infrastructure for implementing i64 math in asm.js, and implement i64Add and llvm_uadd_with_overflow_i64 that way

5 files changed, 76 insertions, 34 deletions

diff --git a/emscripten.py b/emscripten.py
index bbc4f76d..d7a24fbc 100755
--- a/emscripten.py
+++ b/emscripten.py
@@ -313,6 +313,10 @@ def emscript(infile, settings, outfile, libraries=[], compiler_engine=None,
   last_forwarded_json = json.loads(last_forwarded_data)
 
   if settings.get('ASM_JS'):
+    post_funcs, post_rest = post.split('// EMSCRIPTEN_END_FUNCS\n')
+    post = post_rest
+    funcs_js += '\n' + post_funcs + '// EMSCRIPTEN_END_FUNCS\n'
+
     simple = os.environ.get('EMCC_SIMPLE_ASM')
     class Counter:
       i = 0
diff --git a/src/jsifier.js b/src/jsifier.js
index 3cb5c23e..e7e4706b 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;
         }
 
@@ -1563,16 +1568,23 @@ 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) {
+      ['i64Add'].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'));
diff --git a/src/library.js b/src/library.js
index 12e550ff..0f8b0af5 100644
--- a/src/library.js
+++ b/src/library.js
@@ -5206,13 +5206,6 @@ 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);
@@ -7352,6 +7345,41 @@ 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) }}};
+  },
 };
 
 function autoAddDeps(object, name) {
diff --git a/src/long.js b/src/long.js
index c3651bd9..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);
diff --git a/src/parseTools.js b/src/parseTools.js
index 48274cd5..9a91070c 100644
--- a/src/parseTools.js
+++ b/src/parseTools.js
@@ -1724,13 +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) {
@@ -1971,6 +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': {
@@ -2059,7 +2064,7 @@ 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));