[turbofan] Support for %_DoubleHi, %_DoubleLo and %_ConstructDouble.

[turbofan] Support for %_DoubleHi, %_DoubleLo and %_ConstructDouble.

This adds support for the double bits intrinsics to TurboFan, and is
a first step towards fast Math functions inlined into TurboFan code
or even compiled by themselves with TurboFan.

Committed: https://crrev.com/4436c2642a9b7adc9a1b68577ae9ca9a9a404c06
Cr-Commit-Position: refs/heads/master@{#27006}

[Benedikt Meurer, 2015-03-05 07:21:56]

bmeurer@chromium.org changed reviewers:
+ jarin@chromium.org, svenpanne@chromium.org, yangguo@chromium.org

[Benedikt Meurer, 2015-03-05 07:23:01]

Yang: PTAL at the intrinsics.
Jaro: PTAL at the actual backend/TF stuff.
Sven: FYI new intrinsics.

[Benedikt Meurer, 2015-03-05 07:23:26]

I'll add the missing arm64 now.

[Yang, 2015-03-05 07:36:45]

Only looked at code generation, which LGTM, and test cases, for which I got
comments.

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/constru...
File test/mjsunit/asm/construct-double.js (right):

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/constru...
test/mjsunit/asm/construct-double.js:27: assertEquals(%_ConstructDouble(0, 1),
m.cd2(1));
Probably better to use %ConstructDouble instead of %_ConstructDouble. I'd rather
compare with the C++ implementation than the dying Crankshaft implementation.
Once Crankshaft is replaced by Turbofan, this test would become a tautology.

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/double-...
File test/mjsunit/asm/double-hi.js (right):

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/double-...
test/mjsunit/asm/double-hi.js:35: assertEquals(%_DoubleHi(i), m.hi1(i));
same here.

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/double-...
File test/mjsunit/asm/double-lo.js (right):

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/double-...
test/mjsunit/asm/double-lo.js:35: assertEquals(%_DoubleLo(i), m.lo1(i));
same here.

[Sven Panne, 2015-03-05 07:40:12]

LGTM with a nit.

https://codereview.chromium.org/974313002/diff/40001/src/compiler/machine-ope...
File src/compiler/machine-operator.h (right):

https://codereview.chromium.org/974313002/diff/40001/src/compiler/machine-ope...
src/compiler/machine-operator.h:181: const Operator* Float64ExtractWord32(int
location);
If I see this correctly, location == 0 means "Low" and location == 1 means
"High", and there are no other allowed values. Using int and the name "location"
is a bit confusing here, it is *not* really an offset. A separate enum class
might be more appropriate IMHO.

[Sven Panne, 2015-03-05 07:40:12]

LGTM with a nit.

https://codereview.chromium.org/974313002/diff/40001/src/compiler/machine-ope...
File src/compiler/machine-operator.h (right):

https://codereview.chromium.org/974313002/diff/40001/src/compiler/machine-ope...
src/compiler/machine-operator.h:181: const Operator* Float64ExtractWord32(int
location);
If I see this correctly, location == 0 means "Low" and location == 1 means
"High", and there are no other allowed values. Using int and the name "location"
is a bit confusing here, it is *not* really an offset. A separate enum class
might be more appropriate IMHO.

[Jarin, 2015-03-05 07:40:46]

On 2015/03/05 07:23:01, Benedikt Meurer wrote:
> Yang: PTAL at the intrinsics.
> Jaro: PTAL at the actual backend/TF stuff.
> Sven: FYI new intrinsics.

The TF part looks good to me.

[Benedikt Meurer, 2015-03-05 08:51:46]

https://codereview.chromium.org/974313002/diff/40001/src/compiler/machine-ope...
File src/compiler/machine-operator.h (right):

https://codereview.chromium.org/974313002/diff/40001/src/compiler/machine-ope...
src/compiler/machine-operator.h:181: const Operator* Float64ExtractWord32(int
location);
Turned into dedicated opcodes instead, as per offline dicussion.

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/constru...
File test/mjsunit/asm/construct-double.js (right):

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/constru...
test/mjsunit/asm/construct-double.js:27: assertEquals(%_ConstructDouble(0, 1),
m.cd2(1));
On 2015/03/05 07:36:45, Yang wrote:
> Probably better to use %ConstructDouble instead of %_ConstructDouble. I'd
rather
> compare with the C++ implementation than the dying Crankshaft implementation.
> Once Crankshaft is replaced by Turbofan, this test would become a tautology.

Done.

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/double-...
File test/mjsunit/asm/double-hi.js (right):

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/double-...
test/mjsunit/asm/double-hi.js:35: assertEquals(%_DoubleHi(i), m.hi1(i));
On 2015/03/05 07:36:45, Yang wrote:
> same here.

Done.

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/double-...
File test/mjsunit/asm/double-lo.js (right):

https://codereview.chromium.org/974313002/diff/40001/test/mjsunit/asm/double-...
test/mjsunit/asm/double-lo.js:35: assertEquals(%_DoubleLo(i), m.lo1(i));
On 2015/03/05 07:36:45, Yang wrote:
> same here.

Done.

[Benedikt Meurer, 2015-03-05 08:52:39]

The CQ bit was checked by bmeurer@chromium.org

[Benedikt Meurer, 2015-03-05 08:52:40]

The patchset sent to the CQ was uploaded after l-g-t-m from
svenpanne@chromium.org, yangguo@chromium.org
Link to the patchset: https://codereview.chromium.org/974313002/#ps80001 (title:
"Addressed Svens comment.")

[commit-bot: I haz the power, 2015-03-05 08:53:48]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/974313002/80001

[commit-bot: I haz the power, 2015-03-05 09:22:32]

Committed patchset #5 (id:80001)

[commit-bot: I haz the power, 2015-03-05 09:22:50]

Patchset 5 (id:??) landed as
https://crrev.com/4436c2642a9b7adc9a1b68577ae9ca9a9a404c06
Cr-Commit-Position: refs/heads/master@{#27006}