ARM: Implement sqrt in inline assembly.

ARM: Implement sqrt in inline assembly.

Call VSQRT directly to avoid the tiniest (1ulp) precision
error that occurs in the system-supplied sqrt on QNX/ARM.

All precision tests in SunSpider are now passing on this platform.

BUG=
R=bmeurer@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=18506

[c.truta, 2014-01-06 16:28:54]

Although this is for QNX/ARM, I didn't implement it conditionally on V8_OS_QNX.
I think other ARM platforms may also benefit from avoiding the libm indirection
overhead, and the resulting code is simpler, too.

[Benedikt Meurer, 2014-01-07 08:12:33]

LGTM if you address the comment.

https://codereview.chromium.org/98363010/diff/1/src/arm/codegen-arm.cc
File src/arm/codegen-arm.cc (right):

https://codereview.chromium.org/98363010/diff/1/src/arm/codegen-arm.cc#newcod...
src/arm/codegen-arm.cc:354: if (!CpuFeatures::IsSupported(VFP3)) return
&std::sqrt;
This test is unnecessary.

[c.truta, 2014-01-07 08:21:35]

On 2014/01/07 08:12:33, Benedikt Meurer wrote:
>
https://codereview.chromium.org/98363010/diff/1/src/arm/codegen-arm.cc#newcod...
> src/arm/codegen-arm.cc:354: if (!CpuFeatures::IsSupported(VFP3)) return
> &std::sqrt;
> This test is unnecessary.

Fixed, thank you.

[Sven Panne, 2014-01-07 08:27:56]

Please encapsulate the ABI as proposed. Note that we should probably tweak the
macro assembler instruction names a bit, too, but not in this CL.

https://codereview.chromium.org/98363010/diff/1/src/arm/codegen-arm.cc
File src/arm/codegen-arm.cc (right):

https://codereview.chromium.org/98363010/diff/1/src/arm/codegen-arm.cc#newcod...
src/arm/codegen-arm.cc:363: __ vmov(d0, r0, r1);
GetCFunctionDoubleResult(d0) without the if

https://codereview.chromium.org/98363010/diff/1/src/arm/codegen-arm.cc#newcod...
src/arm/codegen-arm.cc:367: __ vmov(r0, r1, d0);
SetCallCDoubleArguments(d0) without the if

[c.truta, 2014-01-07 16:50:23]

On 2014/01/07 08:27:56, Sven Panne wrote:
> Please encapsulate the ABI as proposed. Note that we should probably tweak the
> macro assembler instruction names a bit, too, but not in this CL.

Done.
What would be, in your opinion, the better masm names?

[Sven Panne, 2014-01-08 07:14:07]

On 2014/01/07 16:50:23, cosmin.truta wrote:
> What would be, in your opinion, the better masm names?

Perhaps something like:

   GetCFunctionDoubleResult => FromABI
   SetCallCDoubleArguments => ToABI

Or a bit longer: {From,To}FloatABI? Perhaps our ARM people have a proposal, too.
As it is, it is very confusing, because the current names only make sense for
the caller of some FP function, not for the implementation of such a function.

[Rodolph Perfetta, 2014-01-08 22:31:34]

On 2014/01/08 07:14:07, Sven Panne wrote:
> On 2014/01/07 16:50:23, cosmin.truta wrote:
> > What would be, in your opinion, the better masm names?
> 
> Perhaps something like:
> 
>    GetCFunctionDoubleResult => FromABI
>    SetCallCDoubleArguments => ToABI
> 
> Or a bit longer: {From,To}FloatABI? Perhaps our ARM people have a proposal,
too.
> As it is, it is very confusing, because the current names only make sense for
> the caller of some FP function, not for the implementation of such a function.

The macro would aim to abstract the two variant (softfp and hard) of the
procedure call standard for the ARM eabi, so how about something on the line of
Mov{From,To}FloatParameter and Mov{From, To}FloatResult. I can write a patch if
that sounds good to you.

[Benedikt Meurer, 2014-01-09 06:36:36]

On 2014/01/08 22:31:34, Rodolph Perfetta wrote:
> On 2014/01/08 07:14:07, Sven Panne wrote:
> > On 2014/01/07 16:50:23, cosmin.truta wrote:
> > > What would be, in your opinion, the better masm names?
> > 
> > Perhaps something like:
> > 
> >    GetCFunctionDoubleResult => FromABI
> >    SetCallCDoubleArguments => ToABI
> > 
> > Or a bit longer: {From,To}FloatABI? Perhaps our ARM people have a proposal,
> too.
> > As it is, it is very confusing, because the current names only make sense
for
> > the caller of some FP function, not for the implementation of such a
function.
> 
> The macro would aim to abstract the two variant (softfp and hard) of the
> procedure call standard for the ARM eabi, so how about something on the line
of
> Mov{From,To}FloatParameter and Mov{From, To}FloatResult. I can write a patch
if
> that sounds good to you.

Sounds good to me.

[Benedikt Meurer, 2014-01-09 07:48:05]

Committed patchset #3 manually as r18506 (presubmit successful).

[Sven Panne, 2014-01-09 08:01:21]

On 2014/01/08 22:31:34, Rodolph Perfetta wrote:
> The macro would aim to abstract the two variant (softfp and hard) of the
> procedure call standard for the ARM eabi, so how about something on the line
of
> Mov{From,To}FloatParameter and Mov{From, To}FloatResult. I can write a patch
if
> that sounds good to you.

Sounds good, and the names are definitely better than the ones proposed by me.
:-)