Reland r10908 (Ensure consistent result of transcendental function0.)

Reland r10908 (Ensure consistent result of transcendental function0.)

BUG=
TEST=

Committed: https://code.google.com/p/v8/source/detail?r=10915

[Yang, 2012-03-03 11:05:01]

I had to revert the first CL on this issue since there were two bugs. Those have
now been fixed. PTAL.

http://codereview.chromium.org/9583037/diff/1/src/ia32/codegen-ia32.cc
File src/ia32/codegen-ia32.cc (right):

http://codereview.chromium.org/9583037/diff/1/src/ia32/codegen-ia32.cc#newcode88
src/ia32/codegen-ia32.cc:88: __ mov(edx, Operand(esp, 5 * kPointerSize));
Had wrong stack offsets in the first version of this CL. It happens that
coincidentally there are in most cases the correct values even at those wrong
offsets. In other cases this bug leads to flaky errors in the test case.

http://codereview.chromium.org/9583037/diff/1/src/x64/code-stubs-x64.cc
File src/x64/code-stubs-x64.cc (right):

http://codereview.chromium.org/9583037/diff/1/src/x64/code-stubs-x64.cc#newco...
src/x64/code-stubs-x64.cc:1729: // NaN is represented by 0x7ff8000000000000.
Loading NaN from the root list won't work if we are calling this piece of code
as a C function. Instead we just load the bit pattern for NaN just like we do in
the ia32 version of this.

[Sven Panne, 2012-03-05 07:45:47]

http://codereview.chromium.org/9583037/diff/1/src/x64/code-stubs-x64.cc
File src/x64/code-stubs-x64.cc (right):

http://codereview.chromium.org/9583037/diff/1/src/x64/code-stubs-x64.cc#newco...
src/x64/code-stubs-x64.cc:1729: // NaN is represented by 0x7ff8000000000000.
On 2012/03/03 11:05:01, Yang wrote:
> Loading NaN from the root list won't work if we are calling this piece of code
> as a C function. Instead we just load the bit pattern for NaN just like we do
in
> the ia32 version of this.

I think the cleaner and more consistent way would be leaving the code here as it
was before, and call InitializeRootRegister in the code generated in
CreateTranscendentalFunction instead.

[Sven Panne, 2012-03-05 08:14:19]

LGTM after short discussion: Loading the root register won't work, because that
would be per-Isolate, and the transcendental code is global.