880 >> 2] = 0;
+  $48 = _malloc(576) | 0;
+  $50 = $world + 102876 | 0;
+  HEAP32[$50 >> 2] = $48;
+  _memset($48 | 0, 0, (HEAP32[$46 >> 2] | 0) * 36 & -1 | 0);
+  $54 = (HEAP32[$46 >> 2] | 0) - 1 | 0;
+  if (($54 | 0) > 0) {
+    $i_05_i_i_i = 0;
+    while (1) {
+      $56 = $i_05_i_i_i + 1 | 0;
+      HEAP32[(HEAP32[$50 >> 2] | 0) + ($i_05_i_i_i * 36 & -1) + 20 >> 2] = $56;
+      HEAP32[(HEAP32[$50 >> 2] | 0) + ($i_05_i_i_i * 36 & -1) + 32 >> 2] = -1;
+      $62 = (HEAP32[$46 >> 2] | 0) - 1 | 0;
+      if (($56 | 0) < ($62 | 0)) {
+        $i_05_i_i_i = $56;
+      } else {
+        $_lcssa_i_i_i = $62;
+        break;
+      }
+    }
+  } else {
+    $_lcssa_i_i_i = $54;
+  }
+  HEAP32[(HEAP32[$50 >> 2] | 0) + ($_lcssa_i_i_i * 36 & -1) + 20 >> 2] = -1;
+  HEAP32[(HEAP32[$50 >> 2] | 0) + (((HEAP32[$46 >> 2] | 0) - 1 | 0) * 36 & -1) + 32 >> 2] = -1;
+  _memset($world + 102888 | 0, 0, 16);
+  HEAP32[$world + 102920 >> 2] = 16;
+  HEAP32[$world + 102924 >> 2] = 0;
+  HEAP32[$world + 102916 >> 2] = _malloc(192) | 0;
+  HEAP32[$world + 102908 >> 2] = 16;
+  HEAP32[$world + 102912 >> 2] = 0;
+  HEAP32[$world + 102904 >> 2] = _malloc(64) | 0;
+  HEAP32[$world + 102932 >> 2] = 0;
+  HEAP32[$world + 102936 >> 2] = 0;
+  HEAP32[$world + 102940 >> 2] = 104;
+  HEAP32[$world + 102944 >> 2] = 96;
+  $87 = $world + 102948 | 0;
+  HEAP32[$world + 102980 >> 2] = 0;
+  HEAP32[$world + 102984 >> 2] = 0;
+  _memset($87 | 0, 0, 20);
+  HEAP8[$world + 102992 | 0] = 1;
+  HEAP8[$world + 102993 | 0] = 1;
+  HEAP8[$world + 102994 | 0] = 0;
+  HEAP8[$world + 102995 | 0] = 1;
+  $96 = $world + 102976 | 0;
+  HEAP8[$96] = 1;
+  $97 = $world + 102968 | 0;
+  HEAP32[$97 >> 2] = 0;
+  HEAP32[$97 + 4 >> 2] = -1054867456;
+  $98 = $world + 102868 | 0;
+  HEAP32[$98 >> 2] = 4;
+  HEAPF32[$world + 102988 >> 2] = +0;
+  HEAP32[$87 >> 2] = $14;
+  _memset($world + 102996 | 0, 0, 32);
+  HEAP8[$96] = 0;
+  HEAP32[$bd + 44 >> 2] = 0;
+  _memset($bd + 4 | 0, 0, 32);
+  HEAP8[$bd + 36 | 0] = 1;
+  HEAP8[$bd + 37 | 0] = 1;
+  HEAP8[$bd + 38 | 0] = 0;
+  HEAP8[$bd + 39 | 0] = 0;
+  HEAP32[$bd >> 2] = 0;
+  HEAP8[$bd + 40 | 0] = 1;
+  HEAPF32[$bd + 48 >> 2] = +1;
+  $112 = __ZN16b2BlockAllocator8AllocateEi($14, 152) | 0;
+  if (($112 | 0) == 0) {
+    $116 = 0;
+  } else {
+    $115 = $112;
+    __ZN6b2BodyC2EPK9b2BodyDefP7b2World($115, $bd, $world);
+    $116 = $115;
+  }
+  HEAP32[$116 + 92 >> 2] = 0;
+  $118 = $world + 102952 | 0;
+  HEAP32[$116 + 96 >> 2] = HEAP32[$118 >> 2];
+  $121 = HEAP32[$118 >> 2] | 0;
+  if (!(($121 | 0) == 0)) {
+    HEAP32[$121 + 92 >> 2] = $116;
+  }
+  HEAP32[$118 >> 2] = $116;
+  $126 = $world + 102960 | 0;
+  HEAP32[$126 >> 2] = (HEAP32[$126 >> 2] | 0) + 1;
+  HEAP32[$shape >> 2] = 8016;
+  HEAP32[$shape + 4 >> 2] = 1;
+  HEAPF32[$shape + 8 >> 2] = +.009999999776482582;
+  _memset($shape + 28 | 0, 0, 18);
+  $135 = $shape + 12 | 0;
+  HEAP32[$135 >> 2] = -1038090240;
+  HEAP32[$135 + 4 >> 2] = 0;
+  $137 = $shape + 20 | 0;
+  HEAP32[$137 >> 2] = 1109393408;
+  HEAP32[$137 + 4 >> 2] = 0;
+  HEAP8[$shape + 44 | 0] = 0;
+  HEAP8[$shape + 45 | 0] = 0;
+  HEAP16[$def_i + 22 >> 1] = 1;
+  HEAP16[$def_i + 24 >> 1] = -1;
+  HEAP16[$def_i + 26 >> 1] = 0;
+  HEAP32[$def_i + 4 >> 2] = 0;
+  HEAPF32[$def_i + 8 >> 2] = +.20000000298023224;
+  HEAPF32[$def_i + 12 >> 2] = +0;
+  HEAP8[$def_i + 20 | 0] = 0;
+  HEAP32[$def_i >> 2] = $shape;
+  HEAPF32[$def_i + 16 >> 2] = +0;
+  __ZN6b2Body13CreateFixtureEPK12b2FixtureDef($116, $def_i);
+  HEAP32[$shape1 >> 2] = 7968;
+  HEAP32[$shape1 + 4 >> 2] = 2;
+  HEAPF32[$shape1 + 8 >> 2] = +.009999999776482582;
+  HEAP32[$shape1 + 148 >> 2] = 4;
+  HEAPF32[$shape1 + 20 >> 2] = +-.5;
+  HEAPF32[$shape1 + 24 >> 2] = +-.5;
+  HEAPF32[$shape1 + 28 >> 2] = +.5;
+  HEAPF32[$shape1 + 32 >> 2] = +-.5;
+  HEAPF32[$shape1 + 36 >> 2] = +.5;
+  HEAPF32[$shape1 + 40 >> 2] = +.5;
+  HEAPF32[$shape1 + 44 >> 2] = +-.5;
+  HEAPF32[$shape1 + 48 >> 2] = +.5;
+  HEAPF32[$shape1 + 84 >> 2] = +0;
+  HEAPF32[$shape1 + 88 >> 2] = +-1;
+  HEAPF32[$shape1 + 92 >> 2] = +1;
+  HEAPF32[$shape1 + 96 >> 2] = +0;
+  HEAPF32[$shape1 + 100 >> 2] = +0;
+  HEAPF32[$shape1 + 104 >> 2] = +1;
+  HEAPF32[$shape1 + 108 >> 2] = +-1;
+  HEAPF32[$shape1 + 112 >> 2] = +0;
+  HEAPF32[$shape1 + 12 >> 2] = +0;
+  HEAPF32[$shape1 + 16 >> 2] = +0;
+  $174 = $bd2 + 44 | 0;
+  $176 = $bd2 + 36 | 0;
+  $177 = $bd2 + 4 | 0;
+  $178 = $bd2 + 37 | 0;
+  $179 = $bd2 + 38 | 0;
+  $180 = $bd2 + 39 | 0;
+  $181 = $bd2 | 0;
+  $182 = $bd2 + 40 | 0;
+  $183 = $bd2 + 48 | 0;
+  $185 = $bd2 + 4 | 0;
+  $186 = $shape1 | 0;
+  $188 = $def_i21 + 22 | 0;
+  $189 = $def_i21 + 24 | 0;
+  $190 = $def_i21 + 26 | 0;
+  $191 = $def_i21 | 0;
+  $192 = $def_i21 + 4 | 0;
+  $193 = $def_i21 + 8 | 0;
+  $194 = $def_i21 + 12 | 0;
+  $195 = $def_i21 + 16 | 0;
+  $196 = $def_i21 + 20 | 0;
+  $x_sroa_1_4_load313755 = +.75;
+  $x_sroa_0_0_load303656 = +-7;
+  $i_057 = 0;
+  L82 : while (1) {
+    $y_sroa_1_4_load293550 = $x_sroa_1_4_load313755;
+    $y_sroa_0_0_load283451 = $x_sroa_0_0_load303656;
+    $j_052 = $i_057;
+    while (1) {
+      HEAP32[$174 >> 2] = 0;
+      _memset($177 | 0, 0, 32);
+      HEAP8[$176] = 1;
+      HEAP8[$178] = 1;
+      HEAP8[$179] = 0;
+      HEAP8[$180] = 0;
+      HEAP8[$182] = 1;
+      HEAPF32[$183 >> 2] = +1;
+      HEAP32[$181 >> 2] = 2;
+      $y_sroa_0_0_insert_insert$1 = +$y_sroa_1_4_load293550;
+      HEAPF32[$185 >> 2] = $y_sroa_0_0_load283451;
+      HEAPF32[$185 + 4 >> 2] = $y_sroa_0_0_insert_insert$1;
+      if (!((HEAP32[$98 >> 2] & 2 | 0) == 0)) {
+        label = 65;
+        break L82;
+      }
+      $205 = __ZN16b2BlockAllocator8AllocateEi($14, 152) | 0;
+      if (($205 | 0) == 0) {
+        $209 = 0;
+      } else {
+        $208 = $205;
+        __ZN6b2BodyC2EPK9b2BodyDefP7b2World($208, $bd2, $world);
+        $209 = $208;
+      }
+      HEAP32[$209 + 92 >> 2] = 0;
+      HEAP32[$209 + 96 >> 2] = HEAP32[$118 >> 2];
+      $213 = HEAP32[$118 >> 2] | 0;
+      if (!(($213 | 0) == 0)) {
+        HEAP32[$213 + 92 >> 2] = $209;
+      }
+      HEAP32[$118 >> 2] = $209;
+      HEAP32[$126 >> 2] = (HEAP32[$126 >> 2] | 0) + 1;
+      HEAP16[$188 >> 1] = 1;
+      HEAP16[$189 >> 1] = -1;
+      HEAP16[$190 >> 1] = 0;
+      HEAP32[$192 >> 2] = 0;
+      HEAPF32[$193 >> 2] = +.20000000298023224;
+      HEAPF32[$194 >> 2] = +0;
+      HEAP8[$196] = 0;
+      HEAP32[$191 >> 2] = $186;
+      HEAPF32[$195 >> 2] = +5;
+      __ZN6b2Body13CreateFixtureEPK12b2FixtureDef($209, $def_i21);
+      $223 = $j_052 + 1 | 0;
+      if (($223 | 0) < 40) {
+        $y_sroa_1_4_load293550 = $y_sroa_1_4_load293550 + +0;
+        $y_sroa_0_0_load283451 = $y_sroa_0_0_load283451 + 1.125;
+        $j_052 = $223;
+      } else {
+        break;
+      }
+    }
+    $236 = $i_057 + 1 | 0;
+    if (($236 | 0) < 40) {
+      $x_sroa_1_4_load313755 = $x_sroa_1_4_load313755 + +1;
+      $x_sroa_0_0_load303656 = $x_sroa_0_0_load303656 + +.5625;
+      $i_057 = $236;
+    } else {
+      $i3_042 = 0;
+      break;
+    }
+  }
+  if ((label | 0) == 65) {
+    ___assert_func(112, 109, 5328, 2520);
+    return 0;
+  }
+  while (1) {
+    __ZN7b2World4StepEfii($world);
+    $i3_042 = $i3_042 + 1 | 0;
+    if (($i3_042 | 0) >= ($WARMUP_0 | 0)) {
+      break;
+    }
+  }
+  $241 = HEAP32[2414] | 0;
+  $242 = _llvm_stacksave() | 0;
+  $243 = STACKTOP;
+  STACKTOP = STACKTOP + ($241 * 4 & -1) | 0;
+  STACKTOP = STACKTOP + 7 >> 3 << 3;
+  if (($241 | 0) > 0) {
+    $i4_038 = 0;
+    while (1) {
+      $245 = _clock() | 0;
+      __ZN7b2World4StepEfii($world);
+      HEAP32[$243 + ($i4_038 << 2) >> 2] = (_clock() | 0) - $245;
+      $i4_038 = $i4_038 + 1 | 0;
+      if (($i4_038 | 0) >= (HEAP32[2414] | 0)) {
+        break;
+      }
+    }
+  }
+  __Z7measurePm($result, $243);
+  $260 = +HEAPF32[$result + 4 >> 2];
+  _printf(3480, (tempInt = STACKTOP, STACKTOP = STACKTOP + 16 | 0, HEAPF64[tempInt >> 3] = +HEAPF32[$result >> 2], HEAPF64[tempInt + 8 >> 3] = $260, tempInt) | 0) | 0;
+  _llvm_stackrestore($242 | 0);
+  __ZN7b2WorldD2Ev($world);
+  $_0 = 0;
+  STACKTOP = __stackBase__;
+  return $_0 | 0;
+}
 
diff --git a/tools/test-js-optimizer-asm-pre.js b/tools/test-js-optimizer-asm-pre.js
index 38487d2f..4f3ba780 100644
--- a/tools/test-js-optimizer-asm-pre.js
+++ b/tools/test-js-optimizer-asm-pre.js
@@ -9,6 +9,17 @@ function a() {
   whee(12, 13) | 0;
   +whee(12, 13);
   f((g = t(), (g+g)|0)|0);
+  // always coerce function calls in asm
+  f() | 0;
+  f((h() | 0) + 5 | 0);
+  f(((x + y) | 0) + z | 0);
+  +f();
+  f(+(+h() + 5));
+  $140 = $p_3_i + (-$mantSize_0_i | 0) | 0;
+  f(g() | 0 | 0);
+  f(g() | 0 & -1);
+  f((g() | 0) >> 2);
+  $56 = (_fcntl() | 0) | 1;
 }
 function b($this, $__n) {
   $this = $this | 0;
@@ -165,4 +176,352 @@ function boxx($this, $aabb, $xf, $childIndex) {
   HEAP32[$51 + 4 >> 2] = $_sroa_0_0_insert_insert$1;
   return;
 }
-// EMSCRIPTEN_GENERATED_FUNCTIONS: ["a", "b", "rett", "ret2t", "retf", "i32_8", "tempDoublePtr", "boxx"]
+function _main($argc, $argv) {
+  $argc = $argc | 0;
+  $argv = $argv | 0;
+  var $def_i21 = 0, $def_i = 0, $world = 0, $bd = 0, $shape = 0, $shape1 = 0, $bd2 = 0, $result = 0, $6 = 0, $WARMUP_0 = 0, $14 = 0, $15 = 0, $17 = 0, $i_09_i_i = 0, $j_08_i_i = 0, $34 = 0, $j_1_i_i = 0, $38 = 0, $46 = 0, $48 = 0, $50 = 0, $54 = 0, $i_05_i_i_i = 0, $56 = 0, $62 = 0, $_lcssa_i_i_i = 0, $87 = 0, $96 = 0, $97 = 0, $98 = 0, $112 = 0, $115 = 0, $116 = 0, $118 = 0, $121 = 0, $126 = 0, $135 = 0, $137 = 0, $174 = 0, $176 = 0, $177 = 0, $178 = 0, $179 = 0, $180 = 0, $181 = 0, $182 = 0, $183 = 0, $185 = 0, $186 = 0, $188 = 0, $189 = 0, $190 = 0, $191 = 0, $192 = 0, $193 = 0, $194 = 0, $195 = 0, $196 = 0, $i_057 = 0, $x_sroa_0_0_load303656 = +0, $x_sroa_1_4_load313755 = +0, $j_052 = 0, $y_sroa_0_0_load283451 = +0, $y_sroa_1_4_load293550 = +0, $y_sroa_0_0_insert_insert$1 = 0, $205 = 0, $208 = 0, $209 = 0, $213 = 0, $223 = 0, $236 = 0, $i3_042 = 0, $241 = 0, $242 = 0, $243 = 0, $i4_038 = 0, $245 = 0, $260 = +0, $_0 = 0, label = 0, __stackBase__ = 0;
+  __stackBase__ = STACKTOP;
+  STACKTOP = STACKTOP + 103416 | 0;
+  $def_i21 = __stackBase__ | 0;
+  $def_i = __stackBase__ + 32 | 0;
+  $world = __stackBase__ + 64 | 0;
+  $bd = __stackBase__ + 103096 | 0;
+  $shape = __stackBase__ + 103152 | 0;
+  $shape1 = __stackBase__ + 103200 | 0;
+  $bd2 = __stackBase__ + 103352 | 0;
+  $result = __stackBase__ + 103408 | 0;
+  do {
+    if (($argc | 0) > 1) {
+      $6 = (HEAP8[HEAP32[($argv + 4 | 0) >> 2] | 0] | 0) << 24 >> 24;
+      if (($6 | 0 | 0) == (49 | 0)) {
+        HEAP32[9656 >> 2] = 35;
+        $WARMUP_0 = 5;
+        break;
+      } else if (($6 | 0 | 0) == (50 | 0)) {
+        HEAP32[9656 >> 2] = 161;
+        $WARMUP_0 = 32;
+        break;
+      } else if (($6 | 0 | 0) == (51 | 0)) {
+        label = 43;
+        break;
+      } else if (($6 | 0 | 0) == (52 | 0)) {
+        HEAP32[9656 >> 2] = 2331;
+        $WARMUP_0 = 320;
+        break;
+      } else if (($6 | 0 | 0) == (53 | 0)) {
+        HEAP32[9656 >> 2] = 5661;
+        $WARMUP_0 = 640;
+        break;
+      } else if (($6 | 0 | 0) == (48 | 0)) {
+        $_0 = 0;
+        STACKTOP = __stackBase__;
+        return $_0 | 0;
+      } else {
+        _printf(3512 | 0 | 0, (tempInt = STACKTOP, STACKTOP = STACKTOP + 8 | 0, HEAP32[tempInt >> 2] = $6 - 48 | 0, tempInt) | 0) | 0;
+        $_0 = -1;
+        STACKTOP = __stackBase__;
+        return $_0 | 0;
+      }
+    } else {
+      label = 43;
+    }
+  } while (0);
+  if ((label | 0) == 43) {
+    HEAP32[9656 >> 2] = 333;
+    $WARMUP_0 = 64;
+  }
+  $14 = $world | 0;
+  $15 = $world + 8 | 0;
+  HEAP32[$15 >> 2] = 128;
+  HEAP32[($world + 4 | 0) >> 2] = 0;
+  $17 = _malloc(1024) | 0;
+  HEAP32[($world | 0) >> 2] = $17;
+  _memset($17 | 0 | 0, 0 | 0 | 0, (HEAP32[$15 >> 2] | 0) << 3 | 0 | 0);
+  _memset($world + 12 | 0 | 0 | 0, 0 | 0 | 0, 56 | 0 | 0);
+  $j_08_i_i = 0;
+  $i_09_i_i = 1;
+  while (1) {
+    if (!(($j_08_i_i | 0) < 14)) {
+      label = 49;
+      break;
+    }
+    if (($i_09_i_i | 0) > (HEAP32[(9600 + ($j_08_i_i << 2) | 0) >> 2] | 0 | 0)) {
+      $34 = $j_08_i_i + 1 | 0;
+      HEAP8[$i_09_i_i + 8952 | 0] = $34 & 255;
+      $j_1_i_i = $34;
+    } else {
+      HEAP8[$i_09_i_i + 8952 | 0] = $j_08_i_i & 255;
+      $j_1_i_i = $j_08_i_i;
+    }
+    $38 = $i_09_i_i + 1 | 0;
+    if (($38 | 0) < 641) {
+      $j_08_i_i = $j_1_i_i;
+      $i_09_i_i = $38;
+    } else {
+      break;
+    }
+  }
+  if ((label | 0) == 49) {
+    ___assert_func(3248 | 0 | 0, 73 | 0, 6448 | 0 | 0, 3360 | 0 | 0);
+    return 0 | 0;
+  }
+  HEAP32[($world + 102468 | 0) >> 2] = 0;
+  HEAP32[($world + 102472 | 0) >> 2] = 0;
+  HEAP32[($world + 102476 | 0) >> 2] = 0;
+  HEAP32[($world + 102864 | 0) >> 2] = 0;
+  HEAP32[($world + 102872 | 0) >> 2] = -1;
+  $46 = $world + 102884 | 0;
+  HEAP32[$46 >> 2] = 16;
+  HEAP32[($world + 102880 | 0) >> 2] = 0;
+  $48 = _malloc(576) | 0;
+  $50 = $world + 102876 | 0;
+  HEAP32[$50 >> 2] = $48;
+  _memset($48 | 0 | 0, 0 | 0 | 0, (HEAP32[$46 >> 2] | 0) * 36 & -1 | 0 | 0);
+  $54 = (HEAP32[$46 >> 2] | 0) - 1 | 0;
+  if (($54 | 0) > 0) {
+    $i_05_i_i_i = 0;
+    while (1) {
+      $56 = $i_05_i_i_i + 1 | 0;
+      HEAP32[((HEAP32[$50 >> 2] | 0) + ($i_05_i_i_i * 36 & -1) + 20 | 0) >> 2] = $56;
+      HEAP32[((HEAP32[$50 >> 2] | 0) + ($i_05_i_i_i * 36 & -1) + 32 | 0) >> 2] = -1;
+      $62 = (HEAP32[$46 >> 2] | 0) - 1 | 0;
+      if (($56 | 0) < ($62 | 0)) {
+        $i_05_i_i_i = $56;
+      } else {
+        $_lcssa_i_i_i = $62;
+        break;
+      }
+    }
+  } else {
+    $_lcssa_i_i_i = $54;
+  }
+  HEAP32[((HEAP32[$50 >> 2] | 0) + ($_lcssa_i_i_i * 36 & -1) + 20 | 0) >> 2] = -1;
+  HEAP32[((HEAP32[$50 >> 2] | 0) + (((HEAP32[$46 >> 2] | 0) - 1 | 0) * 36 & -1) + 32 | 0) >> 2] = -1;
+  _memset($world + 102888 | 0 | 0 | 0, 0 | 0 | 0, 16 | 0 | 0);
+  HEAP32[($world + 102920 | 0) >> 2] = 16;
+  HEAP32[($world + 102924 | 0) >> 2] = 0;
+  HEAP32[($world + 102916 | 0) >> 2] = _malloc(192) | 0;
+  HEAP32[($world + 102908 | 0) >> 2] = 16;
+  HEAP32[($world + 102912 | 0) >> 2] = 0;
+  HEAP32[($world + 102904 | 0) >> 2] = _malloc(64) | 0;
+  HEAP32[($world + 102932 | 0) >> 2] = 0;
+  HEAP32[($world + 102936 | 0) >> 2] = 0;
+  HEAP32[($world + 102940 | 0) >> 2] = 104;
+  HEAP32[($world + 102944 | 0) >> 2] = 96;
+  $87 = $world + 102948 | 0;
+  HEAP32[($world + 102980 | 0) >> 2] = 0;
+  HEAP32[($world + 102984 | 0) >> 2] = 0;
+  _memset($87 | 0 | 0, 0 | 0 | 0, 20 | 0 | 0);
+  HEAP8[$world + 102992 | 0] = 1;
+  HEAP8[$world + 102993 | 0] = 1;
+  HEAP8[$world + 102994 | 0] = 0;
+  HEAP8[$world + 102995 | 0] = 1;
+  $96 = $world + 102976 | 0;
+  HEAP8[$96] = 1;
+  $97 = $world + 102968 | 0;
+  HEAP32[($97 | 0) >> 2] = 0;
+  HEAP32[($97 + 4 | 0) >> 2] = -1054867456;
+  $98 = $world + 102868 | 0;
+  HEAP32[$98 >> 2] = 4;
+  HEAPF32[($world + 102988 | 0) >> 2] = +0;
+  HEAP32[$87 >> 2] = $14;
+  _memset($world + 102996 | 0 | 0 | 0, 0 | 0 | 0, 32 | 0 | 0);
+  HEAP8[$96] = 0;
+  HEAP32[($bd + 44 | 0) >> 2] = 0;
+  _memset($bd + 4 | 0 | 0 | 0, 0 | 0 | 0, 32 | 0 | 0);
+  HEAP8[$bd + 36 | 0] = 1;
+  HEAP8[$bd + 37 | 0] = 1;
+  HEAP8[$bd + 38 | 0] = 0;
+  HEAP8[$bd + 39 | 0] = 0;
+  HEAP32[($bd | 0) >> 2] = 0;
+  HEAP8[$bd + 40 | 0] = 1;
+  HEAPF32[($bd + 48 | 0) >> 2] = +1;
+  $112 = __ZN16b2BlockAllocator8AllocateEi($14, 152) | 0;
+  if (($112 | 0) == 0) {
+    $116 = 0;
+  } else {
+    $115 = $112;
+    __ZN6b2BodyC2EPK9b2BodyDefP7b2World($115, $bd, $world);
+    $116 = $115;
+  }
+  HEAP32[($116 + 92 | 0) >> 2] = 0;
+  $118 = $world + 102952 | 0;
+  HEAP32[($116 + 96 | 0) >> 2] = HEAP32[$118 >> 2] | 0;
+  $121 = HEAP32[$118 >> 2] | 0;
+  if (!(($121 | 0) == 0)) {
+    HEAP32[($121 + 92 | 0) >> 2] = $116;
+  }
+  HEAP32[$118 >> 2] = $116;
+  $126 = $world + 102960 | 0;
+  HEAP32[$126 >> 2] = (HEAP32[$126 >> 2] | 0) + 1 | 0;
+  HEAP32[($shape | 0) >> 2] = 8016 | 0;
+  HEAP32[($shape + 4 | 0) >> 2] = 1;
+  HEAPF32[($shape + 8 | 0) >> 2] = +.009999999776482582;
+  _memset($shape + 28 | 0 | 0 | 0, 0 | 0 | 0, 18 | 0 | 0);
+  $135 = $shape + 12 | 0;
+  HEAP32[($135 | 0) >> 2] = -1038090240;
+  HEAP32[($135 + 4 | 0) >> 2] = 0;
+  $137 = $shape + 20 | 0;
+  HEAP32[($137 | 0) >> 2] = 1109393408;
+  HEAP32[($137 + 4 | 0) >> 2] = 0;