Fix Issue 8259: dart:math pow broken for large integers.

runtime/lib/double.cc

Index: runtime/lib/double.cc
===================================================================
--- runtime/lib/double.cc	(revision 18079)
+++ runtime/lib/double.cc	(working copy)
@@ -79,12 +79,16 @@
     args.SetAt(0, String::Handle(String::New(error_msg)));
     Exceptions::ThrowByType(Exceptions::kUnsupported, args);
   }
-  if ((Smi::kMinValue <= val) && (val <= Smi::kMaxValue)) {
-    return Smi::New(static_cast<intptr_t>(val));
-  } else if ((Mint::kMinValue <= val) && (val <= Mint::kMaxValue)) {
-    return Mint::New(static_cast<int64_t>(val));
+  const int64_t int64_val = static_cast<int64_t>(val);
+  // Check if overflow in double-to-int.
+  if ((int64_val == static_cast<int64_t>(0x8000000000000000LL)) &&

[Vyacheslav Egorov (Google), 2013/02/05 00:34:36]

I don't think we can do this.

If a truncated double is not representable in an integer then behavior is
undefined per C++ Standard. Thus we can't rely on it returning the same marker
as Intel instruction use. (see [conv.fpint] in the Standard)

Here is another idea: Mint/Smi::kMinValue is representable in a double and so is
Mint/Smi::kMaxValue + 1 (cause both are powers of two).

If you declare MaxValuePlusOneAsDouble constants then you should be able to use

(X::kMinValue <= val) && (val < X::kMaxValuePlusOneAsDouble) 

to determine whether result of a conversion fits into an X or not. 

+      ((val < -2.0) || (val > 0.0))) {
+    return BigintOperations::NewFromDouble(val);
+  } else if ((Smi::kMinValue <= int64_val) && (int64_val <= Smi::kMaxValue)) {
+    return Smi::New(static_cast<intptr_t>(int64_val));
   } else {
-    return BigintOperations::NewFromDouble(val);
+    ASSERT((Mint::kMinValue <= int64_val) && (int64_val <= Mint::kMaxValue));
+    return Mint::New(int64_val);
   }
 }