diff --git a/emcc b/emcc
index 2a7e10d0fcad6..9849e881b55bd 100755
--- a/emcc
+++ b/emcc
@@ -53,8 +53,6 @@ from tools import shared
 from tools.shared import Compression, execute, suffix, unsuffixed, unsuffixed_basename
 from tools.response_file import read_response_file
 
-logging = logging.getLogger('emcc')
-
 # Mapping of emcc opt levels to llvm opt levels. We use llvm opt level 3 in emcc opt
 # levels 2 and 3 (emcc 3 is unsafe opts, so unsuitable for the only level to get
 # llvm opt level 3, and speed-wise emcc level 2 is already the slowest/most optimizing
@@ -819,8 +817,12 @@ try:
       newargs[i] = ''
     elif newargs[i] == '-v':
       shared.COMPILER_OPTS += ['-v']
-      DEBUG = 1
       os.environ['EMCC_DEBUG'] = '1' # send to child processes too
+      if DEBUG != 1:
+        # swap in debug logging
+        DEBUG = 1
+        shared.set_logging()
+        logging.debug('invocation: ' + ' '.join(sys.argv))
       newargs[i] = ''
     elif newargs[i].startswith('--shell-file'):
       check_bad_eq(newargs[i])
diff --git a/emscripten.py b/emscripten.py
index 248d0ce420929..d831285589007 100755
--- a/emscripten.py
+++ b/emscripten.py
@@ -127,6 +127,9 @@ def emscript(infile, settings, outfile, libraries=[], compiler_engine=None,
   if DEBUG and len(meta) > 1024*1024: print >> sys.stderr, 'emscript warning: large amounts of metadata, will slow things down'
   if DEBUG: print >> sys.stderr, '  emscript: split took %s seconds' % (time.time() - t)
 
+  if len(funcs) == 0:
+    raise RuntimeError('No functions to process. Make sure you prevented LLVM from eliminating them as dead (use EXPORTED_FUNCTIONS if necessary, see the FAQ)')
+
   #if DEBUG:
   #  print >> sys.stderr, '========= pre ================\n'
   #  print >> sys.stderr, ''.join(pre)
diff --git a/src/analyzer.js b/src/analyzer.js
index 2cc46ab69823a..de9a7940b89c2 100644
--- a/src/analyzer.js
+++ b/src/analyzer.js
@@ -20,7 +20,7 @@ var BRANCH_INVOKE = set('branch', 'invoke');
 var LABEL_ENDERS = set('branch', 'return', 'switch');
 var SIDE_EFFECT_CAUSERS = set('call', 'invoke', 'atomic');
 var UNUNFOLDABLE = set('value', 'structvalue', 'type', 'phiparam');
-var I64_DOUBLE_FLIP = { i64: 'double', double: 'i64' };
+var SHADOW_FLIP = { i64: 'double', double: 'i64' }; //, i32: 'float', float: 'i32' };
 
 // Analyzer
 
@@ -124,13 +124,13 @@ function analyzer(data, sidePass) {
           var lines = label.lines;
           for (var i = 0; i < lines.length; i++) {
             var line = lines[i];
-            if (line.intertype == 'bitcast' && line.type in I64_DOUBLE_FLIP) {
+            if (line.intertype == 'bitcast' && line.type in SHADOW_FLIP) {
               has = true;
             }
           }
         });
         if (!has) return;
-        // there are i64<-->double bitcasts, create shadows for everything
+        // there are integer<->floating-point bitcasts, create shadows for everything
         var shadowed = {};
         func.labels.forEach(function(label) {
           var lines = label.lines;
@@ -138,11 +138,11 @@ function analyzer(data, sidePass) {
           while (i < lines.length) {
           var lines = label.lines;
             var line = lines[i];
-            if (line.intertype == 'load' && line.type in I64_DOUBLE_FLIP) {
+            if (line.intertype == 'load' && line.type in SHADOW_FLIP) {
               if (line.pointer.intertype != 'value') { i++; continue } // TODO
               shadowed[line.assignTo] = 1;
               var shadow = line.assignTo + '$$SHADOW';
-              var flip = I64_DOUBLE_FLIP[line.type];
+              var flip = SHADOW_FLIP[line.type];
               lines.splice(i + 1, 0, { // if necessary this element will be legalized in the next phase
                 tokens: null,
                 indent: 2,
@@ -171,7 +171,7 @@ function analyzer(data, sidePass) {
           var lines = label.lines;
           for (var i = 0; i < lines.length; i++) {
             var line = lines[i];
-            if (line.intertype == 'bitcast' && line.type in I64_DOUBLE_FLIP && line.ident in shadowed) {
+            if (line.intertype == 'bitcast' && line.type in SHADOW_FLIP && line.ident in shadowed) {
               var shadow = line.ident + '$$SHADOW';
               line.params[0].ident = shadow;
               line.params[0].type = line.type;
diff --git a/src/headless.js b/src/headless.js
index d81fb5a3639f3..097a42f753bd6 100644
--- a/src/headless.js
+++ b/src/headless.js
@@ -4,6 +4,20 @@
 // TODO: sync from bananabread headless.js
 
 var window = {
+  eventListeners: {},
+  addEventListener: function(id, func) {
+    var listeners = this.eventListeners[id];
+    if (!listeners) {
+      listeners = this.eventListeners[id] = [];
+    }
+    listeners.push(func);
+  },
+  callEventListeners: function(id) {
+    var listeners = this.eventListeners[id];
+    if (listeners) {
+      listeners.forEach(function(listener) { listener() });
+    }
+  },
   location: {
     toString: function() {
       return '%s';
diff --git a/src/jsifier.js b/src/jsifier.js
index 156fd65dee774..faef88d514a01 100644
--- a/src/jsifier.js
+++ b/src/jsifier.js
@@ -328,7 +328,7 @@ function JSify(data, functionsOnly, givenFunctions) {
       var js = (index !== null ? '' : item.ident + '=') + constant;
       if (js) js += ';';
 
-      if (!ASM_JS && (EXPORT_ALL || (item.ident in EXPORTED_GLOBALS))) {
+      if (!ASM_JS && NAMED_GLOBALS && (EXPORT_ALL || (item.ident in EXPORTED_GLOBALS))) {
         js += '\nModule["' + item.ident + '"] = ' + item.ident + ';';
       }
       if (BUILD_AS_SHARED_LIB == 2 && !item.private_) {
diff --git a/src/library.js b/src/library.js
index 2578fda44512e..f958a436a9bac 100644
--- a/src/library.js
+++ b/src/library.js
@@ -6354,7 +6354,8 @@ LibraryManager.library = {
   // Note that we need to emulate functions that use setjmp, and also to create
   // a new label we can return to. Emulation make such functions slower, this
   // can be alleviated by making a new function containing just the setjmp
-  // related functionality so the slowdown is more limited.
+  // related functionality so the slowdown is more limited - you may need
+  // to prevent inlining to keep this isolated, try __attribute__((noinline))
   // ==========================================================================
 
   saveSetjmp__asm: true,
@@ -6374,11 +6375,11 @@ LibraryManager.library = {
     setjmpId = (setjmpId+1)|0;
     {{{ makeSetValueAsm('env', '0', 'setjmpId', 'i32') }}};
     while ((i|0) < {{{ 2*MAX_SETJMPS }}}) {
-      if ({{{ makeGetValueAsm('table', 'i*4', 'i32') }}} == 0) {
-        {{{ makeSetValueAsm('table', 'i*4', 'setjmpId', 'i32') }}};
-        {{{ makeSetValueAsm('table', 'i*4+4', 'label', 'i32') }}};
+      if ({{{ makeGetValueAsm('table', '(i<<2)', 'i32') }}} == 0) {
+        {{{ makeSetValueAsm('table', '(i<<2)', 'setjmpId', 'i32') }}};
+        {{{ makeSetValueAsm('table', '(i<<2)+4', 'label', 'i32') }}};
         // prepare next slot
-        {{{ makeSetValueAsm('table', 'i*4+8', '0', 'i32') }}};
+        {{{ makeSetValueAsm('table', '(i<<2)+8', '0', 'i32') }}};
         return 0;
       }
       i = (i+2)|0;
@@ -6395,10 +6396,10 @@ LibraryManager.library = {
     table = table|0;
     var i = 0, curr = 0;
     while ((i|0) < {{{ MAX_SETJMPS }}}) {
-      curr = {{{ makeGetValueAsm('table', 'i*4', 'i32') }}};
+      curr = {{{ makeGetValueAsm('table', '(i<<2)', 'i32') }}};
       if ((curr|0) == 0) break;
       if ((curr|0) == (id|0)) {
-        return {{{ makeGetValueAsm('table', 'i*4+4', 'i32') }}};
+        return {{{ makeGetValueAsm('table', '(i<<2)+4', 'i32') }}};
       }
       i = (i+2)|0;
     }
@@ -7668,8 +7669,8 @@ LibraryManager.library = {
       nfds = Math.min(64, nfds); // fd sets have 64 bits
 
       for (var fd = 0; fd < nfds; fd++) {
-        var mask = 1 << (fd % 32), int = fd < 32 ? srcLow : srcHigh;
-        if (int & mask) {
+        var mask = 1 << (fd % 32), int_ = fd < 32 ? srcLow : srcHigh;
+        if (int_ & mask) {
           // index is in the set, check if it is ready for read
           var info = FS.streams[fd];
           if (info && can(info)) {
@@ -8057,8 +8058,8 @@ LibraryManager.library = {
       nfds = Math.min(64, nfds); // fd sets have 64 bits
 
       for (var fd = 0; fd < nfds; fd++) {
-        var mask = 1 << (fd % 32), int = fd < 32 ? srcLow : srcHigh;
-        if (int & mask) {
+        var mask = 1 << (fd % 32), int_ = fd < 32 ? srcLow : srcHigh;
+        if (int_ & mask) {
           // index is in the set, check if it is ready for read
           var info = FS.streams[fd];
           if (info && can(info)) {
diff --git a/src/library_sdl.js b/src/library_sdl.js
index 1c8ac52d9799a..b2fea43a959a0 100644
--- a/src/library_sdl.js
+++ b/src/library_sdl.js
@@ -47,6 +47,7 @@ var LibrarySDL = {
 
     startTime: null,
     buttonState: 0,
+    modState: 0,
     DOMButtons: [0, 0, 0],
 
     DOMEventToSDLEvent: {},
@@ -373,7 +374,8 @@ var LibrarySDL = {
             if (event['movementX'] == 0 && event['movementY'] == 0) {
               // ignore a mousemove event if it doesn't contain any movement info
               // (without pointer lock, we infer movement from pageX/pageY, so this check is unnecessary)
-              return false;
+              event.preventDefault();
+              return;
             }
           }
           // fall through
@@ -396,15 +398,20 @@ var LibrarySDL = {
           } else if (event.type == 'mousedown') {
             SDL.DOMButtons[event.button] = 1;
           } else if (event.type == 'mouseup') {
-            if (!SDL.DOMButtons[event.button]) return false; // ignore extra ups, can happen if we leave the canvas while pressing down, then return,
-                                                             // since we add a mouseup in that case
+            // ignore extra ups, can happen if we leave the canvas while pressing down, then return,
+            // since we add a mouseup in that case
+            if (!SDL.DOMButtons[event.button]) {
+              event.preventDefault();
+              return;
+            }
+
             SDL.DOMButtons[event.button] = 0;
           }
 
           if (event.type == 'keypress' && !SDL.textInput) {
             break;
           }
-
+          
           SDL.events.push(event);
           break;
         case 'mouseout':
@@ -438,7 +445,7 @@ var LibrarySDL = {
             // Force-run a main event loop, since otherwise this event will never be caught!
             Browser.mainLoop.runner();
           }
-          return true;
+          return;
         case 'resize':
           SDL.events.push(event);
           break;
@@ -447,7 +454,11 @@ var LibrarySDL = {
         Module.printErr('SDL event queue full, dropping events');
         SDL.events = SDL.events.slice(0, 10000);
       }
-      return false;
+      // manually triggered resize event doesn't have a preventDefault member
+      if (event.preventDefault) {
+        event.preventDefault();
+      }
+      return;
     },
 
     makeCEvent: function(event, ptr) {
@@ -473,15 +484,6 @@ var LibrarySDL = {
           } else {
             scan = SDL.scanCodes[key] || key;
           }
-          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.type', 'SDL.DOMEventToSDLEvent[event.type]', 'i32') }}}
-          //{{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.which', '1', 'i32') }}}
-          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.state', 'down ? 1 : 0', 'i8') }}}
-          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.repeat', '0', 'i8') }}} // TODO
-
-          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.keysym + SDL.structs.keysym.scancode', 'scan', 'i32') }}}
-          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.keysym + SDL.structs.keysym.sym', 'key', 'i32') }}}
-          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.keysym + SDL.structs.keysym.mod', '0', 'i32') }}}
-          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.keysym + SDL.structs.keysym.unicode', 'key', 'i32') }}}
 
           var code = SDL.keyCodes[event.keyCode] || event.keyCode;
           {{{ makeSetValue('SDL.keyboardState', 'code', 'down', 'i8') }}};
@@ -491,6 +493,19 @@ var LibrarySDL = {
             delete SDL.keyboardMap[code];
           }
 
+          // TODO: lmeta, rmeta, numlock, capslock, KMOD_MODE, KMOD_RESERVED
+          SDL.modState = ({{{ makeGetValue('SDL.keyboardState', '1248', 'i8') }}} ? 0x0040 | 0x0080 : 0) | // KMOD_LCTRL & KMOD_RCTRL
+            ({{{ makeGetValue('SDL.keyboardState', '1249', 'i8') }}} ? 0x0001 | 0x0002 : 0) | // KMOD_LSHIFT & KMOD_RSHIFT
+            ({{{ makeGetValue('SDL.keyboardState', '1250', 'i8') }}} ? 0x0100 | 0x0200 : 0); // KMOD_LALT & KMOD_RALT
+
+          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.type', 'SDL.DOMEventToSDLEvent[event.type]', 'i32') }}}
+          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.state', 'down ? 1 : 0', 'i8') }}}
+          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.repeat', '0', 'i8') }}} // TODO
+          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.keysym + SDL.structs.keysym.scancode', 'scan', 'i32') }}}
+          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.keysym + SDL.structs.keysym.sym', 'key', 'i32') }}}
+          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.keysym + SDL.structs.keysym.mod', 'SDL.modState', 'i32') }}}
+          {{{ makeSetValue('ptr', 'SDL.structs.KeyboardEvent.keysym + SDL.structs.keysym.unicode', 'key', 'i32') }}}
+
           break;
         }
         case 'keypress': {
@@ -614,13 +629,13 @@ var LibrarySDL = {
     SDL.startTime = Date.now();
     // capture all key events. we just keep down and up, but also capture press to prevent default actions
     if (!Module['doNotCaptureKeyboard']) {
-      document.onkeydown = SDL.receiveEvent;
-      document.onkeyup = SDL.receiveEvent;
-      document.onkeypress = SDL.receiveEvent;
-      document.onblur = SDL.receiveEvent;
+      document.addEventListener("keydown", SDL.receiveEvent);
+      document.addEventListener("keyup", SDL.receiveEvent);
+      document.addEventListener("keypress", SDL.receiveEvent);
+      document.addEventListener("blur", SDL.receiveEvent);
       document.addEventListener("visibilitychange", SDL.receiveEvent);
     }
-    window.onunload = SDL.receiveEvent;
+    window.addEventListener("unload", SDL.receiveEvent);
     SDL.keyboardState = _malloc(0x10000); // Our SDL needs 512, but 64K is safe for older SDLs
     _memset(SDL.keyboardState, 0, 0x10000);
     // Initialize this structure carefully for closure
@@ -888,12 +903,16 @@ var LibrarySDL = {
   SDL_GetKeyState: function() {
     return _SDL_GetKeyboardState();
   },
+  
+  SDL_GetKeyName: function(key) {
+    if (!SDL.keyName) {
+      SDL.keyName = allocate(intArrayFromString('unknown key'), 'i8', ALLOC_NORMAL);
+    }
+    return SDL.keyName;
+  },
 
   SDL_GetModState: function() {
-    // TODO: numlock, capslock, etc.
-    return (SDL.keyboardState[16] ? 0x0001 | 0x0002 : 0) | // KMOD_LSHIFT & KMOD_RSHIFT
-           (SDL.keyboardState[17] ? 0x0040 | 0x0080 : 0) | // KMOD_LCTRL & KMOD_RCTRL
-           (SDL.keyboardState[18] ? 0x0100 | 0x0200 : 0); // KMOD_LALT & KMOD_RALT
+    return SDL.modState;
   },
 
   SDL_GetMouseState: function(x, y) {
@@ -1103,6 +1122,20 @@ var LibrarySDL = {
     return (a&0xff)+((b&0xff)<<8)+((g&0xff)<<16)+((r&0xff)<<24)
   },
 
+  SDL_GetAppState: function() {
+    var state = 0;
+
+    if (Browser.pointerLock) {
+      state |= 0x01;  // SDL_APPMOUSEFOCUS
+    }
+    if (document.hasFocus()) {
+      state |= 0x02;  // SDL_APPINPUTFOCUS
+    }
+    state |= 0x04;  // SDL_APPACTIVE
+
+    return state;
+  },
+
   SDL_WM_GrabInput: function() {},
   
   SDL_WM_ToggleFullScreen: function(surf) {
@@ -1755,6 +1788,10 @@ var LibrarySDL = {
     return -1;
   },
 
+  SDL_SetGammaRamp: function (redTable, greenTable, blueTable) {
+    return -1;
+  },
+
   // Misc
 
   SDL_InitSubSystem: function(flags) { return 0 },
@@ -1773,7 +1810,7 @@ var LibrarySDL = {
 
   SDL_AddTimer: function(interval, callback, param) {
     return window.setTimeout(function() {
-      Runtime.dynCall('ii', callback, [interval, param]);
+      Runtime.dynCall('iii', callback, [interval, param]);
     }, interval);
   },
   SDL_RemoveTimer: function(id) {
@@ -1792,7 +1829,7 @@ var LibrarySDL = {
   SDL_FreeRW: function() { throw 'SDL_FreeRW: TODO' },
   SDL_CondBroadcast: function() { throw 'SDL_CondBroadcast: TODO' },
   SDL_CondWaitTimeout: function() { throw 'SDL_CondWaitTimeout: TODO' },
-  SDL_WM_ToggleFullScreen: function() { throw 'SDL_WM_ToggleFullScreen: TODO' },
+  SDL_WM_IconifyWindow: function() { throw 'SDL_WM_IconifyWindow TODO' },
 
   Mix_SetPostMix: function() { throw 'Mix_SetPostMix: TODO' },
   Mix_QuerySpec: function() { throw 'Mix_QuerySpec: TODO' },
@@ -1802,6 +1839,16 @@ var LibrarySDL = {
   Mix_Linked_Version: function() { throw 'Mix_Linked_Version: TODO' },
   SDL_CreateRGBSurfaceFrom: function() { throw 'SDL_CreateRGBSurfaceFrom: TODO' },
   SDL_SaveBMP_RW: function() { throw 'SDL_SaveBMP_RW: TODO' },
+
+  SDL_WM_SetIcon: function() { /* This function would set the application window icon surface, which doesn't apply for web canvases, so a no-op. */ },
+  SDL_HasRDTSC: function() { return 0; },
+  SDL_HasMMX: function() { return 0; },
+  SDL_HasMMXExt: function() { return 0; },
+  SDL_Has3DNow: function() { return 0; },
+  SDL_Has3DNowExt: function() { return 0; },
+  SDL_HasSSE: function() { return 0; },
+  SDL_HasSSE2: function() { return 0; },
+  SDL_HasAltiVec: function() { return 0; }
 };
 
 autoAddDeps(LibrarySDL, '$SDL');
diff --git a/src/modules.js b/src/modules.js
index e40930788dc8a..b13ab3c508dd5 100644
--- a/src/modules.js
+++ b/src/modules.js
@@ -341,6 +341,7 @@ var Functions = {
                 call += (j > 1 ? ',' : '') + asmCoercion('a' + j, t[j] != 'i' ? 'float' : 'i32');
               }
               call += ')';
+              if (curr == '_setjmp') printErr('WARNING: setjmp used via a function pointer. If this is for libc setjmp (not something of your own with the same name), it will break things');
               tables.pre += 'function ' + curr + '__wrapper(' + args + ') { ' + arg_coercions + ' ; ' + retPre + call + retPost + ' }\n';
               wrapped[curr] = 1;
             }
diff --git a/src/parseTools.js b/src/parseTools.js
index 3949491edde4c..687faaa8820bd 100644
--- a/src/parseTools.js
+++ b/src/parseTools.js
@@ -2204,7 +2204,7 @@ function processMathop(item) {
     case 'sub': return handleOverflow(getFastValue(idents[0], '-', idents[1], item.type), bits);
     case 'sdiv': case 'udiv': return makeRounding(getFastValue(idents[0], '/', idents[1], item.type), bits, op[0] === 's');
     case 'mul': return getFastValue(idents[0], '*', idents[1], item.type); // overflow handling is already done in getFastValue for '*'
-    case 'urem': case 'srem': return getFastValue(idents[0], '%', idents[1], item.type);
+    case 'urem': case 'srem': return makeRounding(getFastValue(idents[0], '%', idents[1], item.type), bits, op[0] === 's');
     case 'or': {
       if (bits > 32) {
         assert(bits === 64, 'Too many bits for or: ' + bits);
diff --git a/src/relooper/Relooper.cpp b/src/relooper/Relooper.cpp
index 8a6e18b8e5296..7ceeb2f829f8e 100644
--- a/src/relooper/Relooper.cpp
+++ b/src/relooper/Relooper.cpp
@@ -897,6 +897,7 @@ void Relooper::Calculate(Block *Entry) {
   BlockSet Entries;
   Entries.insert(Entry);
   Root = Analyzer(this).Process(AllBlocks, Entries, NULL);
+  assert(Root);
 
   // Post optimizations
 
@@ -1091,6 +1092,7 @@ void Relooper::Calculate(Block *Entry) {
 
 void Relooper::Render() {
   OutputBuffer = OutputBufferRoot;
+  assert(Root);
   Root->Render(false);
 }
 
diff --git a/src/settings.js b/src/settings.js
index 7e6079966bad3..3a91b4886fa78 100644
--- a/src/settings.js
+++ b/src/settings.js
@@ -245,7 +245,8 @@ var EXPORTED_FUNCTIONS = ['_main', '_malloc'];
                                     // through LLVM dead code elimination, and also made accessible outside of
                                     // the generated code even after running closure compiler (on "Module").
                                     // Note the necessary prefix of "_".
-var EXPORT_ALL = 0; // If true, we export all the symbols
+var EXPORT_ALL = 0; // If true, we export all the symbols. Note that this does *not* affect LLVM, so it can
+                    // still eliminate functions as dead. This just exports them on the Module object.
 var EXPORT_BINDINGS = 0; // Export all bindings generator functions (prefixed with emscripten_bind_). This
                          // is necessary to use the bindings generator with asm.js
 
diff --git a/system/include/libc/sys/dirent.h b/system/include/libc/sys/dirent.h
index 9dcf34d1ec72c..0d8b02b5fbb1c 100644
--- a/system/include/libc/sys/dirent.h
+++ b/system/include/libc/sys/dirent.h
@@ -34,8 +34,22 @@ int   scandir(const char *dirp,
 enum {
   DT_UNKNOWN = 0,
 #define DT_UNKNOWN DT_UNKNOWN
-  DT_DIR = 4
+  DT_FIFO = 1,
+#define DT_FIFO DT_FIFO
+  DT_CHR = 2,
+#define DT_CHR DT_CHR
+  DT_DIR = 4,
 #define DT_DIR DT_DIR
+  DT_BLK = 6,
+#define DT_BLK DT_BLK
+  DT_REG = 8,
+#define DT_REG DT_REG
+  DT_LNK = 10,
+#define DT_LNK DT_LNK
+  DT_SOCK = 12,
+#define DT_SOCK DT_SOCK
+  DT_WHT = 14
+#define DT_WHT DT_WHT
 };
 
 #ifdef __cplusplus
diff --git a/system/include/sys/socket.h b/system/include/sys/socket.h
index 56a37375247e2..b83ce89a9c94d 100644
--- a/system/include/sys/socket.h
+++ b/system/include/sys/socket.h
@@ -44,7 +44,7 @@ typedef unsigned int sa_family_t;
 
 struct sockaddr {
   sa_family_t sa_family; 
-  char        sa_data[];
+  char        sa_data[16];
 };
 
 struct sockaddr_storage {
diff --git a/tests/runner.py b/tests/runner.py
index 9214ae2d520c4..73dbab135b430 100755
--- a/tests/runner.py
+++ b/tests/runner.py
@@ -126,15 +126,6 @@ def get_dir(self):
   def in_dir(self, *pathelems):
     return os.path.join(self.get_dir(), *pathelems)
 
-  def get_shared_library_name(self, linux_name):
-    if platform.system() == 'Linux':
-      return linux_name
-    elif platform.system() == 'Darwin':
-      return linux_name.replace('.so', '') + '.dylib'
-    else:
-      print >> sys.stderr, 'get_shared_library_name needs to be implemented on %s' % platform.system()
-      return linux_name
-
   def get_stdout_path(self):
     return os.path.join(self.get_dir(), 'stdout')
 
@@ -8422,7 +8413,7 @@ def process(filename):
                              [os.path.sep.join('codec/CMakeFiles/j2k_to_image.dir/index.c.o'.split('/')),
                               os.path.sep.join('codec/CMakeFiles/j2k_to_image.dir/convert.c.o'.split('/')),
                               os.path.sep.join('codec/CMakeFiles/j2k_to_image.dir/__/common/color.c.o'.split('/')),
-                              os.path.join('bin', self.get_shared_library_name('libopenjpeg.so.1.4.0'))],
+                              os.path.join('bin', 'libopenjpeg.so.1.4.0')],
                              configure=['cmake', '.'],
                              #configure_args=['--enable-tiff=no', '--enable-jp3d=no', '--enable-png=no'],
                              make_args=[]) # no -j 2, since parallel builds can fail
@@ -10505,6 +10496,24 @@ def test_redundant_link(self):
 
       self.assertContained('result: 1', run_js(os.path.join(self.get_dir(), 'a.out.js')))
 
+    def test_export_all(self):
+      lib = r'''
+        #include <stdio.h>
+        void libf1() { printf("libf1\n"); }
+        void libf2() { printf("libf2\n"); }
+      '''
+      lib_name = os.path.join(self.get_dir(), 'lib.c')
+      open(lib_name, 'w').write(lib)
+
+      open('main.js', 'w').write('''
+        _libf1();
+        _libf2();
+      ''')
+
+      Building.emcc(lib_name, ['-s', 'EXPORT_ALL=1', '--post-js', 'main.js'], output_filename='a.out.js')
+
+      self.assertContained('libf1\nlibf2\n', run_js(os.path.join(self.get_dir(), 'a.out.js')))
+
     def test_abspaths(self):
       # Includes with absolute paths are generally dangerous, things like -I/usr/.. will get to system local headers, not our portable ones.
 
@@ -12121,23 +12130,26 @@ def test_sdl_key(self):
           setTimeout(doOne, 1000/60);
         }
 
-        function simulateKeyEvent(c) {
+        function keydown(c) {
           var event = document.createEvent("KeyboardEvent");
           event.initKeyEvent("keydown", true, true, window,
                              0, 0, 0, 0,
                              c, c);
           document.dispatchEvent(event);
-          var event2 = document.createEvent("KeyboardEvent");
-          event2.initKeyEvent("keyup", true, true, window,
+        }
+
+        function keyup(c) {
+          var event = document.createEvent("KeyboardEvent");
+          event.initKeyEvent("keyup", true, true, window,
                              0, 0, 0, 0,
                              c, c);
-          document.dispatchEvent(event2);
+          document.dispatchEvent(event);
         }
       ''')
       open(os.path.join(self.get_dir(), 'sdl_key.c'), 'w').write(self.with_report_result(open(path_from_root('tests', 'sdl_key.c')).read()))
 
       Popen([PYTHON, EMCC, os.path.join(self.get_dir(), 'sdl_key.c'), '-o', 'page.html', '--pre-js', 'pre.js', '-s', '''EXPORTED_FUNCTIONS=['_main', '_one']''']).communicate()
-      self.run_browser('page.html', '', '/report_result?510510')
+      self.run_browser('page.html', '', '/report_result?223092870')
 
     def test_sdl_text(self):
       open(os.path.join(self.get_dir(), 'pre.js'), 'w').write('''
@@ -13537,6 +13549,7 @@ def output_parser(output):
       self.lua('scimark', '[small problem sizes]', output_parser=output_parser)
 
     def test_zzz_lua_binarytrees(self):
+      # js version: ['binarytrees.lua', {0: 0, 1: 9.5, 2: 11.99, 3: 12.85, 4: 14.72, 5: 15.82}[arguments[0]]]
       def args_processor(args):
         arg = int(DEFAULT_ARG)
         if arg == 0:
@@ -13989,7 +14002,7 @@ def test_relooper(self):
           assert os.path.exists(RELOOPER) == (i >= 2), 'have relooper on O2: ' + output
           src = open('a.out.js').read()
           main = src.split('function _main()')[1].split('\n}\n')[0]
-          assert ('while (1) {' in main or 'while(1){' in main) == (i >= 2), 'reloop code on O2: ' + main
+          assert ('while (1) {' in main or 'while(1){' in main or '} while ($' in main or '}while($' in main) == (i >= 2), 'reloop code on O2: ' + main
           assert ('switch' not in main) == (i >= 2), 'reloop code on O2: ' + main
 
     def test_jcache(self):
diff --git a/tests/sdl_key.c b/tests/sdl_key.c
index 19b0a3d6bf83b..7a304fc1dd780 100644
--- a/tests/sdl_key.c
+++ b/tests/sdl_key.c
@@ -12,16 +12,30 @@ void one() {
       case SDL_KEYDOWN:
         break;
       case SDL_KEYUP:
+        // don't handle the modifier key events
+        if (event.key.keysym.sym == SDLK_LCTRL ||
+            event.key.keysym.sym == SDLK_LSHIFT ||
+            event.key.keysym.sym == SDLK_LALT) {
+          return;
+        }
+        if ((event.key.keysym.mod & KMOD_LCTRL) || (event.key.keysym.mod & KMOD_RCTRL)) {
+          result *= 2;
+        }
+        if ((event.key.keysym.mod & KMOD_LSHIFT) || (event.key.keysym.mod & KMOD_RSHIFT)) {
+          result *= 3;
+        }
+        if ((event.key.keysym.mod & KMOD_LALT) || (event.key.keysym.mod & KMOD_RALT)) {
+          result *= 5;
+        }
         switch (event.key.keysym.sym) {
-          case SDLK_RIGHT: printf("right\n"); result *= 2; break;
-          case SDLK_LEFT: printf("left\n"); result *= 3; break;
-          case SDLK_DOWN: printf("down\n"); result *= 5; break;
-          case SDLK_UP: printf("up\n"); result *= 7; break;
-          case SDLK_SPACE: printf("space\n"); result *= 11; break;
-          case SDLK_a: printf("a\n"); result *= 13; break;
+          case SDLK_RIGHT: printf("right\n"); result *= 7; break;
+          case SDLK_LEFT: printf("left\n"); result *= 11; break;
+          case SDLK_DOWN: printf("down\n"); result *= 13; break;
+          case SDLK_UP: printf("up\n"); result *= 17; break;
+          case SDLK_a: printf("a\n"); result *= 19; break;
           default: {
             if (event.key.keysym.scancode == SDL_SCANCODE_B) {
-              printf("b scancode\n"); result *= 17; break;
+              printf("b scancode\n"); result *= 23; break;
             }
             printf("unknown key: sym %d scancode %d\n", event.key.keysym.sym, event.key.keysym.scancode);
             REPORT_RESULT();
@@ -40,14 +54,13 @@ int main(int argc, char **argv) {
   SDL_Init(SDL_INIT_VIDEO);
   SDL_Surface *screen = SDL_SetVideoMode(600, 450, 32, SDL_HWSURFACE);
 
-  emscripten_run_script("simulateKeyEvent(38)"); // up
-  emscripten_run_script("simulateKeyEvent(40)"); // down
-  emscripten_run_script("simulateKeyEvent(37)"); // left
-  emscripten_run_script("simulateKeyEvent(39)"); // right
-  emscripten_run_script("simulateKeyEvent(32)"); // space
-  emscripten_run_script("simulateKeyEvent(65)"); // a
-  emscripten_run_script("simulateKeyEvent(66)"); // b
-  emscripten_run_script("simulateKeyEvent(100)"); // trigger the end
+  emscripten_run_script("keydown(1250);keydown(38);keyup(38);keyup(1250);"); // alt, up
+  emscripten_run_script("keydown(1248);keydown(1249);keydown(40);keyup(40);keyup(1249);keyup(1248);"); // ctrl, shift, down
+  emscripten_run_script("keydown(37);keyup(37);"); // left
+  emscripten_run_script("keydown(39);keyup(39);"); // right
+  emscripten_run_script("keydown(65);keyup(65);"); // a
+  emscripten_run_script("keydown(66);keyup(66);"); // b
+  emscripten_run_script("keydown(100);keyup(100);"); // trigger the end
 
   if (argc == 1337) one(); // keep it alive
 
diff --git a/tools/eliminator/asm-eliminator-test-output.js b/tools/eliminator/asm-eliminator-test-output.js
index 52c91ce1a4652..2ce77b789124d 100644
--- a/tools/eliminator/asm-eliminator-test-output.js
+++ b/tools/eliminator/asm-eliminator-test-output.js
@@ -5094,4 +5094,18 @@ function multiloop2($n_0, $35) {
     }
   }
 }
+function tempDouble2($46, $14, $28, $42, $20, $32, $45) {
+  $46 = $46 | 0;
+  $14 = $14 | 0;
+  $28 = $28 | 0;
+  $42 = $42 | 0;
+  $20 = $20 | 0;
+  $32 = $32 | 0;
+  $45 = $45 | 0;
+  var $_sroa_06_0_insert_insert$1 = 0;
+  $_sroa_06_0_insert_insert$1 = (HEAPF32[tempDoublePtr >> 2] = ($20 < $32 ? $20 : $32) - $42, HEAP32[tempDoublePtr >> 2] | 0) | 0;
+  HEAP32[$45 >> 2] = 0 | (HEAPF32[tempDoublePtr >> 2] = ($14 < $28 ? $14 : $28) - $42, HEAP32[tempDoublePtr >> 2] | 0);
+  HEAP32[$45 + 4 >> 2] = $_sroa_06_0_insert_insert$1;
+  HEAP32[$45 + 8 >> 2] = $_sroa_06_0_insert_insert$1;
+}
 
diff --git a/tools/eliminator/asm-eliminator-test.js b/tools/eliminator/asm-eliminator-test.js
index e01a89ea4fca0..f45f082bf4fca 100644
--- a/tools/eliminator/asm-eliminator-test.js
+++ b/tools/eliminator/asm-eliminator-test.js
@@ -6823,5 +6823,20 @@ function multiloop2($n_0, $35) {
     }
   }
 }
-// EMSCRIPTEN_GENERATED_FUNCTIONS: ["asm", "__Z11printResultPiS_j", "_segment_holding", "__ZN5identC2EiPKcPci", "_vec2Length", "exc", "label", "confuusion", "tempDouble", "_org_apache_harmony_luni_util_NumberConverter_freeFormat__", "__ZN23b2EdgeAndPolygonContact8EvaluateEP10b2ManifoldRK11b2TransformS4_", "_java_nio_charset_Charset_forNameInternal___java_lang_String", "looop2", "looop3", "looop4", "looop5", "looop6", "looop7", "multiloop", "multiloop2"]
+function tempDouble2($46, $14, $28, $42, $20, $32, $45) {
+  $46 = $46 | 0;
+  $14 = $14 | 0;
+  $28 = $28 | 0;
+  $42 = $42 | 0;
+  $20 = $20 | 0;
+  $32 = $32 | 0;
+  $45 = $45 | 0;
+  var $46 = 0, $_sroa_06_0_insert_insert$1 = 0;
+  $46 = (HEAPF32[tempDoublePtr >> 2] = ($14 < $28 ? $14 : $28) - $42, HEAP32[tempDoublePtr >> 2] | 0);
+  $_sroa_06_0_insert_insert$1 = (HEAPF32[tempDoublePtr >> 2] = ($20 < $32 ? $20 : $32) - $42, HEAP32[tempDoublePtr >> 2] | 0) | 0;
+  HEAP32[$45 >> 2] = 0 | $46;
+  HEAP32[$45 + 4 >> 2] = $_sroa_06_0_insert_insert$1;
+  HEAP32[$45 + 8 >> 2] = $_sroa_06_0_insert_insert$1;
+}
+// EMSCRIPTEN_GENERATED_FUNCTIONS: ["asm", "__Z11printResultPiS_j", "_segment_holding", "__ZN5identC2EiPKcPci", "_vec2Length", "exc", "label", "confuusion", "tempDouble", "_org_apache_harmony_luni_util_NumberConverter_freeFormat__", "__ZN23b2EdgeAndPolygonContact8EvaluateEP10b2ManifoldRK11b2TransformS4_", "_java_nio_charset_Charset_forNameInternal___java_lang_String", "looop2", "looop3", "looop4", "looop5", "looop6", "looop7", "multiloop", "multiloop2", "tempDouble2"]
 
diff --git a/tools/file_packager.py b/tools/file_packager.py
index 6a8390ad17101..1443d1658e0fb 100644
--- a/tools/file_packager.py
+++ b/tools/file_packager.py
@@ -240,7 +240,7 @@ def was_seen(name):
       #if not os.path.exists(os.path.basename(crunch_name)):
       #  print >> sys.stderr, 'Failed to crunch, perhaps a weird dxt format? Looking for a source PNG for the DDS'
       #  Popen([CRUNCH, '-file', unsuffixed(file_['srcpath']) + '.png', '-quality', crunch] + format, stdout=sys.stderr).communicate()
-      assert os.path.exists(os.path.basename(src_crunch_name)), 'crunch failed to generate output'
+      assert os.path.exists(src_crunch_name), 'crunch failed to generate output'
       # prepend the dds header
       crunched = open(src_crunch_name, 'rb').read()
       c = open(src_crunch_name, 'wb')
diff --git a/tools/js-optimizer.js b/tools/js-optimizer.js
index 7e1d374fa08f0..07317e0a22d69 100644
--- a/tools/js-optimizer.js
+++ b/tools/js-optimizer.js
@@ -402,10 +402,12 @@ function removeUnneededLabelSettings(ast) {
 
 // Various expression simplifications. Pre run before closure (where we still have metadata), Post run after.
 
+var USEFUL_BINARY_OPS = set('<<', '>>', '|', '&', '^');
+
 function simplifyExpressionsPre(ast) {
   // Look for (x&A)<<B>>B and replace it with X&A if possible.
   function simplifySignExtends(ast) {
-    traverseGenerated(ast, function(node, type) {
+    traverse(ast, function(node, type) {
       if (type == 'binary'       && node[1]    == '>>' && node[3][0] == 'num' &&
           node[2][0] == 'binary' && node[2][1] == '<<' && node[2][3][0] == 'num' && node[3][1] == node[2][3][1]) {
         var innerNode = node[2][2];
@@ -427,13 +429,12 @@ function simplifyExpressionsPre(ast) {
   // 'useful' mathops already |0 anyhow.
 
   function simplifyBitops(ast) {
-    var USEFUL_BINARY_OPS = set('<<', '>>', '|', '&', '^');
     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 ZERO = ['num', 0];
     var rerun = true;
     while (rerun) {
       rerun = false;
-      traverseGenerated(ast, function process(node, type, stack) {
+      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]];
@@ -479,8 +480,12 @@ function simplifyExpressionsPre(ast) {
       return true;
     }
 
-    traverseGenerated(ast, function(node, type) {
-      if (type == 'binary' && node[1] == '&' && node[3][0] == 'num') {
+    var hasTempDoublePtr = false;
+
+    traverse(ast, function(node, type) {
+      if (type == 'name') {
+        if (node[1] == 'tempDoublePtr') hasTempDoublePtr = true;
+      } else if (type == 'binary' && node[1] == '&' && node[3][0] == 'num') {
         if (node[2][0] == 'num') return ['num', node[2][1] & node[3][1]];
         var input = node[2];
         var amount = node[3][1];
@@ -522,12 +527,133 @@ function simplifyExpressionsPre(ast) {
             return node;
           }
         }
+      } else if (type == 'assign') {
+        // optimizations for assigning into HEAP32 specifically
+        if (node[1] === true && node[2][0] == 'sub' && node[2][1][0] == 'name' && node[2][1][1] == 'HEAP32') {
+          // HEAP32[..] = x | 0 does not need the | 0 (unless it is a mandatory |0 of a call)
+          if (node[3][0] == 'binary' && node[3][1] == '|') {
+            if (node[3][2][0] == 'num' && node[3][2][1] == 0 && node[3][3][0] != 'call') {
+              node[3] = node[3][3];
+            } else if (node[3][3][0] == 'num' && node[3][3][1] == 0 && node[3][2][0] != 'call') {
+              node[3] = node[3][2];
+            }
+          }
+        }
+        var value = node[3];
+        if (value[0] == 'binary' && value[1] == '|') {
+          // canonicalize order of |0 to end
+          if (value[2][0] == 'num' && value[2][1] == 0) {
+            var temp = value[2];
+            value[2] = value[3];
+            value[3] = temp;
+          }
+          // if a seq ends in an |0, remove an external |0
+          // note that it is only safe to do this in assigns, like we are doing here (return (x, y|0); is not valid)
+          if (value[2][0] == 'seq' && value[2][2][0] == 'binary' && value[2][2][1] in USEFUL_BINARY_OPS) {
+            node[3] = value[2];
+          }
+        }
       }
     });
 
     if (asm) {
+      if (hasTempDoublePtr) {
+        traverse(ast, function(node, type) {
+          if (type == 'assign') {
+            if (node[1] === true && node[2][0] == 'sub' && node[2][1][0] == 'name' && node[2][1][1] == 'HEAP32') {
+              // remove bitcasts that are now obviously pointless, e.g.
+              // HEAP32[$45 >> 2] = HEAPF32[tempDoublePtr >> 2] = ($14 < $28 ? $14 : $28) - $42, HEAP32[tempDoublePtr >> 2] | 0;
+              var value = node[3];
+              if (value[0] == 'seq' && value[1][0] == 'assign' && value[1][2][0] == 'sub' && value[1][2][1][0] == 'name' && value[1][2][1][1] == 'HEAPF32' &&
+                  value[1][2][2][0] == 'binary' && value[1][2][2][2][0] == 'name' && value[1][2][2][2][1] == 'tempDoublePtr') {
+                // transform to HEAPF32[$45 >> 2] = ($14 < $28 ? $14 : $28) - $42;
+                node[2][1][1] = 'HEAPF32';
+                node[3] = value[1][3];
+              }
+            }
+          } else if (type == 'seq') {
+            // (HEAP32[tempDoublePtr >> 2] = HEAP32[$37 >> 2], +HEAPF32[tempDoublePtr >> 2])
+            //   ==>
+            // +HEAPF32[$37 >> 2]
+            if (node[0] == 'seq' && node[1][0] == 'assign' && node[1][2][0] == 'sub' && node[1][2][1][0] == 'name' &&
+                (node[1][2][1][1] == 'HEAP32' || node[1][2][1][1] == 'HEAPF32') &&
+                node[1][2][2][0] == 'binary' && node[1][2][2][2][0] == 'name' && node[1][2][2][2][1] == 'tempDoublePtr' &&
+                node[1][3][0] == 'sub' && node[1][3][1][0] == 'name' && (node[1][3][1][1] == 'HEAP32' || node[1][3][1][1] == 'HEAPF32')) {
+              if (node[1][2][1][1] == 'HEAP32') {
+                node[1][3][1][1] = 'HEAPF32';
+                return ['unary-prefix', '+', node[1][3]];
+              } else {
+                node[1][3][1][1] = 'HEAP32';
+                return ['binary', '|', node[1][3], ['num', 0]];
+              }
+            }
+          }
+        });
+
+        // finally, wipe out remaining ones by finding cases where all assignments to X are bitcasts, and all uses are writes to
+        // the other heap type, then eliminate the bitcast
+        var bitcastVars = {};
+        traverse(ast, function(node, type) {
+          if (type == 'assign' && node[1] === true && node[2][0] == 'name') {
+            var value = node[3];
+            if (value[0] == 'seq' && value[1][0] == 'assign' && value[1][2][0] == 'sub' && value[1][2][1][0] == 'name' &&
+                (value[1][2][1][1] == 'HEAP32' || value[1][2][1][1] == 'HEAPF32') &&
+                value[1][2][2][0] == 'binary' && value[1][2][2][2][0] == 'name' && value[1][2][2][2][1] == 'tempDoublePtr') {
+              var name = node[2][1];
+              if (!bitcastVars[name]) bitcastVars[name] = {
+                define_HEAP32: 0, define_HEAPF32: 0, use_HEAP32: 0, use_HEAPF32: 0, bad: false, namings: 0, defines: [], uses: []
+              };
+              bitcastVars[name]['define_' + value[1][2][1][1]]++;
+              bitcastVars[name].defines.push(node);
+            }
+          }
+        });
+        traverse(ast, function(node, type) {
+          if (type == 'name' && bitcastVars[node[1]]) {
+            bitcastVars[node[1]].namings++;
+          } else if (type == 'assign' && node[1] === true) {
+            var value = node[3];
+            if (value[0] == 'name') {
+              var name = value[1];
+              if (bitcastVars[name]) {
+                var target = node[2];
+                if (target[0] == 'sub' && target[1][0] == 'name' && (target[1][1] == 'HEAP32' || target[1][1] == 'HEAPF32')) {
+                  bitcastVars[name]['use_' + target[1][1]]++;
+                  bitcastVars[name].uses.push(node);
+                }
+              }
+            }
+          }
+        });
+        var asmData = normalizeAsm(ast);
+        for (var v in bitcastVars) {
+          var info = bitcastVars[v];
+          // good variables define only one type, use only one type, have definitions and uses, and define as a different type than they use
+          if (info.define_HEAP32*info.define_HEAPF32 == 0 && info.use_HEAP32*info.use_HEAPF32 == 0 &&
+              info.define_HEAP32+info.define_HEAPF32 > 0  && info.use_HEAP32+info.use_HEAPF32 > 0 &&
+              info.define_HEAP32*info.use_HEAP32 == 0 && info.define_HEAPF32*info.use_HEAPF32 == 0 &&
+              v in asmData.vars && info.namings == info.define_HEAP32+info.define_HEAPF32+info.use_HEAP32+info.use_HEAPF32) {
+            var correct = info.use_HEAP32 ? 'HEAPF32' : 'HEAP32';
+            info.defines.forEach(function(define) {
+              define[3] = define[3][1][3];
+              if (correct == 'HEAP32') {
+                define[3] = ['binary', '|', define[3], ['num', 0]];
+              } else {
+                define[3] = ['unary-prefix', '+', define[3]];
+              }
+              // do we want a simplifybitops on the new values here?
+            });
+            info.uses.forEach(function(use) {
+              use[2][1][1] = correct;
+            });
+            asmData.vars[v] = 1 - asmData.vars[v];
+          }
+        }
+        denormalizeAsm(ast, asmData);
+      }
+
       // optimize num >> num, in asm we need this here since we do not run optimizeShifts
-      traverseGenerated(ast, function(node, type) {
+      traverse(ast, function(node, type) {
         if (type == 'binary' && node[1] == '>>' && node[2][0] == 'num' && node[3][0] == 'num') {
           node[0] = 'num';
           node[1] = node[2][1] >> node[3][1];
@@ -543,7 +669,7 @@ function simplifyExpressionsPre(ast) {
     var rerun = true;
     while (rerun) {
       rerun = false;
-      traverseGenerated(ast, function(node, type) {
+      traverse(ast, function(node, type) {
         if (type == 'binary' && node[1] == '+') {
           if (node[2][0] == 'num' && node[3][0] == 'num') {
             rerun = true;
@@ -566,7 +692,7 @@ function simplifyExpressionsPre(ast) {
 
   // if (x == 0) can be if (!x), etc.
   function simplifyZeroComp(ast) {
-    traverseGenerated(ast, function(node, type) {
+    traverse(ast, function(node, type) {
       var binary;
       if (type == 'if' && (binary = node[1])[0] == 'binary') {
         if ((binary[1] == '!=' || binary[1] == '!==') && binary[3][0] == 'num' && binary[3][1] == 0) {
@@ -580,40 +706,40 @@ function simplifyExpressionsPre(ast) {
     });
   }
 
-  function asmOpts(ast) {
+  function asmOpts(fun) {
     // 1. Add final returns when necessary
     // 2. Remove unneeded coercions on function calls that have no targets (eliminator removed it)
-    traverseGeneratedFunctions(ast, function(fun) {
-      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.
-      if (returnType !== null) {
-        var stats = getStatements(fun);
-        var last = stats[stats.length-1];
-        if (last[0] != 'return') {
-          var returnValue = ['num', 0];
-          if (returnType == ASM_DOUBLE) returnValue = ['unary-prefix', '+', returnValue];
-          stats.push(['return', returnValue]);
+    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.
+    if (returnType !== null) {
+      var stats = getStatements(fun);
+      var last = stats[stats.length-1];
+      if (last[0] != 'return') {
+        var returnValue = ['num', 0];
+        if (returnType == ASM_DOUBLE) returnValue = ['unary-prefix', '+', returnValue];
+        stats.push(['return', returnValue]);
+      }
+    }
   }
 
-  simplifySignExtends(ast);
-  simplifyBitops(ast);
-  joinAdditions(ast);
-  // simplifyZeroComp(ast); TODO: investigate performance
-  if (asm) asmOpts(ast);
+  traverseGeneratedFunctions(ast, function(func) {
+    simplifySignExtends(func);
+    simplifyBitops(func);
+    joinAdditions(func);
+    // simplifyZeroComp(func); TODO: investigate performance
+    if (asm) asmOpts(func);
+  });
 }
 
 // In typed arrays mode 2, we can have
@@ -953,18 +1079,15 @@ function optimizeShiftsAggressive(ast) {
 function simplifyNotCompsDirect(node) {
   if (node[0] == 'unary-prefix' && node[1] == '!') {
     if (node[2][0] == 'binary') {
-      if (node[2][1] == '<') {
-        return ['binary', '>=', node[2][2], node[2][3]];
-      } else if (node[2][1] == '>') {
-        return ['binary', '<=', node[2][2], node[2][3]];
-      } else if (node[2][1] == '==') {
-        return ['binary', '!=', node[2][2], node[2][3]];
-      } else if (node[2][1] == '!=') {
-        return ['binary', '==', node[2][2], node[2][3]];
-      } else if (node[2][1] == '===') {
-        return ['binary', '!==', node[2][2], node[2][3]];
-      } else if (node[2][1] == '!==') {
-        return ['binary', '===', node[2][2], node[2][3]];
+      switch(node[2][1]) {
+        case '<': return ['binary', '>=', node[2][2], node[2][3]];
+        case '>': return ['binary', '<=', node[2][2], node[2][3]];
+        case '<=': return ['binary', '>', node[2][2], node[2][3]];
+        case '>=': return ['binary', '<', node[2][2], node[2][3]];
+        case '==': return ['binary', '!=', node[2][2], node[2][3]];
+        case '!=': return ['binary', '==', node[2][2], node[2][3]];
+        case '===': return ['binary', '!==', node[2][2], node[2][3]];
+        case '!==': return ['binary', '===', node[2][2], node[2][3]];
       }
     } else if (node[2][0] == 'unary-prefix' && node[2][1] == '!') {
       return node[2][2];
@@ -1850,6 +1973,13 @@ var NODES_WITHOUT_ELIMINATION_SIDE_EFFECTS = set('name', 'num', 'string', 'binar
 var IGNORABLE_ELIMINATOR_SCAN_NODES = set('num', 'toplevel', 'string', 'break', 'continue', 'dot'); // dot can only be STRING_TABLE.*
 var ABORTING_ELIMINATOR_SCAN_NODES = set('new', 'object', 'function', 'defun', 'for', 'while', 'array', 'throw'); // we could handle some of these, TODO, but nontrivial (e.g. for while, the condition is hit multiple times after the body)
 
+function isTempDoublePtrAccess(node) { // these are used in bitcasts; they are not really affecting memory, and should cause no invalidation
+  assert(node[0] == 'sub');
+  return (node[2][0] == 'name' && node[2][1] == 'tempDoublePtr') ||
+         (node[2][0] == 'binary' && ((node[2][2][0] == 'name' && node[2][2][1] == 'tempDoublePtr') ||
+                                     (node[2][3][0] == 'name' && node[2][3][1] == 'tempDoublePtr')));
+}
+
 function eliminate(ast, memSafe) {
   // Find variables that have a single use, and if they can be eliminated, do so
   traverseGeneratedFunctions(ast, function(func, type) {
@@ -2134,7 +2264,7 @@ function eliminate(ast, memSafe) {
               if (allowTracking) track(name, node[3], node);
             }
           } else if (target[0] == 'sub') {
-            if (!memoryInvalidated) {
+            if (!isTempDoublePtrAccess(target) && !memoryInvalidated) {
               invalidateMemory();
               memoryInvalidated = true;
             }
@@ -2142,7 +2272,8 @@ function eliminate(ast, memSafe) {
         } else if (type == 'sub') {
           traverseInOrder(node[1], false, !memSafe); // evaluate inner
           traverseInOrder(node[2]); // evaluate outer
-          if (!ignoreSub) { // ignoreSub means we are a write (happening later), not a read
+          // ignoreSub means we are a write (happening later), not a read
+          if (!ignoreSub && !isTempDoublePtrAccess(node)) {
             // do the memory access
             if (!callsInvalidated) {
               invalidateCalls();
@@ -2341,9 +2472,9 @@ function eliminate(ast, memSafe) {
 
     var seenUses = {}, helperReplacements = {}; // for looper-helper optimization
 
-    // clean up vars
-    //printErr('cleaning up ' + JSON.stringify(varsToRemove));
+    // clean up vars, and loop variable elimination
     traverse(func, function(node, type) {
+      // pre
       if (type === 'var') {
         node[1] = node[1].filter(function(pair) { return !varsToRemove[pair[0]] });
         if (node[1].length == 0) {
@@ -2353,6 +2484,7 @@ function eliminate(ast, memSafe) {
         }
       }
     }, function(node, type) {
+      // post
       if (type == 'name') {
         var name = node[1];
         if (name in helperReplacements) {
diff --git a/tools/shared.py b/tools/shared.py
index 8a172d9cac994..b212a9cc8234d 100644
--- a/tools/shared.py
+++ b/tools/shared.py
@@ -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 @@ def get_temp_files(self):
 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 @@ def emscripten(filename, append_ext=True, extra_args=[]):
 
   @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():
diff --git a/tools/test-js-optimizer-asm-pre-output.js b/tools/test-js-optimizer-asm-pre-output.js
index ab953e5dfa588..25d521abb93f4 100644
--- a/tools/test-js-optimizer-asm-pre-output.js
+++ b/tools/test-js-optimizer-asm-pre-output.js
@@ -8,6 +8,7 @@ function a() {
   HEAP[1024] = 5;
   whee(12, 13);
   whee(12, 13);
+  f((g = t(), g + g | 0) | 0);
 }
 function b($this, $__n) {
   $this = $this | 0;
@@ -107,4 +108,57 @@ function sign_extension_simplification() {
     print(5);
   }
 }
+function tempDoublePtr($45, $14, $28, $42) {
+  $45 = $45 | 0;
+  $14 = $14 | 0;
+  $28 = $28 | 0;
+  $42 = $42 | 0;
+  var unelim = +0, bad = 0, unelim2 = +0;
+  unelim = +(127.5 * +$14);
+  HEAPF32[$45 >> 2] = ($14 < $28 ? $14 : $28) - $42;
+  HEAP32[$world + 102916 >> 2] = _malloc(192) | 0;
+  f(+HEAPF32[$45 >> 2]);
+  g(HEAP32[$14 >> 2] | 0);
+  $42 = +HEAPF32[$42 >> 2];
+  ch($42);
+  HEAPF32[$45 >> 2] = unelim;
+  moar();
+  bad = (HEAPF32[tempDoublePtr >> 2] = 127.5 * +$14, HEAP32[tempDoublePtr >> 2] | 0);
+  func();
+  HEAP32[4] = bad;
+  HEAP32[5] = bad + 1;
+  moar();
+  unelim2 = 127 + $14 | 0;
+  func();
+  HEAP32[4] = unelim2;
+}
+function boxx($this, $aabb, $xf, $childIndex) {
+  $this = $this | 0;
+  $aabb = $aabb | 0;
+  $xf = $xf | 0;
+  $childIndex = $childIndex | 0;
+  var $2 = +0, $4 = +0, $7 = +0, $9 = +0, $13 = +0, $14 = +0, $19 = +0, $20 = +0, $22 = +0, $25 = +0, $28 = +0, $32 = +0, $42 = +0, $45 = 0, $_sroa_06_0_insert_insert$1 = +0, $51 = 0, $_sroa_0_0_insert_insert$1 = +0;
+  $2 = +HEAPF32[$xf + 12 >> 2];
+  $4 = +HEAPF32[$this + 12 >> 2];
+  $7 = +HEAPF32[$xf + 8 >> 2];
+  $9 = +HEAPF32[$this + 16 >> 2];
+  $13 = +HEAPF32[$xf >> 2];
+  $14 = $13 + ($2 * $4 - $7 * $9);
+  $19 = +HEAPF32[$xf + 4 >> 2];
+  $20 = $4 * $7 + $2 * $9 + $19;
+  $22 = +HEAPF32[$this + 20 >> 2];
+  $25 = +HEAPF32[$this + 24 >> 2];
+  $28 = $13 + ($2 * $22 - $7 * $25);
+  $32 = $19 + ($7 * $22 + $2 * $25);
+  $42 = +HEAPF32[$this + 8 >> 2];
+  $45 = $aabb;
+  $_sroa_06_0_insert_insert$1 = +(($20 < $32 ? $20 : $32) - $42);
+  HEAPF32[$45 >> 2] = ($14 < $28 ? $14 : $28) - $42;
+  HEAPF32[$45 + 4 >> 2] = $_sroa_06_0_insert_insert$1;
+  $51 = $aabb + 8 | 0;
+  $_sroa_0_0_insert_insert$1 = +($42 + ($20 > $32 ? $20 : $32));
+  HEAPF32[$51 >> 2] = $42 + ($14 > $28 ? $14 : $28);
+  HEAPF32[$51 + 4 >> 2] = $_sroa_0_0_insert_insert$1;
+  return;
+}
 
diff --git a/tools/test-js-optimizer-asm-pre.js b/tools/test-js-optimizer-asm-pre.js
index 264587d296cda..38487d2fff217 100644
--- a/tools/test-js-optimizer-asm-pre.js
+++ b/tools/test-js-optimizer-asm-pre.js
@@ -8,6 +8,7 @@ function a() {
   HEAP[(4096 & 8191) >> 2] = 5;
   whee(12, 13) | 0;
   +whee(12, 13);
+  f((g = t(), (g+g)|0)|0);
 }
 function b($this, $__n) {
   $this = $this | 0;
@@ -109,4 +110,59 @@ function sign_extension_simplification() {
     print(5);
   }
 }
-// EMSCRIPTEN_GENERATED_FUNCTIONS: ["a", "b", "rett", "ret2t", "retf", "i32_8"]
+function tempDoublePtr($45, $14, $28, $42) {
+  $45 = $45 | 0;
+  $14 = $14 | 0;
+  $28 = $28 | 0;
+  $42 = $42 | 0;
+  var unelim = 0; // only used as assign to int heap, so can avoid bitcast in definition
+  var bad = 0;
+  var unelim2 = 0; // opposite types
+  unelim = (HEAPF32[tempDoublePtr >> 2] = 127.5 * +$14, HEAP32[tempDoublePtr >> 2] | 0);
+  HEAP32[$45 >> 2] = 0 | (HEAPF32[tempDoublePtr >> 2] = ($14 < $28 ? $14 : $28) - $42, HEAP32[tempDoublePtr >> 2] | 0);
+  HEAP32[$world + 102916 >> 2] = _malloc(192) | 0;
+  f((HEAP32[tempDoublePtr >> 2] = HEAP32[$45 >> 2], +HEAPF32[tempDoublePtr >> 2]));
+  g((HEAPF32[tempDoublePtr >> 2] = HEAPF32[$14 >> 2], HEAP32[tempDoublePtr >> 2] | 0));
+  $42 = (HEAP32[tempDoublePtr >> 2] = HEAP32[$42 >> 2] | 0, +HEAPF32[tempDoublePtr >> 2]);
+  ch($42);
+  HEAP32[$45 >> 2] = unelim;
+  moar();
+  bad = (HEAPF32[tempDoublePtr >> 2] = 127.5 * +$14, HEAP32[tempDoublePtr >> 2] | 0);
+  func();
+  HEAP32[4] = bad;
+  HEAP32[5] = (bad + 1) | 0;
+  moar();
+  unelim2 = (HEAP32[tempDoublePtr >> 2] = 127 + $14, +HEAPF32[tempDoublePtr >> 2]);
+  func();
+  HEAPF32[4] = unelim2;
+}
+function boxx($this, $aabb, $xf, $childIndex) {
+  $this = $this | 0;
+  $aabb = $aabb | 0;
+  $xf = $xf | 0;
+  $childIndex = $childIndex | 0;
+  var $2 = +0, $4 = +0, $7 = +0, $9 = +0, $13 = +0, $14 = +0, $19 = +0, $20 = +0, $22 = +0, $25 = +0, $28 = +0, $32 = +0, $42 = +0, $45 = 0, $_sroa_06_0_insert_insert$1 = 0, $51 = 0, $_sroa_0_0_insert_insert$1 = 0;
+  $2 = +HEAPF32[$xf + 12 >> 2];
+  $4 = +HEAPF32[$this + 12 >> 2];
+  $7 = +HEAPF32[$xf + 8 >> 2];
+  $9 = +HEAPF32[$this + 16 >> 2];
+  $13 = +HEAPF32[$xf >> 2];
+  $14 = $13 + ($2 * $4 - $7 * $9);
+  $19 = +HEAPF32[$xf + 4 >> 2];
+  $20 = $4 * $7 + $2 * $9 + $19;
+  $22 = +HEAPF32[$this + 20 >> 2];
+  $25 = +HEAPF32[$this + 24 >> 2];
+  $28 = $13 + ($2 * $22 - $7 * $25);
+  $32 = $19 + ($7 * $22 + $2 * $25);
+  $42 = +HEAPF32[$this + 8 >> 2];
+  $45 = $aabb;
+  $_sroa_06_0_insert_insert$1 = (HEAPF32[tempDoublePtr >> 2] = ($20 < $32 ? $20 : $32) - $42, HEAP32[tempDoublePtr >> 2] | 0) | 0;
+  HEAPF32[$45 >> 2] = ($14 < $28 ? $14 : $28) - $42;
+  HEAP32[$45 + 4 >> 2] = $_sroa_06_0_insert_insert$1;
+  $51 = $aabb + 8 | 0;
+  $_sroa_0_0_insert_insert$1 = (HEAPF32[tempDoublePtr >> 2] = $42 + ($20 > $32 ? $20 : $32), HEAP32[tempDoublePtr >> 2] | 0) | 0;
+  HEAPF32[$51 >> 2] = $42 + ($14 > $28 ? $14 : $28);
+  HEAP32[$51 + 4 >> 2] = $_sroa_0_0_insert_insert$1;
+  return;
+}
+// EMSCRIPTEN_GENERATED_FUNCTIONS: ["a", "b", "rett", "ret2t", "retf", "i32_8", "tempDoublePtr", "boxx"]