Enables many language tests for arm64.

Enables many language tests for arm64.

- Adds msub and smulh instructions for implementing
BinarySmiOp.
- Adds fmovdd to move doubles between V registers.
- Implements many stubs and IL instructions. These are
ported from ARM with small fix-ups. Deopt code comes
from x64 because the stack frame layout is the same.

R=regis@google.com

Committed: https://code.google.com/p/dart/source/detail?r=35753

[zra, 2014-05-05 16:40:10]

Sorry about the size of this CL. I thought it would be best to push forward
until the number of language test failures had been whittled down a bit.

[regis, 2014-05-05 17:27:56]

LGTM

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
File runtime/vm/intermediate_language_arm64.cc (right):

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:169: __ csel(result, TMP, result,
true_condition);
Can't you use ZR instead of result as input? And remove the eor above?
I guess your TODO will take care of this, but why is result also specified as
input in cinc?

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:474: EmitBranchOnCondition(compiler, 
true_condition, labels);
extra space

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:483: __ Bind(&done);
Maybe not for this changelist, but it seems that better code could be emitted.
There is a lot of branching going on.

Also, I suspect that the addresses of Bool::False() and Bool::True() only differ
in the least significant byte. We could try to make use of that.

Instead of
cond branch
load false (4 instructions)
branch
load true (4 instructions)

we could have
load false (4 instructions)
cond branch
fix lsbyte (1 instruction)

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:609: __ Bind(&done);
ditto

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:662: __ Bind(&done);
ditto

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:2747: __ csel(result, TMP2, result,
GE);
Can't you save an instruction here?

       __ sub(TMP, result, Operand(right));
       __ add(/*TMP2*/ result, result, Operand(right));
       __ csel(result, TMP, result, LT);
       /*__ csel(result, TMP2, result, GE);*/

[regis, 2014-05-05 17:48:20]

On 2014/05/05 17:27:56, regis wrote:
> LGTM
> 
>
https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
> File runtime/vm/intermediate_language_arm64.cc (right):
> 
>
https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
> runtime/vm/intermediate_language_arm64.cc:169: __ csel(result, TMP, result,
> true_condition);
> Can't you use ZR instead of result as input? And remove the eor above?
> I guess your TODO will take care of this, but why is result also specified as
> input in cinc?
> 
>
https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
> runtime/vm/intermediate_language_arm64.cc:474: EmitBranchOnCondition(compiler,

> true_condition, labels);
> extra space
> 
>
https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
> runtime/vm/intermediate_language_arm64.cc:483: __ Bind(&done);
> Maybe not for this changelist, but it seems that better code could be emitted.
> There is a lot of branching going on.
> 
> Also, I suspect that the addresses of Bool::False() and Bool::True() only
differ
> in the least significant byte. We could try to make use of that.
> 
> Instead of
> cond branch
> load false (4 instructions)
> branch
> load true (4 instructions)
> 
> we could have
> load false (4 instructions)
> cond branch
> fix lsbyte (1 instruction)
>
Forget my comments about instruction counts. Since we load from the pool, it is
not 4 instructions, but only one.

 
>
https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
> runtime/vm/intermediate_language_arm64.cc:609: __ Bind(&done);
> ditto
> 
>
https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
> runtime/vm/intermediate_language_arm64.cc:662: __ Bind(&done);
> ditto
> 
>
https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
> runtime/vm/intermediate_language_arm64.cc:2747: __ csel(result, TMP2, result,
> GE);
> Can't you save an instruction here?
> 
>        __ sub(TMP, result, Operand(right));
>        __ add(/*TMP2*/ result, result, Operand(right));
>        __ csel(result, TMP, result, LT);
>        /*__ csel(result, TMP2, result, GE);*/

[zra, 2014-05-05 17:52:11]

Thanks!

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
File runtime/vm/intermediate_language_arm64.cc (right):

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:169: __ csel(result, TMP, result,
true_condition);
On 2014/05/05 17:27:56, regis wrote:
> Can't you use ZR instead of result as input? And remove the eor above?
> I guess your TODO will take care of this, but why is result also specified as
> input in cinc? 

Ah, yes. I've eliminated the eor by replacing result with ZR in csel, and in
cinc in the TODO.

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:474: EmitBranchOnCondition(compiler, 
true_condition, labels);
On 2014/05/05 17:27:56, regis wrote:
> extra space

Done.

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:483: __ Bind(&done);
On 2014/05/05 17:27:56, regis wrote:
> Maybe not for this changelist, but it seems that better code could be emitted.
> There is a lot of branching going on.
> 
> Also, I suspect that the addresses of Bool::False() and Bool::True() only
differ
> in the least significant byte. We could try to make use of that.
> 
> Instead of
> cond branch
> load false (4 instructions)
> branch
> load true (4 instructions)
> 
> we could have
> load false (4 instructions)
> cond branch
> fix lsbyte (1 instruction)

Loads of Bool::True and Bool::False will be only one instruction each, loading
from the object pool. I'd still like to eliminate the branches here using the
csel instruction. Added a TODO for that.

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:609: __ Bind(&done);
On 2014/05/05 17:27:56, regis wrote:
> ditto

TODO added.

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:662: __ Bind(&done);
On 2014/05/05 17:27:56, regis wrote:
> ditto

TODO added.

https://codereview.chromium.org/263243002/diff/1/runtime/vm/intermediate_lang...
runtime/vm/intermediate_language_arm64.cc:2747: __ csel(result, TMP2, result,
GE);
On 2014/05/05 17:27:56, regis wrote:
> Can't you save an instruction here?
> 
>        __ sub(TMP, result, Operand(right));
>        __ add(/*TMP2*/ result, result, Operand(right));
>        __ csel(result, TMP, result, LT);
>        /*__ csel(result, TMP2, result, GE);*/

Done.

[zra, 2014-05-05 18:09:16]

Committed patchset #2 manually as r35753 (presubmit successful).