Merge branch 'incoming'

23 files changed, 1730 insertions, 558 deletions

diff --git a/src/analyzer.js b/src/analyzer.js
index ecb5ea6b..92b7d8cf 100644
--- a/src/analyzer.js
+++ b/src/analyzer.js
@@ -18,7 +18,7 @@ function recomputeLines(func) {
 // Handy sets
 
 var BRANCH_INVOKE = set('branch', 'invoke');
-var LABEL_ENDERS = set('branch', 'return');
+var LABEL_ENDERS = set('branch', 'return', 'switch');
 var SIDE_EFFECT_CAUSERS = set('call', 'invoke', 'atomic');
 var UNUNFOLDABLE = set('value', 'structvalue', 'type', 'phiparam');
 
@@ -231,9 +231,10 @@ function analyzer(data, sidePass) {
                 }
                 if (isIllegalType(item.valueType) || isIllegalType(item.type)) {
                   isIllegal = true;
-                }
-                if ((item.intertype == 'load' || item.intertype == 'store') && isStructType(item.valueType)) {
+                } else if ((item.intertype == 'load' || item.intertype == 'store') && isStructType(item.valueType)) {
                   isIllegal = true; // storing an entire structure is illegal
+                } else if (item.intertype == 'mathop' && item.op == 'trunc' && isIllegalType(item.params[1].ident)) { // trunc stores target value in second ident
+                  isIllegal = true;
                 }
               });
               if (!isIllegal) {
@@ -290,7 +291,7 @@ function analyzer(data, sidePass) {
                   var elements = getLegalParams([item.value], bits)[0];
                   var j = 0;
                   elements.forEach(function(element) {
-                    var tempVar = '$st$' + i + '$' + j;
+                    var tempVar = '$st$' + (tempId++) + '$' + j;
                     toAdd.push({
                       intertype: 'getelementptr',
                       assignTo: tempVar,
@@ -401,7 +402,7 @@ function analyzer(data, sidePass) {
                   var j = 0;
                   var toAdd = [];
                   elements.forEach(function(element) {
-                    var tempVar = '$st$' + i + '$' + j;
+                    var tempVar = '$ld$' + (tempId++) + '$' + j;
                     toAdd.push({
                       intertype: 'getelementptr',
                       assignTo: tempVar,
@@ -652,13 +653,14 @@ function analyzer(data, sidePass) {
                   if (!isNumber(shifts)) {
                     // We can't statically legalize this, do the operation at runtime TODO: optimize
                     assert(sourceBits == 64, 'TODO: handle nonconstant shifts on != 64 bits');
+                    assert(PRECISE_I64_MATH, 'Must have precise i64 math for non-constant 64-bit shifts');
+                    Types.preciseI64MathUsed = 1;
                     value.intertype = 'value';
-                    value.ident = 'Runtime' + (ASM_JS ? '_' : '.') + 'bitshift64(' + 
+                    value.ident = 'var ' + value.assignTo + '$0 = _bitshift64' + value.op[0].toUpperCase() + value.op.substr(1) + '(' + 
                         asmCoercion(sourceElements[0].ident, 'i32') + ',' +
                         asmCoercion(sourceElements[1].ident, 'i32') + ',' +
-                        Runtime['BITSHIFT64_' + value.op.toUpperCase()] + ',' +
                         asmCoercion(value.params[1].ident + '$0', 'i32') + ');' +
-                      'var ' + value.assignTo + '$0 = ' + makeGetTempDouble(0, 'i32') + ', ' + value.assignTo + '$1 = ' + makeGetTempDouble(1, 'i32') + ';';
+                        'var ' + value.assignTo + '$1 = tempRet0;';
                     value.assignTo = null;
                     i++;
                     continue;
@@ -952,6 +954,7 @@ function analyzer(data, sidePass) {
         // Function parameters
         func.params.forEach(function(param) {
           if (param.intertype !== 'varargs') {
+            if (func.variables[param.ident]) warn('cannot have duplicate variable names: ' + param.ident); // toNiceIdent collisions?
             func.variables[param.ident] = {
               ident: param.ident,
               type: param.type,
@@ -965,6 +968,7 @@ function analyzer(data, sidePass) {
         // Normal variables
         func.lines.forEach(function(item, i) {
           if (item.assignTo) {
+            if (func.variables[item.assignTo]) warn('cannot have duplicate variable names: ' + item.assignTo); // toNiceIdent collisions?
             var variable = func.variables[item.assignTo] = {
               ident: item.assignTo,
               type: item.type,
@@ -1380,7 +1384,7 @@ function analyzer(data, sidePass) {
           var label = func.labels[i];
           for (var j = 0; j < label.lines.length; j++) {
             var line = label.lines[j];
-            if (line.intertype == 'call' && line.ident == setjmp) {
+            if ((line.intertype == 'call' || line.intertype == 'invoke') && line.ident == setjmp) {
               // Add a new label
               var oldIdent = label.ident;
               var newIdent = func.labelIdCounter++;
diff --git a/src/compiler.js b/src/compiler.js
index 3047daf1..bb72c7dd 100644
--- a/src/compiler.js
+++ b/src/compiler.js
@@ -141,8 +141,13 @@ if (phase == 'pre') {
 
 if (settings_file) {
   var settings = JSON.parse(read(settings_file));
-  for (setting in settings) {
-    eval(setting + ' = ' + JSON.stringify(settings[setting]));
+  for (key in settings) {
+    var value = settings[key];
+    if (value[0] == '@') {
+      // response file type thing, workaround for large inputs: value is @path-to-file
+      value = JSON.parse(read(value.substr(1)));
+    }
+    eval(key + ' = ' + JSON.stringify(value));
   }
 }
 
@@ -163,6 +168,7 @@ if (SAFE_HEAP >= 2) {
 EXPORTED_FUNCTIONS = set(EXPORTED_FUNCTIONS);
 EXPORTED_GLOBALS = set(EXPORTED_GLOBALS);
 EXCEPTION_CATCHING_WHITELIST = set(EXCEPTION_CATCHING_WHITELIST);
+DEAD_FUNCTIONS = numberedSet(DEAD_FUNCTIONS);
 
 RUNTIME_DEBUG = LIBRARY_DEBUG || GL_DEBUG;
 
diff --git a/src/experimental/allow_loopvars_from_memsetcpy_inasm.diff b/src/experimental/allow_loopvars_from_memsetcpy_inasm.diff
new file mode 100644
index 00000000..a2dde7da
--- /dev/null
+++ b/src/experimental/allow_loopvars_from_memsetcpy_inasm.diff
@@ -0,0 +1,97 @@
+commit a61ef3dbbaf7333ad67fca29c0aad5bcc99b653a
+Author: Alon Zakai <alonzakai@gmail.com>
+Date:   Wed Mar 6 18:18:03 2013 -0800
+
+    handle new vars in asm code, such as the loop vars from memset/memcpy loops
+
+diff --git a/tools/js-optimizer.js b/tools/js-optimizer.js
+index f2dc516..65059e8 100644
+--- a/tools/js-optimizer.js
++++ b/tools/js-optimizer.js
+@@ -1321,7 +1321,7 @@ function normalizeAsm(func) {
+       var name = v[0];
+       var value = v[1];
+       if (!(name in data.vars)) {
+-        assert(value[0] == 'num' || (value[0] == 'unary-prefix' && value[2][0] == 'num')); // must be valid coercion no-op
++        if (!(value[0] == 'num' || (value[0] == 'unary-prefix' && value[2][0] == 'num'))) break outer; // must be valid coercion no-op
+         data.vars[name] = detectAsmCoercion(value);
+         v.length = 1; // make an un-assigning var
+       } else {
+@@ -1331,6 +1331,7 @@ function normalizeAsm(func) {
+     i++;
+   }
+   // finally, look for other var definitions and collect them
++  var extra = [];
+   while (i < stats.length) {
+     traverse(stats[i], function(node, type) {
+       if (type == 'var') {
+@@ -1340,6 +1341,7 @@ function normalizeAsm(func) {
+           var value = v[1];
+           if (!(name in data.vars)) {
+             data.vars[name] = detectAsmCoercion(value);
++            extra.push(['var', [[name]]]); // add a 'var' for normal JS
+           }
+         }
+         unVarify(node[1], node);
+@@ -1353,6 +1355,7 @@ function normalizeAsm(func) {
+     });
+     i++;
+   }
++  if (extra.length > 0) stats.splice.apply(stats, [0, 0].concat(extra));
+   //printErr('normalized \n\n' + astToSrc(func) + '\n\nwith: ' + JSON.stringify(data));
+   return data;
+ }
+diff --git a/tools/test-js-optimizer-asm-regs-output.js b/tools/test-js-optimizer-asm-regs-output.js
+index 99bccd2..f84b8d5 100644
+--- a/tools/test-js-optimizer-asm-regs-output.js
++++ b/tools/test-js-optimizer-asm-regs-output.js
+@@ -9,6 +9,18 @@ function asm(d1, i2) {
+   d4 = d1 * 5;
+   return d4;
+ }
++function asm2(d1, i2) {
++  d1 = +d1;
++  i2 = i2 | 0;
++  var i3 = 0, d4 = +0;
++  i3 = i2;
++  i2 = d1 + i3 | 0;
++  d1 = d(Math_max(10, Math_min(5, f())));
++  i3 = i2 + 2 | 0;
++  print(i3);
++  d4 = d1 * 5;
++  return d4;
++}
+ function _doit(i1, i2, i3) {
+   i1 = i1 | 0;
+   i2 = i2 | 0;
+diff --git a/tools/test-js-optimizer-asm-regs.js b/tools/test-js-optimizer-asm-regs.js
+index 0afced2..fbaa7c4 100644
+--- a/tools/test-js-optimizer-asm-regs.js
++++ b/tools/test-js-optimizer-asm-regs.js
+@@ -10,6 +10,19 @@ function asm(x, y) {
+   double2 = double1*5;
+   return double2;
+ }
++function asm2(x, y) {
++  x = +x;
++  y = y | 0;
++  var int1 = 0, int2 = 0; // do not mix the types!
++  var double1 = +0, double2 = +0;
++  var tempy = y;
++  int1 = (x+tempy)|0;
++  double1 = d(Math.max(10, Math_min(5, f())));
++  int2 = (int1+2)|0;
++  print(int2);
++  double2 = double1*5;
++  return double2;
++}
+ function _doit($x, $y$0, $y$1) {
+   $x = $x | 0;
+   $y$0 = $y$0 | 0;
+@@ -41,5 +54,5 @@ function retf() {
+   }
+   // missing final return, need it as a float
+ }
+-// EMSCRIPTEN_GENERATED_FUNCTIONS: ["asm", "_doit", "rett", "ret2t", "retf"]
++// EMSCRIPTEN_GENERATED_FUNCTIONS: ["asm", "asm2", "_doit", "rett", "ret2t", "retf"]
+ 
diff --git a/src/experimental/simplifyGeneratedFunctionsDetection.diff b/src/experimental/simplifyGeneratedFunctionsDetection.diff
new file mode 100644
index 00000000..09e0ebcd
--- /dev/null
+++ b/src/experimental/simplifyGeneratedFunctionsDetection.diff
@@ -0,0 +1,336 @@
+diff --git a/emscripten.py b/emscripten.py
+index b698654..a7843c7 100755
+--- a/emscripten.py
++++ b/emscripten.py
+@@ -294,8 +294,6 @@ def emscript(infile, settings, outfile, libraries=[], compiler_engine=None,
+   outfile.write(blockaddrsize(indexize(pre)))
+   pre = None
+ 
+-  #if DEBUG: outfile.write('// funcs\n')
+-
+   # forward
+   forwarded_data = json.dumps(forwarded_json)
+   forwarded_file = temp_files.get('.2.json').name
+@@ -483,9 +481,14 @@ Runtime.stackRestore = function(top) { asm.stackRestore(top) };
+   else:
+     function_tables_defs = '\n'.join([table for table in last_forwarded_json['Functions']['tables'].itervalues()])
+     outfile.write(function_tables_defs)
++
++  outfile.write('// EMSCRIPTEN_START_FUNCS\n')
++
+   outfile.write(blockaddrsize(indexize(funcs_js)))
+   funcs_js = None
+ 
++  outfile.write('// EMSCRIPTEN_END_FUNCS\n')
++
+   outfile.write(indexize(post))
+   if DEBUG: print >> sys.stderr, '  emscript: phase 3 took %s seconds' % (time.time() - t)
+ 
+diff --git a/src/jsifier.js b/src/jsifier.js
+index ff58ece..d6aa110 100644
+--- a/src/jsifier.js
++++ b/src/jsifier.js
+@@ -1587,12 +1587,6 @@ function JSify(data, functionsOnly, givenFunctions) {
+ 
+     var shellParts = read(shellFile).split('{{BODY}}');
+     print(shellParts[1]);
+-    // Print out some useful metadata (for additional optimizations later, like the eliminator)
+-    if (EMIT_GENERATED_FUNCTIONS) {
+-      print('// EMSCRIPTEN_GENERATED_FUNCTIONS: ' + JSON.stringify(keys(Functions.implementedFunctions).filter(function(func) {
+-        return IGNORED_FUNCTIONS.indexOf(func.ident) < 0;
+-      })) + '\n');
+-    }
+ 
+     PassManager.serialize();
+ 
+diff --git a/src/settings.js b/src/settings.js
+index 1bfcf92..99c83f4 100644
+--- a/src/settings.js
++++ b/src/settings.js
+@@ -336,8 +336,6 @@ var EXPLICIT_ZEXT = 0; // If 1, generate an explicit conversion of zext i1 to i3
+ 
+ var NECESSARY_BLOCKADDRS = []; // List of (function, block) for all block addresses that are taken.
+ 
+-var EMIT_GENERATED_FUNCTIONS = 0; // whether to emit the list of generated functions, needed for external JS optimization passes
+-
+ // Compiler debugging options
+ var DEBUG_TAGS_SHOWING = [];
+   // Some useful items:
+diff --git a/tools/js-optimizer.js b/tools/js-optimizer.js
+index f2dc516..9fa038c 100644
+--- a/tools/js-optimizer.js
++++ b/tools/js-optimizer.js
+@@ -140,16 +140,6 @@ var UNDEFINED_NODE = ['unary-prefix', 'void', ['num', 0]];
+ var TRUE_NODE = ['unary-prefix', '!', ['num', 0]];
+ var FALSE_NODE = ['unary-prefix', '!', ['num', 1]];
+ 
+-var GENERATED_FUNCTIONS_MARKER = '// EMSCRIPTEN_GENERATED_FUNCTIONS:';
+-var generatedFunctions = null;
+-function setGeneratedFunctions(metadata) {
+-  var start = metadata.indexOf(GENERATED_FUNCTIONS_MARKER);
+-  generatedFunctions = set(eval(metadata.substr(start + GENERATED_FUNCTIONS_MARKER.length)));
+-}
+-function isGenerated(ident) {
+-  return ident in generatedFunctions;
+-}
+-
+ function srcToAst(src) {
+   return uglify.parser.parse(src);
+ }
+@@ -212,21 +202,9 @@ function traverse(node, pre, post, stack) {
+   return result;
+ }
+ 
+-// Only walk through the generated functions
+-function traverseGenerated(ast, pre, post, stack) {
+-  assert(generatedFunctions);
+-  traverse(ast, function(node) {
+-    if (node[0] == 'defun' && isGenerated(node[1])) {
+-      traverse(node, pre, post, stack);
+-      return null;
+-    }
+-  });
+-}
+-
+-function traverseGeneratedFunctions(ast, callback) {
+-  assert(generatedFunctions);
++function traverseFunctions(ast, callback) {
+   traverse(ast, function(node) {
+-    if (node[0] == 'defun' && isGenerated(node[1])) {
++    if (node[0] == 'defun') {
+       callback(node);
+       return null;
+     }
+@@ -418,7 +396,7 @@ function simplifyExpressionsPre(ast) {
+     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]];
+@@ -455,7 +433,7 @@ function simplifyExpressionsPre(ast) {
+     }
+ 
+     // &-related optimizations
+-    traverseGenerated(ast, function(node, type) {
++    traverse(ast, function(node, type) {
+       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];
+@@ -489,7 +467,7 @@ function simplifyExpressionsPre(ast) {
+ 
+     if (asm) {
+       // 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];
+@@ -505,7 +483,7 @@ 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;
+@@ -528,7 +506,7 @@ 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) {
+@@ -554,7 +532,7 @@ function simplifyExpressionsPre(ast) {
+ // TODO: when shifting a variable, if there are other uses, keep an unshifted version too, to prevent slowdowns?
+ function optimizeShiftsInternal(ast, conservative) {
+   var MAX_SHIFTS = 3;
+-  traverseGeneratedFunctions(ast, function(fun) {
++  traverseFunctions(ast, function(fun) {
+     var funMore = true;
+     var funFinished = {};
+     while (funMore) {
+@@ -999,7 +977,7 @@ function vacuum(ast) {
+       } break;
+     }
+   }
+-  traverseGeneratedFunctions(ast, function(node) {
++  traverseFunctions(ast, function(node) {
+     vacuumInternal(node);
+     simplifyNotComps(node);
+   });
+@@ -1021,7 +999,7 @@ function getStatements(node) {
+ //   if (condition) { label == x } else ..
+ // We can hoist the multiple block into the condition, thus removing code and one 'if' check
+ function hoistMultiples(ast) {
+-  traverseGeneratedFunctions(ast, function(node) {
++  traverseFunctions(ast, function(node) {
+     traverse(node, function(node, type) {
+       var statements = getStatements(node);
+       if (!statements) return;
+@@ -1135,7 +1113,7 @@ function hoistMultiples(ast) {
+   //    if (..) { .. break|continue } else { .. }
+   // to
+   //    if (..) { .. break|continue } ..
+-  traverseGenerated(ast, function(container, type) {
++  traverse(ast, function(container, type) {
+     var statements = getStatements(container);
+     if (!statements) return;
+     for (var i = 0; i < statements.length; i++) {
+@@ -1168,7 +1146,7 @@ function loopOptimizer(ast) {
+   function passTwo(ast) {
+     var neededDos = [];
+     // Find unneeded labels
+-    traverseGenerated(ast, function(node, type, stack) {
++    traverse(ast, function(node, type, stack) {
+       if (type == 'label' && node[2][0] in LOOP) {
+         // this is a labelled loop. we don't know if it's needed yet. Mark its label for removal for now now.
+         stack.push(node);
+@@ -1212,7 +1190,7 @@ function loopOptimizer(ast) {
+     // We return whether another pass is necessary
+     var more = false;
+     // Remove unneeded labels
+-    traverseGenerated(ast, function(node, type) {
++    traverse(ast, function(node, type) {
+       if (type == 'label' && node[1][0] == '+') {
+         more = true;
+         var ident = node[1].substr(1);
+@@ -1227,13 +1205,13 @@ function loopOptimizer(ast) {
+     });
+     // Remove unneeded one-time loops. We need such loops if (1) they have a label, or (2) they have a direct break so they are in neededDos.
+     // First, add all labeled loops of this nature to neededDos
+-    traverseGenerated(ast, function(node, type) {
++    traverse(ast, function(node, type) {
+       if (type == 'label' && node[2][0] == 'do') {
+         neededDos.push(node[2]);
+       }
+     });
+     // Remove unneeded dos, we know who they are now
+-    traverseGenerated(ast, function(node, type) {
++    traverse(ast, function(node, type) {
+       if (type == 'do' && neededDos.indexOf(node) < 0) {
+         assert(jsonCompare(node[1], ['num', 0]), 'Trying to remove a one-time do loop that is not one of our generated ones.;');
+         more = true;
+@@ -1407,7 +1385,7 @@ function denormalizeAsm(func, data) {
+ //       we still need the eliminator? Closure? And in what order? Perhaps just
+ //       closure simple?
+ function registerize(ast) {
+-  traverseGeneratedFunctions(ast, function(fun) {
++  traverseFunctions(ast, function(fun) {
+     if (asm) var asmData = normalizeAsm(fun);
+     // Add parameters as a first (fake) var (with assignment), so they get taken into consideration
+     var params = {}; // note: params are special, they can never share a register between them (see later)
+@@ -1671,7 +1649,7 @@ var ABORTING_ELIMINATOR_SCAN_NODES = set('new', 'object', 'function', 'defun', '
+ 
+ function eliminate(ast, memSafe) {
+   // Find variables that have a single use, and if they can be eliminated, do so
+-  traverseGeneratedFunctions(ast, function(func, type) {
++  traverseFunctions(ast, function(func, type) {
+     if (asm) var asmData = normalizeAsm(func);
+     //printErr('eliminate in ' + func[1]);
+ 
+@@ -2218,9 +2196,6 @@ var passes = {
+ var src = read(arguments_[0]);
+ var ast = srcToAst(src);
+ //printErr(JSON.stringify(ast)); throw 1;
+-var metadata = src.split('\n').filter(function(line) { return line.indexOf(GENERATED_FUNCTIONS_MARKER) >= 0 })[0];
+-//assert(metadata, 'Must have EMSCRIPTEN_GENERATED_FUNCTIONS metadata');
+-if (metadata) setGeneratedFunctions(metadata);
+ 
+ arguments_.slice(1).forEach(function(arg) {
+   passes[arg](ast);
+diff --git a/tools/js_optimizer.py b/tools/js_optimizer.py
+index 2fd2211..a253afb 100644
+--- a/tools/js_optimizer.py
++++ b/tools/js_optimizer.py
+@@ -41,17 +41,17 @@ def run_on_js(filename, passes, js_engine, jcache):
+   if os.linesep != '\n':
+     js = js.replace(os.linesep, '\n') # we assume \n in the splitting code
+ 
+-  # Find suffix
+-  suffix_marker = '// EMSCRIPTEN_GENERATED_FUNCTIONS'
+-  suffix_start = js.find(suffix_marker)
+-  suffix = ''
+-  if suffix_start >= 0:
+-    suffix = js[suffix_start:js.find('\n', suffix_start)] + '\n'
+-    # if there is metadata, we will run only on the generated functions. If there isn't, we will run on everything.
+-    generated = set(eval(suffix[len(suffix_marker)+1:]))
+-
+-  if not suffix and jcache:
+-    # JCache cannot be used without metadata, since it might reorder stuff, and that's dangerous since only generated can be reordered
++  # Find markers
++  start_marker = '// EMSCRIPTEN_START_FUNCS\n'
++  end_marker = '// EMSCRIPTEN_END_FUNCS\n'
++  start = js.find(start_marker)
++  end = js.find(end_marker)
++  assert (start >= 0) == (end >= 0), 'must have both markers or neither'
++  have_markers = start >= 0
++  # if there are markers, we will run only on the generated functions. If there isn't, we will run on everything.
++
++  if not have_markers and jcache:
++    # JCache cannot be used without markers, since it might reorder stuff, and that's dangerous since only generated can be reordered
+     # This means jcache does not work after closure compiler runs, for example. But you won't get much benefit from jcache with closure
+     # anyhow (since closure is likely the longest part of the build).
+     if DEBUG: print >>sys.stderr, 'js optimizer: no metadata, so disabling jcache'
+@@ -59,27 +59,10 @@ def run_on_js(filename, passes, js_engine, jcache):
+ 
+   # If we process only generated code, find that and save the rest on the side
+   func_sig = re.compile('( *)function (_[\w$]+)\(')
+-  if suffix:
+-    pos = 0
+-    gen_start = 0
+-    gen_end = 0
+-    while 1:
+-      m = func_sig.search(js, pos)
+-      if not m: break
+-      pos = m.end()
+-      indent = m.group(1)
+-      ident = m.group(2)
+-      if ident in generated:
+-        if not gen_start:
+-          gen_start = m.start()
+-          assert gen_start
+-        gen_end = js.find('\n%s}\n' % indent, m.end()) + (3 + len(indent))
+-    assert gen_end > gen_start
+-    pre = js[:gen_start]
+-    post = js[gen_end:]
+-    if 'last' in passes:
+-      post = post.replace(suffix, '') # no need to write out the metadata - nothing after us needs it
+-    js = js[gen_start:gen_end]
++  if have_markers:
++    pre = js[:start + len(start_marker)] # includes start marker
++    post = js[end:] # includes end marker
++    js = js[start + len(start_marker):end]
+   else:
+     pre = ''
+     post = ''
+@@ -95,7 +78,7 @@ def run_on_js(filename, passes, js_engine, jcache):
+     if m:
+       ident = m.group(2)
+     else:
+-      if suffix: continue # ignore whitespace
++      if have_markers: continue # ignore whitespace
+       ident = 'anon_%d' % i
+     assert ident
+     funcs.append((ident, func))
+@@ -131,7 +114,6 @@ def run_on_js(filename, passes, js_engine, jcache):
+       temp_file = temp_files.get('.jsfunc_%d.js' % i).name
+       f = open(temp_file, 'w')
+       f.write(chunk)
+-      f.write(suffix)
+       f.close()
+       return temp_file
+     filenames = [write_chunk(chunks[i], i) for i in range(len(chunks))]
+@@ -169,7 +151,6 @@ def run_on_js(filename, passes, js_engine, jcache):
+       f.write(cached); # TODO: preserve order
+       f.write('\n')
+   f.write(post);
+-  # No need to write suffix: if there was one, it is inside post which exists when suffix is there
+   f.write('\n')
+   f.close()
+ 
diff --git a/src/headless.js b/src/headless.js
index 8e847d27..d81fb5a3 100644
--- a/src/headless.js
+++ b/src/headless.js
@@ -537,7 +537,7 @@ var document = {
                       case /* GL_MAX_FRAGMENT_UNIFORM_VECTORS     */ 0x8DFD: return 4096;
                       case /* GL_MAX_VARYING_VECTORS              */ 0x8DFC: return 32;
                       case /* GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS */ 0x8B4D: return 32;
-                      default: throw 'getParameter ' + pname;
+                      default: console.log('getParameter ' + pname + '?'); return 0;
                     }
                   },
                   getSupportedExtensions: function() {
@@ -686,6 +686,7 @@ var document = {
               document.callEventListeners('pointerlockchange');
             });
           },
+          exitPointerLock: function(){},
           style: {},
           eventListeners: {},
           addEventListener: document.addEventListener,
@@ -748,6 +749,8 @@ var document = {
   body: {
     appendChild: function(){},
   },
+  exitPointerLock: function(){},
+  cancelFullScreen: function(){},
 };
 var alert = function(x) {
   print(x);
diff --git a/src/intertyper.js b/src/intertyper.js
index 2103ecfa..57e3011d 100644
--- a/src/intertyper.js
+++ b/src/intertyper.js
@@ -122,16 +122,19 @@ function intertyper(data, sidePass, baseLineNums) {
               SKIP_STACK_IN_SMALL = 0;
             }
 
-            unparsedBundles.push({
-              intertype: 'unparsedFunction',
-              // We need this early, to know basic function info - ident, params, varargs
-              ident: toNiceIdent(func.ident),
-              params: func.params,
-              returnType: func.returnType,
-              hasVarArgs: func.hasVarArgs,
-              lineNum: currFunctionLineNum,
-              lines: currFunctionLines
-            });
+            var ident = toNiceIdent(func.ident);
+            if (!(ident in DEAD_FUNCTIONS)) {
+              unparsedBundles.push({
+                intertype: 'unparsedFunction',
+                // We need this early, to know basic function info - ident, params, varargs
+                ident: ident,
+                params: func.params,
+                returnType: func.returnType,
+                hasVarArgs: func.hasVarArgs,
+                lineNum: currFunctionLineNum,
+                lines: currFunctionLines
+              });
+            }
             currFunctionLines = [];
           }
         }
diff --git a/src/jsifier.js b/src/jsifier.js
index ff58ece2..4263618a 100644
--- a/src/jsifier.js
+++ b/src/jsifier.js
@@ -397,6 +397,20 @@ function JSify(data, functionsOnly, givenFunctions) {
     }
   });
 
+  function processLibraryFunction(snippet, ident) {
+    snippet = snippet.toString();
+    assert(snippet.indexOf('XXX missing C define') == -1,
+           'Trying to include a library function with missing C defines: ' + ident + ' | ' + snippet);
+
+    // name the function; overwrite if it's already named
+    snippet = snippet.replace(/function(?:\s+([^(]+))?\s*\(/, 'function _' + ident + '(');
+    if (LIBRARY_DEBUG) {
+      snippet = snippet.replace('{', '{ var ret = (function() { if (Runtime.debug) Module.printErr("[library call:' + ident + ': " + Array.prototype.slice.call(arguments).map(Runtime.prettyPrint) + "]"); ');
+      snippet = snippet.substr(0, snippet.length-1) + '}).apply(this, arguments); if (Runtime.debug && typeof ret !== "undefined") Module.printErr("  [     return:" + Runtime.prettyPrint(ret)); return ret; \n}';
+    }
+    return snippet;
+  }
+
   // functionStub
   substrate.addActor('FunctionStub', {
     processItem: function(item) {
@@ -434,16 +448,7 @@ function JSify(data, functionsOnly, givenFunctions) {
           snippet = stringifyWithFunctions(snippet);
         } else if (typeof snippet === 'function') {
           isFunction = true;
-          snippet = snippet.toString();
-          assert(snippet.indexOf('XXX missing C define') == -1,
-                 'Trying to include a library function with missing C defines: ' + ident + ' | ' + snippet);
-
-          // name the function; overwrite if it's already named
-          snippet = snippet.replace(/function(?:\s+([^(]+))?\s*\(/, 'function _' + ident + '(');
-          if (LIBRARY_DEBUG) {
-            snippet = snippet.replace('{', '{ var ret = (function() { if (Runtime.debug) Module.printErr("[library call:' + ident + ': " + Array.prototype.slice.call(arguments).map(Runtime.prettyPrint) + "]"); ');
-            snippet = snippet.substr(0, snippet.length-1) + '}).apply(this, arguments); if (Runtime.debug && typeof ret !== "undefined") Module.printErr("  [     return:" + Runtime.prettyPrint(ret)); return ret; \n}';
-          }
+          snippet = processLibraryFunction(snippet, ident);
           if (ASM_JS) Functions.libraryFunctions[ident] = 1;
         }
 
@@ -600,6 +605,10 @@ function JSify(data, functionsOnly, givenFunctions) {
       
       func.JS += 'function ' + func.ident + '(' + paramIdents.join(', ') + ') {\n';
 
+      if (PGO) {
+        func.JS += '  PGOMonitor.called["' + func.ident + '"] = 1;\n';
+      }
+
       if (ASM_JS) {
         // spell out argument types
         func.params.forEach(function(param) {
@@ -731,7 +740,7 @@ function JSify(data, functionsOnly, givenFunctions) {
               return indent + '  case ' + getLabelId(label.ident) + ': ' + (SHOW_LABELS ? '// ' + getOriginalLabelId(label.ident) : '') + '\n'
                             + getLabelLines(label, indent + '    ');
             }).join('\n') + '\n';
-            if (ASSERTIONS) ret += indent + '  default: assert(0, "bad label: " + label);\n';
+            if (ASSERTIONS) ret += indent + '  default: assert(0' + (ASM_JS ? '' : ', "bad label: " + label') + ');\n';
             ret += indent + '}\n';
             if (func.setjmpTable) {
               ret += ' } catch(e) { if (!e.longjmp || !(e.id in mySetjmpIds)) throw(e); setjmpTable[setjmpLabels[e.id]](e.value) }';
@@ -1312,6 +1321,8 @@ function JSify(data, functionsOnly, givenFunctions) {
     } else {
       callIdent = ident;
     }
+    if (callIdent == '0') return 'abort(-2)';
+
     var args = [];
     var argsTypes = [];
     var varargs = [];
@@ -1414,6 +1425,12 @@ function JSify(data, functionsOnly, givenFunctions) {
       returnType = getReturnType(type);
     }
 
+    if (callIdent in DEAD_FUNCTIONS) {
+      var ret = 'abort(' + DEAD_FUNCTIONS[callIdent] + ')';
+      if (ASM_JS) ret = asmCoercion(ret, returnType);
+      return ret;
+    }
+
     if (byPointer) {
       var sig = Functions.getSignature(returnType, argsTypes, hasVarArgs);
       if (ASM_JS) {
@@ -1443,7 +1460,7 @@ function JSify(data, functionsOnly, givenFunctions) {
 
   makeFuncLineActor('unreachable', function(item) {
     if (ASSERTIONS) {
-      return 'throw "Reached an unreachable!"';
+      return ASM_JS ? 'abort()' : 'throw "Reached an unreachable!"';
     } else {
       return ';';
     }
@@ -1552,16 +1569,25 @@ function JSify(data, functionsOnly, givenFunctions) {
     // This is the main 'post' pass. Print out the generated code that we have here, together with the
     // rest of the output that we started to print out earlier (see comment on the
     // "Final shape that will be created").
-    if (CORRUPTION_CHECK) {
-      assert(!ASM_JS); // cannot monkeypatch asm!
-      print(processMacros(read('corruptionCheck.js')));
-    }
     if (PRECISE_I64_MATH && Types.preciseI64MathUsed) {
+      if (!INCLUDE_FULL_LIBRARY) {
+        ['i64Add', 'bitshift64Shl', 'bitshift64Lshr', 'bitshift64Ashr'].forEach(function(func) {
+          print(processLibraryFunction(LibraryManager.library[func], func)); // must be first to be close to generated code
+          Functions.implementedFunctions['_' + func] = LibraryManager.library[func + '__sig'];
+        });
+      }
+      print('// EMSCRIPTEN_END_FUNCS\n');
       print(read('long.js'));
     } else {
+      print('// EMSCRIPTEN_END_FUNCS\n');
       print('// Warning: printing of i64 values may be slightly rounded! No deep i64 math used, so precise i64 code not included');
       print('var i64Math = null;');
     }
+
+    if (CORRUPTION_CHECK) {
+      assert(!ASM_JS); // cannot monkeypatch asm!
+      print(processMacros(read('corruptionCheck.js')));
+    }
     if (HEADLESS) {
       print('if (!ENVIRONMENT_IS_WEB) {');
       print(read('headless.js').replace("'%s'", "'http://emscripten.org'").replace("'?%s'", "''").replace('%s,', 'null,').replace('%d', '0'));
@@ -1587,11 +1613,17 @@ function JSify(data, functionsOnly, givenFunctions) {
 
     var shellParts = read(shellFile).split('{{BODY}}');
     print(shellParts[1]);
-    // Print out some useful metadata (for additional optimizations later, like the eliminator)
-    if (EMIT_GENERATED_FUNCTIONS) {
-      print('// EMSCRIPTEN_GENERATED_FUNCTIONS: ' + JSON.stringify(keys(Functions.implementedFunctions).filter(function(func) {
+    // Print out some useful metadata
+    if (EMIT_GENERATED_FUNCTIONS || PGO) {
+      var generatedFunctions = JSON.stringify(keys(Functions.implementedFunctions).filter(function(func) {
         return IGNORED_FUNCTIONS.indexOf(func.ident) < 0;
-      })) + '\n');
+      }));
+      if (PGO) {
+        print('PGOMonitor.allGenerated = ' + generatedFunctions + ';\nremoveRunDependency("pgo");\n');
+      }
+      if (EMIT_GENERATED_FUNCTIONS) {
+        print('// EMSCRIPTEN_GENERATED_FUNCTIONS: ' + generatedFunctions + '\n');
+      }
     }
 
     PassManager.serialize();
diff --git a/src/library.js b/src/library.js
index 1cd7e3f5..45187d8d 100644
--- a/src/library.js
+++ b/src/library.js
@@ -504,7 +504,7 @@ LibraryManager.library = {
       }
       var utf8 = new Runtime.UTF8Processor();
       function simpleOutput(val) {
-        if (val === null || val === '\n'.charCodeAt(0)) {
+        if (val === null || val === {{{ charCode('\n') }}}) {
           output.printer(output.buffer.join(''));
           output.buffer = [];
         } else {
@@ -600,8 +600,8 @@ LibraryManager.library = {
     quit: function() {
       if (!FS.init.initialized) return;
       // Flush any partially-printed lines in stdout and stderr. Careful, they may have been closed
-      if (FS.streams[2] && FS.streams[2].object.output.buffer.length > 0) FS.streams[2].object.output('\n'.charCodeAt(0));
-      if (FS.streams[3] && FS.streams[3].object.output.buffer.length > 0) FS.streams[3].object.output('\n'.charCodeAt(0));
+      if (FS.streams[2] && FS.streams[2].object.output.buffer.length > 0) FS.streams[2].object.output({{{ charCode('\n') }}});
+      if (FS.streams[3] && FS.streams[3].object.output.buffer.length > 0) FS.streams[3].object.output({{{ charCode('\n') }}});
     },
 
     // Standardizes a path. Useful for making comparisons of pathnames work in a consistent manner.
@@ -828,11 +828,11 @@ LibraryManager.library = {
       // Null or empty results in '.'.
       var me = ___libgenSplitName;
       if (!me.ret) {
-        me.ret = allocate(['.'.charCodeAt(0), 0], 'i8', ALLOC_NORMAL);
+        me.ret = allocate([{{{ charCode('.') }}}, 0], 'i8', ALLOC_NORMAL);
       }
       return [me.ret, -1];
     } else {
-      var slash = '/'.charCodeAt(0);
+      var slash = {{{ charCode('/') }}};
       var allSlashes = true;
       var slashPositions = [];
       for (var i = 0; {{{ makeGetValue('path', 'i', 'i8') }}} !== 0; i++) {
@@ -1730,7 +1730,12 @@ LibraryManager.library = {
       }
       var contents = stream.object.contents;
       var size = Math.min(contents.length - offset, nbyte);
-      if (contents.subarray || contents.slice) { // typed array or normal array
+#if USE_TYPED_ARRAYS == 2
+      if (contents.subarray) { // typed array
+        HEAPU8.set(contents.subarray(offset, offset+size), buf);
+      } else
+#endif
+      if (contents.slice) { // normal array
         for (var i = 0; i < size; i++) {
           {{{ makeSetValue('buf', 'i', 'contents[offset + i]', 'i8') }}}
         }
@@ -2403,6 +2408,7 @@ LibraryManager.library = {
       case {{{ cDefine('_SC_STREAM_MAX') }}}: return 16;
       case {{{ cDefine('_SC_TZNAME_MAX') }}}: return 6;
       case {{{ cDefine('_SC_THREAD_DESTRUCTOR_ITERATIONS') }}}: return 4;
+      case {{{ cDefine('_SC_NPROCESSORS_ONLN') }}}: return 1;
     }
     ___setErrNo(ERRNO_CODES.EINVAL);
     return -1;
@@ -2451,9 +2457,9 @@ LibraryManager.library = {
   _scanString: function(format, get, unget, varargs) {
     if (!__scanString.whiteSpace) {
       __scanString.whiteSpace = {};
-      __scanString.whiteSpace[' '.charCodeAt(0)] = 1;
-      __scanString.whiteSpace['\t'.charCodeAt(0)] = 1;
-      __scanString.whiteSpace['\n'.charCodeAt(0)] = 1;
+      __scanString.whiteSpace[{{{ charCode(' ') }}}] = 1;
+      __scanString.whiteSpace[{{{ charCode('\t') }}}] = 1;
+      __scanString.whiteSpace[{{{ charCode('\n') }}}] = 1;
       __scanString.whiteSpace[' '] = 1;
       __scanString.whiteSpace['\t'] = 1;
       __scanString.whiteSpace['\n'] = 1;
@@ -2513,8 +2519,8 @@ LibraryManager.library = {
       if (format[formatIndex] === '%') {
         formatIndex++;
         var maxSpecifierStart = formatIndex;
-        while (format[formatIndex].charCodeAt(0) >= '0'.charCodeAt(0) &&
-               format[formatIndex].charCodeAt(0) <= '9'.charCodeAt(0)) {
+        while (format[formatIndex].charCodeAt(0) >= {{{ charCode('0') }}} &&
+               format[formatIndex].charCodeAt(0) <= {{{ charCode('9') }}}) {
           formatIndex++;
         }
         var max_;
@@ -2560,11 +2566,11 @@ LibraryManager.library = {
           while ((curr < max_ || isNaN(max_)) && next > 0) {
             if (!(next in __scanString.whiteSpace) && // stop on whitespace
                 (type == 's' ||
-                 ((type === 'd' || type == 'u' || type == 'i') && ((next >= '0'.charCodeAt(0) && next <= '9'.charCodeAt(0)) ||
-                                                                   (first && next == '-'.charCodeAt(0)))) ||
-                 (type === 'x' && (next >= '0'.charCodeAt(0) && next <= '9'.charCodeAt(0) ||
-                                   next >= 'a'.charCodeAt(0) && next <= 'f'.charCodeAt(0) ||
-                                   next >= 'A'.charCodeAt(0) && next <= 'F'.charCodeAt(0)))) &&
+                 ((type === 'd' || type == 'u' || type == 'i') && ((next >= {{{ charCode('0') }}} && next <= {{{ charCode('9') }}}) ||
+                                                                   (first && next == {{{ charCode('-') }}}))) ||
+                 (type === 'x' && (next >= {{{ charCode('0') }}} && next <= {{{ charCode('9') }}} ||
+                                   next >= {{{ charCode('a') }}} && next <= {{{ charCode('f') }}} ||
+                                   next >= {{{ charCode('A') }}} && next <= {{{ charCode('F') }}}))) &&
                 (formatIndex >= format.length || next !== format[formatIndex].charCodeAt(0))) { // Stop when we read something that is coming up
               buffer.push(String.fromCharCode(next));
               next = get();
@@ -2669,7 +2675,7 @@ LibraryManager.library = {
       curr = {{{ makeGetValue(0, 'textIndex', 'i8') }}};
       if (curr === 0) break;
       next = {{{ makeGetValue(0, 'textIndex+1', 'i8') }}};
-      if (curr == '%'.charCodeAt(0)) {
+      if (curr == {{{ charCode('%') }}}) {
         // Handle flags.
         var flagAlwaysSigned = false;
         var flagLeftAlign = false;
@@ -2677,16 +2683,16 @@ LibraryManager.library = {
         var flagZeroPad = false;
         flagsLoop: while (1) {
           switch (next) {
-            case '+'.charCodeAt(0):
+            case {{{ charCode('+') }}}:
               flagAlwaysSigned = true;
               break;
-            case '-'.charCodeAt(0):
+            case {{{ charCode('-') }}}:
               flagLeftAlign = true;
               break;
-            case '#'.charCodeAt(0):
+            case {{{ charCode('#') }}}:
               flagAlternative = true;
               break;
-            case '0'.charCodeAt(0):
+            case {{{ charCode('0') }}}:
               if (flagZeroPad) {
                 break flagsLoop;
               } else {
@@ -2702,13 +2708,13 @@ LibraryManager.library = {
 
         // Handle width.
         var width = 0;
-        if (next == '*'.charCodeAt(0)) {
+        if (next == {{{ charCode('*') }}}) {
           width = getNextArg('i32');
           textIndex++;
           next = {{{ makeGetValue(0, 'textIndex+1', 'i8') }}};
         } else {
-          while (next >= '0'.charCodeAt(0) && next <= '9'.charCodeAt(0)) {
-            width = width * 10 + (next - '0'.charCodeAt(0));
+          while (next >= {{{ charCode('0') }}} && next <= {{{ charCode('9') }}}) {
+            width = width * 10 + (next - {{{ charCode('0') }}});
             textIndex++;
             next = {{{ makeGetValue(0, 'textIndex+1', 'i8') }}};
           }
@@ -2716,20 +2722,20 @@ LibraryManager.library = {
 
         // Handle precision.
         var precisionSet = false;
-        if (next == '.'.charCodeAt(0)) {
+        if (next == {{{ charCode('.') }}}) {
           var precision = 0;
           precisionSet = true;
           textIndex++;
           next = {{{ makeGetValue(0, 'textIndex+1', 'i8') }}};
-          if (next == '*'.charCodeAt(0)) {
+          if (next == {{{ charCode('*') }}}) {
             precision = getNextArg('i32');
             textIndex++;
           } else {
             while(1) {
               var precisionChr = {{{ makeGetValue(0, 'textIndex+1', 'i8') }}};
-              if (precisionChr < '0'.charCodeAt(0) ||
-                  precisionChr > '9'.charCodeAt(0)) break;
-              precision = precision * 10 + (precisionChr - '0'.charCodeAt(0));
+              if (precisionChr < {{{ charCode('0') }}} ||
+                  precisionChr > {{{ charCode('9') }}}) break;
+              precision = precision * 10 + (precisionChr - {{{ charCode('0') }}});
               textIndex++;
             }
           }
@@ -2743,7 +2749,7 @@ LibraryManager.library = {
         switch (String.fromCharCode(next)) {
           case 'h':
             var nextNext = {{{ makeGetValue(0, 'textIndex+2', 'i8') }}};
-            if (nextNext == 'h'.charCodeAt(0)) {
+            if (nextNext == {{{ charCode('h') }}}) {
               textIndex++;
               argSize = 1; // char (actually i32 in varargs)
             } else {
@@ -2752,7 +2758,7 @@ LibraryManager.library = {
             break;
           case 'l':
             var nextNext = {{{ makeGetValue(0, 'textIndex+2', 'i8') }}};
-            if (nextNext == 'l'.charCodeAt(0)) {
+            if (nextNext == {{{ charCode('l') }}}) {
               textIndex++;
               argSize = 8; // long long
             } else {
@@ -2778,7 +2784,7 @@ LibraryManager.library = {
         // Handle type specifier.
         if (['d', 'i', 'u', 'o', 'x', 'X', 'p'].indexOf(String.fromCharCode(next)) != -1) {
           // Integer.
-          var signed = next == 'd'.charCodeAt(0) || next == 'i'.charCodeAt(0);
+          var signed = next == {{{ charCode('d') }}} || next == {{{ charCode('i') }}};
           argSize = argSize || 4;
           var currArg = getNextArg('i' + (argSize * 8));
 #if PRECISE_I64_MATH
@@ -2788,7 +2794,7 @@ LibraryManager.library = {
 #if USE_TYPED_ARRAYS == 2
           // Flatten i64-1 [low, high] into a (slightly rounded) double
           if (argSize == 8) {
-            currArg = Runtime.makeBigInt(currArg[0], currArg[1], next == 'u'.charCodeAt(0));
+            currArg = Runtime.makeBigInt(currArg[0], currArg[1], next == {{{ charCode('u') }}});
           }
 #endif
           // Truncate to requested size.
@@ -2799,20 +2805,20 @@ LibraryManager.library = {
           // Format the number.
           var currAbsArg = Math.abs(currArg);
           var prefix = '';
-          if (next == 'd'.charCodeAt(0) || next == 'i'.charCodeAt(0)) {
+          if (next == {{{ charCode('d') }}} || next == {{{ charCode('i') }}}) {
 #if PRECISE_I64_MATH
             if (argSize == 8 && i64Math) argText = i64Math.stringify(origArg[0], origArg[1], null); else
 #endif
             argText = reSign(currArg, 8 * argSize, 1).toString(10);
-          } else if (next == 'u'.charCodeAt(0)) {
+          } else if (next == {{{ charCode('u') }}}) {
 #if PRECISE_I64_MATH
             if (argSize == 8 && i64Math) argText = i64Math.stringify(origArg[0], origArg[1], true); else
 #endif
             argText = unSign(currArg, 8 * argSize, 1).toString(10);
             currArg = Math.abs(currArg);
-          } else if (next == 'o'.charCodeAt(0)) {
+          } else if (next == {{{ charCode('o') }}}) {
             argText = (flagAlternative ? '0' : '') + currAbsArg.toString(8);
-          } else if (next == 'x'.charCodeAt(0) || next == 'X'.charCodeAt(0)) {
+          } else if (next == {{{ charCode('x') }}} || next == {{{ charCode('X') }}}) {
             prefix = flagAlternative ? '0x' : '';
 #if PRECISE_I64_MATH
             if (argSize == 8 && i64Math) argText = (origArg[1]>>>0).toString(16) + (origArg[0]>>>0).toString(16); else
@@ -2830,11 +2836,11 @@ LibraryManager.library = {
             } else {
               argText = currAbsArg.toString(16);
             }
-            if (next == 'X'.charCodeAt(0)) {
+            if (next == {{{ charCode('X') }}}) {
               prefix = prefix.toUpperCase();
               argText = argText.toUpperCase();
             }
-          } else if (next == 'p'.charCodeAt(0)) {
+          } else if (next == {{{ charCode('p') }}}) {
             if (currAbsArg === 0) {
               argText = '(nil)';
             } else {
@@ -2892,27 +2898,27 @@ LibraryManager.library = {
 
             // Convert g/G to f/F or e/E, as per:
             // http://pubs.opengroup.org/onlinepubs/9699919799/functions/printf.html
-            if (next == 'g'.charCodeAt(0) || next == 'G'.charCodeAt(0)) {
+            if (next == {{{ charCode('g') }}} || next == {{{ charCode('G') }}}) {
               isGeneral = true;
               precision = precision || 1;
               var exponent = parseInt(currArg.toExponential(effectivePrecision).split('e')[1], 10);
               if (precision > exponent && exponent >= -4) {
-                next = ((next == 'g'.charCodeAt(0)) ? 'f' : 'F').charCodeAt(0);
+                next = ((next == {{{ charCode('g') }}}) ? 'f' : 'F').charCodeAt(0);
                 precision -= exponent + 1;
               } else {
-                next = ((next == 'g'.charCodeAt(0)) ? 'e' : 'E').charCodeAt(0);
+                next = ((next == {{{ charCode('g') }}}) ? 'e' : 'E').charCodeAt(0);
                 precision--;
               }
               effectivePrecision = Math.min(precision, 20);
             }
 
-            if (next == 'e'.charCodeAt(0) || next == 'E'.charCodeAt(0)) {
+            if (next == {{{ charCode('e') }}} || next == {{{ charCode('E') }}}) {
               argText = currArg.toExponential(effectivePrecision);
               // Make sure the exponent has at least 2 digits.
               if (/[eE][-+]\d$/.test(argText)) {
                 argText = argText.slice(0, -1) + '0' + argText.slice(-1);
               }
-            } else if (next == 'f'.charCodeAt(0) || next == 'F'.charCodeAt(0)) {
+            } else if (next == {{{ charCode('f') }}} || next == {{{ charCode('F') }}}) {
               argText = currArg.toFixed(effectivePrecision);
             }
 
@@ -2932,7 +2938,7 @@ LibraryManager.library = {
             argText = parts[0] + (parts.length > 1 ? 'e' + parts[1] : '');
 
             // Capitalize 'E' if needed.
-            if (next == 'E'.charCodeAt(0)) argText = argText.toUpperCase();
+            if (next == {{{ charCode('E') }}}) argText = argText.toUpperCase();
 
             // Add sign.
             if (flagAlwaysSigned && currArg >= 0) {
@@ -2954,20 +2960,20 @@ LibraryManager.library = {
           }
 
           // Adjust case.
-          if (next < 'a'.charCodeAt(0)) argText = argText.toUpperCase();
+          if (next < {{{ charCode('a') }}}) argText = argText.toUpperCase();
 
           // Insert the result into the buffer.
           argText.split('').forEach(function(chr) {
             ret.push(chr.charCodeAt(0));
           });
-        } else if (next == 's'.charCodeAt(0)) {
+        } else if (next == {{{ charCode('s') }}}) {
           // String.
           var arg = getNextArg('i8*') || nullString;
           var argLength = _strlen(arg);
           if (precisionSet) argLength = Math.min(argLength, precision);
           if (!flagLeftAlign) {
             while (argLength < width--) {
-              ret.push(' '.charCodeAt(0));
+              ret.push({{{ charCode(' ') }}});
             }
           }
           for (var i = 0; i < argLength; i++) {
@@ -2975,21 +2981,21 @@ LibraryManager.library = {
           }
           if (flagLeftAlign) {
             while (argLength < width--) {
-              ret.push(' '.charCodeAt(0));
+              ret.push({{{ charCode(' ') }}});
             }
           }
-        } else if (next == 'c'.charCodeAt(0)) {
+        } else if (next == {{{ charCode('c') }}}) {
           // Character.
           if (flagLeftAlign) ret.push(getNextArg('i8'));
           while (--width > 0) {
-            ret.push(' '.charCodeAt(0));
+            ret.push({{{ charCode(' ') }}});
           }
           if (!flagLeftAlign) ret.push(getNextArg('i8'));
-        } else if (next == 'n'.charCodeAt(0)) {
+        } else if (next == {{{ charCode('n') }}}) {
           // Write the length written so far to the next parameter.
           var ptr = getNextArg('i32*');
           {{{ makeSetValue('ptr', '0', 'ret.length', 'i32') }}}
-        } else if (next == '%'.charCodeAt(0)) {
+        } else if (next == {{{ charCode('%') }}}) {
           // Literal percent sign.
           ret.push(curr);
         } else {
@@ -3138,7 +3144,7 @@ LibraryManager.library = {
     var streamObj = FS.streams[stream];
     if (streamObj.error || streamObj.eof) return 0;
     var byte_;
-    for (var i = 0; i < n - 1 && byte_ != '\n'.charCodeAt(0); i++) {
+    for (var i = 0; i < n - 1 && byte_ != {{{ charCode('\n') }}}; i++) {
       byte_ = _fgetc(stream);
       if (byte_ == -1) {
         if (streamObj.error) return 0;
@@ -3244,7 +3250,7 @@ LibraryManager.library = {
     if (ret < 0) {
       return ret;
     } else {
-      var newlineRet = _fputc('\n'.charCodeAt(0), stdout);
+      var newlineRet = _fputc({{{ charCode('\n') }}}, stdout);
       return (newlineRet < 0) ? -1 : ret + 1;
     }
   },
@@ -3370,8 +3376,8 @@ LibraryManager.library = {
     var stdout = {{{ makeGetValue(makeGlobalUse('_stdout'), '0', 'void*') }}};
     if (s) {
       _fputs(s, stdout);
-      _fputc(':'.charCodeAt(0), stdout);
-      _fputc(' '.charCodeAt(0), stdout);
+      _fputc({{{ charCode(':') }}}, stdout);
+      _fputc({{{ charCode(' ') }}}, stdout);
     }
     var errnum = {{{ makeGetValue('___errno_location()', '0', 'i32') }}};
     _puts(_strerror(errnum));
@@ -3669,6 +3675,17 @@ LibraryManager.library = {
 
   abs: 'Math.abs',
   labs: 'Math.abs',
+#if USE_TYPED_ARRAYS == 2
+  llabs__deps: [function() { Types.preciseI64MathUsed = 1 }],
+  llabs: function(lo, hi) {
+    i64Math.abs(lo, hi);
+    {{{ makeStructuralReturn([makeGetTempDouble(0, 'i32'), makeGetTempDouble(1, 'i32')]) }}};
+  },
+#else
+  llabs: function(lo, hi) {
+    throw 'unsupported llabs';
+  },
+#endif
 
   exit__deps: ['_exit'],
   exit: function(status) {
@@ -3726,93 +3743,6 @@ LibraryManager.library = {
     return ret;
   },
 
-  strtod__deps: ['isspace', 'isdigit'],
-  strtod: function(str, endptr) {
-    var origin = str;
-
-    // Skip space.
-    while (_isspace({{{ makeGetValue('str', 0, 'i8') }}})) str++;
-
-    // Check for a plus/minus sign.
-    var multiplier = 1;
-    if ({{{ makeGetValue('str', 0, 'i8') }}} == '-'.charCodeAt(0)) {
-      multiplier = -1;
-      str++;
-    } else if ({{{ makeGetValue('str', 0, 'i8') }}} == '+'.charCodeAt(0)) {
-      str++;
-    }
-
-    var chr;
-    var ret = 0;
-
-    // Get whole part.
-    var whole = false;
-    while(1) {
-      chr = {{{ makeGetValue('str', 0, 'i8') }}};
-      if (!_isdigit(chr)) break;
-      whole = true;
-      ret = ret*10 + chr - '0'.charCodeAt(0);
-      str++;
-    }
-
-    // Get fractional part.
-    var fraction = false;
-    if ({{{ makeGetValue('str', 0, 'i8') }}} == '.'.charCodeAt(0)) {
-      str++;
-      var mul = 1/10;
-      while(1) {
-        chr = {{{ makeGetValue('str', 0, 'i8') }}};
-        if (!_isdigit(chr)) break;
-        fraction = true;
-        ret += mul*(chr - '0'.charCodeAt(0));
-        mul /= 10;
-        str++;
-      }
-    }
-
-    if (!whole && !fraction) {
-      if (endptr) {
-        {{{ makeSetValue('endptr', 0, 'origin', '*') }}}
-      }
-      return 0;
-    }
-
-    // Get exponent part.
-    chr = {{{ makeGetValue('str', 0, 'i8') }}};
-    if (chr == 'e'.charCodeAt(0) || chr == 'E'.charCodeAt(0)) {
-      str++;
-      var exponent = 0;
-      var expNegative = false;
-      chr = {{{ makeGetValue('str', 0, 'i8') }}};
-      if (chr == '-'.charCodeAt(0)) {
-        expNegative = true;
-        str++;
-      } else if (chr == '+'.charCodeAt(0)) {
-        str++;
-      }
-      chr = {{{ makeGetValue('str', 0, 'i8') }}};
-      while(1) {
-        if (!_isdigit(chr)) break;
-        exponent = exponent*10 + chr - '0'.charCodeAt(0);
-        str++;
-        chr = {{{ makeGetValue('str', 0, 'i8') }}};
-      }
-      if (expNegative) exponent = -exponent;
-      ret *= Math.pow(10, exponent);
-    }
-
-    // Set end pointer.
-    if (endptr) {
-      {{{ makeSetValue('endptr', 0, 'str', '*') }}}
-    }
-
-    return ret * multiplier;
-  },
-  strtod_l: 'strtod', // no locale support yet
-  strtold: 'strtod', // XXX add real support for long double
-  strtold_l: 'strtold', // no locale support yet
-  strtof: 'strtod', // use stdtod to handle strtof
-
   _parseInt__deps: ['isspace', '__setErrNo', '$ERRNO_CODES'],
   _parseInt: function(str, endptr, base, min, max, bits, unsign) {
     // Skip space.
@@ -3820,19 +3750,19 @@ LibraryManager.library = {
 
     // Check for a plus/minus sign.
     var multiplier = 1;
-    if ({{{ makeGetValue('str', 0, 'i8') }}} == '-'.charCodeAt(0)) {
+    if ({{{ makeGetValue('str', 0, 'i8') }}} == {{{ charCode('-') }}}) {
       multiplier = -1;
       str++;
-    } else if ({{{ makeGetValue('str', 0, 'i8') }}} == '+'.charCodeAt(0)) {
+    } else if ({{{ makeGetValue('str', 0, 'i8') }}} == {{{ charCode('+') }}}) {
       str++;
     }
 
     // Find base.
     var finalBase = base;
     if (!finalBase) {
-      if ({{{ makeGetValue('str', 0, 'i8') }}} == '0'.charCodeAt(0)) {
-        if ({{{ makeGetValue('str+1', 0, 'i8') }}} == 'x'.charCodeAt(0) ||
-            {{{ makeGetValue('str+1', 0, 'i8') }}} == 'X'.charCodeAt(0)) {
+      if ({{{ makeGetValue('str', 0, 'i8') }}} == {{{ charCode('0') }}}) {
+        if ({{{ makeGetValue('str+1', 0, 'i8') }}} == {{{ charCode('x') }}} ||
+            {{{ makeGetValue('str+1', 0, 'i8') }}} == {{{ charCode('X') }}}) {
           finalBase = 16;
           str += 2;
         } else {
@@ -3896,9 +3826,9 @@ LibraryManager.library = {
     while (_isspace({{{ makeGetValue('str', 0, 'i8') }}})) str++;
 
     // Check for a plus/minus sign.
-    if ({{{ makeGetValue('str', 0, 'i8') }}} == '-'.charCodeAt(0)) {
+    if ({{{ makeGetValue('str', 0, 'i8') }}} == {{{ charCode('-') }}}) {
       str++;
-    } else if ({{{ makeGetValue('str', 0, 'i8') }}} == '+'.charCodeAt(0)) {
+    } else if ({{{ makeGetValue('str', 0, 'i8') }}} == {{{ charCode('+') }}}) {
       str++;
     }
 
@@ -3906,9 +3836,9 @@ LibraryManager.library = {
     var ok = false;
     var finalBase = base;
     if (!finalBase) {
-      if ({{{ makeGetValue('str', 0, 'i8') }}} == '0'.charCodeAt(0)) {
-        if ({{{ makeGetValue('str+1', 0, 'i8') }}} == 'x'.charCodeAt(0) ||
-            {{{ makeGetValue('str+1', 0, 'i8') }}} == 'X'.charCodeAt(0)) {
+      if ({{{ makeGetValue('str', 0, 'i8') }}} == {{{ charCode('0') }}}) {
+        if ({{{ makeGetValue('str+1', 0, 'i8') }}} == {{{ charCode('x') }}} ||
+            {{{ makeGetValue('str+1', 0, 'i8') }}} == {{{ charCode('X') }}}) {
           finalBase = 16;
           str += 2;
         } else {
@@ -3971,11 +3901,6 @@ LibraryManager.library = {
   },
   strtoull_l: 'strtoull', // no locale support yet
 
-  atof__deps: ['strtod'],
-  atof: function(ptr) {
-    return _strtod(ptr, null);
-  },
-
   atoi__deps: ['strtol'],
   atoi: function(ptr) {
     return _strtol(ptr, null, 10);
@@ -4235,13 +4160,15 @@ LibraryManager.library = {
   memcpy__inline: function (dest, src, num, align) {
     var ret = '';
 #if ASSERTIONS
+#if ASM_JS == 0
     ret += "assert(" + num + " % 1 === 0);"; //, 'memcpy given ' + " + num + " + ' bytes to copy. Problem with quantum=1 corrections perhaps?');";
 #endif
+#endif
     ret += makeCopyValues(dest, src, num, 'null', null, align);
     return ret;
   },
 
-  memcpy__asm: 'true',
+  memcpy__asm: true,
   memcpy__sig: 'iiii',
   memcpy: function (dest, src, num) {
     dest = dest|0; src = src|0; num = num|0;
@@ -4314,7 +4241,7 @@ LibraryManager.library = {
     ptr = ptr|0; value = value|0; num = num|0;
     var stop = 0, value4 = 0, stop4 = 0, unaligned = 0;
     stop = (ptr + num)|0;
-    if ((num|0) >= {{{ SEEK_OPTIMAL_ALIGN_MIN }}}) {
+    if ((num|0) >= {{{ Math.round(2.5*UNROLL_LOOP_MAX) }}}) {
       // This is unaligned, but quite large, so work hard to get to aligned settings
       value = value & 0xff;
       unaligned = ptr & 3;
@@ -4402,14 +4329,18 @@ LibraryManager.library = {
     }
   },
 
+  strcpy__asm: true,
+  strcpy__sig: 'iii',
   strcpy: function(pdest, psrc) {
+    pdest = pdest|0; psrc = psrc|0;
     var i = 0;
     do {
-      {{{ makeCopyValues('pdest+i', 'psrc+i', 1, 'i8', null, 1) }}};
-      i ++;
-    } while ({{{ makeGetValue('psrc', 'i-1', 'i8') }}} != 0);
-    return pdest;
+      {{{ makeCopyValues('(pdest+i)|0', '(psrc+i)|0', 1, 'i8', null, 1) }}};
+      i = (i+1)|0;
+    } while (({{{ makeGetValue('psrc', 'i-1', 'i8') }}})|0 != 0);
+    return pdest|0;
   },
+
   stpcpy: function(pdest, psrc) {
     var i = 0;
     do {
@@ -4419,14 +4350,18 @@ LibraryManager.library = {
     return pdest + i - 1;
   },
 
+  strncpy__asm: true,
+  strncpy__sig: 'iiii',
   strncpy: function(pdest, psrc, num) {
-    var padding = false, curr;
-    for (var i = 0; i < num; i++) {
-      curr = padding ? 0 : {{{ makeGetValue('psrc', 'i', 'i8') }}};
+    pdest = pdest|0; psrc = psrc|0; num = num|0;
+    var padding = 0, curr = 0, i = 0;
+    while ((i|0) < (num|0)) {
+      curr = padding ? 0 : {{{ makeGetValueAsm('psrc', 'i', 'i8') }}};
       {{{ makeSetValue('pdest', 'i', 'curr', 'i8') }}}
-      padding = padding || {{{ makeGetValue('psrc', 'i', 'i8') }}} == 0;
+      padding = padding ? 1 : ({{{ makeGetValueAsm('psrc', 'i', 'i8') }}} == 0);
+      i = (i+1)|0;
     }
-    return pdest;
+    return pdest|0;
   },
   
   strlwr__deps:['tolower'],
@@ -4451,15 +4386,18 @@ LibraryManager.library = {
     }
   },
 
+  strcat__asm: true,
+  strcat__sig: 'iii',
   strcat__deps: ['strlen'],
   strcat: function(pdest, psrc) {
-    var len = _strlen(pdest);
+    pdest = pdest|0; psrc = psrc|0;
     var i = 0;
+    pdest = (pdest + _strlen(pdest))|0;
     do {
-      {{{ makeCopyValues('pdest+len+i', 'psrc+i', 1, 'i8', null, 1) }}};
-      i ++;
-    } while ({{{ makeGetValue('psrc', 'i-1', 'i8') }}} != 0);
-    return pdest;
+      {{{ makeCopyValues('pdest+i', 'psrc+i', 1, 'i8', null, 1) }}};
+      i = (i+1)|0;
+    } while ({{{ makeGetValueAsm('psrc', 'i-1', 'i8') }}} != 0);
+    return pdest|0;
   },
 
   strncat__deps: ['strlen'],
@@ -4485,9 +4423,12 @@ LibraryManager.library = {
   // We always assume ASCII locale.
   strcoll: 'strcmp',
 
+  strcasecmp__asm: true,
+  strcasecmp__sig: 'iii',
   strcasecmp__deps: ['strncasecmp'],
   strcasecmp: function(px, py) {
-    return _strncasecmp(px, py, TOTAL_MEMORY);
+    px = px|0; py = py|0;
+    return _strncasecmp(px, py, -1)|0;
   },
 
   strncmp: function(px, py, n) {
@@ -4508,26 +4449,29 @@ LibraryManager.library = {
     return 0;
   },
 
+  strncasecmp__asm: true,
+  strncasecmp__sig: 'iiii',
   strncasecmp__deps: ['tolower'],
   strncasecmp: function(px, py, n) {
-    var i = 0;
-    while (i < n) {
-      var x = _tolower({{{ makeGetValue('px', 'i', 'i8', 0, 1) }}});
-      var y = _tolower({{{ makeGetValue('py', 'i', 'i8', 0, 1) }}});
-      if (x == y && x == 0) return 0;
-      if (x == 0) return -1;
-      if (y == 0) return 1;
-      if (x == y) {
-        i ++;
+    px = px|0; py = py|0; n = n|0;
+    var i = 0, x = 0, y = 0;
+    while ((i>>>0) < (n>>>0)) {
+      x = _tolower({{{ makeGetValueAsm('px', 'i', 'i8', 0, 1) }}});
+      y = _tolower({{{ makeGetValueAsm('py', 'i', 'i8', 0, 1) }}});
+      if (((x|0) == (y|0)) & ((x|0) == 0)) return 0;
+      if ((x|0) == 0) return -1;
+      if ((y|0) == 0) return 1;
+      if ((x|0) == (y|0)) {
+        i = (i + 1)|0;
         continue;
       } else {
-        return x > y ? 1 : -1;
+        return ((x>>>0) > (y>>>0) ? 1 : -1)|0;
       }
     }
     return 0;
   },
 
-  memcmp__asm: 'true',
+  memcmp__asm: true,
   memcmp__sig: 'iiii',
   memcmp: function(p1, p2, num) {
     p1 = p1|0; p2 = p2|0; num = num|0;
@@ -4729,58 +4673,61 @@ LibraryManager.library = {
     return chr & 0x7F;
   },
   toupper: function(chr) {
-    if (chr >= 'a'.charCodeAt(0) && chr <= 'z'.charCodeAt(0)) {
-      return chr - 'a'.charCodeAt(0) + 'A'.charCodeAt(0);
+    if (chr >= {{{ charCode('a') }}} && chr <= {{{ charCode('z') }}}) {
+      return chr - {{{ charCode('a') }}} + {{{ charCode('A') }}};
     } else {
       return chr;
     }
   },
   _toupper: 'toupper',
+
+  tolower__asm: true,
+  tolower__sig: 'ii',
   tolower: function(chr) {
-    if (chr >= 'A'.charCodeAt(0) && chr <= 'Z'.charCodeAt(0)) {
-      return chr - 'A'.charCodeAt(0) + 'a'.charCodeAt(0);
-    } else {
-      return chr;
-    }
+    chr = chr|0;
+    if ((chr|0) < {{{ charCode('A') }}}) return chr|0;
+    if ((chr|0) > {{{ charCode('Z') }}}) return chr|0;
+    return (chr - {{{ charCode('A') }}} + {{{ charCode('a') }}})|0;
   },
   _tolower: 'tolower',
+
   // The following functions are defined as macros in glibc.
   islower: function(chr) {
-    return chr >= 'a'.charCodeAt(0) && chr <= 'z'.charCodeAt(0);
+    return chr >= {{{ charCode('a') }}} && chr <= {{{ charCode('z') }}};
   },
   isupper: function(chr) {
-    return chr >= 'A'.charCodeAt(0) && chr <= 'Z'.charCodeAt(0);
+    return chr >= {{{ charCode('A') }}} && chr <= {{{ charCode('Z') }}};
   },
   isalpha: function(chr) {
-    return (chr >= 'a'.charCodeAt(0) && chr <= 'z'.charCodeAt(0)) ||
-           (chr >= 'A'.charCodeAt(0) && chr <= 'Z'.charCodeAt(0));
+    return (chr >= {{{ charCode('a') }}} && chr <= {{{ charCode('z') }}}) ||
+           (chr >= {{{ charCode('A') }}} && chr <= {{{ charCode('Z') }}});
   },
   isdigit: function(chr) {
-    return chr >= '0'.charCodeAt(0) && chr <= '9'.charCodeAt(0);
+    return chr >= {{{ charCode('0') }}} && chr <= {{{ charCode('9') }}};
   },
   isdigit_l: 'isdigit', // no locale support yet
   isxdigit: function(chr) {
-    return (chr >= '0'.charCodeAt(0) && chr <= '9'.charCodeAt(0)) ||
-           (chr >= 'a'.charCodeAt(0) && chr <= 'f'.charCodeAt(0)) ||
-           (chr >= 'A'.charCodeAt(0) && chr <= 'F'.charCodeAt(0));
+    return (chr >= {{{ charCode('0') }}} && chr <= {{{ charCode('9') }}}) ||
+           (chr >= {{{ charCode('a') }}} && chr <= {{{ charCode('f') }}}) ||
+           (chr >= {{{ charCode('A') }}} && chr <= {{{ charCode('F') }}});
   },
   isxdigit_l: 'isxdigit', // no locale support yet
   isalnum: function(chr) {
-    return (chr >= '0'.charCodeAt(0) && chr <= '9'.charCodeAt(0)) ||
-           (chr >= 'a'.charCodeAt(0) && chr <= 'z'.charCodeAt(0)) ||
-           (chr >= 'A'.charCodeAt(0) && chr <= 'Z'.charCodeAt(0));
+    return (chr >= {{{ charCode('0') }}} && chr <= {{{ charCode('9') }}}) ||
+           (chr >= {{{ charCode('a') }}} && chr <= {{{ charCode('z') }}}) ||
+           (chr >= {{{ charCode('A') }}} && chr <= {{{ charCode('Z') }}});
   },
   ispunct: function(chr) {
-    return (chr >= '!'.charCodeAt(0) && chr <= '/'.charCodeAt(0)) ||
-           (chr >= ':'.charCodeAt(0) && chr <= '@'.charCodeAt(0)) ||
-           (chr >= '['.charCodeAt(0) && chr <= '`'.charCodeAt(0)) ||
-           (chr >= '{'.charCodeAt(0) && chr <= '~'.charCodeAt(0));
+    return (chr >= {{{ charCode('!') }}} && chr <= {{{ charCode('/') }}}) ||
+           (chr >= {{{ charCode(':') }}} && chr <= {{{ charCode('@') }}}) ||
+           (chr >= {{{ charCode('[') }}} && chr <= {{{ charCode('`') }}}) ||
+           (chr >= {{{ charCode('{') }}} && chr <= {{{ charCode('~') }}});
   },
   isspace: function(chr) {
     return chr in { 32: 0, 9: 0, 10: 0, 11: 0, 12: 0, 13: 0 };
   },
   isblank: function(chr) {
-    return chr == ' '.charCodeAt(0) || chr == '\t'.charCodeAt(0);
+    return chr == {{{ charCode(' ') }}} || chr == {{{ charCode('\t') }}};
   },
   iscntrl: function(chr) {
     return (0 <= chr && chr <= 0x1F) || chr === 0x7F;
@@ -4976,6 +4923,20 @@ LibraryManager.library = {
 #endif
   },
 
+  llvm_ctpop_i32: function(x) {
+    var ret = 0;
+    while (x) {
+      if (x&1) ret++;
+      x >>= 1;
+    }
+    return ret;
+  },
+
+  llvm_ctpop_i64__deps: ['llvm_ctpop_i32'],
+  llvm_ctpop_i64: function(l, h) {
+    return _llvm_ctpop_i32(l) + _llvm_ctpop_i32(h);
+  },
+
   llvm_trap: function() {
     throw 'trap! ' + new Error().stack;
   },
@@ -5270,13 +5231,6 @@ LibraryManager.library = {
     {{{ makeStructuralReturn(['(x*y)>>>0', 'x*y > 4294967295']) }}};
   },
 
-  llvm_uadd_with_overflow_i64__deps: [function() { Types.preciseI64MathUsed = 1 }],
-  llvm_uadd_with_overflow_i64: function(xl, xh, yl, yh) {
-    i64Math.add(xl, xh, yl, yh);
-    {{{ makeStructuralReturn([makeGetTempDouble(0, 'i32'), makeGetTempDouble(1, 'i32'), '0']) }}};
-    // XXX Need to hack support for second param in long.js
-  },
-
   llvm_umul_with_overflow_i64__deps: [function() { Types.preciseI64MathUsed = 1 }],
   llvm_umul_with_overflow_i64: function(xl, xh, yl, yh) {
     i64Math.multiply(xl, xh, yl, yh);
@@ -5506,6 +5460,14 @@ LibraryManager.library = {
       return -a;
   },
   copysignf: 'copysign',
+  __signbit__deps: ['copysign'],
+  __signbit: function(x) {
+    // We implement using copysign so that we get support
+    // for negative zero (once copysign supports that).
+    return _copysign(1.0, x) < 0;
+  },
+  __signbitf: '__signbit',
+  __signbitd: '__signbit',
   hypot: function(a, b) {
      return Math.sqrt(a*a + b*b);
   },
@@ -6064,6 +6026,15 @@ LibraryManager.library = {
   __timespec_struct_layout: Runtime.generateStructInfo([
     ['i32', 'tv_sec'],
     ['i32', 'tv_nsec']]),
+  nanosleep__deps: ['usleep', '__timespec_struct_layout'],
+  nanosleep: function(rqtp, rmtp) {
+    // int nanosleep(const struct timespec  *rqtp, struct timespec *rmtp);
+    var seconds = {{{ makeGetValue('rqtp', '___timespec_struct_layout.tv_sec', 'i32') }}};
+    var nanoseconds = {{{ makeGetValue('rqtp', '___timespec_struct_layout.tv_nsec', 'i32') }}};
+    {{{ makeSetValue('rmtp', '___timespec_struct_layout.tv_sec', '0', 'i32') }}}
+    {{{ makeSetValue('rmtp', '___timespec_struct_layout.tv_nsec', '0', 'i32') }}}
+    return _usleep((seconds * 1e6) + (nanoseconds / 1000));
+  },
   // TODO: Implement these for real.
   clock_gettime__deps: ['__timespec_struct_layout'],
   clock_gettime: function(clk_id, tp) {
@@ -6170,7 +6141,7 @@ LibraryManager.library = {
 
   setjmp__inline: function(env) {
     // Save the label
-    return '(tempInt = setjmpId++, mySetjmpIds[tempInt] = 1, setjmpLabels[tempInt] = label,' + makeSetValue(env, '0', 'tempInt', 'i32') + ', 0)';
+    return '(tempInt = setjmpId++, mySetjmpIds[tempInt] = 1, setjmpLabels[tempInt] = label,' + makeSetValue(env, '0', 'tempInt', 'i32', undefined, undefined, undefined, undefined,  ',') + ', 0)';
   },
 
   longjmp: function(env, value) {
@@ -6693,6 +6664,13 @@ LibraryManager.library = {
   },
 
   // ==========================================================================
+  // sched.h (stubs only - no thread support yet!)
+  // ==========================================================================
+  sched_yield: function() {
+    return 0;
+  },
+
+  // ==========================================================================
   // pthread.h (stubs for mutexes only - no thread support yet!)
   // ==========================================================================
 
@@ -6708,8 +6686,15 @@ LibraryManager.library = {
   },
   pthread_cond_init: function() {},
   pthread_cond_destroy: function() {},
-  pthread_cond_broadcast: function() {},
-  pthread_cond_wait: function() {},
+  pthread_cond_broadcast: function() {
+    return 0;
+  },
+  pthread_cond_wait: function() {
+    return 0;
+  },
+  pthread_cond_timedwait: function() {
+    return 0;
+  },
   pthread_self: function() {
     //FIXME: assumes only a single thread
     return 0;
@@ -7385,6 +7370,81 @@ LibraryManager.library = {
     Module.print(intArrayToString(__formatString(_emscripten_jcache_printf_.buffer, varargs + i*4)).replace('\\n', ''));
     Runtime.stackAlloc(-4*i); // free up the stack space we know is ok to free
   },
+
+  //============================
+  // i64 math
+  //============================
+
+  i64Add__asm: true,
+  i64Add__sig: 'iiiii',
+  i64Add: function(a, b, c, d) {
+    /*
+      x = a + b*2^32
+      y = c + d*2^32
+      result = l + h*2^32
+    */
+    a = a|0; b = b|0; c = c|0; d = d|0;
+    var l = 0, h = 0;
+    l = (a + c)>>>0;
+    h = (b + d)>>>0;
+    if ((l>>>0) < (a>>>0)) { // iff we overflowed
+      h = (h+1)>>>0;
+    }
+    {{{ makeStructuralReturn(['l|0', 'h'], true) }}};
+  },
+  llvm_uadd_with_overflow_i64__asm: true,
+  llvm_uadd_with_overflow_i64__sig: 'iiiii',
+  llvm_uadd_with_overflow_i64: function(a, b, c, d) {
+    a = a|0; b = b|0; c = c|0; d = d|0;
+    var l = 0, h = 0, overflow = 0;
+    l = (a + c)>>>0;
+    h = (b + d)>>>0;
+    if ((l>>>0) < (a>>>0)) { // iff we overflowed
+      h = (h+1)>>>0;
+      overflow = 1;
+    }
+    {{{ makeStructuralReturn(['l|0', 'h', 'overflow'], true) }}};
+  },
+
+  bitshift64Shl__asm: true,
+  bitshift64Shl__sig: 'iiii',
+  bitshift64Shl: function(low, high, bits) {
+    low = low|0; high = high|0; bits = bits|0;
+    var ander = 0;
+    if ((bits|0) < 32) {
+      ander = ((1 << bits) - 1)|0;
+      tempRet0 = (high << bits) | ((low&(ander << (32 - bits))) >>> (32 - bits));
+      return low << bits;
+    }
+    tempRet0 = low << (bits - 32);
+    return 0;
+  },
+  bitshift64Ashr__asm: true,
+  bitshift64Ashr__sig: 'iiii',
+  bitshift64Ashr: function(low, high, bits) {
+    low = low|0; high = high|0; bits = bits|0;
+    var ander = 0;
+    if ((bits|0) < 32) {
+      ander = ((1 << bits) - 1)|0;
+      tempRet0 = high >> bits;
+      return (low >>> bits) | ((high&ander) << (32 - bits));
+    }
+    tempRet0 = (high|0) < 0 ? -1 : 0;
+    return (high >> (bits - 32))|0;
+  },
+  bitshift64Lshr__asm: true,
+  bitshift64Lshr__sig: 'iiii',
+  bitshift64Lshr: function(low, high, bits) {
+    low = low|0; high = high|0; bits = bits|0;
+    var ander = 0;
+    if ((bits|0) < 32) {
+      ander = ((1 << bits) - 1)|0;
+      tempRet0 = high >>> bits;
+      return (low >>> bits) | ((high&ander) << (32 - bits));
+    }
+    tempRet0 = 0;
+    return (high >>> (bits - 32))|0;
+  },
 };
 
 function autoAddDeps(object, name) {
@@ -7396,3 +7456,4 @@ function autoAddDeps(object, name) {
   }
 }
 
+
diff --git a/src/library_browser.js b/src/library_browser.js
index 5b19a360..bdd94bac 100644
--- a/src/library_browser.js
+++ b/src/library_browser.js
@@ -3,7 +3,7 @@
 // Utilities for browser environments
 
 mergeInto(LibraryManager.library, {
-  $Browser__postset: 'Module["requestFullScreen"] = function() { Browser.requestFullScreen() };\n' + // exports
+  $Browser__postset: 'Module["requestFullScreen"] = function(lockPointer, resizeCanvas) { Browser.requestFullScreen(lockPointer, resizeCanvas) };\n' + // exports
                      'Module["requestAnimationFrame"] = function(func) { Browser.requestAnimationFrame(func) };\n' +
                      'Module["pauseMainLoop"] = function() { Browser.mainLoop.pause() };\n' +
                      'Module["resumeMainLoop"] = function() { Browser.mainLoop.resume() };\n',
@@ -40,6 +40,7 @@ mergeInto(LibraryManager.library, {
         }
       }
     },
+    isFullScreen: false,
     pointerLock: false,
     moduleContextCreatedCallbacks: [],
     workers: [],
@@ -205,10 +206,10 @@ mergeInto(LibraryManager.library, {
       try {
         if (useWebGL) {
           ctx = canvas.getContext('experimental-webgl', {
-            alpha: false,
 #if GL_TESTING
-            preserveDrawingBuffer: true
+            preserveDrawingBuffer: true,
 #endif
+            alpha: false
           });
         } else {
           ctx = canvas.getContext('2d');
@@ -273,36 +274,60 @@ mergeInto(LibraryManager.library, {
       }
       return ctx;
     },
+
     destroyContext: function(canvas, useWebGL, setInModule) {},
-    requestFullScreen: function() {
+
+    fullScreenHandlersInstalled: false,
+    lockPointer: undefined,
+    resizeCanvas: undefined,
+    requestFullScreen: function(lockPointer, resizeCanvas) {
+      this.lockPointer = lockPointer;
+      this.resizeCanvas = resizeCanvas;
+      if (typeof this.lockPointer === 'undefined') this.lockPointer = true;
+      if (typeof this.resizeCanvas === 'undefined') this.resizeCanvas = false;
+
       var canvas = Module['canvas'];
       function fullScreenChange() {
-        var isFullScreen = false;
+        Browser.isFullScreen = false;
         if ((document['webkitFullScreenElement'] || document['webkitFullscreenElement'] ||
              document['mozFullScreenElement'] || document['mozFullscreenElement'] ||
              document['fullScreenElement'] || document['fullscreenElement']) === canvas) {
           canvas.requestPointerLock = canvas['requestPointerLock'] ||
                                       canvas['mozRequestPointerLock'] ||
                                       canvas['webkitRequestPointerLock'];
-          canvas.requestPointerLock();
-          isFullScreen = true;
+          canvas.exitPointerLock = document['exitPointerLock'] ||
+                                   document['mozExitPointerLock'] ||
+                                   document['webkitExitPointerLock'];
+          canvas.exitPointerLock = canvas.exitPointerLock.bind(document);
+          canvas.cancelFullScreen = document['cancelFullScreen'] ||
+                                    document['mozCancelFullScreen'] ||
+                                    document['webkitCancelFullScreen'];
+          canvas.cancelFullScreen = canvas.cancelFullScreen.bind(document);
+          if (Browser.lockPointer) canvas.requestPointerLock();
+          Browser.isFullScreen = true;
+          if (Browser.resizeCanvas) Browser.setFullScreenCanvasSize();
+        } else if (Browser.resizeCanvas){
+          Browser.setWindowedCanvasSize();
         }
-        if (Module['onFullScreen']) Module['onFullScreen'](isFullScreen);
+        if (Module['onFullScreen']) Module['onFullScreen'](Browser.isFullScreen);
       }
 
-      document.addEventListener('fullscreenchange', fullScreenChange, false);
-      document.addEventListener('mozfullscreenchange', fullScreenChange, false);
-      document.addEventListener('webkitfullscreenchange', fullScreenChange, false);
-
       function pointerLockChange() {
         Browser.pointerLock = document['pointerLockElement'] === canvas ||
                               document['mozPointerLockElement'] === canvas ||
                               document['webkitPointerLockElement'] === canvas;
       }
 
-      document.addEventListener('pointerlockchange', pointerLockChange, false);
-      document.addEventListener('mozpointerlockchange', pointerLockChange, false);
-      document.addEventListener('webkitpointerlockchange', pointerLockChange, false);
+      if (!this.fullScreenHandlersInstalled) {
+        this.fullScreenHandlersInstalled = true;
+        document.addEventListener('fullscreenchange', fullScreenChange, false);
+        document.addEventListener('mozfullscreenchange', fullScreenChange, false);
+        document.addEventListener('webkitfullscreenchange', fullScreenChange, false);
+
+        document.addEventListener('pointerlockchange', pointerLockChange, false);
+        document.addEventListener('mozpointerlockchange', pointerLockChange, false);
+        document.addEventListener('webkitpointerlockchange', pointerLockChange, false);
+      }
 
       canvas.requestFullScreen = canvas['requestFullScreen'] ||
                                  canvas['mozRequestFullScreen'] ||
@@ -380,7 +405,32 @@ mergeInto(LibraryManager.library, {
       canvas.width = width;
       canvas.height = height;
       if (!noUpdates) Browser.updateResizeListeners();
+    },
+
+    windowedWidth: 0,
+    windowedHeight: 0,
+    setFullScreenCanvasSize: function() {
+      var canvas = Module['canvas'];
+      this.windowedWidth = canvas.width;
+      this.windowedHeight = canvas.height;
+      canvas.width = screen.width;
+      canvas.height = screen.height;
+      var flags = {{{ makeGetValue('SDL.screen+Runtime.QUANTUM_SIZE*0', '0', 'i32', 0, 1) }}};
+      flags = flags | 0x00800000; // set SDL_FULLSCREEN flag
+      {{{ makeSetValue('SDL.screen+Runtime.QUANTUM_SIZE*0', '0', 'flags', 'i32') }}}
+      Browser.updateResizeListeners();
+    },
+    
+    setWindowedCanvasSize: function() {
+      var canvas = Module['canvas'];
+      canvas.width = this.windowedWidth;
+      canvas.height = this.windowedHeight;
+      var flags = {{{ makeGetValue('SDL.screen+Runtime.QUANTUM_SIZE*0', '0', 'i32', 0, 1) }}};
+      flags = flags & ~0x00800000; // clear SDL_FULLSCREEN flag
+      {{{ makeSetValue('SDL.screen+Runtime.QUANTUM_SIZE*0', '0', 'flags', 'i32') }}}
+      Browser.updateResizeListeners();
     }
+    
   },
 
   emscripten_async_wget: function(url, file, onload, onerror) {
diff --git a/src/library_gl.js b/src/library_gl.js
index 4977d2e9..9e12e4ee 100644
--- a/src/library_gl.js
+++ b/src/library_gl.js
@@ -21,7 +21,6 @@ var LibraryGL = {
 
 #if FULL_ES2
     clientBuffers: [],
-    enabledClientBuffers: [],
 #endif
     currArrayBuffer: 0,
     currElementArrayBuffer: 0,
@@ -59,6 +58,65 @@ var LibraryGL = {
       return ret;
     },
 
+    // Temporary buffers
+    MAX_TEMP_BUFFER_SIZE: {{{ GL_MAX_TEMP_BUFFER_SIZE }}},
+    tempBufferIndexLookup: null,
+    tempVertexBuffers: null,
+    tempIndexBuffers: null,
+    tempQuadIndexBuffer: null,
+
+    generateTempBuffers: function(quads) {
+      this.tempBufferIndexLookup = new Uint8Array(this.MAX_TEMP_BUFFER_SIZE+1);
+      this.tempVertexBuffers = [];
+      this.tempIndexBuffers = [];
+      var last = -1, curr = -1;
+      var size = 1;
+      for (var i = 0; i <= this.MAX_TEMP_BUFFER_SIZE; i++) {
+        if (i > size) {
+          size <<= 1;
+        }
+        if (size != last) {
+          curr++;
+          this.tempVertexBuffers[curr] = Module.ctx.createBuffer();
+          Module.ctx.bindBuffer(Module.ctx.ARRAY_BUFFER, this.tempVertexBuffers[curr]);
+          Module.ctx.bufferData(Module.ctx.ARRAY_BUFFER, size, Module.ctx.DYNAMIC_DRAW);
+          Module.ctx.bindBuffer(Module.ctx.ARRAY_BUFFER, null);
+          this.tempIndexBuffers[curr] = Module.ctx.createBuffer();
+          Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, this.tempIndexBuffers[curr]);
+          Module.ctx.bufferData(Module.ctx.ELEMENT_ARRAY_BUFFER, size, Module.ctx.DYNAMIC_DRAW);
+          Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, null);
+          last = size;
+        }
+        this.tempBufferIndexLookup[i] = curr;
+      }
+
+      if (quads) {
+        // GL_QUAD indexes can be precalculated
+        this.tempQuadIndexBuffer = Module.ctx.createBuffer();
+        Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, this.tempQuadIndexBuffer);
+        var numIndexes = this.MAX_TEMP_BUFFER_SIZE >> 1;
+        var quadIndexes = new Uint16Array(numIndexes);
+        var i = 0, v = 0;
+        while (1) {
+          quadIndexes[i++] = v;
+          if (i >= numIndexes) break;
+          quadIndexes[i++] = v+1;
+          if (i >= numIndexes) break;
+          quadIndexes[i++] = v+2;
+          if (i >= numIndexes) break;
+          quadIndexes[i++] = v;
+          if (i >= numIndexes) break;
+          quadIndexes[i++] = v+2;
+          if (i >= numIndexes) break;
+          quadIndexes[i++] = v+3;
+          if (i >= numIndexes) break;
+          v += 4;
+        }
+        Module.ctx.bufferData(Module.ctx.ELEMENT_ARRAY_BUFFER, quadIndexes, Module.ctx.STATIC_DRAW);
+        Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, null);
+      }
+    },
+
     // Linear lookup in one of the tables (buffers, programs, etc.). TODO: consider using a weakmap to make this faster, if it matters
     scan: function(table, object) {
       for (var item in table) {
@@ -207,20 +265,35 @@ var LibraryGL = {
       return size * typeSize * count;
     },
 
+    usedTempBuffers: [],
+
     preDrawHandleClientVertexAttribBindings: function(count) {
       GL.resetBufferBinding = false;
+
+      var used = GL.usedTempBuffers;
+      used.length = 0;
+
+      // TODO: initial pass to detect ranges we need to upload, might not need an upload per attrib
       for (var i = 0; i < GL.maxVertexAttribs; ++i) {
-        if (!GL.enabledClientBuffers[i] || !GL.clientBuffers[i]) continue;
-  
-        GL.resetBufferBinding = true;
-  
         var cb = GL.clientBuffers[i];
-  
-        var buf = Module.ctx.createBuffer();
+        if (!cb.clientside || !cb.enabled) continue;
+
+        GL.resetBufferBinding = true;
+
+        var size = GL.calcBufLength(cb.size, cb.type, cb.stride, count);
+        var index = GL.tempBufferIndexLookup[size];
+        var buf;
+        do {
+#if ASSERTIONS
+          assert(index < GL.tempVertexBuffers.length);
+#endif
+          buf = GL.tempVertexBuffers[index++];
+        } while (used.indexOf(buf) >= 0);
+        used.push(buf);
         Module.ctx.bindBuffer(Module.ctx.ARRAY_BUFFER, buf);
-        Module.ctx.bufferData(Module.ctx.ARRAY_BUFFER,
-                              HEAPU8.subarray(cb.ptr, cb.ptr + GL.calcBufLength(cb.size, cb.type, cb.stride, count)),
-                              Module.ctx.DYNAMIC_DRAW);
+        Module.ctx.bufferSubData(Module.ctx.ARRAY_BUFFER,
+                                 0,
+                                 HEAPU8.subarray(cb.ptr, cb.ptr + size));
         Module.ctx.vertexAttribPointer(i, cb.size, cb.type, cb.normalized, cb.stride, 0);
       }
     },
@@ -239,6 +312,13 @@ var LibraryGL = {
       if (!Module.useWebGL) return; // an app might link both gl and 2d backends
 
       GL.maxVertexAttribs = Module.ctx.getParameter(Module.ctx.MAX_VERTEX_ATTRIBS);
+#if FULL_ES2
+      for (var i = 0; i < GL.maxVertexAttribs; i++) {
+        GL.clientBuffers[i] = { enabled: false, clientside: false, size: 0, type: 0, normalized: 0, stride: 0, ptr: 0 };
+      }
+
+      GL.generateTempBuffers();
+#endif
 
       GL.compressionExt = Module.ctx.getExtension('WEBGL_compressed_texture_s3tc') ||
                           Module.ctx.getExtension('MOZ_WEBGL_compressed_texture_s3tc') ||
@@ -612,7 +692,7 @@ var LibraryGL = {
 
   glGetVertexAttribfv: function(index, pname, params) {
 #if FULL_ES2
-    if (GL.clientBuffers[index]) {
+    if (GL.clientBuffers[index].enabled) {
       Module.printErr("glGetVertexAttribfv on client-side array: not supported, bad data returned");
     }
 #endif
@@ -628,7 +708,7 @@ var LibraryGL = {
 
   glGetVertexAttribiv: function(index, pname, params) {
 #if FULL_ES2
-    if (GL.clientBuffers[index]) {
+    if (GL.clientBuffers[index].enabled) {
       Module.printErr("glGetVertexAttribiv on client-side array: not supported, bad data returned");
     }
 #endif
@@ -644,7 +724,7 @@ var LibraryGL = {
 
   glGetVertexAttribPointerv: function(index, pname, pointer) {
 #if FULL_ES2
-    if (GL.clientBuffers[index]) {
+    if (GL.clientBuffers[index].enabled) {
       Module.printErr("glGetVertexAttribPointer on client-side array: not supported, bad data returned");
     }
 #endif
@@ -1693,62 +1773,6 @@ var LibraryGL = {
       this.modifiedClientAttributes = true;
     },
 
-    // Temporary buffers
-    MAX_TEMP_BUFFER_SIZE: {{{ GL_MAX_TEMP_BUFFER_SIZE }}},
-    tempBufferIndexLookup: null,
-    tempVertexBuffers: null,
-    tempIndexBuffers: null,
-    tempQuadIndexBuffer: null,
-
-    generateTempBuffers: function() {
-      this.tempBufferIndexLookup = new Uint8Array(this.MAX_TEMP_BUFFER_SIZE+1);
-      this.tempVertexBuffers = [];
-      this.tempIndexBuffers = [];
-      var last = -1, curr = -1;
-      var size = 1;
-      for (var i = 0; i <= this.MAX_TEMP_BUFFER_SIZE; i++) {
-        if (i > size) {
-          size <<= 1;
-        }
-        if (size != last) {
-          curr++;
-          this.tempVertexBuffers[curr] = Module.ctx.createBuffer();
-          Module.ctx.bindBuffer(Module.ctx.ARRAY_BUFFER, this.tempVertexBuffers[curr]);
-          Module.ctx.bufferData(Module.ctx.ARRAY_BUFFER, size, Module.ctx.DYNAMIC_DRAW);
-          Module.ctx.bindBuffer(Module.ctx.ARRAY_BUFFER, null);
-          this.tempIndexBuffers[curr] = Module.ctx.createBuffer();
-          Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, this.tempIndexBuffers[curr]);
-          Module.ctx.bufferData(Module.ctx.ELEMENT_ARRAY_BUFFER, size, Module.ctx.DYNAMIC_DRAW);
-          Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, null);
-          last = size;
-        }
-        this.tempBufferIndexLookup[i] = curr;
-      }
-      // GL_QUAD indexes can be precalculated
-      this.tempQuadIndexBuffer = Module.ctx.createBuffer();
-      Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, this.tempQuadIndexBuffer);
-      var numIndexes = this.MAX_TEMP_BUFFER_SIZE >> 1;
-      var quadIndexes = new Uint16Array(numIndexes);
-      var i = 0, v = 0;
-      while (1) {
-        quadIndexes[i++] = v;
-        if (i >= numIndexes) break;
-        quadIndexes[i++] = v+1;
-        if (i >= numIndexes) break;
-        quadIndexes[i++] = v+2;
-        if (i >= numIndexes) break;
-        quadIndexes[i++] = v;
-        if (i >= numIndexes) break;
-        quadIndexes[i++] = v+2;
-        if (i >= numIndexes) break;
-        quadIndexes[i++] = v+3;
-        if (i >= numIndexes) break;
-        v += 4;
-      }
-      Module.ctx.bufferData(Module.ctx.ELEMENT_ARRAY_BUFFER, quadIndexes, Module.ctx.STATIC_DRAW);
-      Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, null);
-    },
-
     // Renderers
     addRendererComponent: function(name, size, type) {
       if (!this.rendererComponents[name]) {
@@ -1969,8 +1993,8 @@ var LibraryGL = {
           if (!GL.currArrayBuffer) {
             var start = GL.immediate.firstVertex*GL.immediate.stride;
             var end = GL.immediate.lastVertex*GL.immediate.stride;
-            assert(end <= GL.immediate.MAX_TEMP_BUFFER_SIZE, 'too much vertex data');
-            arrayBuffer = GL.immediate.tempVertexBuffers[GL.immediate.tempBufferIndexLookup[end]];
+            assert(end <= GL.MAX_TEMP_BUFFER_SIZE, 'too much vertex data');
+            arrayBuffer = GL.tempVertexBuffers[GL.tempBufferIndexLookup[end]];
             // TODO: consider using the last buffer we bound, if it was larger. downside is larger buffer, but we might avoid rebinding and preparing
           } else {
             arrayBuffer = GL.currArrayBuffer;
@@ -2161,12 +2185,12 @@ var LibraryGL = {
       this.rendererCache = this.rendererCacheItemTemplate.slice();
 
       // Buffers for data
-      this.tempData = new Float32Array(this.MAX_TEMP_BUFFER_SIZE >> 2);
-      this.indexData = new Uint16Array(this.MAX_TEMP_BUFFER_SIZE >> 1);
+      this.tempData = new Float32Array(GL.MAX_TEMP_BUFFER_SIZE >> 2);
+      this.indexData = new Uint16Array(GL.MAX_TEMP_BUFFER_SIZE >> 1);
 
       this.vertexDataU8 = new Uint8Array(this.tempData.buffer);
 
-      this.generateTempBuffers();
+      GL.generateTempBuffers(true);
 
       this.clientColor = new Float32Array([1, 1, 1, 1]);
     },
@@ -2209,7 +2233,7 @@ var LibraryGL = {
 #if ASSERTIONS
         Runtime.warnOnce('Unpacking/restriding attributes, this is not fast');
 #endif
-        if (!GL.immediate.restrideBuffer) GL.immediate.restrideBuffer = _malloc(GL.immediate.MAX_TEMP_BUFFER_SIZE);
+        if (!GL.immediate.restrideBuffer) GL.immediate.restrideBuffer = _malloc(GL.MAX_TEMP_BUFFER_SIZE);
         start = GL.immediate.restrideBuffer;
 #if ASSERTIONS
         assert(start % 4 == 0);
@@ -2224,7 +2248,7 @@ var LibraryGL = {
           bytes += size;
         }
 #if ASSERTIONS
-        assert(count*bytes <= GL.immediate.MAX_TEMP_BUFFER_SIZE);
+        assert(count*bytes <= GL.MAX_TEMP_BUFFER_SIZE);
 #endif
         // copy out the data (we need to know the stride for that, and define attribute.pointer
         for (var i = 0; i < attributes.length; i++) {
@@ -2298,8 +2322,8 @@ var LibraryGL = {
         }
         if (!GL.currElementArrayBuffer) {
           // If no element array buffer is bound, then indices is a literal pointer to clientside data
-          assert(numProvidedIndexes << 1 <= GL.immediate.MAX_TEMP_BUFFER_SIZE, 'too many immediate mode indexes (a)');
-          var indexBuffer = GL.immediate.tempIndexBuffers[GL.immediate.tempBufferIndexLookup[numProvidedIndexes << 1]];
+          assert(numProvidedIndexes << 1 <= GL.MAX_TEMP_BUFFER_SIZE, 'too many immediate mode indexes (a)');
+          var indexBuffer = GL.tempIndexBuffers[GL.tempBufferIndexLookup[numProvidedIndexes << 1]];
           Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, indexBuffer);
           Module.ctx.bufferSubData(Module.ctx.ELEMENT_ARRAY_BUFFER, 0, {{{ makeHEAPView('U16', 'ptr', 'ptr + (numProvidedIndexes << 1)') }}});
           ptr = 0;
@@ -2314,8 +2338,8 @@ var LibraryGL = {
         ptr = GL.immediate.firstVertex*3;
         var numQuads = numVertexes / 4;
         numIndexes = numQuads * 6; // 0 1 2, 0 2 3 pattern
-        assert(ptr + (numIndexes << 1) <= GL.immediate.MAX_TEMP_BUFFER_SIZE, 'too many immediate mode indexes (b)');
-        Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, this.tempQuadIndexBuffer);
+        assert(ptr + (numIndexes << 1) <= GL.MAX_TEMP_BUFFER_SIZE, 'too many immediate mode indexes (b)');
+        Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, GL.tempQuadIndexBuffer);
         emulatedElementArrayBuffer = true;
       }
 
@@ -2369,7 +2393,7 @@ var LibraryGL = {
     GL.immediate.vertexData[GL.immediate.vertexCounter++] = y;
     GL.immediate.vertexData[GL.immediate.vertexCounter++] = z || 0;
 #if ASSERTIONS
-    assert(GL.immediate.vertexCounter << 2 < GL.immediate.MAX_TEMP_BUFFER_SIZE);
+    assert(GL.immediate.vertexCounter << 2 < GL.MAX_TEMP_BUFFER_SIZE);
 #endif
     GL.immediate.addRendererComponent(GL.immediate.VERTEX, 3, Module.ctx.FLOAT);
   },
@@ -2849,15 +2873,25 @@ var LibraryGL = {
 
   glShadeModel: function() { Runtime.warnOnce('TODO: glShadeModel') },
 
+  // GLES2 emulation
+
   glVertexAttribPointer__sig: 'viiiiii',
   glVertexAttribPointer: function(index, size, type, normalized, stride, ptr) {
 #if FULL_ES2
+    var cb = GL.clientBuffers[index];
+#if ASSERTIONS
+    assert(cb, index);
+#endif
     if (!GL.currArrayBuffer) {
-      GL.clientBuffers[index] = { size: size, type: type, normalized: normalized, stride: stride, ptr: ptr };
+      cb.size = size;
+      cb.type = type;
+      cb.normalized = normalized;
+      cb.stride = stride;
+      cb.ptr = ptr;
+      cb.clientside = true;
       return;
     }
-
-    GL.clientBuffers[index] = null;
+    cb.clientside = false;
 #endif
     Module.ctx.vertexAttribPointer(index, size, type, normalized, stride, ptr);
   },
@@ -2865,7 +2899,11 @@ var LibraryGL = {
   glEnableVertexAttribArray__sig: 'vi',
   glEnableVertexAttribArray: function(index) {
 #if FULL_ES2
-    GL.enabledClientBuffers[index] = true;
+    var cb = GL.clientBuffers[index];
+#if ASSERTIONS
+    assert(cb, index);
+#endif
+    cb.enabled = true;
 #endif
     Module.ctx.enableVertexAttribArray(index);
   },
@@ -2873,7 +2911,11 @@ var LibraryGL = {
   glDisableVertexAttribArray__sig: 'vi',
   glDisableVertexAttribArray: function(index) {
 #if FULL_ES2
-    GL.enabledClientBuffers[index] = false;
+    var cb = GL.clientBuffers[index];
+#if ASSERTIONS
+    assert(cb, index);
+#endif
+    cb.enabled = false;
 #endif
     Module.ctx.disableVertexAttribArray(index);
   },
@@ -2881,7 +2923,7 @@ var LibraryGL = {
   glDrawArrays: function(mode, first, count) {
 #if FULL_ES2
     // bind any client-side buffers
-    GL.preDrawHandleClientVertexAttribBindings(count);
+    GL.preDrawHandleClientVertexAttribBindings(first + count);
 #endif
 
     Module.ctx.drawArrays(mode, first, count);
@@ -2895,11 +2937,12 @@ var LibraryGL = {
 #if FULL_ES2
     var buf;
     if (!GL.currElementArrayBuffer) {
-      buf = Module.ctx.createBuffer();
+      var size = GL.calcBufLength(1, type, 0, count);
+      buf = GL.tempIndexBuffers[GL.tempBufferIndexLookup[size]];
       Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, buf);
-      Module.ctx.bufferData(Module.ctx.ELEMENT_ARRAY_BUFFER,
-                            HEAPU8.subarray(indices, indices + GL.calcBufLength(1, type, 0, count)),
-                            Module.ctx.DYNAMIC_DRAW);
+      Module.ctx.bufferSubData(Module.ctx.ELEMENT_ARRAY_BUFFER,
+                               0,
+                               HEAPU8.subarray(indices, indices + size));
       // the index is now 0
       indices = 0;
     }
@@ -2915,17 +2958,18 @@ var LibraryGL = {
 
     if (!GL.currElementArrayBuffer) {
       Module.ctx.bindBuffer(Module.ctx.ELEMENT_ARRAY_BUFFER, null);
-      Module.ctx.deleteBuffer(buf);
     }
 #endif
   },
 
   // signatures of simple pass-through functions, see later
+
   glActiveTexture__sig: 'vi',
   glCheckFramebufferStatus__sig: 'ii',
   glRenderbufferStorage__sig: 'viiii',
 
   // Open GLES1.1 compatibility
+
   glGenFramebuffersOES : 'glGenFramebuffers',
   glGenRenderbuffersOES : 'glGenRenderbuffers',
   glBindFramebufferOES : 'glBindFramebuffer',
diff --git a/src/library_openal.js b/src/library_openal.js
new file mode 100644
index 00000000..5511ccb7
--- /dev/null
+++ b/src/library_openal.js
@@ -0,0 +1,471 @@
+//"use strict";
+
+var LibraryOpenAL = {
+  $AL__deps: ['$Browser'],
+  $AL: {
+    contexts: [],
+    currentContext: null,
+  },
+
+  alcProcessContext: function(context) {},
+  alcSuspendContext: function(context) {},
+
+  alcMakeContextCurrent: function(context) {
+    if (context == 0) {
+      AL.currentContext = null;
+    } else {
+      AL.currentContext = AL.contexts[context - 1];
+    }
+  },
+
+  alcDestroyContext: function(context) {
+    // Stop playback, etc
+  },
+
+  alcCloseDevice: function(device) {
+    // Stop playback, etc
+  },
+
+  alcOpenDevice: function(deviceName) {
+    if (typeof(AudioContext) == "function" ||
+        typeof(webkitAudioContext) == "function") {
+      return 1; // non-null pointer -- we just simulate one device
+    } else {
+      return 0;
+    }
+  },
+
+  alcCreateContext: function(device, attrList) {
+    if (device != 1) {
+      return 0;
+    }
+
+    if (attrList) {
+      console.log("The attrList argument of alcCreateContext is not supported yet");
+      return 0;
+    }
+
+    var ctx;
+    try {
+      ctx = new AudioContext();
+    } catch (e) {
+      try {
+        ctx = new webkitAudioContext();
+      } catch (e) {}
+    }
+
+    if (ctx) {
+      AL.contexts.push({ctx: ctx, err: 0, src: [], buf: []});
+      return AL.contexts.length;
+    } else {
+      return 0;
+    }
+  },
+
+  alGetError: function() {
+    if (!AL.currentContext) {
+      return 0xA004 /* AL_INVALID_OPERATION */;
+    } else {
+      return AL.currentContext.err;
+    }
+  },
+
+  alDeleteSources: function(count, sources)
+  {
+    if (!AL.currentContext) {
+      console.error("alDeleteSources called without a valid context");
+      return;
+    }
+    for (var i = 0; i < count; ++i) {
+      var sourceIdx = {{{ makeGetValue('sources', 'i', 'i32') }}} - 1;
+      delete AL.currentContext.src[sourceIdx];
+    }
+  },
+
+  alGenSources: function(count, sources) {
+    if (!AL.currentContext) {
+      console.error("alGenSources called without a valid context");
+      return;
+    }
+    for (var i = 0; i < count; ++i) {
+      var gain = AL.currentContext.ctx.createGain();
+      var panner = AL.currentContext.ctx.createPanner();
+      if (typeof(webkitAudioContext) == 'function') {
+        gain.connect(panner);
+        panner.connect(AL.currentContext.ctx.destination);
+      } else {
+        // Work around a Firefox bug (bug 849916)
+        gain.connect(AL.currentContext.ctx.destination);
+      }
+      gain.gain.value = 1; // work around a Firefox bug (bug 850970)
+      AL.currentContext.src.push({
+        loop: false,
+        buffer: null,
+        gain: gain,
+        panner: panner,
+        paused: false,
+        playTime: -1,
+        pausedTime: 0
+      });
+      {{{ makeSetValue('sources', 'i', 'AL.currentContext.src.length', 'i32') }}};
+    }
+  },
+
+  alSourcei: function(source, param, value) {
+    if (!AL.currentContext) {
+      console.error("alSourcei called without a valid context");
+      return;
+    }
+    if (source > AL.currentContext.src.length) {
+      console.error("alSourcei called with an invalid source");
+      return;
+    }
+    switch (param) {
+    case 0x1007 /* AL_LOOPING */:
+      AL.currentContext.src[source - 1].loop = (value != 0 /* AL_FALSE */);
+      break;
+    case 0x1009 /* AL_BUFFER */:
+      if (value == 0) {
+        AL.currentContext.src[source - 1].buffer = null;
+      } else {
+        AL.currentContext.src[source - 1].buffer = AL.currentContext.buf[value - 1].buf;
+      }
+      break;
+    default:
+      console.log("alSourcei with param " + param + " not implemented yet");
+      break;
+    }
+  },
+
+  alSourcef: function(source, param, value) {
+    if (!AL.currentContext) {
+      consoue.error("alSourcef called without a valid context");
+      return;
+    }
+    if (source > AL.currentContext.src.length) {
+      console.error("alSourcef called with an invalid source");
+      return;
+    }
+    switch (param) {
+    case 0x100A /* AL_GAIN */:
+      AL.currentContext.src[source - 1].gain.gain.value = value;
+      break;
+    case 0x1003 /* AL_PITCH */:
+      console.log("alSourcef was called with AL_PITCH, but Web Audio does not support static pitch changes");
+      break;
+    default:
+      console.log("alSourcef with param " + param + " not implemented yet");
+      break;
+    }
+  },
+
+  alSourcefv: function(source, param, value) {
+    if (!AL.currentContext) {
+      consoue.error("alSourcefv called without a valid context");
+      return;
+    }
+    if (source > AL.currentContext.src.length) {
+      console.error("alSourcefv called with an invalid source");
+      return;
+    }
+    switch (param) {
+    case 0x1004 /* AL_POSITION */:
+      AL.currentContext.src[source - 1].panner.setPosition(
+          {{{ makeGetValue('value', '0', 'float') }}},
+          {{{ makeGetValue('value', '1', 'float') }}},
+          {{{ makeGetValue('value', '2', 'float') }}}
+        );
+      break;
+    case 0x1006 /* AL_VELOCITY */:
+      AL.currentContext.src[source - 1].panner.setVelocity(
+          {{{ makeGetValue('value', '0', 'float') }}},
+          {{{ makeGetValue('value', '1', 'float') }}},
+          {{{ makeGetValue('value', '2', 'float') }}}
+        );
+      break;
+    default:
+      console.log("alSourcefv with param " + param + " not implemented yet");
+      break;
+    }
+  },
+
+  alSourceQueueBuffers: function(source, count, buffers) {
+    if (!AL.currentContext) {
+      console.error("alSourceQueueBuffers called without a valid context");
+      return;
+    }
+    if (source > AL.currentContext.src.length) {
+      console.error("alSourceQueueBuffers called with an invalid source");
+      return;
+    }
+    if (count != 1) {
+      console.error("Queuing multiple buffers using alSourceQueueBuffers is not supported yet");
+      return;
+    }
+    for (var i = 0; i < count; ++i) {
+      var buffer = {{{ makeGetValue('buffers', 'i', 'i32') }}};
+      if (buffer > AL.currentContext.buf.length) {
+        console.error("alSourceQueueBuffers called with an invalid buffer");
+        return;
+      }
+      AL.currentContext.src[source - 1].buffer = AL.currentContext.buf[buffer - 1].buf;
+    }
+  },
+
+  alSourceUnqueueBuffers: function(source, count, buffers)
+  {
+    if (!AL.currentContext) {
+      console.error("alSourceUnqueueBuffers called without a valid context");
+      return;
+    }
+    if (source > AL.currentContext.src.length) {
+      console.error("alSourceUnqueueBuffers called with an invalid source");
+      return;
+    }
+    if (count != 1) {
+      console.error("Queuing multiple buffers using alSourceUnqueueBuffers is not supported yet");
+      return;
+    }
+    for (var i = 0; i < count; ++i) {
+      var buffer = AL.currentContext.src[source - 1].buffer;
+      for (var j = 0; j < AL.currentContext.buf.length; ++j) {
+        if (buffer == AL.currentContext.buf[j].buf) {
+          {{{ makeSetValue('buffers', 'i', 'j+1', 'i32') }}};
+          AL.currentContext.src[source - 1].buffer = null;
+          break;
+        }
+      }
+    }
+  },
+
+  alDeleteBuffers: function(count, buffers)
+  {
+    if (!AL.currentContext) {
+      console.error("alDeleteBuffers called without a valid context");
+      return;
+    }
+    for (var i = 0; i < count; ++i) {
+      var bufferIdx = {{{ makeGetValue('buffers', 'i', 'i32') }}} - 1;
+      var buffer = AL.currentContext.buf[bufferIdx].buf;
+      for (var j = 0; j < AL.currentContext.src.length; ++j) {
+        if (buffer == AL.currentContext.src[j].buffer) {
+          AL.currentContext.err = 0xA004 /* AL_INVALID_OPERATION  */;
+          return;
+        }
+      }
+      delete AL.currentContext.buf[bufferIdx];
+    }
+  },
+
+  alGenBuffers: function(count, buffers) {
+    if (!AL.currentContext) {
+      console.error("alGenBuffers called without a valid context");
+      return;
+    }
+    for (var i = 0; i < count; ++i) {
+      AL.currentContext.buf.push({buf: null});
+      {{{ makeSetValue('buffers', 'i', 'AL.currentContext.buf.length', 'i32') }}};
+    }
+  },
+
+  alBufferData: function(buffer, format, data, size, freq) {
+    if (!AL.currentContext) {
+      console.error("alBufferData called without a valid context");
+      return;
+    }
+    if (buffer > AL.currentContext.buf.length) {
+      console.error("alBufferData called with an invalid buffer");
+      return;
+    }
+    var channels, bytes;
+    switch (format) {
+    case 0x1100 /* AL_FORMAT_MONO8 */:
+      bytes = 1;
+      channels = 1;
+      break;
+    case 0x1101 /* AL_FORMAT_MONO16 */:
+      bytes = 2;
+      channels = 1;
+      break;
+    case 0x1102 /* AL_FORMAT_STEREO8 */:
+      bytes = 1;
+      channels = 2;
+      break;
+    case 0x1103 /* AL_FORMAT_STEREO16 */:
+      bytes = 2;
+      channels = 2;
+      break;
+    default:
+      console.error("alBufferData called with invalid format " + format);
+      return;
+    }
+    AL.currentContext.buf[buffer - 1].buf = AL.currentContext.ctx.createBuffer(channels, size / (bytes * channels), freq);
+    var buf = new Array(channels);
+    for (var i = 0; i < channels; ++i) {
+      buf[i] = AL.currentContext.buf[buffer - 1].buf.getChannelData(i);
+    }
+    for (var i = 0; i < size / (bytes * channels); ++i) {
+      for (var j = 0; j < channels; ++j) {
+        switch (bytes) {
+        case 1:
+          var val = {{{ makeGetValue('data', 'i*channels+j', 'i8') }}};
+          buf[j][i] = -1.0 + val * (2/256);
+          break;
+        case 2:
+          var val = {{{ makeGetValue('data', '2*(i*channels+j)', 'i16') }}};
+          buf[j][i] = val/32768;
+          break;
+        }
+      }
+    }
+  },
+
+  alSourcePlay__deps: ["alSourceStop"],
+  alSourcePlay: function(source) {
+    if (!AL.currentContext) {
+      console.error("alSourcePlay called without a valid context");
+      return;
+    }
+    if (source > AL.currentContext.src.length) {
+      console.error("alSourcePlay called with an invalid source");
+      return;
+    }
+    var offset = 0;
+    if ("src" in AL.currentContext.src[source - 1]) {
+      // If the source is already playing, we need to resume from beginning.
+      // We do that by stopping the current source and replaying it.
+      _alSourceStop(source);
+    } else if (AL.currentContext.src[source - 1].paused) {
+      // So now we have to resume playback, remember the offset here.
+      offset = AL.currentContext.src[source - 1].pausedTime -
+               AL.currentContext.src[source - 1].playTime;
+    }
+    var src = AL.currentContext.ctx.createBufferSource();
+    src.loop = AL.currentContext.src[source - 1].loop;
+    src.buffer = AL.currentContext.src[source - 1].buffer;
+    src.connect(AL.currentContext.src[source - 1].gain);
+    src.start(0, offset);
+    // Work around Firefox bug 851338
+    AL.currentContext.src[source - 1].playTime = AL.currentContext.ctx.currentTime || 0;
+    AL.currentContext.src[source - 1].paused = false;
+    AL.currentContext.src[source - 1]['src'] = src;
+  },
+
+  alSourceStop: function(source) {
+    if (!AL.currentContext) {
+      console.error("alSourceStop called without a valid context");
+      return;
+    }
+    if (source > AL.currentContext.src.length) {
+      console.error("alSourceStop called with an invalid source");
+      return;
+    }
+    if ("src" in AL.currentContext.src[source - 1]) {
+      AL.currentContext.src[source - 1]["src"].stop(0);
+      delete AL.currentContext.src[source - 1]["src"];
+    }
+  },
+
+  alSourcePause: function(source) {
+    if (!AL.currentContext) {
+      console.error("alSourcePause called without a valid context");
+      return;
+    }
+    if (source > AL.currentContext.src.length) {
+      console.error("alSourcePause called with an invalid source");
+      return;
+    }
+    if ("src" in AL.currentContext.src[source - 1] &&
+        !AL.currentContext.src[source - 1].paused) {
+      AL.currentContext.src[source - 1].paused = true;
+      // Work around Firefox bug 851338
+      AL.currentContext.src[source - 1].pausedTime = AL.currentContext.ctx.currentTime || 0;
+      AL.currentContext.src[source - 1]["src"].stop(0);
+      delete AL.currentContext.src[source - 1].src;
+    }
+  },
+
+  alGetSourcei: function(source, param, value) {
+    if (!AL.currentContext) {
+      console.error("alGetSourcei called without a valid context");
+      return;
+    }
+    if (source > AL.currentContext.src.length) {
+      console.error("alGetSourcei called with an invalid source");
+      return;
+    }
+    switch (param) {
+    case 0x202 /* AL_SOURCE_RELATIVE */:
+      // Always return 1
+      {{{ makeSetValue('value', '0', '1', 'i32') }}};
+      break;
+    case 0x1009 /* AL_BUFFER */:
+      if (AL.currentContext.src[source - 1].buffer == null) {
+        {{{ makeSetValue('value', '0', '0', 'i32') }}};
+      } else {
+        var buf = AL.currentContext.src[source - 1].buffer;
+        for (var i = 0; i < AL.currentContext.buf.length; ++i) {
+          if (buf == AL.currentContext.buf[i].buf) {
+            {{{ makeSetValue('value', '0', 'i+1', 'i32') }}};
+            return;
+          }
+        }
+        {{{ makeSetValue('value', '0', '0', 'i32') }}};
+      }
+      break;
+    case 0x1010 /* AL_SOURCE_STATE */:
+      if ("src" in AL.currentContext.src[source - 1]) {
+        {{{ makeSetValue('value', '0', '0x1012', 'i32') }}} /* AL_PLAYING */;
+      } else if (AL.currentContext.src[source - 1].paused) {
+        {{{ makeSetValue('value', '0', '0x1013', 'i32') }}} /* AL_PAUSED */;
+      } else if (AL.currentContext.src[source - 1].playTime == -1) {
+        {{{ makeSetValue('value', '0', '0x1011', 'i32') }}} /* AL_INITIAL */;
+      } else {
+        {{{ makeSetValue('value', '0', '0x1014', 'i32') }}} /* AL_STOPPED */;
+      }
+      break;
+    case 0x1015 /* AL_BUFFERS_QUEUED */:
+      if (AL.currentContext.src[source - 1].buffer) {
+        {{{ makeSetValue('value', '0', '1', 'i32') }}}
+      } else {
+        {{{ makeSetValue('value', '0', '0', 'i32') }}}
+      }
+      break;
+    case 0x1016 /* AL_BUFFERS_PROCESSED */:
+      // Always return 1
+      {{{ makeSetValue('value', '0', '1', 'i32') }}}
+      break;
+    }
+  },
+
+  alDistanceModel: function(model) {
+    if (model != 0 /* AL_NONE */) {
+      console.log("Only alDistanceModel(AL_NONE) is currently supported");
+    }
+  },
+
+  alListenerfv: function(param, values) {
+    console.log("alListenerfv is not supported yet");
+  },
+
+  alIsExtensionPresent: function(extName) {
+    return 0;
+  },
+
+  alcIsExtensionPresent: function(device, extName) {
+    return 0;
+  },
+
+  alGetProcAddress: function(fname) {
+    return 0;
+  },
+
+  alcGetProcAddress: function(device, fname) {
+    return 0;
+  },
+
+};
+
+autoAddDeps(LibraryOpenAL, '$AL');
+mergeInto(LibraryManager.library, LibraryOpenAL);
+
diff --git a/src/library_sdl.js b/src/library_sdl.js
index 96ae6fa2..9fc979d2 100644
--- a/src/library_sdl.js
+++ b/src/library_sdl.js
@@ -35,6 +35,7 @@ var LibrarySDL = {
     mixerFormat: 0x8010, // AUDIO_S16LSB
     mixerNumChannels: 2,
     mixerChunkSize: 1024,
+    channelMinimumNumber: 0,
 
     GL: false, // Set to true if we call SDL_SetVideoMode with SDL_OPENGL, and if so, we do not create 2D canvases&contexts for blitting
                // Note that images loaded before SDL_SetVideoMode will not get this optimization
@@ -661,8 +662,8 @@ var LibrarySDL = {
     {{{ makeSetValue('ret+Runtime.QUANTUM_SIZE*0', '0', '0', 'i32') }}} // TODO
     {{{ makeSetValue('ret+Runtime.QUANTUM_SIZE*1', '0', '0', 'i32') }}} // TODO
     {{{ makeSetValue('ret+Runtime.QUANTUM_SIZE*2', '0', '0', 'void*') }}}
-    {{{ makeSetValue('ret+Runtime.QUANTUM_SIZE*3', '0', 'SDL.defaults.width', 'i32') }}}
-    {{{ makeSetValue('ret+Runtime.QUANTUM_SIZE*4', '0', 'SDL.defaults.height', 'i32') }}}
+    {{{ makeSetValue('ret+Runtime.QUANTUM_SIZE*3', '0', 'Module["canvas"].width', 'i32') }}}
+    {{{ makeSetValue('ret+Runtime.QUANTUM_SIZE*4', '0', 'Module["canvas"].height', 'i32') }}}
     return ret;
   },
 
@@ -919,7 +920,26 @@ var LibrarySDL = {
   },
 
   SDL_ShowCursor: function(toggle) {
-    // TODO
+    switch (toggle) {
+      case 0: // SDL_DISABLE
+        if (Browser.isFullScreen) { // only try to lock the pointer when in full screen mode
+          Module['canvas'].requestPointerLock();
+          return 0;
+        } else { // else return SDL_ENABLE to indicate the failure
+          return 1;
+        }
+        break;
+      case 1: // SDL_ENABLE
+        Module['canvas'].exitPointerLock();
+        return 1;
+        break;
+      case -1: // SDL_QUERY
+        return !Browser.pointerLock;
+        break;
+      default:
+        console.log( "SDL_ShowCursor called with unknown toggle parameter value: " + toggle + "." );
+        break;
+    }
   },
 
   SDL_GetError: function() {
@@ -1075,6 +1095,15 @@ var LibrarySDL = {
   },
 
   SDL_WM_GrabInput: function() {},
+  
+  SDL_WM_ToggleFullScreen: function(surf) {
+    if (Browser.isFullScreen) {
+      Module['canvas'].cancelFullScreen();
+      return 1;
+    } else {
+      return 0;
+    }
+  },
 
   // SDL_Image
 
@@ -1289,7 +1318,9 @@ var LibrarySDL = {
   Mix_FreeChunk: function(id) {
     SDL.audios[id] = null;
   },
-
+  Mix_ReserveChannels: function(num) {
+    SDL.channelMinimumNumber = num;
+  },
   Mix_PlayChannel: function(channel, id, loops) {
     // TODO: handle loops
 
@@ -1302,8 +1333,8 @@ var LibrarySDL = {
     // If the user asks us to allocate a channel automatically, get the first
     // free one.
     if (channel == -1) {
-      channel = 0;
-      for (var i = 0; i < SDL.numChannels; i++) {
+      channel = SDL.channelMinimumNumber;
+      for (var i = SDL.channelMinimumNumber; i < SDL.numChannels; i++) {
         if (!SDL.channels[i].audio) {
           channel = i;
           break;
@@ -1375,6 +1406,7 @@ var LibrarySDL = {
       audio.play();
     }
     audio.volume = channelInfo.volume;
+    audio.paused = false;
     return channel;
   },
   Mix_PlayChannelTimed: 'Mix_PlayChannel', // XXX ignore Timing
@@ -1463,10 +1495,71 @@ var LibrarySDL = {
     return (SDL.music.audio && !SDL.music.audio.paused) ? 1 : 0;
   },
 
+  // http://www.libsdl.org/projects/SDL_mixer/docs/SDL_mixer_38.html#SEC38
+  // "Note: Does not check if the channel has been paused."
+  Mix_Playing: function(channel) {
+    if (channel === -1) {
+      var count = 0;
+      for (var i = 0; i < SDL.channels.length; i++) {
+        count += _Mix_Playing(i);
+      }
+      return count;
+    }
+    var info = SDL.channels[channel];
+    if (info && info.audio && !info.audio.paused) {
+      return 1;
+    }
+    return 0;
+  },
+  
+  Mix_Pause: function(channel) {
+    if (channel === -1) {
+      for (var i = 0; i<SDL.channels.length;i++) {
+        _Mix_Pause(i);
+      }
+      return;
+    }
+    var info = SDL.channels[channel];
+    if (info && info.audio) {
+      info.audio.pause();
+      info.audio.paused = true;
+    }
+  },
+  
+  // http://www.libsdl.org/projects/SDL_mixer/docs/SDL_mixer_39.html#SEC39
+  Mix_Paused: function(channel) {
+    if (channel === -1) {
+      var pausedCount = 0;
+      for (var i = 0; i<SDL.channels.length;i++) {
+        pausedCount += _Mix_Paused(i);
+      }
+      return pausedCount;
+    }
+    var info = SDL.channels[channel];
+    if (info && info.audio && info.audio.paused) {
+      return 1;
+    }
+    return 0;
+  },
+
   Mix_PausedMusic: function() {
     return (SDL.music.audio && SDL.music.audio.paused) ? 1 : 0;
   },
 
+  // http://www.libsdl.org/projects/SDL_mixer/docs/SDL_mixer_33.html#SEC33
+  Mix_Resume: function(channel) {
+    if (channel === -1) {
+      for (var i = 0; i<SDL.channels.length;i++) {
+        _Mix_Resume(i);
+      }
+      return;
+    }
+    var info = SDL.channels[channel];
+    if (info && info.audio) {
+      info.audio.play();
+    }
+  },
+
   // SDL TTF
 
   TTF_Init: function() { return 0 },
diff --git a/src/long.js b/src/long.js
index c3b0e605..6d0e873d 100644
--- a/src/long.js
+++ b/src/long.js
@@ -1530,13 +1530,6 @@ var i64Math = (function() { // Emscripten wrapper
 
   // Emscripten wrapper
   var Wrapper = {
-    add: function(xl, xh, yl, yh) {
-      var x = new goog.math.Long(xl, xh);
-      var y = new goog.math.Long(yl, yh);
-      var ret = x.add(y);
-      HEAP32[tempDoublePtr>>2] = ret.low_;
-      HEAP32[tempDoublePtr+4>>2] = ret.high_;
-    },
     subtract: function(xl, xh, yl, yh) {
       var x = new goog.math.Long(xl, xh);
       var y = new goog.math.Long(yl, yh);
@@ -1551,6 +1544,17 @@ var i64Math = (function() { // Emscripten wrapper
       HEAP32[tempDoublePtr>>2] = ret.low_;
       HEAP32[tempDoublePtr+4>>2] = ret.high_;
     },
+    abs: function(l, h) {
+      var x = new goog.math.Long(l, h);
+      var ret;
+      if (x.isNegative()) {
+        ret = x.negate();
+      } else {
+        ret = x;
+      }
+      HEAP32[tempDoublePtr>>2] = ret.low_;
+      HEAP32[tempDoublePtr+4>>2] = ret.high_;
+    },
     ensureTemps: function() {
       if (Wrapper.ensuredTemps) return;
       Wrapper.ensuredTemps = true;
diff --git a/src/modules.js b/src/modules.js
index afdbc21e..bda8a605 100644
--- a/src/modules.js
+++ b/src/modules.js
@@ -90,6 +90,7 @@ var Debugging = {
         lines[i] = ';';
         continue;
       }
+      if (line[0] == '!') skipLine = true;
       lines[i] = skipLine ? ';' : line;
     }
 
@@ -290,7 +291,7 @@ var Functions = {
       var sig = ASM_JS ? Functions.implementedFunctions[ident] || Functions.unimplementedFunctions[ident] || LibraryManager.library[ident.substr(1) + '__sig'] : 'x';
       assert(sig, ident);
       if (!tables[sig]) tables[sig] = emptyTable(sig); // TODO: make them compact
-      tables[sig][this.indexedFunctions[ident]] = ident;
+      tables[sig][this.indexedFunctions[ident]] = ident in DEAD_FUNCTIONS ? '0' : ident;
     }
     var generated = false;
     var wrapped = {};
@@ -314,7 +315,7 @@ var Functions = {
         }
         if (ASM_JS) {
           var curr = table[i];
-          if (curr && !Functions.implementedFunctions[curr]) {
+          if (curr && curr != '0' && !Functions.implementedFunctions[curr]) {
             // This is a library function, we can't just put it in the function table, need a wrapper
             if (!wrapped[curr]) {
               var args = '', arg_coercions = '', call = curr + '(', retPre = '', retPost = '';
@@ -373,7 +374,7 @@ var LibraryManager = {
   load: function() {
     if (this.library) return;
 
-    var libraries = ['library.js', 'library_browser.js', 'library_sdl.js', 'library_gl.js', 'library_glut.js', 'library_xlib.js', 'library_egl.js', 'library_gc.js', 'library_jansson.js'].concat(additionalLibraries);
+    var libraries = ['library.js', 'library_browser.js', 'library_sdl.js', 'library_gl.js', 'library_glut.js', 'library_xlib.js', 'library_egl.js', 'library_gc.js', 'library_jansson.js', 'library_openal.js'].concat(additionalLibraries);
     for (var i = 0; i < libraries.length; i++) {
       eval(processMacros(preprocess(read(libraries[i]))));
     }
diff --git a/src/parseTools.js b/src/parseTools.js
index 7f4f3a18..2664baed 100644
--- a/src/parseTools.js
+++ b/src/parseTools.js
@@ -172,6 +172,7 @@ function getBits(type) {
   }
   if (isStructType(type)) {
     var typeData = Types.types[type];
+    if (typeData === undefined) return 0;
     return typeData.flatSize*8;
   }
   return 0;
@@ -583,7 +584,7 @@ function cleanOutTokens(filterOut, tokens, indexes) {
   if (typeof indexes !== 'object') indexes = [indexes];
   for (var i = indexes.length-1; i >=0; i--) {
     var index = indexes[i];
-    while (tokens[index].text in filterOut) {
+    while (index < tokens.length && tokens[index].text in filterOut) {
       tokens.splice(index, 1);
     }
   }
@@ -1248,7 +1249,6 @@ function makeSetValueAsm(ptr, pos, value, type, noNeedFirst, ignore, align, noSa
   return makeSetValue(ptr, pos, value, type, noNeedFirst, ignore, align, noSafe, sep, forcedAlign, true);
 }
 
-var SEEK_OPTIMAL_ALIGN_MIN = 20;
 var UNROLL_LOOP_MAX = 8;
 
 function makeSetValues(ptr, pos, value, type, num, align) {
@@ -1268,7 +1268,7 @@ function makeSetValues(ptr, pos, value, type, num, align) {
   } else { // USE_TYPED_ARRAYS == 2
     // If we don't know how to handle this at compile-time, or handling it is best done in a large amount of code, call memset
     // TODO: optimize the case of numeric num but non-numeric value
-    if (!isNumber(num) || !isNumber(value) || (align < 4 && parseInt(num) >= SEEK_OPTIMAL_ALIGN_MIN)) {
+    if (!isNumber(num) || !isNumber(value) || (parseInt(num)/align >= UNROLL_LOOP_MAX)) {
       return '_memset(' + asmCoercion(getFastValue(ptr, '+', pos), 'i32') + ', ' + asmCoercion(value, 'i32') + ', ' + asmCoercion(num, 'i32') + ')';
     }
     num = parseInt(num);
@@ -1283,13 +1283,7 @@ function makeSetValues(ptr, pos, value, type, num, align) {
     [4, 2, 1].forEach(function(possibleAlign) {
       if (num == 0) return;
       if (align >= possibleAlign) {
-        if (num <= UNROLL_LOOP_MAX*possibleAlign || ASM_JS) { // XXX test asm performance
-          ret.push(unroll('i' + (possibleAlign*8), Math.floor(num/possibleAlign), possibleAlign, values[possibleAlign]));
-        } else {
-          ret.push('for (var $$dest = ' + getFastValue(ptr, '+', pos) + (possibleAlign > 1 ? '>>' + log2(possibleAlign) : '') + ', ' +
-                            '$$stop = $$dest + ' + Math.floor(num/possibleAlign) + '; $$dest < $$stop; $$dest++) {\n' +
-                   '  HEAP' + (possibleAlign*8) + '[$$dest] = ' + values[possibleAlign] + '\n}');
-        }
+        ret.push(unroll('i' + (possibleAlign*8), Math.floor(num/possibleAlign), possibleAlign, values[possibleAlign]));
         pos = getFastValue(pos, '+', Math.floor(num/possibleAlign)*possibleAlign);
         num %= possibleAlign;
       }
@@ -1326,7 +1320,7 @@ function makeCopyValues(dest, src, num, type, modifier, align, sep) {
             unroll(type, 1) + ' }';
   } else { // USE_TYPED_ARRAYS == 2
     // If we don't know how to handle this at compile-time, or handling it is best done in a large amount of code, call memset
-    if (!isNumber(num) || (align < 4 && parseInt(num) >= SEEK_OPTIMAL_ALIGN_MIN)) {
+    if (!isNumber(num) || (parseInt(num)/align >= UNROLL_LOOP_MAX)) {
       return '_memcpy(' + dest + ', ' + src + ', ' + num + ')';
     }
     num = parseInt(num);
@@ -1334,16 +1328,7 @@ function makeCopyValues(dest, src, num, type, modifier, align, sep) {
     [4, 2, 1].forEach(function(possibleAlign) {
       if (num == 0) return;
       if (align >= possibleAlign) {
-        // If we can unroll the loop, do so. Also do so if we must unroll it (we do not create real loops when inlined)
-        if (num <= UNROLL_LOOP_MAX*possibleAlign || sep == ',' || ASM_JS) { // XXX test asm performance
-          ret.push(unroll('i' + (possibleAlign*8), Math.floor(num/possibleAlign), possibleAlign));
-        } else {
-          assert(sep == ';');
-          ret.push('for (var $$src = ' + src + (possibleAlign > 1 ? '>>' + log2(possibleAlign) : '') + ', ' +
-                            '$$dest = ' + dest + (possibleAlign > 1 ? '>>' + log2(possibleAlign) : '') + ', ' +
-                            '$$stop = $$src + ' + Math.floor(num/possibleAlign) + '; $$src < $$stop; $$src++, $$dest++) {\n' +
-                   '  HEAP' + (possibleAlign*8) + '[$$dest] = HEAP' + (possibleAlign*8) + '[$$src]\n}');
-        }
+        ret.push(unroll('i' + (possibleAlign*8), Math.floor(num/possibleAlign), possibleAlign));
         src = getFastValue(src, '+', Math.floor(num/possibleAlign)*possibleAlign);
         dest = getFastValue(dest, '+', Math.floor(num/possibleAlign)*possibleAlign);
         num %= possibleAlign;
@@ -1739,13 +1724,14 @@ function makeLLVMStruct(values) {
   }
 }
 
-function makeStructuralReturn(values) {
+function makeStructuralReturn(values, inAsm) {
   if (USE_TYPED_ARRAYS == 2) {
-    var i = 0;
-    return 'return (' + values.slice(1).map(function(value) {
-      return ASM_JS ? 'asm.setTempRet' + (i++) + '(' + value + ')'
+    var i = -1;
+    return 'return ' + asmCoercion(values.slice(1).map(function(value) {
+      i++;
+      return ASM_JS ? (inAsm ? 'tempRet' + i + ' = ' + value : 'asm.setTempRet' + i + '(' + value + ')')
                     : 'tempRet' + (i++) + ' = ' + value;
-    }).concat([values[0]]).join(',') + ')';
+    }).concat([values[0]]).join(','), 'i32');
   } else {
     var i = 0;
     return 'return { ' + values.map(function(value) {
@@ -1986,6 +1972,10 @@ function processMathop(item) {
       return finish(['(i64Math' + (ASM_JS ? '_' : '.') + type + '(' + asmCoercion(low1, 'i32') + ',' + asmCoercion(high1, 'i32') + ',' + asmCoercion(low2, 'i32') + ',' + asmCoercion(high2, 'i32') +
                      (lastArg ? ',' + asmCoercion(+lastArg, 'i32') : '') + '),' + makeGetValue('tempDoublePtr', 0, 'i32') + ')', makeGetValue('tempDoublePtr', Runtime.getNativeTypeSize('i32'), 'i32')]);
     }
+    function i64PreciseLib(type) {
+      Types.preciseI64MathUsed = true;
+      return finish(['_i64' + type[0].toUpperCase() + type.substr(1) + '(' + low1 + ',' + high1 + ',' + low2 + ',' + high2 + ')', 'tempRet0']);
+    }
     switch (op) {
       // basic integer ops
       case 'or': {
@@ -2000,43 +1990,7 @@ function processMathop(item) {
       case 'shl':
       case 'ashr':
       case 'lshr': {
-        if (!isNumber(idents[1])) {
-          return '(Runtime' + (ASM_JS ? '_' : '.') + 'bitshift64(' + idents[0] + '[0], ' + idents[0] + '[1],' + Runtime['BITSHIFT64_' + op.toUpperCase()] + ',' + stripCorrections(idents[1]) + '[0]|0),' +
-            '[' + makeGetTempDouble(0, 'i32') + ',' + makeGetTempDouble(1, 'i32') + '])';
-        }
-        bits = parseInt(idents[1]);
-        var ander = Math.pow(2, bits)-1;
-        if (bits < 32) {
-          switch (op) {
-            case 'shl':
-              return '[' + idents[0] + '[0] << ' + idents[1] + ', ' +
-                       '('+idents[0] + '[1] << ' + idents[1] + ') | ((' + idents[0] + '[0]&(' + ander + '<<' + (32 - bits) + ')) >>> (32-' + idents[1] + '))]';
-            case 'ashr':
-              return '[((('+idents[0] + '[0] >>> ' + idents[1] + ') | ((' + idents[0] + '[1]&' + ander + ')<<' + (32 - bits) + ')) >> 0) >>> 0,' +
-                          '(' + idents[0] + '[1] >> ' + idents[1] + ') >>> 0]';
-            case 'lshr':
-              return '[(('+idents[0] + '[0] >>> ' + idents[1] + ') | ((' + idents[0] + '[1]&' + ander + ')<<' + (32 - bits) + ')) >>> 0,' +
-                          idents[0] + '[1] >>> ' + idents[1] + ']';
-          }
-        } else if (bits == 32) {
-          switch (op) {
-            case 'shl':
-              return '[0, ' + idents[0] + '[0]]';
-            case 'ashr':
-              return '[' + idents[0] + '[1], (' + idents[0] + '[1]|0) < 0 ? ' + ander + ' : 0]';
-            case 'lshr':
-              return '[' + idents[0] + '[1], 0]';
-          }
-        } else { // bits > 32
-          switch (op) {
-            case 'shl':
-              return '[0, ' + idents[0] + '[0] << ' + (bits - 32) + ']';
-            case 'ashr':
-              return '[(' + idents[0] + '[1] >> ' + (bits - 32) + ') >>> 0, (' + idents[0] + '[1]|0) < 0 ? ' + ander + ' : 0]';
-            case 'lshr':
-              return '[' + idents[0] + '[1] >>> ' + (bits - 32) + ', 0]';
-          }
-        }
+				throw 'shifts should have been legalized!';
       }
       case 'uitofp': case 'sitofp': return RuntimeGenerator.makeBigInt(low1, high1, op[0] == 'u');
       case 'fptoui': case 'fptosi': return finish(splitI64(idents[0], true));
@@ -2074,7 +2028,7 @@ function processMathop(item) {
       // Dangerous, rounded operations. TODO: Fully emulate
       case 'add': {
         if (PRECISE_I64_MATH) {
-          return i64PreciseOp('add');
+          return i64PreciseLib('add');
         } else {
           warnI64_1();
           return finish(splitI64(mergeI64(idents[0]) + '+' + mergeI64(idents[1]), true));
@@ -2365,3 +2319,7 @@ function getImplementationType(varInfo) {
   return varInfo.type;
 }
 
+function charCode(char) {
+  return char.charCodeAt(0);
+}
+
diff --git a/src/preamble.js b/src/preamble.js
index 9bc68d8f..6f3a969c 100644
--- a/src/preamble.js
+++ b/src/preamble.js
@@ -503,6 +503,7 @@ function allocate(slab, types, allocator, ptr) {
 Module['allocate'] = allocate;
 
 function Pointer_stringify(ptr, /* optional */ length) {
+#if UTF_STRING_SUPPORT
   var utf8 = new Runtime.UTF8Processor();
   var nullTerminated = typeof(length) == "undefined";
   var ret = "";
@@ -519,18 +520,16 @@ function Pointer_stringify(ptr, /* optional */ length) {
     if (!nullTerminated && i == length) break;
   }
   return ret;
+#else
+#if USE_TYPED_ARRAYS == 2
+  return String.fromCharCode.apply(String, HEAPU8.subarray(ptr, ptr + (length || _strlen(ptr))));
+#else
+  throw 'unsupported combination';
+#endif
+#endif
 }
 Module['Pointer_stringify'] = Pointer_stringify;
 
-function Array_stringify(array) {
-  var ret = "";
-  for (var i = 0; i < array.length; i++) {
-    ret += String.fromCharCode(array[i]);
-  }
-  return ret;
-}
-Module['Array_stringify'] = Array_stringify;
-
 // Memory management
 
 var PAGE_SIZE = 4096;
@@ -848,5 +847,22 @@ Module['removeRunDependency'] = removeRunDependency;
 Module["preloadedImages"] = {}; // maps url to image data
 Module["preloadedAudios"] = {}; // maps url to audio data
 
+#if PGO
+var PGOMonitor = {
+  called: {},
+  dump: function() {
+    var dead = [];
+    for (var i = 0; i < this.allGenerated.length; i++) {
+      var func = this.allGenerated[i];
+      if (!this.called[func]) dead.push(func);
+    }
+    Module.print('-s DEAD_FUNCTIONS=\'' + JSON.stringify(dead) + '\'\n');
+  }
+};
+__ATEXIT__.push({ func: function() { PGOMonitor.dump() } });
+if (!Module.preRun) Module.preRun = [];
+Module.preRun.push(function() { addRunDependency('pgo') });
+#endif
+
 // === Body ===
 
diff --git a/src/relooper/Relooper.cpp b/src/relooper/Relooper.cpp
index 1a7acc15..ae393de3 100644
--- a/src/relooper/Relooper.cpp
+++ b/src/relooper/Relooper.cpp
@@ -347,17 +347,25 @@ struct RelooperRecursor {
   RelooperRecursor(Relooper *ParentInit) : Parent(ParentInit) {}
 };
 
+typedef std::list<Block*> BlockList;
+
 void Relooper::Calculate(Block *Entry) {
   // Scan and optimize the input
   struct PreOptimizer : public RelooperRecursor {
     PreOptimizer(Relooper *Parent) : RelooperRecursor(Parent) {}
     BlockSet Live;
 
-    void FindLive(Block *Curr) {
-      if (Live.find(Curr) != Live.end()) return;
-      Live.insert(Curr);
-      for (BlockBranchMap::iterator iter = Curr->BranchesOut.begin(); iter != Curr->BranchesOut.end(); iter++) {
-        FindLive(iter->first);
+    void FindLive(Block *Root) {
+      BlockList ToInvestigate;
+      ToInvestigate.push_back(Root);
+      while (ToInvestigate.size() > 0) {
+        Block *Curr = ToInvestigate.front();
+        ToInvestigate.pop_front();
+        if (Live.find(Curr) != Live.end()) continue;
+        Live.insert(Curr);
+        for (BlockBranchMap::iterator iter = Curr->BranchesOut.begin(); iter != Curr->BranchesOut.end(); iter++) {
+          ToInvestigate.push_back(iter->first);
+        }
       }
     }
 
@@ -529,7 +537,6 @@ void Relooper::Calculate(Block *Entry) {
     // ignore directly reaching the entry itself by another entry.
     void FindIndependentGroups(BlockSet &Blocks, BlockSet &Entries, BlockBlockSetMap& IndependentGroups) {
       typedef std::map<Block*, Block*> BlockBlockMap;
-      typedef std::list<Block*> BlockList;
 
       struct HelperClass {
         BlockBlockSetMap& IndependentGroups;
@@ -872,33 +879,38 @@ void Relooper::Calculate(Block *Entry) {
     // A flow operation is trivially unneeded if the shape we naturally get to by normal code
     // execution is the same as the flow forces us to.
     void RemoveUnneededFlows(Shape *Root, Shape *Natural=NULL) {
-      SHAPE_SWITCH(Root, {
-        // If there is a next block, we already know at Simple creation time to make direct branches,
-        // and we can do nothing more. If there is no next however, then Natural is where we will
-        // go to by doing nothing, so we can potentially optimize some branches to direct.
-        if (Simple->Next) {
-          RemoveUnneededFlows(Simple->Next, Natural);
-        } else {
-          for (BlockBranchMap::iterator iter = Simple->Inner->ProcessedBranchesOut.begin(); iter != Simple->Inner->ProcessedBranchesOut.end(); iter++) {
-            Block *Target = iter->first;
-            Branch *Details = iter->second;
-            if (Details->Type != Branch::Direct && Target->Parent == Natural) {
-              Details->Type = Branch::Direct;
-              if (MultipleShape *Multiple = Shape::IsMultiple(Details->Ancestor)) {
-                Multiple->NeedLoop--;
+      Shape *Next = Root;
+      while (Next) {
+        Root = Next;
+        Next = NULL;
+        SHAPE_SWITCH(Root, {
+          // If there is a next block, we already know at Simple creation time to make direct branches,
+          // and we can do nothing more. If there is no next however, then Natural is where we will
+          // go to by doing nothing, so we can potentially optimize some branches to direct.
+          if (Simple->Next) {
+            Next = Simple->Next;
+          } else {
+            for (BlockBranchMap::iterator iter = Simple->Inner->ProcessedBranchesOut.begin(); iter != Simple->Inner->ProcessedBranchesOut.end(); iter++) {
+              Block *Target = iter->first;
+              Branch *Details = iter->second;
+              if (Details->Type != Branch::Direct && Target->Parent == Natural) {
+                Details->Type = Branch::Direct;
+                if (MultipleShape *Multiple = Shape::IsMultiple(Details->Ancestor)) {
+                  Multiple->NeedLoop--;
+                }
               }
             }
           }
-        }
-      }, {
-        for (BlockShapeMap::iterator iter = Multiple->InnerMap.begin(); iter != Multiple->InnerMap.end(); iter++) {
-          RemoveUnneededFlows(iter->second, Multiple->Next);
-        }
-        RemoveUnneededFlows(Multiple->Next, Natural);
-      }, {
-        RemoveUnneededFlows(Loop->Inner, Loop->Inner);
-        RemoveUnneededFlows(Loop->Next, Natural);
-      });
+        }, {
+          for (BlockShapeMap::iterator iter = Multiple->InnerMap.begin(); iter != Multiple->InnerMap.end(); iter++) {
+            RemoveUnneededFlows(iter->second, Multiple->Next);
+          }
+          Next = Multiple->Next;
+        }, {
+          RemoveUnneededFlows(Loop->Inner, Loop->Inner);
+          Next = Loop->Next;
+        });
+      }
     }
 
     // After we know which loops exist, we can calculate which need to be labeled
diff --git a/src/relooper/fuzzer.py b/src/relooper/fuzzer.py
index 887eab3b..96929028 100644
--- a/src/relooper/fuzzer.py
+++ b/src/relooper/fuzzer.py
@@ -98,14 +98,14 @@ int main() {
   open('fuzz.slow.js', 'w').write(slow)
   open('fuzz.cpp', 'w').write(fast)
   print '_'
-  slow_out = subprocess.Popen(['/home/alon/Dev/mozilla-central/js/src/fast/js', '-m', '-n', 'fuzz.slow.js'], stdout=subprocess.PIPE).communicate()[0]
+  slow_out = subprocess.Popen(['/home/alon/Dev/odinmonkey/js/src/fast/js', '-m', '-n', 'fuzz.slow.js'], stdout=subprocess.PIPE).communicate()[0]
 
   print '.'
   subprocess.call(['g++', 'fuzz.cpp', 'Relooper.o', '-o', 'fuzz', '-g'])
   print '*'
   subprocess.call(['./fuzz'], stdout=open('fuzz.fast.js', 'w'))
   print '-'
-  fast_out = subprocess.Popen(['/home/alon/Dev/mozilla-central/js/src/fast/js', '-m', '-n', 'fuzz.fast.js'], stdout=subprocess.PIPE).communicate()[0]
+  fast_out = subprocess.Popen(['/home/alon/Dev/odinmonkey/js/src/fast/js', '-m', '-n', 'fuzz.fast.js'], stdout=subprocess.PIPE).communicate()[0]
   print
 
   if slow_out != fast_out:
diff --git a/src/relooper/test.txt b/src/relooper/test.txt
index b7c8794d..12d0ef39 100644
--- a/src/relooper/test.txt
+++ b/src/relooper/test.txt
@@ -54,7 +54,7 @@ while(1) {
   // code 2
   if (!($2)) {
     var $x_1 = $x_0; 
-    label = 19;
+    label = 18;
     break;
   }
   // code 3
@@ -64,7 +64,7 @@ while(1) {
     var $i_0 = $7;var $x_0 = $5; 
   }
 }
-if (label == 19) {
+if (label == 18) {
   // code 7
 }
 // code 4
diff --git a/src/runtime.js b/src/runtime.js
index 7f97da35..2a26db28 100644
--- a/src/runtime.js
+++ b/src/runtime.js
@@ -27,7 +27,7 @@ var RuntimeGenerator = {
     //                                                               The stack is always QUANTUM SIZE aligned, so we may not need to force alignment here
     var ret = RuntimeGenerator.alloc(size, 'STACK', false, sep, USE_TYPED_ARRAYS != 2 || (isNumber(size) && parseInt(size) % {{{ QUANTUM_SIZE }}} == 0));
     if (ASSERTIONS) {
-      ret += sep + 'assert(STACKTOP|0 < STACK_MAX|0)';
+      ret += sep + 'assert(' + asmCoercion('(STACKTOP|0) < (STACK_MAX|0)', 'i32') + ')';
     }
     return ret;
   },
@@ -38,12 +38,12 @@ var RuntimeGenerator = {
     if (initial > 0) ret += '; STACKTOP = (STACKTOP + ' + initial + ')|0';
     if (USE_TYPED_ARRAYS == 2) {
       assert(initial % QUANTUM_SIZE == 0);
-      if (ASSERTIONS) {
-        ret += '; assert(STACKTOP|0 % {{{ QUANTUM_SIZE }}} == 0)';
+      if (ASSERTIONS && QUANTUM_SIZE == 4) {
+        ret += '; assert(' + asmCoercion('!(STACKTOP&3)', 'i32') + ')';
       }
     }
     if (ASSERTIONS) {
-      ret += '; assert(STACKTOP < STACK_MAX)';
+      ret += '; assert(' + asmCoercion('(STACKTOP|0) < (STACK_MAX|0)', 'i32') + ')';
     }
     if (false) {
       ret += '; _memset(' + asmCoercion('__stackBase__', 'i32') + ', 0, ' + initial + ')';
@@ -87,7 +87,7 @@ var RuntimeGenerator = {
 };
 
 function unInline(name_, params) {
-  var src = '(function ' + name_ + '(' + params + ') { var ret = ' + RuntimeGenerator[name_].apply(null, params) + '; return ret; })';
+  var src = '(function(' + params + ') { var ret = ' + RuntimeGenerator[name_].apply(null, params) + '; return ret; })';
   var ret = eval(src);
   return ret;
 }
@@ -122,57 +122,6 @@ var Runtime = {
   INT_TYPES: set('i1', 'i8', 'i16', 'i32', 'i64'),
   FLOAT_TYPES: set('float', 'double'),
 
-  // Mirrors processMathop's treatment of constants (which we optimize directly)
-  BITSHIFT64_SHL: 0,
-  BITSHIFT64_ASHR: 1,
-  BITSHIFT64_LSHR: 2,
-  bitshift64: function(low, high, op, bits) {
-    var ret;
-    var ander = Math.pow(2, bits)-1;
-    if (bits < 32) {
-      switch (op) {
-        case Runtime.BITSHIFT64_SHL:
-          ret = [low << bits, (high << bits) | ((low&(ander << (32 - bits))) >>> (32 - bits))];
-          break;
-        case Runtime.BITSHIFT64_ASHR:
-          ret = [(((low >>> bits ) | ((high&ander) << (32 - bits))) >> 0) >>> 0, (high >> bits) >>> 0];
-          break;
-        case Runtime.BITSHIFT64_LSHR:
-          ret = [((low >>> bits) | ((high&ander) << (32 - bits))) >>> 0, high >>> bits];
-          break;
-      }
-    } else if (bits == 32) {
-      switch (op) {
-        case Runtime.BITSHIFT64_SHL:
-          ret = [0, low];
-          break;
-        case Runtime.BITSHIFT64_ASHR:
-          ret = [high, (high|0) < 0 ? ander : 0];
-          break;
-        case Runtime.BITSHIFT64_LSHR:
-          ret = [high, 0];
-          break;
-      }
-    } else { // bits > 32
-      switch (op) {
-        case Runtime.BITSHIFT64_SHL:
-          ret = [0, low << (bits - 32)];
-          break;
-        case Runtime.BITSHIFT64_ASHR:
-          ret = [(high >> (bits - 32)) >>> 0, (high|0) < 0 ? ander : 0];
-          break;
-        case Runtime.BITSHIFT64_LSHR:
-          ret = [high >>>  (bits - 32) , 0];
-          break;
-      }
-    }
-#if ASSERTIONS
-    assert(ret);
-#endif
-    HEAP32[tempDoublePtr>>2] = ret[0]; // cannot use utility functions since we are in runtime itself
-    HEAP32[tempDoublePtr+4>>2] = ret[1];
-  },
-
   // Imprecise bitops utilities
   or64: function(x, y) {
     var l = (x | 0) | (y | 0);
@@ -365,6 +314,11 @@ var Runtime = {
     return ret;
   },
 
+  removeFunction: function(index) {
+    var table = FUNCTION_TABLE; // TODO: support asm
+    table[index] = null;
+  },
+
   warnOnce: function(text) {
     if (!Runtime.warnOnce.shown) Runtime.warnOnce.shown = {};
     if (!Runtime.warnOnce.shown[text]) {
diff --git a/src/settings.js b/src/settings.js
index 1bfcf92a..36f53c3c 100644
--- a/src/settings.js
+++ b/src/settings.js
@@ -15,7 +15,7 @@ var QUANTUM_SIZE = 4; // This is the size of an individual field in a structure.
                       // the normal value of 4 means all fields take 4 memory addresses,
                       // as per the norm on a 32-bit machine.
                       //
-                      // 1 is somewhat faster than 4, but dangerous.
+                      // Changing this from the default of 4 is deprecated.
 
 var CORRECT_SIGNS = 1; // Whether we make sure to convert unsigned values to signed values.
                        // Decreases performance with additional runtime checks. Might not be
@@ -63,7 +63,7 @@ var RELOOPER = 'relooper.js'; // Loads the relooper from this path relative to c
 var USE_TYPED_ARRAYS = 2; // Use typed arrays for the heap. See https://github.com/kripken/emscripten/wiki/Code-Generation-Modes/
                           // 0 means no typed arrays are used.
                           // 1 has two heaps, IHEAP (int32) and FHEAP (double),
-                          // and addresses there are a match for normal addresses.
+                          // and addresses there are a match for normal addresses. This is deprecated.
                           // 2 is a single heap, accessible through views as int8, int32, etc. This is
                           //   the recommended mode both for performance and for compatibility.
 var USE_FHEAP = 1; // Relevant in USE_TYPED_ARRAYS == 1. If this is disabled, only IHEAP will be used, and FHEAP
@@ -171,6 +171,8 @@ var GL_UNSAFE_OPTS = 1; // Enables some potentially-unsafe optimizations in GL e
 var FULL_ES2 = 0; // Forces support for all GLES2 features, not just the WebGL-friendly subset.
 var FORCE_GL_EMULATION = 0; // Forces inclusion of full GL emulation code.
 
+var UTF_STRING_SUPPORT = 1; // Perform utf-8 conversion between C and JS strings (adds overhead in such conversions)
+
 var DISABLE_EXCEPTION_CATCHING = 0; // Disables generating code to actually catch exceptions. If the code you
                                     // are compiling does not actually rely on catching exceptions (but the
                                     // compiler generates code for it, maybe because of stdlibc++ stuff),
@@ -230,13 +232,14 @@ var EXPORTED_FUNCTIONS = ['_main']; // Functions that are explicitly exported. T
                                     // the generated code even after running closure compiler (on "Module").
                                     // Note the necessary prefix of "_".
 
-var DEFAULT_LIBRARY_FUNCS_TO_INCLUDE = ['memcpy', 'memset', 'malloc', 'free', '$Browser']; // JS library functions (C functions implemented in JS)
-                                                                                           // that we include by default. If you want to make sure
-                                                                                           // something is included by the JS compiler, add it here.
-                                                                                           // For example, if you do not use some emscripten_*
-                                                                                           // C API call from C, but you want to call it from JS,
-                                                                                           // add it here (and in EXPORTED FUNCTIONS with prefix
-                                                                                           // "_", for closure).
+// JS library functions (C functions implemented in JS)
+// that we include by default. If you want to make sure
+// something is included by the JS compiler, add it here.
+// For example, if you do not use some emscripten_*
+// C API call from C, but you want to call it from JS,
+// add it here (and in EXPORTED FUNCTIONS with prefix
+// "_", for closure).
+var DEFAULT_LIBRARY_FUNCS_TO_INCLUDE = ['memcpy', 'memset', 'malloc', 'free', 'strlen', '$Browser'];
 
 var LIBRARY_DEPS_TO_AUTOEXPORT = ['memcpy']; // This list is also used to determine
                                              // auto-exporting of library dependencies (i.e., functions that
@@ -332,6 +335,14 @@ var ASM_JS = 0; // If 1, generate code in asm.js format. XXX This is highly expe
                 // and will not work on most codebases yet. It is NOT recommended that you
                 // try this yet.
 
+var PGO = 0; // Enables profile-guided optimization in the form of runtime checks for
+             // which functions are actually called. Emits a list during shutdown that you
+             // can pass to DEAD_FUNCTIONS (you can also emit the list manually by
+             // calling PGOMonitor.dump());
+var DEAD_FUNCTIONS = []; // A list of functions that no code will be emitted for, and
+                         // a runtime abort will happen if they are called
+                         // TODO: options to lazily load such functions
+
 var EXPLICIT_ZEXT = 0; // If 1, generate an explicit conversion of zext i1 to i32, using ?:
 
 var NECESSARY_BLOCKADDRS = []; // List of (function, block) for all block addresses that are taken.
diff --git a/src/shell.html b/src/shell.html
index 4f39b26a..8743d403 100644
--- a/src/shell.html
+++ b/src/shell.html
@@ -19,7 +19,14 @@
     </div>
     <canvas class="emscripten" id="canvas" oncontextmenu="event.preventDefault()"></canvas>
     <hr/>
-    <div class="emscripten"><input type="button" value="fullscreen" onclick="Module.requestFullScreen()"></div>
+    <div class="emscripten">
+      <input type="checkbox" id="resize">Resize canvas
+      <input type="checkbox" id="pointerLock" checked>Lock/hide mouse pointer
+      &nbsp;&nbsp;&nbsp;
+      <input type="button" value="Fullscreen" onclick="Module.requestFullScreen(document.getElementById('pointerLock').checked, 
+                                                                                document.getElementById('resize').checked)">
+    </div>
+    
     <hr/>
     <textarea class="emscripten" id="output" rows="8"></textarea>
     <hr>
diff --git a/src/utility.js b/src/utility.js
index 8db37c61..19444675 100644
--- a/src/utility.js
+++ b/src/utility.js
@@ -267,6 +267,15 @@ function set() {
 }
 var unset = keys;
 
+function numberedSet() {
+  var args = typeof arguments[0] === 'object' ? arguments[0] : arguments;
+  var ret = {};
+  for (var i = 0; i < args.length; i++) {
+    ret[args[i]] = i;
+  }
+  return ret;
+}
+
 function setSub(x, y) {
   var ret = set(keys(x));
   for (yy in y) {