Intrinsify OnebyteString's substringUnchecked. Significant performance improvement on Json benchmar…

Intrinsify OnebyteString's substringUnchecked. Significant performance improvement on Json benchmarks.
Committed: https://code.google.com/p/dart/source/detail?r=21232

[Florian Schneider, 2013-04-10 16:09:51]

LGTM.

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
File runtime/vm/intrinsifier_ia32.cc (right):

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:1504: static void
AllocateOnebyteString(Assembler* assembler,
I think TryAllocateOnebyteString would be a better name.

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:1505: Label* fall_through,
May rename the label to "failure".

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:1549: __ movl(EDI, Immediate(0));
__ xorl(EDI, EDI);

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:1600: __ movb(FieldAddress(EAX, ECX, TIMES_1,
OneByteString::data_offset()), BL);
Consider replacing this loop using

rep movsb

According to the Intel optimization manual, new CPUs have efficient support for
string moves using the rep-prefix.

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_x6...
File runtime/vm/intrinsifier_x64.cc (right):

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_x6...
runtime/vm/intrinsifier_x64.cc:1441: 
Extra new line?

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_x6...
runtime/vm/intrinsifier_x64.cc:1511: // EDI: Start address to copy from
(untagged).
s/EDI/RDI/g

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_x6...
runtime/vm/intrinsifier_x64.cc:1519: // the byte copies into word copies or even
call memcpy.
Same as on ia32: Consider using rep movsb instead of a loop.

[srdjan, 2013-04-10 20:08:05]

Thanks.

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
File runtime/vm/intrinsifier_ia32.cc (right):

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:1504: static void
AllocateOnebyteString(Assembler* assembler,
On 2013/04/10 16:09:51, Florian Schneider wrote:
> I think TryAllocateOnebyteString would be a better name.

Done.

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:1505: Label* fall_through,
On 2013/04/10 16:09:51, Florian Schneider wrote:
> May rename the label to "failure".

Done.

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:1549: __ movl(EDI, Immediate(0));
On 2013/04/10 16:09:51, Florian Schneider wrote:
> __ xorl(EDI, EDI);

Done.

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_ia...
runtime/vm/intrinsifier_ia32.cc:1600: __ movb(FieldAddress(EAX, ECX, TIMES_1,
OneByteString::data_offset()), BL);
On 2013/04/10 16:09:51, Florian Schneider wrote:
> Consider replacing this loop using
> 
> rep movsb
> 
> According to the Intel optimization manual, new CPUs have efficient support
for
> string moves using the rep-prefix.
> 

Added TODO, will do in next CL

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_x6...
File runtime/vm/intrinsifier_x64.cc (right):

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_x6...
runtime/vm/intrinsifier_x64.cc:1441: 
On 2013/04/10 16:09:51, Florian Schneider wrote:
> Extra new line?

Removed

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_x6...
runtime/vm/intrinsifier_x64.cc:1511: // EDI: Start address to copy from
(untagged).
On 2013/04/10 16:09:51, Florian Schneider wrote:
> s/EDI/RDI/g

Done.

https://codereview.chromium.org/13964002/diff/5001/runtime/vm/intrinsifier_x6...
runtime/vm/intrinsifier_x64.cc:1519: // the byte copies into word copies or even
call memcpy.
On 2013/04/10 16:09:51, Florian Schneider wrote:
> Same as on ia32: Consider using rep movsb instead of a loop.

Added TODO.

[srdjan, 2013-04-10 20:37:08]

Committed patchset #4 manually as r21232 (presubmit successful).