V8: Add SIMD functions.

V8: Add SIMD functions for Phase 1.

Float32x4:
  abs, neg, sqrt, reciprocalApproximation, reciprocalSqrtApproximation, add, sub, mul, div,
  min, max, minNum, maxNum, lessThan, lessThanOrEqual, greaterThan, greaterThanOrEqual,
  equal, notEqual, select, swizzle, shuffle, fromInt32x4, fromInt32x4Bits, fromInt16x8Bits,
  fromInt8x16Bits.

Int32x4:
  neg, add, sub, mul, min, max, and, or, xor, not, shiftLeftByScalar, shiftRightLogicalByScalar,
  shiftRightArithmeticByScalar, lessThan, lessThanOrEqual, greaterThan, greaterThanOrEqual,
  equal, notEqual, select, swizzle, shuffle, fromFloat32x4, fromFloat32x4Bits, fromInt16x8Bits,
  fromInt8x16Bits.

Int16x8:
  neg, add, sub, mul, min, max, and, or, xor, not, shiftLeftByScalar, shiftRightLogicalByScalar,
  shiftRightArithmeticByScalar, lessThan, lessThanOrEqual, greaterThan, greaterThanOrEqual,
  equal, notEqual, select, swizzle, shuffle, fromFloat32x4Bits, fromInt32x4Bits, fromInt8x16Bits.

Int8x16:
  neg, add, sub, mul, min, max, and, or, xor, not, shiftLeftByScalar, shiftRightLogicalByScalar,
  shiftRightArithmeticByScalar, lessThan, lessThanOrEqual, greaterThan, greaterThanOrEqual,
  equal, notEqual, select, swizzle, shuffle, fromFloat32x4Bits, fromInt32x4Bits, fromInt16x8Bitss.

Bool32x4, Bool16x8, Bool8x16:
  and, or, not, anyTrue, allTrue, select, swizzle, shuffle.

I might have forgotten a few.

LOG=N
BUG=v8:4124
Committed: https://crrev.com/6378f57b30f1141850ef7c260972ae109450b023
Cr-Commit-Position: refs/heads/master@{#30051}

[bbudge, 2015-07-15 23:06:56]

Patchset #1 (id:1) has been deleted

[bbudge, 2015-08-01 04:58:26]

Patchset #2 (id:40001) has been deleted

[bbudge, 2015-08-01 18:34:26]

Patchset #2 (id:60001) has been deleted

[bbudge, 2015-08-01 19:38:00]

Patchset #2 (id:80001) has been deleted

[bbudge, 2015-08-03 23:58:16]

Patchset #2 (id:100001) has been deleted

[bbudge, 2015-08-03 23:58:24]

Patchset #1 (id:20001) has been deleted

[bbudge, 2015-08-04 00:35:48]

bbudge@chromium.org changed reviewers:
+ rossberg@chromium.org

[bbudge, 2015-08-04 00:35:49]

So sorry about sending you this. PTAL

https://codereview.chromium.org/1230343003/diff/120001/src/heap/heap.cc
File src/heap/heap.cc (right):

https://codereview.chromium.org/1230343003/diff/120001/src/heap/heap.cc#newco...
src/heap/heap.cc:3378: NameDictionary::New(isolate(), 2 *
Runtime::kNumFunctions, TENURED);
This is just a hack until yangguo lands this:
https://codereview.chromium.org/1265983006

[bbudge, 2015-08-04 00:48:27]

On 2015/08/04 00:35:49, bbudge wrote:
> So sorry about sending you this. PTAL
> 
> https://codereview.chromium.org/1230343003/diff/120001/src/heap/heap.cc
> File src/heap/heap.cc (right):
> 
>
https://codereview.chromium.org/1230343003/diff/120001/src/heap/heap.cc#newco...
> src/heap/heap.cc:3378: NameDictionary::New(isolate(), 2 *
> Runtime::kNumFunctions, TENURED);
> This is just a hack until yangguo lands this:
> https://codereview.chromium.org/1265983006

A few random notes:
- I'm leaving some functions that take typed arrays (load, load1/2/3, store,
store1/2/3) for another CL to keep this manageable.
- The spec may change to add unsigned integer types, so I skipped most of the
unsigned integer functions for now. The unsigned extract functions are needed
for the tests to pass.

[rossberg, 2015-08-04 13:13:02]

rossberg@chromium.org changed reviewers:
+ littledan@chromium.org

[rossberg, 2015-08-04 13:13:12]

rossberg@chromium.org changed required reviewers:
+ littledan@chromium.org

[rossberg, 2015-08-04 13:13:33]

LGTM, but Dan should have a look as well.

[Dan Ehrenberg, 2015-08-06 00:12:25]

Sorry for the long turnaround time. Just a couple changes needed; the
denormal/reciprocal approximation stuff can wait for now. I'm happy with how you
were able to use C macros to get rid of most of the duplication here.

https://codereview.chromium.org/1230343003/diff/140001/src/harmony-simd.js
File src/harmony-simd.js (right):

https://codereview.chromium.org/1230343003/diff/140001/src/harmony-simd.js#ne...
src/harmony-simd.js:11: var $bool8x16ToString;
You're supposed to use the new export/import system these days rather than
properties of the global object. See src/harmony-arrays.js for an example.

https://codereview.chromium.org/1230343003/diff/140001/src/harmony-simd.js#ne...
src/harmony-simd.js:106: return %NAMESelect(selector, a, b);
The spec v0.7.3 defines select only on numeric types, but equal and notEqual on
all types. You did the reverse here. Do you think the spec should be changed? I
don't feel so strongly about it.

https://codereview.chromium.org/1230343003/diff/140001/src/runtime/runtime-si...
File src/runtime/runtime-simd.cc (right):

https://codereview.chromium.org/1230343003/diff/140001/src/runtime/runtime-si...
src/runtime/runtime-simd.cc:62: inline float RecipSqrtApprox(float a) { return
1.0f / std::sqrt(a); }
We may have to revisit these at some point, to create consistent results for
optimized code vs fullcodegen.

https://codereview.chromium.org/1230343003/diff/140001/src/runtime/runtime-si...
src/runtime/runtime-simd.cc:429: shift = shift & 0x3f;                          
                   \
Looks like you're implementing wrap-around behavior here (like JS <<). However, 
SIMD operations max out at 0 for shifting too far left; see
http://littledan.github.io/simd.html#simd-shift-left-by-scalar . This is done to
match what the processors apparently do. Analogous behavior exists on the right
shifts.

https://codereview.chromium.org/1230343003/diff/140001/src/runtime/runtime-si...
src/runtime/runtime-simd.cc:564: #define BINARY_ADD(a, b) (a) + (b)
We'll have to revisit ops like this for proper denormal behavior on floats on
ARMv7, to be consistent with optimized code. Fine for now though.

[bbudge, 2015-08-06 14:07:32]

https://codereview.chromium.org/1230343003/diff/140001/src/harmony-simd.js
File src/harmony-simd.js (right):

https://codereview.chromium.org/1230343003/diff/140001/src/harmony-simd.js#ne...
src/harmony-simd.js:11: var $bool8x16ToString;
On 2015/08/06 00:12:25, Dan Ehrenberg wrote:
> You're supposed to use the new export/import system these days rather than
> properties of the global object. See src/harmony-arrays.js for an example.

Done.

https://codereview.chromium.org/1230343003/diff/140001/src/harmony-simd.js#ne...
src/harmony-simd.js:106: return %NAMESelect(selector, a, b);
On 2015/08/06 00:12:25, Dan Ehrenberg wrote:
> The spec v0.7.3 defines select only on numeric types, but equal and notEqual
on
> all types. You did the reverse here. Do you think the spec should be changed?
I
> don't feel so strongly about it.

No, select on boolean types doesn't seem useful. Done. (Equal and NotEqual are
OK)

https://codereview.chromium.org/1230343003/diff/140001/src/runtime/runtime-si...
File src/runtime/runtime-simd.cc (right):

https://codereview.chromium.org/1230343003/diff/140001/src/runtime/runtime-si...
src/runtime/runtime-simd.cc:62: inline float RecipSqrtApprox(float a) { return
1.0f / std::sqrt(a); }
On 2015/08/06 00:12:25, Dan Ehrenberg wrote:
> We may have to revisit these at some point, to create consistent results for
> optimized code vs fullcodegen.

I'll add a TODO.

https://codereview.chromium.org/1230343003/diff/140001/src/runtime/runtime-si...
src/runtime/runtime-simd.cc:429: shift = shift & 0x3f;                          
                   \
On 2015/08/06 00:12:25, Dan Ehrenberg wrote:
> Looks like you're implementing wrap-around behavior here (like JS <<).
However, 
> SIMD operations max out at 0 for shifting too far left; see
> http://littledan.github.io/simd.html#simd-shift-left-by-scalar . This is done
to
> match what the processors apparently do. Analogous behavior exists on the
right
> shifts.

Done.

[bbudge, 2015-08-06 15:13:54]

It looks like CalculateFirstPageSizes needs to be updated.
https://code.google.com/p/chromium/codesearch#chromium/src/v8/src/snapshot/sn...

Example bot failure:
http://build.chromium.org/p/tryserver.v8/builders/v8_linux64_rel/builds/8429

[bbudge, 2015-08-06 17:55:44]

The failures are because I forgot to remove a hack in heap.cc that requests
twice the amount of memory for the name table.

PTAL

[Dan Ehrenberg, 2015-08-06 20:02:14]

lgtm

[bbudge, 2015-08-06 20:23:02]

The CQ bit was checked by bbudge@chromium.org

[bbudge, 2015-08-06 20:23:02]

The patchset sent to the CQ was uploaded after l-g-t-m from
rossberg@chromium.org
Link to the patchset: https://codereview.chromium.org/1230343003/#ps180001
(title: "Remove name table size hack.")

[commit-bot: I haz the power, 2015-08-06 20:23:14]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1230343003/180001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1230343003/180001

[commit-bot: I haz the power, 2015-08-06 20:24:52]

Committed patchset #4 (id:180001)

[commit-bot: I haz the power, 2015-08-06 20:25:05]

Patchset 4 (id:??) landed as
https://crrev.com/6378f57b30f1141850ef7c260972ae109450b023
Cr-Commit-Position: refs/heads/master@{#30051}