[vm] Implement more efficient CAS in simarm/simarm64 modes

[vm] Implement more efficient CAS in simarm/simarm64 modes

When running on the simulator for arm/arm64, compare-and-swap operations
used in VM were synchronized with simulator using mutex.
It heavily impacts performance when doing parallel marking due to high
contention (see #30317).
This synchronization was implemented in order to make simulated
LDREX/STREX instructions aware of CAS performed in VM.

This CL drops this synchronization between VM and simulator:
CAS operations in VM become regular, and simulator remembers value loaded
with load-exclusive and performs CAS when doing store-exclusive to catch
any concurrent modifications.

Speeds up gen_snapshot of flutter benchmark complex_layout from 9.2s
to 8.0s on my MacBook.

R=asiva@google.com, rmacnak@google.com

Committed: https://github.com/dart-lang/sdk/commit/db26c8934799de6205b1c9c9dc498d8fc58a0d07

[alexmarkov, 2017-08-11 23:35:22]

Description was changed from

==========
[vm] Implement more efficient CAS in simarm/simarm64 modes

When running on the simulator for arm/arm64, compare-and-swap operations
used in VM were synchronized with simulator using mutex.
It heavily impacts performance when doing parallel marking due to high
contention (see #30317).
This synchronization was implemented in order to make simulated
LDREX/STREX instructions aware of CAS performed in VM.

This CL drops this synchronization between VM and simulator:
CAS operations in VM become regular, and simulator remembers value loaded
with load-exclusive and performs CAS when doing store-exclusive to catch
any concurrent modifications.
==========

to

==========
[vm] Implement more efficient CAS in simarm/simarm64 modes

When running on the simulator for arm/arm64, compare-and-swap operations
used in VM were synchronized with simulator using mutex.
It heavily impacts performance when doing parallel marking due to high
contention (see #30317).
This synchronization was implemented in order to make simulated
LDREX/STREX instructions aware of CAS performed in VM.

This CL drops this synchronization between VM and simulator:
CAS operations in VM become regular, and simulator remembers value loaded
with load-exclusive and performs CAS when doing store-exclusive to catch
any concurrent modifications.

Speeds up gen_snapshot of flutter benchmark complex_layout from 9.2s
to 8.0s on my MacBook.
==========

[alexmarkov, 2017-08-11 23:36:07]

alexmarkov@google.com changed reviewers:
+ asiva@google.com, rmacnak@google.com, zra@google.com

[rmacnak, 2017-08-11 23:44:46]

LGTM

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/atomic.h
File runtime/vm/atomic.h (right):

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/atomic.h#new...
runtime/vm/atomic.h:19: // NOTE: Not to be used for any atomic operations
involving memory locations
These comments about generated code can be removed now.

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/simulator_arm.h
File runtime/vm/simulator_arm.h (right):

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/simulator_ar...
runtime/vm/simulator_arm.h:215: // changes.
// Assumes we don't need ABA detection.

[siva, 2017-08-11 23:58:44]

lgtm

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/simulator_arm.h
File runtime/vm/simulator_arm.h (right):

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/simulator_ar...
runtime/vm/simulator_arm.h:215: // changes.
On 2017/08/11 23:44:46, rmacnak wrote:
> // Assumes we don't need ABA detection.

Will that affect our result here, because we could have a similar situation with
atomic CAS too?

[alexmarkov, 2017-08-12 00:34:55]

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/atomic.h
File runtime/vm/atomic.h (right):

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/atomic.h#new...
runtime/vm/atomic.h:19: // NOTE: Not to be used for any atomic operations
involving memory locations
On 2017/08/11 23:44:46, rmacnak wrote:
> These comments about generated code can be removed now.

Done.

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/simulator_arm.h
File runtime/vm/simulator_arm.h (right):

https://codereview.chromium.org/2995803002/diff/20001/runtime/vm/simulator_ar...
runtime/vm/simulator_arm.h:215: // changes.
On 2017/08/11 23:58:44, siva wrote:
> On 2017/08/11 23:44:46, rmacnak wrote:
> > // Assumes we don't need ABA detection.
> 
> Will that affect our result here, because we could have a similar situation
with
> atomic CAS too?

It looks like in our uses of ldrex/strex instructions the ABA detection is not
needed. Added a clarifying comment.

[alexmarkov, 2017-08-12 00:36:15]

Description was changed from

==========
[vm] Implement more efficient CAS in simarm/simarm64 modes

When running on the simulator for arm/arm64, compare-and-swap operations
used in VM were synchronized with simulator using mutex.
It heavily impacts performance when doing parallel marking due to high
contention (see #30317).
This synchronization was implemented in order to make simulated
LDREX/STREX instructions aware of CAS performed in VM.

This CL drops this synchronization between VM and simulator:
CAS operations in VM become regular, and simulator remembers value loaded
with load-exclusive and performs CAS when doing store-exclusive to catch
any concurrent modifications.

Speeds up gen_snapshot of flutter benchmark complex_layout from 9.2s
to 8.0s on my MacBook.
==========

to

==========
[vm] Implement more efficient CAS in simarm/simarm64 modes

When running on the simulator for arm/arm64, compare-and-swap operations
used in VM were synchronized with simulator using mutex.
It heavily impacts performance when doing parallel marking due to high
contention (see #30317).
This synchronization was implemented in order to make simulated
LDREX/STREX instructions aware of CAS performed in VM.

This CL drops this synchronization between VM and simulator:
CAS operations in VM become regular, and simulator remembers value loaded
with load-exclusive and performs CAS when doing store-exclusive to catch
any concurrent modifications.

Speeds up gen_snapshot of flutter benchmark complex_layout from 9.2s
to 8.0s on my MacBook.

R=asiva@google.com, rmacnak@google.com

Committed:
https://github.com/dart-lang/sdk/commit/db26c8934799de6205b1c9c9dc498d8fc58a0d07
==========

[alexmarkov, 2017-08-12 00:36:16]

Committed patchset #3 (id:40001) manually as
db26c8934799de6205b1c9c9dc498d8fc58a0d07 (presubmit successful).