ARM64: improve Adr patching

ARM64: improve Adr patching

R=ulan@chromium.org, bmeurer@chromium.org

BUG=

[vincent.belliard, 2014-07-03 10:20:30]

Reduce the patch size for Adr and remove the add instruction in the most found
case.

Sunspider/dna 9% faster
Kraken/pkbdf2 2.5% faster
Octane/regexp 1.7% better

Other benchmarks unchanged or slightly improved.

[ulan, 2014-07-03 10:36:21]

lgtm

https://codereview.chromium.org/363293004/diff/1/src/arm64/assembler-arm64.cc
File src/arm64/assembler-arm64.cc (left):

https://codereview.chromium.org/363293004/diff/1/src/arm64/assembler-arm64.cc...
src/arm64/assembler-arm64.cc:3115: //   add  rd, rd, scratch
Why not remove (nop, nop) instead of (nop, add)?

[vincent.belliard, 2014-07-03 10:46:06]

https://codereview.chromium.org/363293004/diff/1/src/arm64/assembler-arm64.cc
File src/arm64/assembler-arm64.cc (left):

https://codereview.chromium.org/363293004/diff/1/src/arm64/assembler-arm64.cc...
src/arm64/assembler-arm64.cc:3115: //   add  rd, rd, scratch
The add has a cost (the processor doesn't know that the register holds zero) but
the nop is optimized away. Removing the add gives most of the improvement.

[ulan, 2014-07-03 11:01:48]

https://codereview.chromium.org/363293004/diff/1/src/arm64/assembler-arm64.cc
File src/arm64/assembler-arm64.cc (left):

https://codereview.chromium.org/363293004/diff/1/src/arm64/assembler-arm64.cc...
src/arm64/assembler-arm64.cc:3115: //   add  rd, rd, scratch
On 2014/07/03 10:46:06, vincent.belliard.arm wrote:
> The add has a cost (the processor doesn't know that the register holds zero)
but
> the nop is optimized away. Removing the add gives most of the improvement.

ok, thanks.

[vincent.belliard, 2014-07-03 13:20:16]

committed r22198