the idea is that often times in code, you will see something like this:...

src/parser.cc

 // Copyright 2006-2008 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 2780 matching lines @@
 
 // Precedence >= 4
 Expression* Parser::ParseBinaryExpression(int prec, bool accept_IN, bool* ok) {
   ASSERT(prec >= 4);
   Expression* x = ParseUnaryExpression(CHECK_OK);
   for (int prec1 = Precedence(peek(), accept_IN); prec1 >= prec; prec1--) {
     // prec1 >= 4
     while (Precedence(peek(), accept_IN) == prec1) {
       Token::Value op = Next();
       Expression* y = ParseBinaryExpression(prec1 + 1, accept_IN, CHECK_OK);
-
       // Compute some expressions involving only number literals.
-      if (x && x->AsLiteral() && x->AsLiteral()->handle()->IsNumber() &&
-          y && y->AsLiteral() && y->AsLiteral()->handle()->IsNumber()) {
-        double x_val = x->AsLiteral()->handle()->Number();
-        double y_val = y->AsLiteral()->handle()->Number();
-
-        switch (op) {
-          case Token::ADD:
-            x = NewNumberLiteral(x_val + y_val);
-            continue;
-          case Token::SUB:
-            x = NewNumberLiteral(x_val - y_val);
-            continue;
-          case Token::MUL:
-            x = NewNumberLiteral(x_val * y_val);
-            continue;
-          case Token::DIV:
-            x = NewNumberLiteral(x_val / y_val);
-            continue;
-          case Token::BIT_OR:
-            x = NewNumberLiteral(DoubleToInt32(x_val) | DoubleToInt32(y_val));
-            continue;
-          case Token::BIT_AND:
-            x = NewNumberLiteral(DoubleToInt32(x_val) & DoubleToInt32(y_val));
-            continue;
-          case Token::BIT_XOR:
-            x = NewNumberLiteral(DoubleToInt32(x_val) ^ DoubleToInt32(y_val));
-            continue;
-          case Token::SHL: {
-            int value = DoubleToInt32(x_val) << (DoubleToInt32(y_val) & 0x1f);
-            x = NewNumberLiteral(value);
+      if (y && y->AsLiteral()) {
+        Handle<Object> x_handle;
+        Handle<Object> y_handle = y->AsLiteral()->handle();
+        if (x && x->AsLiteral()) {
+          x_handle = x->AsLiteral()->handle();
+          if (x_handle->IsNumber() && y_handle->IsNumber()) {
+            double x_val = x_handle->Number();
+            double y_val = y_handle->Number();
+            double new_number = NAN;

[Kevin Millikin (Chromium), 2009/10/09 04:23:24]

I find this new code less clear than the old because it's less direct.  If you
want to jump to the top of the loop ('continue'), why bother setting a variable,
jumping to the bottom of the switch, and then checking the variable?

+            
+            switch (op) {
+              case Token::ADD: {
+                new_number = x_val + y_val;
+                break;
+              }
+              case Token::SUB: {
+                new_number = x_val - y_val;
+                break;
+              }
+              case Token::MUL: {
+                new_number = x_val * y_val;
+                break;
+              }
+              case Token::DIV: {
+                new_number = x_val / y_val;
+                break;
+              }
+              case Token::BIT_OR: {
+                new_number = DoubleToInt32(x_val) | DoubleToInt32(y_val);
+                break;
+              }
+              case Token::BIT_AND: {
+                new_number = DoubleToInt32(x_val) & DoubleToInt32(y_val);
+                break;
+              }
+              case Token::BIT_XOR: {
+                new_number = DoubleToInt32(x_val) ^ DoubleToInt32(y_val);
+                break;
+              }
+              case Token::SHL: {
+                new_number = DoubleToInt32(x_val) << (DoubleToInt32(y_val) & 0x1f);
+                break;
+              }
+              case Token::SHR: {
+                uint32_t shift = DoubleToInt32(y_val) & 0x1f;
+                new_number = DoubleToUint32(x_val) >> shift;
+                break;
+              }
+              case Token::SAR: {
+                uint32_t shift = DoubleToInt32(y_val) & 0x1f;
+                new_number = ArithmeticShiftRight(DoubleToInt32(x_val), shift);
+                break;
+              }
+              default:
+                break;
+            }
+            
+            if (new_number != NAN) {

[Kevin Millikin (Chromium), 2009/10/09 04:23:24]

You have introduced a bug.  Because of the way NaN comparisons work,
*everything* is not equal to NaN (even NaN != NaN).  If you don't hit one of the
cases in the switch you will constant-fold to NaN.

1==1 should be true, not NaN.

+              x = NewNumberLiteral(new_number);
+              continue;
+            }
+          } else if (x_handle->IsString() && y_handle->IsString()) {

[Kevin Millikin (Chromium), 2009/10/09 04:23:24]

No, you cannot do this optimization unless op is Token::ADD.

+            Handle<String> str1 = Handle<String>::cast(x_handle);
+            Handle<String> str2 = Handle<String>::cast(y_handle);
+            Handle<String> str3 = Factory::NewConsString(str1, str2, TENURED);
+            x = NEW(Literal(str3));
             continue;
           }
-          case Token::SHR: {
-            uint32_t shift = DoubleToInt32(y_val) & 0x1f;
-            uint32_t value = DoubleToUint32(x_val) >> shift;
-            x = NewNumberLiteral(value);
-            continue;
+        } else if (x && x->AsBinaryOperation() && y_handle->IsString()) {

[Kevin Millikin (Chromium), 2009/10/09 04:23:24]

You cannot do this optimization unless op is Token::Add.

+          // is a binary operation
+          BinaryOperation* bin_op = x->AsBinaryOperation();
+          Expression* right = bin_op->right();
+          if (bin_op->op() == Token::ADD && right && right->AsLiteral()) {
+            x_handle = right->AsLiteral()->handle();
+            if (x_handle->IsString()) {
+              Handle<String> str1 = Handle<String>::cast(x_handle);
+              Handle<String> str2 = Handle<String>::cast(y_handle);
+              Handle<String> str3 = Factory::NewConsString(str1, str2, TENURED);
+              right = NEW(Literal(str3));
+              x = NEW(BinaryOperation(bin_op->op(), bin_op->left(), right));
+              continue;
+            }
           }
-          case Token::SAR: {
-            uint32_t shift = DoubleToInt32(y_val) & 0x1f;
-            int value = ArithmeticShiftRight(DoubleToInt32(x_val), shift);
-            x = NewNumberLiteral(value);
-            continue;
-          }
-          default:
-            break;
         }
       }
 
       // Convert constant divisions to multiplications for speed.
       if (op == Token::DIV &&
           y && y->AsLiteral() && y->AsLiteral()->handle()->IsNumber()) {
         double y_val = y->AsLiteral()->handle()->Number();
         int64_t y_int = static_cast<int64_t>(y_val);
         // There are rounding issues with this optimization, but they don't
         // apply if the number to be divided with has a reciprocal that can be
@@ skipping 1967 matching lines @@
                        start_position,
                        is_expression);
   return result;
 }
 
 
 #undef NEW
 
 
 } }  // namespace v8::internal