fix registerize semantics to treat each arm of ifs and switches separately

3 files changed, 186 insertions, 11 deletions

diff --git a/tools/js-optimizer.js b/tools/js-optimizer.js
index a52ad6b7..9e0307da 100644
--- a/tools/js-optimizer.js
+++ b/tools/js-optimizer.js
@@ -1596,7 +1596,15 @@ function registerize(ast) {
     var varLevels = {};
     var possibles = {};
     var unoptimizables = {};
-    traverse(fun, function(node, type) {
+    function purgeLevel() {
+      // Invalidate all dominating on this level, further users make it unoptimizable
+      for (var name in levelDominations[level]) {
+        varLevels[name] = 0;
+      }
+      levelDominations[level] = null;
+      level--;
+    }
+    traverse(fun, function possibilifier(node, type) {
       if (type == 'name') {
         var name = node[1];
         if (localVars[name]) {
@@ -1617,16 +1625,49 @@ function registerize(ast) {
           }
         }
       } else if (type in CONTROL_FLOW) {
-        level++;
-      }
-    }, function(node, type) {
-      if (type in CONTROL_FLOW) {
-        // Invalidate all dominating on this level, further users make it unoptimizable
-        for (var name in levelDominations[level]) {
-          varLevels[name] = 0;
+        // recurse children, in the context of a loop
+        switch(type) {
+          case 'while': case 'do': {
+            traverse(node[1], possibilifier);
+            level++;
+            traverse(node[2], possibilifier);
+            purgeLevel();
+            break;
+          }
+          case 'for': {
+            traverse(node[1], possibilifier);
+            for (var i = 2; i <= 4; i++) {
+              level++;
+              traverse(node[i], possibilifier);
+              purgeLevel();
+            }
+            break;
+          }
+          case 'if': {
+            traverse(node[1], possibilifier);
+            level++;
+            traverse(node[2], possibilifier);
+            purgeLevel();
+            if (node[3]) {
+              level++;
+              traverse(node[3], possibilifier);
+              purgeLevel();
+            }
+            break;
+          }
+          case 'switch': {
+            traverse(node[1], possibilifier);
+            var cases = node[2];
+            for (var i = 0; i < cases.length; i++) {
+              level++;
+              traverse(cases[i][1], possibilifier);
+              purgeLevel();
+            }
+            break;
+          }
+          default: throw dumpAst(node);
         }
-        levelDominations[level] = null;
-        level--;
+        return null; // prevent recursion into children, which we already did
       }
     });
     var optimizables = {};
@@ -1635,6 +1676,10 @@ function registerize(ast) {
         if (!unoptimizables[possible]) optimizables[possible] = 1;
       }
     }
+
+    //printErr('optimizables: ' + JSON.stringify(optimizables));
+    //printErr('unoptimizables: ' + JSON.stringify(unoptimizables));
+
     // Go through the function's code, assigning 'registers'.
     // The only tricky bit is to keep variables locked on a register through loops,
     // since they can potentially be returned to. Optimizable variables lock onto
diff --git a/tools/test-js-optimizer-asm-regs-output.js b/tools/test-js-optimizer-asm-regs-output.js
index ba2316ea..c2efa506 100644
--- a/tools/test-js-optimizer-asm-regs-output.js
+++ b/tools/test-js-optimizer-asm-regs-output.js
@@ -38,4 +38,69 @@ function switchey(d1, i2) {
     return 20;
   }
 }
+function switchey2() {
+  var i1 = 0, i2 = 0, i3 = 0, i4 = 0, i5 = 0, d6 = +0, d7 = +0, i8 = 0, i9 = 0;
+  i1 = STACKTOP;
+  STACKTOP = STACKTOP + 8 | 0;
+  i2 = 1;
+  while (1) switch (i2 | 0) {
+   case 1:
+    i3 = i1 | 0;
+    __ZN6RandomC1Ev(i3);
+    i4 = 0;
+    i5 = 0;
+    i2 = 2;
+    break;
+   case 2:
+    d6 = +__ZN6Random3getEf(8, +1);
+    d7 = +__ZN6Random3getEf(i3, +1);
+    _printf(24, (tempInt = STACKTOP, STACKTOP = STACKTOP + 16 | 0, HEAPF64[CHECK_ALIGN_8(tempInt | 0) >> 3] = d6, HEAPF64[CHECK_ALIGN_8(tempInt + 8 | 0) >> 3] = d7, tempInt) | 0);
+    i8 = (d6 != d7 & 1) + i5 | 0;
+    i9 = i4 + 1 | 0;
+    if ((i9 | 0) < 100) {
+      i4 = i9;
+      i5 = i8;
+      i2 = 2;
+      break;
+    } else {
+      i2 = 3;
+      break;
+    }
+   case 3:
+    _printf(16, (tempInt = STACKTOP, STACKTOP = STACKTOP + 8 | 0, HEAP32[CHECK_ALIGN_4(tempInt | 0) >> 2] = i8, tempInt) | 0);
+    STACKTOP = i1;
+    return 0;
+  }
+  return 0;
+}
+function iffey() {
+  var i1 = 0, i2 = 0, i3 = 0, i4 = 0, i5 = 0, d6 = +0, d7 = +0, i8 = 0, i9 = 0;
+  i1 = STACKTOP;
+  STACKTOP = STACKTOP + 8 | 0;
+  i2 = 1;
+  while (1) {
+    if (i2 | 0) {
+      i3 = i1 | 0;
+      __ZN6RandomC1Ev(i3);
+      i4 = 0;
+      i5 = 0;
+      i2 = 2;
+    } else {
+      d6 = +__ZN6Random3getEf(8, +1);
+      d7 = +__ZN6Random3getEf(i3, +1);
+      _printf(24, (tempInt = STACKTOP, STACKTOP = STACKTOP + 16 | 0, HEAPF64[CHECK_ALIGN_8(tempInt | 0) >> 3] = d6, HEAPF64[CHECK_ALIGN_8(tempInt + 8 | 0) >> 3] = d7, tempInt) | 0);
+      i8 = (d6 != d7 & 1) + i5 | 0;
+      i9 = i4 + 1 | 0;
+      if ((i9 | 0) < 100) {
+        i4 = i9;
+        i5 = i8;
+        i2 = 2;
+      } else {
+        i2 = 3;
+        return 10;
+      }
+    }
+  }
+  return 0;
+}
 
diff --git a/tools/test-js-optimizer-asm-regs.js b/tools/test-js-optimizer-asm-regs.js
index 64797e38..8c3a1c98 100644
--- a/tools/test-js-optimizer-asm-regs.js
+++ b/tools/test-js-optimizer-asm-regs.js
@@ -41,5 +41,70 @@ function switchey(x, y) {
       return 20;
   }
 }
-// EMSCRIPTEN_GENERATED_FUNCTIONS: ["asm", "_doit", "stackRestore", "switchey"]
+function switchey2() {
+  var $rng2 = 0, $count_06 = 0, $i_05 = 0, $2 = +0, $3 = +0, $count_1 = 0, $9 = 0, label = 0, __stackBase__ = 0;
+  __stackBase__ = STACKTOP;
+  STACKTOP = STACKTOP + 8 | 0;
+  label = 1;
+  while (1) switch (label | 0) {
+   case 1:
+    $rng2 = __stackBase__ | 0;
+    __ZN6RandomC1Ev($rng2);
+    $i_05 = 0;
+    $count_06 = 0;
+    label = 2;
+    break;
+   case 2:
+    $2 = +__ZN6Random3getEf(8, +1);
+    $3 = +__ZN6Random3getEf($rng2, +1);
+    _printf(24, (tempInt = STACKTOP, STACKTOP = STACKTOP + 16 | 0, HEAPF64[CHECK_ALIGN_8(tempInt | 0) >> 3] = $2, HEAPF64[CHECK_ALIGN_8(tempInt + 8 | 0) >> 3] = $3, tempInt) | 0);
+    $count_1 = ($2 != $3 & 1) + $count_06 | 0;
+    $9 = $i_05 + 1 | 0;
+    if (($9 | 0) < 100) {
+      $i_05 = $9;
+      $count_06 = $count_1;
+      label = 2;
+      break;
+    } else {
+      label = 3;
+      break;
+    }
+   case 3:
+    _printf(16, (tempInt = STACKTOP, STACKTOP = STACKTOP + 8 | 0, HEAP32[CHECK_ALIGN_4(tempInt | 0) >> 2] = $count_1, tempInt) | 0);
+    STACKTOP = __stackBase__;
+    return 0;
+  }
+  return 0;
+}
+function iffey() {
+  var $rng2 = 0, $count_06 = 0, $i_05 = 0, $2 = +0, $3 = +0, $count_1 = 0, $9 = 0, label = 0, __stackBase__ = 0;
+  __stackBase__ = STACKTOP;
+  STACKTOP = STACKTOP + 8 | 0;
+  label = 1;
+  while (1) {
+    if (label | 0) {
+      $rng2 = __stackBase__ | 0;
+      __ZN6RandomC1Ev($rng2);
+      $i_05 = 0;
+      $count_06 = 0;
+      label = 2;
+    } else {
+      $2 = +__ZN6Random3getEf(8, +1);
+      $3 = +__ZN6Random3getEf($rng2, +1);
+      _printf(24, (tempInt = STACKTOP, STACKTOP = STACKTOP + 16 | 0, HEAPF64[CHECK_ALIGN_8(tempInt | 0) >> 3] = $2, HEAPF64[CHECK_ALIGN_8(tempInt + 8 | 0) >> 3] = $3, tempInt) | 0);
+      $count_1 = ($2 != $3 & 1) + $count_06 | 0;
+      $9 = $i_05 + 1 | 0;
+      if (($9 | 0) < 100) {
+        $i_05 = $9;
+        $count_06 = $count_1;
+        label = 2;
+      } else {
+        label = 3;
+        return 10;
+      }
+    }
+  }
+  return 0;
+}
+// EMSCRIPTEN_GENERATED_FUNCTIONS: ["asm", "_doit", "stackRestore", "switchey", "switchey2", "iffey"]