DBC: Enables unboxed doubles

DBC: Enables unboxed doubles

This CL enables unboxed doubles for simdbc64. Unboxed
doubles are stored in regular dbc registers, which are
really 64-bit stack slots. It also implements binary
and unary double operations and comparisons.

R=vegorov@google.com

Committed: https://github.com/dart-lang/sdk/commit/fe2b7b35c4dee55ce2dd236b1b94b6fd833f25aa

[zra, 2016-07-12 17:25:30]

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

[zra, 2016-07-14 15:02:54]

ping

[Vyacheslav Egorov (Google), 2016-07-14 16:06:07]

Hi Zach,

Is this ready for review? 

I was a bit confused by the description. 

If it's ready for review please update description to reflect the changes that
this CL introduces.

[zra, 2016-07-14 16:11:18]

Description was changed from

==========
DBC: Start work on unboxing doubles
==========

to

==========
DBC: Enables unboxed doubles

This CL enables unboxed doubles for simdbc64. Unboxed
doubles are stored in regular dbc registers, which are
really 64-bit stack slots. It also implements binary
and unary double operations and comparisons.
==========

[zra, 2016-07-14 16:11:47]

On 2016/07/14 16:06:07, Vyacheslav Egorov (Google) wrote:
> Hi Zach,
> 
> Is this ready for review? 
> 
> I was a bit confused by the description. 
> 
> If it's ready for review please update description to reflect the changes that
> this CL introduces.

Oh yah. Sorry about that. I've added a better description. PTAL.

[Vyacheslav Egorov (Google), 2016-07-14 16:26:24]

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/constants_d...
File runtime/vm/constants_dbc.h (right):

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/constants_d...
runtime/vm/constants_dbc.h:218: //    Unboxes FP[rD] into FP[rA] and skips the
following instruction unless
I guess the next instruction is always a deoptimization which means instead of
actually embedding deoptimization into the bytecode we could make
CheckUnboxDouble instruction itself deoptimize.

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/deopt_instr...
File runtime/vm/deopt_instructions.cc (right):

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/deopt_instr...
runtime/vm/deopt_instructions.cc:1130: deopt_instr = new(zone()) DeoptWordInstr(
This seems incorrect. DeoptDoubleInstr takes care of materializing Double object
on deoptimization. DeoptWorkInstr would not do it. It just copies things as is.

Please write some microtest that deoptimizes with unboxed double to verify that
it is materialized correctly.

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/simulator_d...
File runtime/vm/simulator_dbc.cc (right):

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/simulator_d...
runtime/vm/simulator_dbc.cc:1699: BYTECODE(BoxDouble, A_D);
I would call this bytecode WriteIntoDouble or something like that.

It can be used when implementing mutable double fields to update boxes.

[zra, 2016-07-14 21:12:10]

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/constants_d...
File runtime/vm/constants_dbc.h (right):

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/constants_d...
runtime/vm/constants_dbc.h:218: //    Unboxes FP[rD] into FP[rA] and skips the
following instruction unless
On 2016/07/14 16:26:24, Vyacheslav Egorov (Google) wrote:
> I guess the next instruction is always a deoptimization which means instead of
> actually embedding deoptimization into the bytecode we could make
> CheckUnboxDouble instruction itself deoptimize.

For now I'd prefer to follow the pattern we have going with the other places we
can deopt.

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/deopt_instr...
File runtime/vm/deopt_instructions.cc (right):

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/deopt_instr...
runtime/vm/deopt_instructions.cc:1130: deopt_instr = new(zone()) DeoptWordInstr(
On 2016/07/14 16:26:24, Vyacheslav Egorov (Google) wrote:
> This seems incorrect. DeoptDoubleInstr takes care of materializing Double
object
> on deoptimization. DeoptWorkInstr would not do it. It just copies things as
is.
> 
> Please write some microtest that deoptimizes with unboxed double to verify
that
> it is materialized correctly.

Done.

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/simulator_d...
File runtime/vm/simulator_dbc.cc (right):

https://codereview.chromium.org/2120703002/diff/100001/runtime/vm/simulator_d...
runtime/vm/simulator_dbc.cc:1699: BYTECODE(BoxDouble, A_D);
On 2016/07/14 16:26:24, Vyacheslav Egorov (Google) wrote:
> I would call this bytecode WriteIntoDouble or something like that.
> 
> It can be used when implementing mutable double fields to update boxes.

Done.

[Vyacheslav Egorov (Google), 2016-07-14 21:33:01]

lgtm

[zra, 2016-07-14 23:05:22]

Description was changed from

==========
DBC: Enables unboxed doubles

This CL enables unboxed doubles for simdbc64. Unboxed
doubles are stored in regular dbc registers, which are
really 64-bit stack slots. It also implements binary
and unary double operations and comparisons.
==========

to

==========
DBC: Enables unboxed doubles

This CL enables unboxed doubles for simdbc64. Unboxed
doubles are stored in regular dbc registers, which are
really 64-bit stack slots. It also implements binary
and unary double operations and comparisons.

R=vegorov@google.com

Committed:
https://github.com/dart-lang/sdk/commit/fe2b7b35c4dee55ce2dd236b1b94b6fd833f25aa
==========

[zra, 2016-07-14 23:05:24]

Committed patchset #8 (id:140001) manually as
fe2b7b35c4dee55ce2dd236b1b94b6fd833f25aa (presubmit successful).