A64: Implement and use FillFields

A64: Implement and use FillFields

BUG=
R=ulan@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=19895

[ulan, 2014-03-12 12:17:26]

https://codereview.chromium.org/190763012/diff/20001/src/a64/builtins-a64.cc
File src/a64/builtins-a64.cc (right):

https://codereview.chromium.org/190763012/diff/20001/src/a64/builtins-a64.cc#...
src/a64/builtins-a64.cc:456: Register obj_end = x5;
Maybe set first_prop = NoReg to show that it is not used anymore? The same for
prop_fields below.

https://codereview.chromium.org/190763012/diff/20001/src/a64/builtins-a64.cc#...
src/a64/builtins-a64.cc:483: __ Ldrb(x3, FieldMemOperand(init_map,
Map::kUnusedPropertyFieldsOffset));
x3 => element_count, here and below.

https://codereview.chromium.org/190763012/diff/20001/src/a64/macro-assembler-...
File src/a64/macro-assembler-a64.cc (right):

https://codereview.chromium.org/190763012/diff/20001/src/a64/macro-assembler-...
src/a64/macro-assembler-a64.cc:2479: Sub(field_ptr, dst, Operand(field_ptr, LSL,
kPointerSizeLog2));
I am convinced that this implementation is correct.

Any reason why you didn't choose the algorithm that is used in other places? 

In pseudocode:

counter = field_count - 2;
while (counter >= 0) {
  store pair of fillers;
  counter -= 2;
}
if (counter & 1) {
  store one remaining filler;
}

[m.m.capewell, 2014-03-12 17:06:05]

On 2014/03/12 12:17:26, ulan wrote:
> Any reason why you didn't choose the algorithm that is used in other places? 
> 
> In pseudocode:
> 
> counter = field_count - 2;
> while (counter >= 0) {
>   store pair of fillers;
>   counter -= 2;
> }
> if (counter & 1) {
>   store one remaining filler;
> }

As we don't have conditional stores on A64, I was trying to avoid the 50/50
predicted taken branch over the final 'odd' store. I decided a (half) repeated
store was cheaper than a branch misprediction.

[ulan, 2014-03-13 08:47:19]

On 2014/03/12 17:06:05, m.m.capewell wrote:
> On 2014/03/12 12:17:26, ulan wrote:
> > Any reason why you didn't choose the algorithm that is used in other places?

> > 
> > In pseudocode:
> > 
> > counter = field_count - 2;
> > while (counter >= 0) {
> >   store pair of fillers;
> >   counter -= 2;
> > }
> > if (counter & 1) {
> >   store one remaining filler;
> > }
> 
> As we don't have conditional stores on A64, I was trying to avoid the 50/50
> predicted taken branch over the final 'odd' store. I decided a (half) repeated
> store was cheaper than a branch misprediction.
I see, lgtm

[m.m.capewell, 2014-03-13 13:18:56]

Committed patchset #3 manually as r19895 (presubmit successful).