[turbofan] Implement rounding of floats on x64 and ia32 without sse4.1.

src/assembler.cc

Index: src/assembler.cc
diff --git a/src/assembler.cc b/src/assembler.cc
index 784331738ee23f47cb9bc5cf9b796304f3e9a6d5..6bb8bc73eecab7e298d23375664d089696651b58 100644
--- a/src/assembler.cc
+++ b/src/assembler.cc
@@ -149,6 +149,7 @@ double minus_one_half;
 double negative_infinity;
 double the_hole_nan;
 double uint32_bias;
+double rounding_limit;
 };
 
 static DoubleConstant double_constants;
@@ -950,6 +951,8 @@ void ExternalReference::SetUp() {
   double_constants.negative_infinity = -V8_INFINITY;
   double_constants.uint32_bias =
     static_cast<double>(static_cast<uint32_t>(0xFFFFFFFF)) + 1;
+  // 2^52;
+  double_constants.rounding_limit = 4503599627370496.0;
 
   math_exp_data_mutex = new base::Mutex();
 }
@@ -1277,6 +1280,12 @@ ExternalReference ExternalReference::address_of_uint32_bias() {
 }
 
 
+ExternalReference ExternalReference::address_of_rounding_limit() {
+  return ExternalReference(
+      reinterpret_cast<void*>(&double_constants.rounding_limit));
+}
+
+
 ExternalReference ExternalReference::is_profiling_address(Isolate* isolate) {
   return ExternalReference(isolate->cpu_profiler()->is_profiling_address());
 }