[ARM] Implement irregular vector shuffles for SIMD.

[ARM] Implement irregular vector shuffles for SIMD.
- S32x4Shuffle by decomposing into s-register moves if no patterns match.
- S16x8Shuffle, S8x16Shuffle implemented with vtbl if no patterns match.

LOG=N
BUG=v8:6020
Review-Url: https://codereview.chromium.org/2856363003
Cr-Commit-Position: refs/heads/master@{#45210}
Committed: https://chromium.googlesource.com/v8/v8/+/a459f188fab8e026cc7f19bb5eacd01b12699da1

[bbudge, 2017-05-04 17:29:49]

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

[commit-bot: I haz the power, 2017-05-04 17:29:52]

Dry run: CQ is trying da patch.

Follow status at:
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[commit-bot: I haz the power, 2017-05-04 18:02:01]

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

[commit-bot: I haz the power, 2017-05-04 18:02:02]

Dry run: This issue passed the CQ dry run.

[bbudge, 2017-05-04 19:31:45]

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

[commit-bot: I haz the power, 2017-05-04 19:31:50]

Dry run: CQ is trying da patch.

Follow status at:
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[commit-bot: I haz the power, 2017-05-04 19:59:46]

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

[commit-bot: I haz the power, 2017-05-04 19:59:47]

Dry run: This issue passed the CQ dry run.

[bbudge, 2017-05-04 20:14:14]

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

[commit-bot: I haz the power, 2017-05-04 20:14:15]

Dry run: CQ is trying da patch.

Follow status at:
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[bbudge, 2017-05-04 20:16:34]

bbudge@chromium.org changed reviewers:
+ bmeurer@chromium.org, martyn.capewell@arm.com

[bbudge, 2017-05-04 20:16:35]

+Benedikt for compiler OWNERS.

[commit-bot: I haz the power, 2017-05-04 20:39:22]

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

[commit-bot: I haz the power, 2017-05-04 20:39:23]

Dry run: This issue passed the CQ dry run.

[martyn.capewell, 2017-05-05 14:28:49]

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
File src/compiler/arm/code-generator-arm.cc (right):

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
src/compiler/arm/code-generator-arm.cc:2222: __ VmovExtended(dst_code + i,
src_code + lane, kScratchReg);
This will become expensive when each s-register move goes via kScratchReg -
core-vfp and vfp-core moves can take longer than core-core or vfp-vfp.

A follow up patch to improve VmovExtended could try using combinations of vdup,
vsli and vsri to eliminate the need for a core scratch register.

e.g. move s40 (a) to s43: d20 = [ X | a ], d21 = [ Y | Z ], use "vsli d21, d20,
#32" to give d21 = [ a | Z ].

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
src/compiler/arm/code-generator-arm.cc:2307: bit_cast<float>(mask));
This may cause a problem - if your mask looks like a signalling NaN, it could be
quietened on passing it to vmov. I had a similar problem in the ARM64 simulator.
The outcome was some of the test results differed between hardware and
simulation.

A vmov-immediate function that accepts uint32_t would help.

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
src/compiler/arm/code-generator-arm.cc:2313: __ vmov(dst, kScratchQuadReg);
If you know dst doesn't alias src0 or src1, vtbl can write into dst directly.

[bbudge, 2017-05-05 20:36:05]

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

[commit-bot: I haz the power, 2017-05-05 20:36:07]

Dry run: CQ is trying da patch.

Follow status at:
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[bbudge, 2017-05-05 20:36:28]

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
File src/compiler/arm/code-generator-arm.cc (right):

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
src/compiler/arm/code-generator-arm.cc:2222: __ VmovExtended(dst_code + i,
src_code + lane, kScratchReg);
On 2017/05/05 14:28:49, martyn.capewell wrote:
> This will become expensive when each s-register move goes via kScratchReg -
> core-vfp and vfp-core moves can take longer than core-core or vfp-vfp.
> 
> A follow up patch to improve VmovExtended could try using combinations of
vdup,
> vsli and vsri to eliminate the need for a core scratch register.
> 
> e.g. move s40 (a) to s43: d20 = [ X | a ], d21 = [ Y | Z ], use "vsli d21,
d20,
> #32" to give d21 = [ a | Z ].

Yes, I think I have a TODO to improve VMovExtended. It's also potentially used a
lot for gap moves. I'll get on that.

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
src/compiler/arm/code-generator-arm.cc:2307: bit_cast<float>(mask));
On 2017/05/05 14:28:49, martyn.capewell wrote:
> This may cause a problem - if your mask looks like a signalling NaN, it could
be
> quietened on passing it to vmov. I had a similar problem in the ARM64
simulator.
> The outcome was some of the test results differed between hardware and
> simulation.
> 
> A vmov-immediate function that accepts uint32_t would help.

I could also finesse the code so NaNs can't be generated. We can guarantee that
byte indices are in [0, 31] - but I'll have to mask them above. This way there's
a very slight chance (depending on the mask) that we can leverage the more
efficient encoding this move supports.

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
src/compiler/arm/code-generator-arm.cc:2313: __ vmov(dst, kScratchQuadReg);
On 2017/05/05 14:28:49, martyn.capewell wrote:
> If you know dst doesn't alias src0 or src1, vtbl can write into dst directly.

Nice, done.

[commit-bot: I haz the power, 2017-05-05 21:12:06]

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

[commit-bot: I haz the power, 2017-05-05 21:12:07]

Dry run: This issue passed the CQ dry run.

[martyn.capewell, 2017-05-08 13:33:53]

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
File src/compiler/arm/code-generator-arm.cc (right):

https://codereview.chromium.org/2856363003/diff/60001/src/compiler/arm/code-g...
src/compiler/arm/code-generator-arm.cc:2307: bit_cast<float>(mask));
That will fix the NaN problem. However, the assembler doesn't support some of
the other non-FP vmov immediate forms that might be useful in this context. See
"A7.4.6 One register and a modified immediate value" in the ARM ARM for more
info.

Not for this patch, though.

[martyn.capewell, 2017-05-08 13:34:09]

lgtm

[bbudge, 2017-05-09 12:38:00]

+Benedikt for OWNERS on compiler.

[Benedikt Meurer, 2017-05-09 13:05:43]

LGTM (rubber-stamped)

[bbudge, 2017-05-09 20:33:52]

The CQ bit was checked by bbudge@chromium.org

[commit-bot: I haz the power, 2017-05-09 20:34:00]

CQ is trying da patch.

Follow status at:
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[commit-bot: I haz the power, 2017-05-09 21:04:28]

CQ is committing da patch.

Bot data: {"patchset_id": 80001, "attempt_start_ts": 1494362032281930,
"parent_rev": "e9a05114a868c240ead8d6b3c23ca4e97f4ed476", "commit_rev":
"a459f188fab8e026cc7f19bb5eacd01b12699da1"}

[commit-bot: I haz the power, 2017-05-09 21:04:35]

Description was changed from

==========
[ARM] Implement irregular vector shuffles for SIMD.
- S32x4Shuffle by decomposing into s-register moves if no patterns match.
- S16x8Shuffle, S8x16Shuffle implemented with vtbl if no patterns match.

LOG=N
BUG=v8:6020
==========

to

==========
[ARM] Implement irregular vector shuffles for SIMD.
- S32x4Shuffle by decomposing into s-register moves if no patterns match.
- S16x8Shuffle, S8x16Shuffle implemented with vtbl if no patterns match.

LOG=N
BUG=v8:6020

Review-Url: https://codereview.chromium.org/2856363003
Cr-Commit-Position: refs/heads/master@{#45210}
Committed:
https://chromium.googlesource.com/v8/v8/+/a459f188fab8e026cc7f19bb5eacd01b126...
==========

[commit-bot: I haz the power, 2017-05-09 21:04:37]

Committed patchset #5 (id:80001) as
https://chromium.googlesource.com/v8/v8/+/a459f188fab8e026cc7f19bb5eacd01b126...