A64: Fix CallApiFunctionAndReturn.

A64: Fix CallApiFunctionAndReturn.

TEST=cctest/test-api/FunctionTemplateWithProfiler
R=jochen@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=19257

[ulan, 2014-02-10 20:21:23]

PTAL

https://codereview.chromium.org/145073006/diff/1/src/a64/macro-assembler-a64.cc
File src/a64/macro-assembler-a64.cc (right):

https://codereview.chromium.org/145073006/diff/1/src/a64/macro-assembler-a64....
src/a64/macro-assembler-a64.cc:1350: Mov(x3, Operand(thunk_ref));
I took this from the ARM port.

[jochen (gone - plz use gerrit), 2014-02-10 20:40:17]

lgtm

[ulan, 2014-02-11 08:36:01]

Committed patchset #1 manually as r19257 (tree was closed).

[jbramley, 2014-02-11 10:45:25]

https://codereview.chromium.org/145073006/diff/1/src/a64/macro-assembler-a64.cc
File src/a64/macro-assembler-a64.cc (right):

https://codereview.chromium.org/145073006/diff/1/src/a64/macro-assembler-a64....
src/a64/macro-assembler-a64.cc:1350: Mov(x3, Operand(thunk_ref));
On 2014/02/10 20:21:23, ulan wrote:
> I took this from the ARM port.

Ah, I must have overlooked it in the merge.

Out of curiosity, what is the common case here? If the profiler is usually
turned on, this might be faster:

Cmp(w10, 0);
Csel(x3, function_address, Operand(thunk_ref), eq);

It's smaller by one instruction (and doesn't branch), but if the profiler is
disabled we waste a few cycles loading the ExternalReference.

[ulan, 2014-02-11 14:56:36]

[+Dan]

Dan is planning to clean-up this code after we merge to the bleeding_edge.

> Out of curiosity, what is the common case here? If the profiler is usually
> turned on, this might be faster:

I think the profiler is usually turned off, but I didn't check this.

[dcarney, 2014-02-11 15:19:29]

https://codereview.chromium.org/145073006/diff/1/src/a64/macro-assembler-a64.cc
File src/a64/macro-assembler-a64.cc (right):

https://codereview.chromium.org/145073006/diff/1/src/a64/macro-assembler-a64....
src/a64/macro-assembler-a64.cc:1350: Mov(x3, Operand(thunk_ref));
the profiler is almost always disabled
after cleanup, the code to enable profiling will be deferred to the end of this
function on most architectures