VM: Fix a couple of issues in the AOT optimizer, add fast path smi multiply.

VM: Fix a couple of issues in the AOT optimizer, add fast path smi multiply.

1. Add fast path smi multiply

2. Recognize more operands as likely smi: If they are the result
  another CheckedSmiOp, or the result of a phi that has a known smi.

3. Fix guessing receiver cids of getters/setters based.

4. Try to specialize calls based on propagated receiver type first before
  using guessed cids (which generates a call with checks + megamorphic slow path)


BUG=

Committed: https://github.com/dart-lang/sdk/commit/6d27476cd1e7892ea7ad4b946d6ab97b1d3b399b

[Florian Schneider, 2016-03-22 15:08:59]

fschneider@google.com changed reviewers:
+ srdjan@google.com

[srdjan, 2016-03-22 16:37:23]

LGTMwC and one question

https://codereview.chromium.org/1824023002/diff/1/runtime/vm/aot_optimizer.cc
File runtime/vm/aot_optimizer.cc (left):

https://codereview.chromium.org/1824023002/diff/1/runtime/vm/aot_optimizer.cc...
runtime/vm/aot_optimizer.cc:230: // Check if getter or setter in function's
class and class is currently leaf.
Removing this guessing will have a considerable negative performance impact on
Richard, DeltaBlue, as far as I remember (aot). Please evaluate benefit vs
performance loss before removing this optimization (it may have beed superseded
by other optimizations).

https://codereview.chromium.org/1824023002/diff/1/runtime/vm/aot_optimizer.cc
File runtime/vm/aot_optimizer.cc (right):

https://codereview.chromium.org/1824023002/diff/1/runtime/vm/aot_optimizer.cc...
runtime/vm/aot_optimizer.cc:2536:
thread()->cha()->AddToLeafClasses(receiver_class);
AddToLeafClasses is I think only needed when deoptimizing (RegisterCHACode), and
could be taken out from aot_optimizer, I believe.

[Florian Schneider, 2016-03-22 16:51:03]

https://codereview.chromium.org/1824023002/diff/1/runtime/vm/aot_optimizer.cc
File runtime/vm/aot_optimizer.cc (left):

https://codereview.chromium.org/1824023002/diff/1/runtime/vm/aot_optimizer.cc...
runtime/vm/aot_optimizer.cc:230: // Check if getter or setter in function's
class and class is currently leaf.
On 2016/03/22 16:37:22, srdjan wrote:
> Removing this guessing will have a considerable negative performance impact on
> Richard, DeltaBlue, as far as I remember (aot). Please evaluate benefit vs
> performance loss before removing this optimization (it may have beed
superseded
> by other optimizations).

One question: Is this only meant for getter/setter invocations with "this" as
receiver or any receiver?

I'll check the numbers before landing. If there is a regression, I can also put
this back in for now. It is independent from the rest of the CL.

https://codereview.chromium.org/1824023002/diff/1/runtime/vm/aot_optimizer.cc
File runtime/vm/aot_optimizer.cc (right):

https://codereview.chromium.org/1824023002/diff/1/runtime/vm/aot_optimizer.cc...
runtime/vm/aot_optimizer.cc:2536:
thread()->cha()->AddToLeafClasses(receiver_class);
On 2016/03/22 16:37:22, srdjan wrote:
> AddToLeafClasses is I think only needed when deoptimizing (RegisterCHACode),
and
> could be taken out from aot_optimizer, I believe.

Done.

[Florian Schneider, 2016-03-23 13:06:52]

I fixed the regression on Tracer by adjusting the inlining decisions for numeric
operators. For these we either specialize in the optimizer, or rely on intrinsic
code for the fast path. Inlining these methods does not give any additional
benefit, but makes the generated code unnecessarily bigger.

[Florian Schneider, 2016-03-25 15:52:19]

Description was changed from

==========
VM: Fix a couple of issues in the AOT optimizer, add fast path smi multiply.

1. Add fast path smi multiply

2. Recognize more operands as likely smi: If they are the result
  another CheckedSmiOp, or the result of a phi that has a known smi.

3. Fix guessing receiver cids of getters/setters based.

4. Try to specialize calls based on propagated receiver type first before
  using guessed cids (which generates a call with checks + megamorphic slow
path)


BUG=
==========

to

==========
VM: Fix a couple of issues in the AOT optimizer, add fast path smi multiply.

1. Add fast path smi multiply

2. Recognize more operands as likely smi: If they are the result
  another CheckedSmiOp, or the result of a phi that has a known smi.

3. Fix guessing receiver cids of getters/setters based.

4. Try to specialize calls based on propagated receiver type first before
  using guessed cids (which generates a call with checks + megamorphic slow
path)


BUG=

Committed:
https://github.com/dart-lang/sdk/commit/6d27476cd1e7892ea7ad4b946d6ab97b1d3b399b
==========

[Florian Schneider, 2016-03-25 15:52:21]

Committed patchset #3 (id:40001) manually as
6d27476cd1e7892ea7ad4b946d6ab97b1d3b399b (presubmit successful).