[x64] Use vmovapd and vmovsd when AVX is enabled.

[x64] Use vmovapd and vmovsd when AVX is enabled.

R=bmeurer@chromium.org
BUG=v8:4406
LOG=N

Committed: https://crrev.com/5cda2bcfabc24f988f57ecfc63abc83006245597
Cr-Commit-Position: refs/heads/master@{#31275}

[alph, 2015-10-14 23:32:05]

alph@chromium.org changed reviewers:
+ bmeurer@chromium.org, danno@chromium.org, yurys@chromium.org

[alph, 2015-10-14 23:32:43]

The CQ bit was checked by alph@chromium.org to run a CQ dry run

[commit-bot: I haz the power, 2015-10-14 23:32:54]

Dry run: CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1391963005/1
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1391963005/1

[commit-bot: I haz the power, 2015-10-14 23:43:12]

The CQ bit was unchecked by commit-bot@chromium.org

[commit-bot: I haz the power, 2015-10-14 23:43:13]

Dry run: Try jobs failed on following builders:
  v8_win64_rel on tryserver.v8 (JOB_FAILED,
http://build.chromium.org/p/tryserver.v8/builders/v8_win64_rel/builds/9580)

[Benedikt Meurer, 2015-10-15 03:51:04]

Nice. LGTM.

[alph, 2015-10-15 03:52:47]

The CQ bit was checked by alph@chromium.org

[commit-bot: I haz the power, 2015-10-15 03:52:54]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1391963005/20001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1391963005/20001

[commit-bot: I haz the power, 2015-10-15 04:14:36]

Committed patchset #2 (id:20001)

[commit-bot: I haz the power, 2015-10-15 04:15:10]

Patchset 2 (id:??) landed as
https://crrev.com/5cda2bcfabc24f988f57ecfc63abc83006245597
Cr-Commit-Position: refs/heads/master@{#31275}

[Weiliang, 2015-10-15 05:39:23]

I wonder if you have measure the performance impact on various benchmarks since
after this patch, the JIT code generated from full codegen and crankshaft also
mix SSE and AVX instructions.

Thanks
-Weiliang

[Weiliang, 2015-10-15 05:41:33]

On 2015/10/15 05:39:23, Weiliang wrote:
> I wonder if you have measure the performance impact on various benchmarks
since
> after this patch, the JIT code generated from full codegen and crankshaft also
> mix SSE and AVX instructions.
> 
> Thanks
> -Weiliang

While originally, we only use AVX for TurboFan. Also TurboFan has internal
switch to turn on/off backend features, which I think is broken now.

[Weiliang, 2015-10-15 05:41:35]

On 2015/10/15 05:39:23, Weiliang wrote:
> I wonder if you have measure the performance impact on various benchmarks
since
> after this patch, the JIT code generated from full codegen and crankshaft also
> mix SSE and AVX instructions.
> 
> Thanks
> -Weiliang

While originally, we only use AVX for TurboFan. Also TurboFan has internal
switch to turn on/off backend features, which I think is broken now.

[alph, 2015-10-16 00:51:55]

On 2015/10/15 05:39:23, Weiliang wrote:
> I wonder if you have measure the performance impact on various benchmarks
since
> after this patch, the JIT code generated from full codegen and crankshaft also
> mix SSE and AVX instructions.
> 
> Thanks
> -Weiliang

Sure. Done some benchmarking. The impact is near zero and within the std
deviation.
https://docs.google.com/spreadsheets/d/1XV4wf49GR5C_OTLFJnfUR0RPgqfqIiHzIimO4...

[alph, 2015-10-16 00:56:16]

On 2015/10/15 05:41:35, Weiliang wrote:
> On 2015/10/15 05:39:23, Weiliang wrote:
> > I wonder if you have measure the performance impact on various benchmarks
> since
> > after this patch, the JIT code generated from full codegen and crankshaft
also
> > mix SSE and AVX instructions.
> > 
> > Thanks
> > -Weiliang
> 
> While originally, we only use AVX for TurboFan. Also TurboFan has internal
> switch to turn on/off backend features, which I think is broken now.

Hmm, what is the flag? Strange if something broke, I seem to be using the same
techniques used for e.g. addsd -> vaddsd switching.

[Benedikt Meurer, 2015-10-16 03:37:47]

On 2015/10/16 00:56:16, alph wrote:
> On 2015/10/15 05:41:35, Weiliang wrote:
> > On 2015/10/15 05:39:23, Weiliang wrote:
> > > I wonder if you have measure the performance impact on various benchmarks
> > since
> > > after this patch, the JIT code generated from full codegen and crankshaft
> also
> > > mix SSE and AVX instructions.
> > > 
> > > Thanks
> > > -Weiliang
> > 
> > While originally, we only use AVX for TurboFan. Also TurboFan has internal
> > switch to turn on/off backend features, which I think is broken now.
> 
> Hmm, what is the flag? Strange if something broke, I seem to be using the same
> techniques used for e.g. addsd -> vaddsd switching.

The CpuFeatures flag passed to the InstructionSelector. But I think your patch
is OK. The AVX flag is used in the InstructionSelector to pick better register
allocation constraints for (v)addsd and friends. There's no difference for
(v)movsd and friends from the POV of the register allocator.

[Weiliang, 2015-10-19 02:59:09]

On 2015/10/16 03:37:47, Benedikt Meurer wrote:
> On 2015/10/16 00:56:16, alph wrote:
> > On 2015/10/15 05:41:35, Weiliang wrote:
> > > On 2015/10/15 05:39:23, Weiliang wrote:
> > > > I wonder if you have measure the performance impact on various
benchmarks
> > > since
> > > > after this patch, the JIT code generated from full codegen and
crankshaft
> > also
> > > > mix SSE and AVX instructions.
> > > > 
> > > > Thanks
> > > > -Weiliang
> > > 
> > > While originally, we only use AVX for TurboFan. Also TurboFan has internal
> > > switch to turn on/off backend features, which I think is broken now.
> > 
> > Hmm, what is the flag? Strange if something broke, I seem to be using the
same
> > techniques used for e.g. addsd -> vaddsd switching.
> 
> The CpuFeatures flag passed to the InstructionSelector. But I think your patch
> is OK. The AVX flag is used in the InstructionSelector to pick better register
> allocation constraints for (v)addsd and friends. There's no difference for
> (v)movsd and friends from the POV of the register allocator.
Well. mixing AVX-128 and legacy SSE is OK only if there is no AVX-256. V8 JIT
code does not have any AVX-256, so the bug 4406 seems more to be the OS issue. 
But I agree to convert all legacy SSE to AVX-128. Thanks for the job and it
would be great if there is also IA32 port.

Thanks
-Weiliang