Optimize double min/max by reducing the code size (reusing left register as result removes the need…

Optimize double min/max by reducing the code size (reusing left register as result removes the need to copy left into result). Add code for two Smi arguments.

R=regis@google.com, zra@google.com

Committed: https://code.google.com/p/dart/source/detail?r=25367

[regis, 2013-07-23 19:45:47]

LGTM

https://codereview.chromium.org/19695007/diff/1/runtime/vm/intermediate_langu...
File runtime/vm/intermediate_language_arm.cc (right):

https://codereview.chromium.org/19695007/diff/1/runtime/vm/intermediate_langu...
runtime/vm/intermediate_language_arm.cc:3428: __ mov(result,
ShifterOperand(right), GE);
GT would be more intuitive than GE, but it does not make any difference in the
number of cycles or in correctness.

[zra, 2013-07-23 19:51:49]

lgtm

[srdjan, 2013-07-23 20:34:01]

https://codereview.chromium.org/19695007/diff/1/runtime/vm/intermediate_langu...
File runtime/vm/intermediate_language_arm.cc (right):

https://codereview.chromium.org/19695007/diff/1/runtime/vm/intermediate_langu...
runtime/vm/intermediate_language_arm.cc:3428: __ mov(result,
ShifterOperand(right), GE);
On 2013/07/23 19:45:47, regis wrote:
> GT would be more intuitive than GE, but it does not make any difference in the
> number of cycles or in correctness.

Done.

[srdjan, 2013-07-23 20:40:45]

Committed patchset #2 manually as r25367 (presubmit successful).

[sra1, 2013-07-23 20:53:27]

https://codereview.chromium.org/19695007/diff/9001/runtime/vm/intermediate_la...
File runtime/vm/intermediate_language_ia32.cc (right):

https://codereview.chromium.org/19695007/diff/9001/runtime/vm/intermediate_la...
runtime/vm/intermediate_language_ia32.cc:3718: __ j(LESS_EQUAL, &done,
Assembler::kNearJump);
CMOVLE result, right

Or is there some problem with CMOVcc?  I know CMOVcc r,[mem] does a load
regardless but this is not that case.

[srdjan, 2013-07-23 21:22:21]

https://codereview.chromium.org/19695007/diff/9001/runtime/vm/intermediate_la...
File runtime/vm/intermediate_language_ia32.cc (right):

https://codereview.chromium.org/19695007/diff/9001/runtime/vm/intermediate_la...
runtime/vm/intermediate_language_ia32.cc:3718: __ j(LESS_EQUAL, &done,
Assembler::kNearJump);
On 2013/07/23 20:53:27, sra1 wrote:
> CMOVLE result, right
> 
> Or is there some problem with CMOVcc?  I know CMOVcc r,[mem] does a load
> regardless but this is not that case.

Good suggestion. Currently we do not have cmovcc implemented in the assembler
(except signed test), but I will do that in a next CL.

[Vyacheslav Egorov (Google), 2013-07-24 09:33:41]

drive by:

I wonder this new instruction can be merged with IfThenElseInstr to keep the
size of the IR down?