Introduce intrinsics for double values in Javascript.

Introduce intrinsics for double values in Javascript.

R=svenpanne@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=19704

[Yang, 2014-02-28 15:04:00]

Only ia32 port for now. Please take a first look.

[Sven Panne, 2014-03-03 11:28:26]

Seems to be OK, just 1 remark and 1 nit so far.

https://codereview.chromium.org/178583006/diff/1/src/full-codegen.cc
File src/full-codegen.cc (right):

https://codereview.chromium.org/178583006/diff/1/src/full-codegen.cc#newcode957
src/full-codegen.cc:957: masm()->CallRuntime(Runtime::kDoubleHi, 1);
Hopefully the usual benchmarks will not force us to do this directly in an
architecture-dependent way. Calls into the runtime are costly and if we run full
code "long" enough... :-/

https://codereview.chromium.org/178583006/diff/1/src/runtime.cc
File src/runtime.cc (right):

https://codereview.chromium.org/178583006/diff/1/src/runtime.cc#newcode7687
src/runtime.cc:7687: static_cast<int32_t>(double_to_uint64(x) & 0xffffffff));
Add an unsigned suffix 'u' to the hexadecimal constant, otherwise it's far from
obvious that things are correct here.

[Yang, 2014-03-04 13:58:40]

On 2014/03/03 11:28:26, Sven Panne wrote:
> Seems to be OK, just 1 remark and 1 nit so far.
> 
> https://codereview.chromium.org/178583006/diff/1/src/full-codegen.cc
> File src/full-codegen.cc (right):
> 
>
https://codereview.chromium.org/178583006/diff/1/src/full-codegen.cc#newcode957
> src/full-codegen.cc:957: masm()->CallRuntime(Runtime::kDoubleHi, 1);
> Hopefully the usual benchmarks will not force us to do this directly in an
> architecture-dependent way. Calls into the runtime are costly and if we run
full
> code "long" enough... :-/
> 
> https://codereview.chromium.org/178583006/diff/1/src/runtime.cc
> File src/runtime.cc (right):
> 
> https://codereview.chromium.org/178583006/diff/1/src/runtime.cc#newcode7687
> src/runtime.cc:7687: static_cast<int32_t>(double_to_uint64(x) & 0xffffffff));
> Add an unsigned suffix 'u' to the hexadecimal constant, otherwise it's far
from
> obvious that things are correct here.

Addressed comments.
Changed some small details (TIL DataEquals doesn't need to care about the
operands).
Platform ports added.

[Sven Panne, 2014-03-05 08:00:52]

LGTM with one (big) nit.

https://codereview.chromium.org/178583006/diff/20001/test/mjsunit/double-intr...
File test/mjsunit/double-intrinsics.js (right):

https://codereview.chromium.org/178583006/diff/20001/test/mjsunit/double-intr...
test/mjsunit/double-intrinsics.js:30: %OptimizeFunctionOnNextCall(test);
Hmmm, this relies on inlining heuristics, otherwise we don't test what we think
we test. It's assertDoubleBits which needs to be optimized.

[Yang, 2014-03-05 08:46:13]

On 2014/03/05 08:00:52, Sven Panne wrote:
> LGTM with one (big) nit.
> 
>
https://codereview.chromium.org/178583006/diff/20001/test/mjsunit/double-intr...
> File test/mjsunit/double-intrinsics.js (right):
> 
>
https://codereview.chromium.org/178583006/diff/20001/test/mjsunit/double-intr...
> test/mjsunit/double-intrinsics.js:30: %OptimizeFunctionOnNextCall(test);
> Hmmm, this relies on inlining heuristics, otherwise we don't test what we
think
> we test. It's assertDoubleBits which needs to be optimized.

Good point. I updated the test to not rely on inlining heuristics.

[Yang, 2014-03-07 09:05:26]

Committed patchset #3 manually as r19704 (presubmit successful).