Write-barrier verification on ARM.

Write-barrier verification on ARM.

Add and use methods on Assembler for writing and verifying object fields.
Enable tests in verified_mem mode for simarm.
Also generalize and use InitializeFieldsNoBarrierUnrolled in one more place.

R=iposva@google.com

Committed: https://code.google.com/p/dart/source/detail?r=42208

[Ivan Posva, 2014-12-08 19:27:56]

LGTM

-Ivan

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/assembler_arm.cc
File runtime/vm/assembler_arm.cc (right):

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/assembler_arm...
runtime/vm/assembler_arm.cc:1710: Register old_value = AllocateRegister(&used);
How do you prevent allocating one of the pre-assigned, non-allocatable registers
(e.g. SP) being used here?

Maybe we should not initialize with 0?

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/intermediate_...
File runtime/vm/intermediate_language_arm.cc (right):

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/intermediate_...
runtime/vm/intermediate_language_arm.cc:2222: }
Can you make sure to clobber R7 with RAW_NULL (0x1) in debug builds?

[koda, 2014-12-08 22:23:42]

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/assembler_arm.cc
File runtime/vm/assembler_arm.cc (right):

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/assembler_arm...
runtime/vm/assembler_arm.cc:1710: Register old_value = AllocateRegister(&used);
On 2014/12/08 19:27:56, Ivan Posva wrote:
> How do you prevent allocating one of the pre-assigned, non-allocatable
registers
> (e.g. SP) being used here?
> 
> Maybe we should not initialize with 0?

Added assertions to UseRegister. This would only be an issue if the args +
temporaries use more than 13 registers in total.

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/intermediate_...
File runtime/vm/intermediate_language_arm.cc (right):

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/intermediate_...
runtime/vm/intermediate_language_arm.cc:2222: }
On 2014/12/08 19:27:56, Ivan Posva wrote:
> Can you make sure to clobber R7 with RAW_NULL (0x1) in debug builds?

Done.

[koda, 2014-12-09 05:12:49]

Sync'ing revealed some further bugs. PTAL at the diff between patch sets 4 and
5:

- Force far branches when emitting edge counter increment in DEBUG mode, to
ensure that Stop has a predictable number of instructions.

- Fix context parent pointer store in allocation stub.

[Vyacheslav Egorov (Google), 2014-12-09 11:14:11]

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

[Vyacheslav Egorov (Google), 2014-12-09 11:14:12]

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/assembler_arm.cc
File runtime/vm/assembler_arm.cc (right):

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/assembler_arm...
runtime/vm/assembler_arm.cc:1711: Register shadow_value =
AllocateRegister(&used);
I suggest you merge your custom register allocator with
ParallelMoveResolver::ScratchRegisterScope. It's somewhat bad to have several of
those.

[koda, 2014-12-09 14:38:28]

Submitting soon if no more comments.

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/assembler_arm.cc
File runtime/vm/assembler_arm.cc (right):

https://codereview.chromium.org/730343005/diff/40001/runtime/vm/assembler_arm...
runtime/vm/assembler_arm.cc:1711: Register shadow_value =
AllocateRegister(&used);
On 2014/12/09 11:14:12, Vyacheslav Egorov (Google) wrote:
> I suggest you merge your custom register allocator with
> ParallelMoveResolver::ScratchRegisterScope. It's somewhat bad to have several
of
> those.

VerifiedWrite needs to be used in lower-level code like stubs, which is why it's
in Assembler. If more code starts using its "allocator", and especially if it's
non-DEBUG code, then I'll look into refactoring to share what is possible to
share.

[koda, 2014-12-09 16:15:53]

Committed patchset #5 (id:80001) manually as r42208 (presubmit successful).