[turbofan] ARM: elide masks on loads and stores

[turbofan] ARM: elide masks on loads and stores

Elide mask operations by merging them with the input load or output store where
possible. Also, use the size of the load or store to simplify the mask immediate
with the aim of encoding it in the instruction.

BUG=

[martyn.capewell, 2016-04-14 12:38:26]

Description was changed from

==========
[turbofan] ARM: elide masks on loads and stores

Elide mask operations by merging them with the input load or output store where
possible. Also, use the size of the load or store to simplify the mask immediate
with the aim of encoding it in the instruction.

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

to

==========
[turbofan] ARM: elide masks on loads and stores

Elide mask operations by merging them with the input load or output store where
possible. Also, use the size of the load or store to simplify the mask immediate
with the aim of encoding it in the instruction.

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

[martyn.capewell, 2016-04-14 12:38:26]

martyn.capewell@arm.com changed reviewers:
+ bmeurer@chromium.org, danno@chromium.org, jarin@chromium.org

[Benedikt Meurer, 2016-04-14 18:39:59]

https://codereview.chromium.org/1883133002/diff/1/src/compiler/arm/instructio...
File src/compiler/arm/instruction-selector-arm.cc (right):

https://codereview.chromium.org/1883133002/diff/1/src/compiler/arm/instructio...
src/compiler/arm/instruction-selector-arm.cc:338: Node*
SimplifyAndForMaskedStore(InstructionSelector* selector, Node* node,
This looks like something that should be done in the MachineOperatorReducer, as
it applies to all platforms. Can you move it there instead?

[martyn.capewell, 2016-04-15 14:05:49]

https://codereview.chromium.org/1883133002/diff/1/src/compiler/arm/instructio...
File src/compiler/arm/instruction-selector-arm.cc (right):

https://codereview.chromium.org/1883133002/diff/1/src/compiler/arm/instructio...
src/compiler/arm/instruction-selector-arm.cc:338: Node*
SimplifyAndForMaskedStore(InstructionSelector* selector, Node* node,
You're right, I should update MachineOperatorReducer::ReduceStore() for elision
on Checked operations.

However, the simplification part is machine specific. For example, on ARM64, the
immediate 0xfffffffd is encodable in the instruction, but the simplified
immediates 0xfffd or 0xfd aren't. ARM64 would then need a "complexify" operation
to make 0xfffd into 0xfffdfffd, which is encodable.

[martyn.capewell, 2016-04-25 10:15:05]

Ping?

[Benedikt Meurer, 2016-04-25 11:16:41]

Can you separate the MachineOperatorReducer change into its own CL? I think
that's uncontroversial, but I'm still not sure about the instruction  selector
changes.

[martyn.capewell, 2016-04-25 17:30:18]

No problem, MachineOperatorReducer change here:
https://codereview.chromium.org/1921673003/