ARM: add AArch32 support and new vcvt instructions

ARM: add AArch32 support and new vcvt instructions

R=ulan@chromium.org, bmeurer@chromium.org

[vincent.belliard, 2014-06-11 14:34:25]

First step to use new AArch32 functions. An other patch will then use these new
instructions.

[ulan, 2014-06-20 11:40:07]

[+Ross for Android perspective]

I wonder about cost vs benefit of this change. At the moment we do not test or
run benchmarks for Aarch32. After this change we will have to. I am not sure of
there are embedders that going to run 32-bit V8 on arm64. Do you have any in
mind?

[rmcilroy, 2014-06-20 13:09:32]

On 2014/06/20 11:40:07, ulan wrote:
> [+Ross for Android perspective]
> 
> I wonder about cost vs benefit of this change. At the moment we do not test or
> run benchmarks for Aarch32. After this change we will have to. I am not sure
of
> there are embedders that going to run 32-bit V8 on arm64. Do you have any in
> mind?

From the point of view of Chrome on Android, we don't have any plans to target
aarch32 currently.  The plan is that we run targeting aarch64 on arm-v8 chips,
so I don't think there would be any situations where we would run on an arm-v8
chip and not be using the arm64 backend (this may change in the future of
course).  The Android WebView will run both on 32bit and 64 bit processes, so we
could conceivably target aarch32 instructions there if we probed and found we
were running on an arm-v8 chip.  I'm not sure if this would be worth it though -
Do these instructions make significant performance improvements over armv7?

[vincent.belliard, 2014-06-23 08:31:13]

If you decide one day to re-open this CL, with this CL which enables AArch32
(which means that gcc can use AArch32 and the snapshot can use 32 dregs and
Neon) and a further CL which optimizes Math.round, the Kraken performances are
improved by a little bit less than 1% and the Octane performances are improved
by a little bit more than 2%.