Merge branch 'incoming'

8 files changed, 69 insertions, 51 deletions

diff --git a/src/fastLong.js b/src/fastLong.js
index d1ce5d39..4f6efd9f 100644
--- a/src/fastLong.js
+++ b/src/fastLong.js
@@ -5,12 +5,12 @@ function ___muldsi3($a, $b) {
   var $1 = 0, $2 = 0, $3 = 0, $6 = 0, $8 = 0, $11 = 0, $12 = 0;
   $1 = $a & 65535;
   $2 = $b & 65535;
-  $3 = Math.imul($2, $1);
+  $3 = Math.imul($2, $1) | 0;
   $6 = $a >>> 16;
-  $8 = ($3 >>> 16) + Math.imul($2, $6) | 0;
+  $8 = ($3 >>> 16) + (Math.imul($2, $6) | 0) | 0;
   $11 = $b >>> 16;
-  $12 = Math.imul($11, $1);
-  return (tempRet0 = (($8 >>> 16) + Math.imul($11, $6) | 0) + ((($8 & 65535) + $12 | 0) >>> 16) | 0, 0 | ($8 + $12 << 16 | $3 & 65535)) | 0;
+  $12 = Math.imul($11, $1) | 0;
+  return (tempRet0 = (($8 >>> 16) + (Math.imul($11, $6) | 0) | 0) + ((($8 & 65535) + $12 | 0) >>> 16) | 0, 0 | ($8 + $12 << 16 | $3 & 65535)) | 0;
 }
 function ___divdi3($a$0, $a$1, $b$0, $b$1) {
   $a$0 = $a$0 | 0;
@@ -47,7 +47,7 @@ function ___remdi3($a$0, $a$1, $b$0, $b$1) {
   $4$0 = _i64Subtract($1$0 ^ $a$0, $1$1 ^ $a$1, $1$0, $1$1) | 0;
   $4$1 = tempRet0;
   $6$0 = _i64Subtract($2$0 ^ $b$0, $2$1 ^ $b$1, $2$0, $2$1) | 0;
-  ___udivmoddi4($4$0, $4$1, $6$0, tempRet0, $rem);
+  ___udivmoddi4($4$0, $4$1, $6$0, tempRet0, $rem) | 0;
   $10$0 = _i64Subtract(HEAP32[$rem >> 2] ^ $1$0, HEAP32[$rem + 4 >> 2] ^ $1$1, $1$0, $1$1) | 0;
   $10$1 = tempRet0;
   STACKTOP = __stackBase__;
@@ -63,8 +63,8 @@ function ___muldi3($a$0, $a$1, $b$0, $b$1) {
   $y_sroa_0_0_extract_trunc = $b$0;
   $1$0 = ___muldsi3($x_sroa_0_0_extract_trunc, $y_sroa_0_0_extract_trunc) | 0;
   $1$1 = tempRet0;
-  $2 = Math.imul($a$1, $y_sroa_0_0_extract_trunc);
-  return (tempRet0 = (Math.imul($b$1, $x_sroa_0_0_extract_trunc) + $2 | 0) + $1$1 | $1$1 & 0, 0 | $1$0 & -1) | 0;
+  $2 = Math.imul($a$1, $y_sroa_0_0_extract_trunc) | 0;
+  return (tempRet0 = ((Math.imul($b$1, $x_sroa_0_0_extract_trunc) | 0) + $2 | 0) + $1$1 | $1$1 & 0, 0 | $1$0 & -1) | 0;
 }
 function ___udivdi3($a$0, $a$1, $b$0, $b$1) {
   $a$0 = $a$0 | 0;
@@ -84,7 +84,7 @@ function ___uremdi3($a$0, $a$1, $b$0, $b$1) {
   __stackBase__ = STACKTOP;
   STACKTOP = STACKTOP + 8 | 0;
   $rem = __stackBase__ | 0;
-  ___udivmoddi4($a$0, $a$1, $b$0, $b$1, $rem);
+  ___udivmoddi4($a$0, $a$1, $b$0, $b$1, $rem) | 0;
   STACKTOP = __stackBase__;
   return (tempRet0 = HEAP32[$rem + 4 >> 2] | 0, HEAP32[$rem >> 2] | 0) | 0;
 }
@@ -258,7 +258,7 @@ function ___udivmoddi4($a$0, $a$1, $b$0, $b$1, $rem) {
       $149 = $carry_0203 | $q_sroa_0_1199 << 1;
       $r_sroa_0_0_insert_insert42$0 = 0 | ($r_sroa_0_1201 << 1 | $q_sroa_1_1198 >>> 31);
       $r_sroa_0_0_insert_insert42$1 = $r_sroa_0_1201 >>> 31 | $r_sroa_1_1200 << 1 | 0;
-      _i64Subtract($137$0, $137$1, $r_sroa_0_0_insert_insert42$0, $r_sroa_0_0_insert_insert42$1);
+      _i64Subtract($137$0, $137$1, $r_sroa_0_0_insert_insert42$0, $r_sroa_0_0_insert_insert42$1) | 0;
       $150$1 = tempRet0;
       $151$0 = $150$1 >> 31 | (($150$1 | 0) < 0 ? -1 : 0) << 1;
       $152 = $151$0 & 1;
diff --git a/src/jsifier.js b/src/jsifier.js
index 156fd65d..e9bc9a79 100644
--- a/src/jsifier.js
+++ b/src/jsifier.js
@@ -328,7 +328,7 @@ function JSify(data, functionsOnly, givenFunctions) {
       var js = (index !== null ? '' : item.ident + '=') + constant;
       if (js) js += ';';
 
-      if (!ASM_JS && (EXPORT_ALL || (item.ident in EXPORTED_GLOBALS))) {
+      if (!ASM_JS && NAMED_GLOBALS && (EXPORT_ALL || (item.ident in EXPORTED_GLOBALS))) {
         js += '\nModule["' + item.ident + '"] = ' + item.ident + ';';
       }
       if (BUILD_AS_SHARED_LIB == 2 && !item.private_) {
@@ -1185,8 +1185,13 @@ function JSify(data, functionsOnly, givenFunctions) {
     if (disabled) {
       ret = call_ + ';';
     } else if (ASM_JS) {
+      if (item.type != 'void') call_ = asmCoercion(call_, item.type); // ensure coercion to ffi in comma operator
       call_ = call_.replace('; return', ''); // we auto-add returns when aborting, but do not need them here
-      ret = '(__THREW__ = 0,' +  call_ + ');';
+      if (item.type == 'void') {
+        ret = '__THREW__ = 0;' +  call_ + ';';
+      } else {
+        ret = '(__THREW__ = 0,' +  call_ + ');';
+      }
     } else {
       ret = '(function() { try { __THREW__ = 0; return '
           + call_ + ' '
@@ -1310,8 +1315,10 @@ function JSify(data, functionsOnly, givenFunctions) {
     assert(TARGET_LE32);
     var ident = item.value.ident;
     var move = Runtime.STACK_ALIGN;
-    return '(tempInt=' + makeGetValue(ident, 4, '*') + ',' +
-                         makeSetValue(ident, 4, 'tempInt + ' + move, '*') + ',' +
+    
+    // store current list offset in tempInt, advance list offset by STACK_ALIGN, return list entry stored at tempInt
+    return '(tempInt=' + makeGetValue(ident, Runtime.QUANTUM_SIZE, '*') + ',' +
+                         makeSetValue(ident, Runtime.QUANTUM_SIZE, 'tempInt + ' + move, '*') + ',' +
                          makeGetValue(makeGetValue(ident, 0, '*'), 'tempInt', item.type) + ')';
   });
 
diff --git a/src/library.js b/src/library.js
index f958a436..01a67804 100644
--- a/src/library.js
+++ b/src/library.js
@@ -4399,7 +4399,7 @@ LibraryManager.library = {
         {{{ makeSetValueAsm('dest', 0, makeGetValueAsm('src', 0, 'i8'), 'i8') }}};
       }
     } else {
-      _memcpy(dest, src, num);
+      _memcpy(dest, src, num) | 0;
     }
   },
   llvm_memmove_i32: 'memmove',
@@ -4657,6 +4657,14 @@ LibraryManager.library = {
     return 0;
   },
 
+  strnlen: function(ptr, num) {
+    for (var i = 0; i < num; i++) {
+      if ({{{ makeGetValue('ptr', 0, 'i8') }}} == 0) return i;
+      ptr++;
+    }
+    return num;
+  },
+
   strstr: function(ptr1, ptr2) {
     var check = 0, start;
     do {
@@ -4991,22 +4999,19 @@ LibraryManager.library = {
     return makeSetValue(ptr, 0, 'varrp', 'void*');
 #endif
 #if TARGET_LE32
-    // 4-word structure: start, current offset
-    return makeSetValue(ptr, 0, 'varrp', 'void*') + ';' + makeSetValue(ptr, 4, 0, 'void*');
+    // 2-word structure: struct { void* start; void* currentOffset; }
+    return makeSetValue(ptr, 0, 'varrp', 'void*') + ';' + makeSetValue(ptr, Runtime.QUANTUM_SIZE, 0, 'void*');
 #endif
   },
 
   llvm_va_end: function() {},
 
   llvm_va_copy: function(ppdest, ppsrc) {
+  	// copy the list start
     {{{ makeCopyValues('ppdest', 'ppsrc', Runtime.QUANTUM_SIZE, 'null', null, 1) }}};
-    /* Alternate implementation that copies the actual DATA; it assumes the va_list is prefixed by its size
-    var psrc = IHEAP[ppsrc]-1;
-    var num = IHEAP[psrc]; // right before the data, is the number of (flattened) values
-    var pdest = _malloc(num+1);
-    _memcpy(pdest, psrc, num+1);
-    IHEAP[ppdest] = pdest+1;
-    */
+    
+    // copy the list's current offset (will be advanced with each call to va_arg)
+    {{{ makeCopyValues('(ppdest+'+Runtime.QUANTUM_SIZE+')', '(ppsrc+'+Runtime.QUANTUM_SIZE+')', Runtime.QUANTUM_SIZE, 'null', null, 1) }}};
   },
 
   llvm_bswap_i16: function(x) {
diff --git a/src/library_browser.js b/src/library_browser.js
index 9800fedf..925b64e2 100644
--- a/src/library_browser.js
+++ b/src/library_browser.js
@@ -47,8 +47,11 @@ mergeInto(LibraryManager.library, {
     workers: [],
 
     init: function() {
-      if (Browser.initted) return;
+      if (!Module["preloadPlugins"]) Module["preloadPlugins"] = []; // needs to exist even in workers
+
+      if (Browser.initted || ENVIRONMENT_IS_WORKER) return;
       Browser.initted = true;
+
       try {
         new Blob();
         Browser.hasBlobConstructor = true;
@@ -79,8 +82,6 @@ mergeInto(LibraryManager.library, {
         }[name.substr(name.lastIndexOf('.')+1)];
       }
 
-      if (!Module["preloadPlugins"]) Module["preloadPlugins"] = [];
-
       var imagePlugin = {};
       imagePlugin['canHandle'] = function(name) {
         return !Module.noImageDecoding && /\.(jpg|jpeg|png|bmp)$/.exec(name);
diff --git a/src/library_sdl.js b/src/library_sdl.js
index a131f424..356c9746 100644
--- a/src/library_sdl.js
+++ b/src/library_sdl.js
@@ -242,7 +242,7 @@ var LibrarySDL = {
     },
 
     translateColorToCSSRGBA: function(rgba) {
-      return 'rgba(' + ((rgba >> 24)&255) + ',' + ((rgba >> 16)&255) + ',' + ((rgba >> 8)&255) + ',' + ((rgba&255)/255) + ')';
+      return 'rgba(' + (rgba&0xff) + ',' + (rgba>>8 & 0xff) + ',' + (rgba>>16 & 0xff) + ',' + (rgba>>>24)/0xff + ')';
     },
 
     translateRGBAToCSSRGBA: function(r, g, b, a) {
@@ -250,7 +250,7 @@ var LibrarySDL = {
     },
 
     translateRGBAToColor: function(r, g, b, a) {
-      return (r << 24) + (g << 16) + (b << 8) + a;
+      return r | g << 8 | b << 16 | a << 24;
     },
 
     makeSurface: function(width, height, flags, usePageCanvas, source, rmask, gmask, bmask, amask) {
@@ -819,7 +819,6 @@ var LibrarySDL = {
     if (surfData.isFlagSet(0x00200000 /* SDL_HWPALETTE */)) {
       SDL.copyIndexedColorData(surfData);
     } else if (!surfData.colors) {
-      var num = surfData.image.data.length;
       var data = surfData.image.data;
       var buffer = surfData.buffer;
 #if USE_TYPED_ARRAYS == 2
@@ -827,17 +826,14 @@ var LibrarySDL = {
       var src = buffer >> 2;
       var dst = 0;
       var isScreen = surf == SDL.screen;
+      var data32 = new Uint32Array(data.buffer);
+      var num = data32.length;
       while (dst < num) {
-        // TODO: access underlying data buffer and write in 32-bit chunks or more
-        var val = HEAP32[src]; // This is optimized. Instead, we could do {{{ makeGetValue('buffer', 'dst', 'i32') }}};
-        data[dst  ] = val & 0xff;
-        data[dst+1] = (val >> 8) & 0xff;
-        data[dst+2] = (val >> 16) & 0xff;
-        data[dst+3] = isScreen ? 0xff : ((val >> 24) & 0xff);
-        src++;
-        dst += 4;
+        // HEAP32[src++] is an optimization. Instead, we could do {{{ makeGetValue('buffer', 'dst', 'i32') }}};
+        data32[dst++] = HEAP32[src++] | (isScreen ? 0xff000000 : 0);
       }
 #else
+      var num = surfData.image.data.length;
       for (var i = 0; i < num; i++) {
         // We may need to correct signs here. Potentially you can hardcode a write of 255 to alpha, say, and
         // the compiler may decide to write -1 in the llvm bitcode...
@@ -965,6 +961,13 @@ var LibrarySDL = {
     return SDL.makeSurface(width, height, flags, false, 'CreateRGBSurface', rmask, gmask, bmask, amask);
   },
 
+  SDL_CreateRGBSurfaceFrom: function(pixels, width, height, depth, pitch, rmask, gmask, bmask, amask) {
+    // TODO: Actually fill pixel data to created surface.
+    // TODO: Take into account depth and pitch parameters.
+    console.log('TODO: Partially unimplemented SDL_CreateRGBSurfaceFrom called!');
+    return SDL.makeSurface(width, height, 0, false, 'CreateRGBSurfaceFrom', rmask, gmask, bmask, amask);
+  },
+
   SDL_DisplayFormatAlpha: function(surf) {
     var oldData = SDL.surfaces[surf];
     var ret = SDL.makeSurface(oldData.width, oldData.height, oldData.flags, false, 'copy:' + oldData.source);
@@ -1113,13 +1116,13 @@ var LibrarySDL = {
   },
 
   SDL_MapRGB: function(fmt, r, g, b) {
-    // Canvas screens are always RGBA
-    return 0xff+((b&0xff)<<8)+((g&0xff)<<16)+((r&0xff)<<24)
+    // Canvas screens are always RGBA. We assume the machine is little-endian.
+    return r&0xff|(g&0xff)<<8|(b&0xff)<<16|0xff000000;
   },
 
   SDL_MapRGBA: function(fmt, r, g, b, a) {
-    // Canvas screens are always RGBA
-    return (a&0xff)+((b&0xff)<<8)+((g&0xff)<<16)+((r&0xff)<<24)
+    // Canvas screens are always RGBA. We assume the machine is little-endian.
+    return r&0xff|(g&0xff)<<8|(b&0xff)<<16|(a&0xff)<<24;
   },
 
   SDL_GetAppState: function() {
@@ -1810,7 +1813,7 @@ var LibrarySDL = {
 
   SDL_AddTimer: function(interval, callback, param) {
     return window.setTimeout(function() {
-      Runtime.dynCall('ii', callback, [interval, param]);
+      Runtime.dynCall('iii', callback, [interval, param]);
     }, interval);
   },
   SDL_RemoveTimer: function(id) {
@@ -1837,9 +1840,9 @@ var LibrarySDL = {
   Mix_FadeOutChannel: function() { throw 'Mix_FadeOutChannel' },
 
   Mix_Linked_Version: function() { throw 'Mix_Linked_Version: TODO' },
-  SDL_CreateRGBSurfaceFrom: function() { throw 'SDL_CreateRGBSurfaceFrom: TODO' },
   SDL_SaveBMP_RW: function() { throw 'SDL_SaveBMP_RW: TODO' },
 
+  SDL_WM_SetIcon: function() { /* This function would set the application window icon surface, which doesn't apply for web canvases, so a no-op. */ },
   SDL_HasRDTSC: function() { return 0; },
   SDL_HasMMX: function() { return 0; },
   SDL_HasMMXExt: function() { return 0; },
diff --git a/src/parseTools.js b/src/parseTools.js
index 687faaa8..0b83a12b 100644
--- a/src/parseTools.js
+++ b/src/parseTools.js
@@ -710,9 +710,9 @@ function splitI64(value, floatConversion) {
   var lowInput = legalizedI64s ? value : 'VALUE';
   if (floatConversion && ASM_JS) lowInput = asmFloatToInt(lowInput);
   if (legalizedI64s) {
-    return [lowInput + '>>>0', 'Math.min(Math.floor((' + value + ')/' + asmEnsureFloat(4294967296, 'float') + '), ' + asmEnsureFloat(4294967295, 'float') + ')>>>0'];
+    return [lowInput + '>>>0', asmCoercion('Math.min(' + asmCoercion('Math.floor((' + value + ')/' + asmEnsureFloat(4294967296, 'float') + ')', 'double') + ', ' + asmEnsureFloat(4294967295, 'float') + ')', 'i32') + '>>>0'];
   } else {
-    return makeInlineCalculation(makeI64(lowInput + '>>>0', 'Math.min(Math.floor(VALUE/' + asmEnsureFloat(4294967296, 'float') + '), ' + asmEnsureFloat(4294967295, 'float') + ')>>>0'), value, 'tempBigIntP');
+    return makeInlineCalculation(makeI64(lowInput + '>>>0', asmCoercion('Math.min(' + asmCoercion('Math.floor(VALUE/' + asmEnsureFloat(4294967296, 'float') + ')', 'double') + ', ' + asmEnsureFloat(4294967295, 'float') + ')', 'i32') + '>>>0'), value, 'tempBigIntP');
   }
 }
 function mergeI64(value, unsigned) {
@@ -1049,7 +1049,7 @@ function getHeapOffset(offset, type, forceAsm) {
   offset = '(' + offset + ')';
   if (shifts != 0) {
     if (CHECK_HEAP_ALIGN) {
-      return '(CHECK_ALIGN_' + sz + '(' + offset + '|0)>>' + shifts + ')';
+      return '((CHECK_ALIGN_' + sz + '(' + offset + '|0)|0)>>' + shifts + ')';
     } else {
       return '(' + offset + '>>' + shifts + ')';
     }
@@ -1383,7 +1383,7 @@ function makeCopyValues(dest, src, num, type, modifier, align, sep) {
     if (!isNumber(num)) num = stripCorrections(num);
     if (!isNumber(align)) align = stripCorrections(align);
     if (!isNumber(num) || (parseInt(num)/align >= UNROLL_LOOP_MAX)) {
-      return '_memcpy(' + dest + ', ' + src + ', ' + num + ')';
+      return '(_memcpy(' + dest + ', ' + src + ', ' + num + ')|0)';
     }
     num = parseInt(num);
     if (ASM_JS) {
@@ -1465,7 +1465,7 @@ function getFastValue(a, op, b, type) {
         if ((isNumber(a) && Math.abs(a) < TWO_TWENTY) || (isNumber(b) && Math.abs(b) < TWO_TWENTY) || (bits < 32 && !ASM_JS)) {
           return '(((' + a + ')*(' + b + '))&' + ((Math.pow(2, bits)-1)|0) + ')'; // keep a non-eliminatable coercion directly on this
         }
-        return 'Math.imul(' + a + ',' + b + ')';
+        return '(Math.imul(' + a + ',' + b + ')|0)';
       }
     } else {
       if (a == '0') {
diff --git a/src/relooper/Relooper.cpp b/src/relooper/Relooper.cpp
index 8a6e18b8..7ceeb2f8 100644
--- a/src/relooper/Relooper.cpp
+++ b/src/relooper/Relooper.cpp
@@ -897,6 +897,7 @@ void Relooper::Calculate(Block *Entry) {
   BlockSet Entries;
   Entries.insert(Entry);
   Root = Analyzer(this).Process(AllBlocks, Entries, NULL);
+  assert(Root);
 
   // Post optimizations
 
@@ -1091,6 +1092,7 @@ void Relooper::Calculate(Block *Entry) {
 
 void Relooper::Render() {
   OutputBuffer = OutputBufferRoot;
+  assert(Root);
   Root->Render(false);
 }
 
diff --git a/src/runtime.js b/src/runtime.js
index 9bedfe68..e6d5f962 100644
--- a/src/runtime.js
+++ b/src/runtime.js
@@ -25,7 +25,7 @@ var RuntimeGenerator = {
     sep = sep || ';';
     var ret = RuntimeGenerator.alloc(size, 'STACK', false, sep, USE_TYPED_ARRAYS != 2 || (isNumber(size) && parseInt(size) % {{{ STACK_ALIGN }}} == 0));
     if (ASSERTIONS) {
-      ret += sep + 'assert(' + asmCoercion('(STACKTOP|0) < (STACK_MAX|0)', 'i32') + ')';
+      ret += sep + '(assert(' + asmCoercion('(STACKTOP|0) < (STACK_MAX|0)', 'i32') + ')|0)';
     }
     return ret;
   },
@@ -37,11 +37,11 @@ var RuntimeGenerator = {
     if (USE_TYPED_ARRAYS == 2) {
       assert(initial % Runtime.STACK_ALIGN == 0);
       if (ASSERTIONS && Runtime.STACK_ALIGN == 4) {
-        ret += '; assert(' + asmCoercion('!(STACKTOP&3)', 'i32') + ')';
+        ret += '; (assert(' + asmCoercion('!(STACKTOP&3)', 'i32') + ')|0)';
       }
     }
     if (ASSERTIONS) {
-      ret += '; assert(' + asmCoercion('(STACKTOP|0) < (STACK_MAX|0)', 'i32') + ')';
+      ret += '; (assert(' + asmCoercion('(STACKTOP|0) < (STACK_MAX|0)', 'i32') + ')|0)';
     }
     if (false) {
       ret += '; _memset(' + asmCoercion('__stackBase__', 'i32') + ', 0, ' + initial + ')';