ARM64: Use default-NaN mode to canonicalize NaNs.

ARM64: Use default-NaN mode to canonicalize NaNs.

BUG=
R=ulan@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=21156

[jbramley, 2014-04-30 14:10:54]

This speeds up Kraken by about 2.3-2.7%. Most of that improvement comes from
desaturate and fft.

[jbramley, 2014-04-30 14:41:10]

not lgtm

I need to update test-assembler-arm64 because some tests check NaN propagation,
which of course doesn't happen now.

[jbramley, 2014-04-30 16:26:45]

On 2014/04/30 14:41:10, jbramley wrote:
> not lgtm
> 
> I need to update test-assembler-arm64 because some tests check NaN
propagation,
> which of course doesn't happen now.

Done.

[ulan, 2014-05-02 09:10:32]

https://codereview.chromium.org/255343004/diff/40001/src/arm64/macro-assemble...
File src/arm64/macro-assembler-arm64.cc (right):

https://codereview.chromium.org/255343004/diff/40001/src/arm64/macro-assemble...
src/arm64/macro-assembler-arm64.cc:1264: Msr(FPCR, fpcr);
AssertFPCRState also checks flush-to-zero and rounding mode, but this function
doesn't. Is this asymmetry because we do not restore FPCR and changing rounding
mode could lead to more incorrect results than changing NAN mode?

Could you please add comment explaining that?

https://codereview.chromium.org/255343004/diff/40001/src/arm64/macro-assemble...
src/arm64/macro-assembler-arm64.cc:1275: // With DN=1 and RMode=FPTieEven,
subtracting 0.0 preserves all inputs expect
s/expect/except

[jbramley, 2014-05-02 09:32:04]

Thanks Ulan!

https://codereview.chromium.org/255343004/diff/40001/src/arm64/macro-assemble...
File src/arm64/macro-assembler-arm64.cc (right):

https://codereview.chromium.org/255343004/diff/40001/src/arm64/macro-assemble...
src/arm64/macro-assembler-arm64.cc:1264: Msr(FPCR, fpcr);
On 2014/05/02 09:10:33, ulan wrote:
> AssertFPCRState also checks flush-to-zero and rounding mode, but this function
> doesn't. Is this asymmetry because we do not restore FPCR and changing
rounding
> mode could lead to more incorrect results than changing NAN mode?
> 
> Could you please add comment explaining that?

The flush-to-zero and rounding mode settings are left at the default value, so
AssertFPCRState checks that the default is appropriate. (The simulator doesn't
support non-default values for those bits anyway). Also, yes, we would get
incorrect results even in C++ code if we changed those bits: JavaScript requires
subnormal support (not-flush-to-zero) and round-to-nearest mode for almost all
arithmetic operations.

I'll add a comment.

https://codereview.chromium.org/255343004/diff/40001/src/arm64/macro-assemble...
src/arm64/macro-assembler-arm64.cc:1275: // With DN=1 and RMode=FPTieEven,
subtracting 0.0 preserves all inputs expect
On 2014/05/02 09:10:33, ulan wrote:
> s/expect/except

Done.

[jbramley, 2014-05-02 10:34:26]

On 2014/05/02 09:32:04, jbramley wrote:
> Thanks Ulan!
> 
>
https://codereview.chromium.org/255343004/diff/40001/src/arm64/macro-assemble...
> File src/arm64/macro-assembler-arm64.cc (right):
> 
>
https://codereview.chromium.org/255343004/diff/40001/src/arm64/macro-assemble...
> src/arm64/macro-assembler-arm64.cc:1264: Msr(FPCR, fpcr);
> On 2014/05/02 09:10:33, ulan wrote:
> > AssertFPCRState also checks flush-to-zero and rounding mode, but this
function
> > doesn't. Is this asymmetry because we do not restore FPCR and changing
> rounding
> > mode could lead to more incorrect results than changing NAN mode?
> > 
> > Could you please add comment explaining that?
> 
> The flush-to-zero and rounding mode settings are left at the default value, so
> AssertFPCRState checks that the default is appropriate. (The simulator doesn't
> support non-default values for those bits anyway). Also, yes, we would get
> incorrect results even in C++ code if we changed those bits: JavaScript
requires
> subnormal support (not-flush-to-zero) and round-to-nearest mode for almost all
> arithmetic operations.
> 
> I'll add a comment.
> 
>
https://codereview.chromium.org/255343004/diff/40001/src/arm64/macro-assemble...
> src/arm64/macro-assembler-arm64.cc:1275: // With DN=1 and RMode=FPTieEven,
> subtracting 0.0 preserves all inputs expect
> On 2014/05/02 09:10:33, ulan wrote:
> > s/expect/except
> 
> Done.

All done. Sorry, there's quite a bit of rebase noise in the delta view.

[ulan, 2014-05-05 07:45:19]

> All done. Sorry, there's quite a bit of rebase noise in the delta view.
No problem, lgtm.

[jbramley, 2014-05-06 08:05:42]

Committed patchset #4 manually as r21156 (presubmit successful).