1 files changed, 86 insertions, 2 deletions

diff --git a/src/analyzer.js b/src/analyzer.js
index 03d44cb7..2cc46ab6 100644
--- a/src/analyzer.js
+++ b/src/analyzer.js
@@ -20,6 +20,7 @@ var BRANCH_INVOKE = set('branch', 'invoke');
 var LABEL_ENDERS = set('branch', 'return', 'switch');
 var SIDE_EFFECT_CAUSERS = set('call', 'invoke', 'atomic');
 var UNUNFOLDABLE = set('value', 'structvalue', 'type', 'phiparam');
+var I64_DOUBLE_FLIP = { i64: 'double', double: 'i64' };
 
 // Analyzer
 
@@ -99,7 +100,86 @@ function analyzer(data, sidePass) {
         }
       }
       delete item.items;
+      this.forwardItem(item, 'CastAway');
+    }
+  });
+
+  // CastAway - try to remove bitcasts of double<-->i64, which LLVM sometimes generates unnecessarily
+  // (load a double, convert to i64, use as i64).
+  // We optimize this by checking if there are such bitcasts. If so we create a shadow
+  // variable that is of the other type, and use that in the relevant places. (As SSA, this is valid, and
+  // variable elimination later will remove the double load if it is no longer needed.)
+  //
+  // Note that aside from being an optimization, this is needed for correctness in some cases: If code
+  // assumes it can bitcast a double to an i64 and back and forth without loss, that may be violated
+  // due to NaN canonicalization.
+  substrate.addActor('CastAway', {
+    processItem: function(item) {
       this.forwardItem(item, 'Legalizer');
+      if (USE_TYPED_ARRAYS != 2) return;
+
+      item.functions.forEach(function(func) {
+        var has = false;
+        func.labels.forEach(function(label) {
+          var lines = label.lines;
+          for (var i = 0; i < lines.length; i++) {
+            var line = lines[i];
+            if (line.intertype == 'bitcast' && line.type in I64_DOUBLE_FLIP) {
+              has = true;
+            }
+          }
+        });
+        if (!has) return;
+        // there are i64<-->double bitcasts, create shadows for everything
+        var shadowed = {};
+        func.labels.forEach(function(label) {
+          var lines = label.lines;
+          var i = 0;
+          while (i < lines.length) {
+          var lines = label.lines;
+            var line = lines[i];
+            if (line.intertype == 'load' && line.type in I64_DOUBLE_FLIP) {
+              if (line.pointer.intertype != 'value') { i++; continue } // TODO
+              shadowed[line.assignTo] = 1;
+              var shadow = line.assignTo + '$$SHADOW';
+              var flip = I64_DOUBLE_FLIP[line.type];
+              lines.splice(i + 1, 0, { // if necessary this element will be legalized in the next phase
+                tokens: null,
+                indent: 2,
+                lineNum: line.lineNum + 0.5,
+                assignTo: shadow,
+                intertype: 'load',
+                pointerType: flip + '*',
+                type: flip,
+                valueType: flip,
+                pointer: {
+                 intertype: 'value',
+                 ident: line.pointer.ident,
+                 type: flip + '*'
+                },
+                align: line.align,
+                ident: line.ident
+              });
+              // note: no need to update func.lines, it is generated in a later pass
+              i++;
+            }
+            i++;
+          }
+        });
+        // use shadows where possible
+        func.labels.forEach(function(label) {
+          var lines = label.lines;
+          for (var i = 0; i < lines.length; i++) {
+            var line = lines[i];
+            if (line.intertype == 'bitcast' && line.type in I64_DOUBLE_FLIP && line.ident in shadowed) {
+              var shadow = line.ident + '$$SHADOW';
+              line.params[0].ident = shadow;
+              line.params[0].type = line.type;
+              line.type2 = line.type;
+            }
+          }
+        });
+      });
     }
   });
 
@@ -174,12 +254,15 @@ function analyzer(data, sidePass) {
           var factor = (next - prev)/(4*toAdd.length+3);
           for (var k = 0; k < toAdd.length; k++) {
             toAdd[k].lineNum = prev + ((k+1)*factor);
+            assert(k == 0 || toAdd[k].lineNum > toAdd[k-1].lineNum);
           }
         }
         function removeAndAdd(lines, i, toAdd) {
           var item = lines[i];
           interpLines(lines, i, toAdd);
           Array.prototype.splice.apply(lines, [i, 1].concat(toAdd));
+          if (i > 0) assert(lines[i].lineNum > lines[i-1].lineNum);
+          if (i + toAdd.length < lines.length) assert(lines[i + toAdd.length - 1].lineNum < lines[i + toAdd.length].lineNum);
           return toAdd.length;
         }
         function legalizeFunctionParameters(params) {
@@ -328,8 +411,9 @@ function analyzer(data, sidePass) {
                   // legalize parameters
                   legalizeFunctionParameters(value.params);
                   // legalize return value, if any
-                  if (value.assignTo && isIllegalType(item.type)) {
-                    bits = getBits(value.type);
+                  var returnType = getReturnType(item.type);
+                  if (value.assignTo && isIllegalType(returnType)) {
+                    bits = getBits(returnType);
                     var elements = getLegalVars(item.assignTo, bits);
                     // legalize return value
                     value.assignTo = elements[0].ident;