Process two 32-bit digits as one 64-bit digit in all bigint intrinsics on ARM64.

Process two 32-bit digits as one 64-bit digit in all bigint intrinsics on ARM64.

R=zra@google.com

Committed: https://code.google.com/p/dart/source/detail?r=42611

[regis, 2015-01-03 03:21:29]

regis@google.com changed reviewers:
+ zra@google.com

[regis, 2015-01-03 03:21:30]

Same change for arm64 as earlier change for x64. However, since arm64 does not
have a 128-bit by 64-bit division instruction, the quotient estimation is quite
more complicated. It requires two 64-bit by 32-bit divisions and two 64-bit by
64-bit multiplications to adjust the two 32-bit digits of the estimated
quotient.

Thanks,
Regis

[zra, 2015-01-05 17:00:28]

https://codereview.chromium.org/817583003/diff/1/runtime/vm/intrinsifier_arm6...
File runtime/vm/intrinsifier_arm64.cc (right):

https://codereview.chromium.org/817583003/diff/1/runtime/vm/intrinsifier_arm6...
runtime/vm/intrinsifier_arm64.cc:1155: void
Intrinsifier::Bigint_estQuotientDigit(Assembler* assembler) {
There are a lot more instruction here. I think I should compare the two
implementation on the device to see which is faster. Would the RSA benchmark be
able to measure that? If not, can you recommend a benchmark I could run that
could measure that?

[regis, 2015-01-05 17:12:09]

On 2015/01/05 17:00:28, zra wrote:
>
https://codereview.chromium.org/817583003/diff/1/runtime/vm/intrinsifier_arm6...
> File runtime/vm/intrinsifier_arm64.cc (right):
> 
>
https://codereview.chromium.org/817583003/diff/1/runtime/vm/intrinsifier_arm6...
> runtime/vm/intrinsifier_arm64.cc:1155: void
> Intrinsifier::Bigint_estQuotientDigit(Assembler* assembler) {
> There are a lot more instruction here. I think I should compare the two
> implementation on the device to see which is faster. Would the RSA benchmark
be
> able to measure that? If not, can you recommend a benchmark I could run that
> could measure that?

Yes, there are many instructions and the adjustment loops may run twice each.
However, this is required in order for the other intrinsics to process digit
pairs (mulAdd, sqrAdd, mulMod).

It would be great if you can measure on arm64 hardware, with and without this
change. RSA is perfect as a benchmark.
Note that you will need to patch in https://codereview.chromium.org/786933009/
first, since it contains necessary changes to bigint.dart.

Thanks!
Regis

[zra, 2015-01-05 19:40:47]

lgtm

Ah, ok, I see. I remember you saying now that the intrinsics must all agree
about the number of digits processed.

~40% improvement on arm64 hardware on RSA with these changes.

[regis, 2015-01-05 20:46:29]

On 2015/01/05 19:40:47, zra wrote:
> lgtm
> 
> Ah, ok, I see. I remember you saying now that the intrinsics must all agree
> about the number of digits processed.
> 
> ~40% improvement on arm64 hardware on RSA with these changes.

Thanks for measuring!
Regis

[regis, 2015-01-05 20:46:53]

Committed patchset #2 (id:20001) manually as r42611.