Accurate function prototypes for native calls from ARM simulator.

Accurate function prototypes for native calls from ARM simulator.

Native method invocation from the arm/simulator-arm.cc previously made
non-portable assumptions about calling conventions. This was okay for 32-bit
stack-based machines, where by-value structs are automatically materialized
on the stack, and where both int and double parameters could be passed on the
stack. However they are not okay for x86-64, which has an elaborate scheme
for passing parameters in registers.

This CL replaces the previous non-portable code paths with portable code,
using call-sites that accurately match the prototype of the callee.

BUG=2614
Committed: https://code.google.com/p/v8/source/detail?r=14230

[Brad Chen, 2013-04-09 16:11:14]

PTAL

[ulan, 2013-04-10 14:19:20]

The change looks good overall.

https://codereview.chromium.org/13818012/diff/3001/src/arguments.h
File src/arguments.h (right):

https://codereview.chromium.org/13818012/diff/3001/src/arguments.h#newcode122
src/arguments.h:122: Type Name(int argslength, Object** argsobject, Isolate*
isolate) {  \
Nit: args_length, args_object would be better.

https://codereview.chromium.org/13818012/diff/3001/src/builtins.cc
File src/builtins.cc (right):

https://codereview.chromium.org/13818012/diff/3001/src/builtins.cc#newcode132
src/builtins.cc:132: int arglength, Object** argobject, Isolate* isolate) { \
Nit: args_length, args_object would be better.

https://codereview.chromium.org/13818012/diff/3001/src/builtins.cc#newcode148
src/builtins.cc:148: name##ArgumentsType args(arglength, argobject);          \
This makes builtin calls more expensive. I am running benchmarks to see the
impact.

[Brad Chen, 2013-04-10 16:53:28]

Thanks for the comments. I'll upload a minor update after a little bit of local
testing.

https://codereview.chromium.org/13818012/diff/3001/src/arguments.h
File src/arguments.h (right):

https://codereview.chromium.org/13818012/diff/3001/src/arguments.h#newcode122
src/arguments.h:122: Type Name(int argslength, Object** argsobject, Isolate*
isolate) {  \
On 2013/04/10 14:19:20, ulan wrote:
> Nit: args_length, args_object would be better.

Done.

https://codereview.chromium.org/13818012/diff/3001/src/builtins.cc
File src/builtins.cc (right):

https://codereview.chromium.org/13818012/diff/3001/src/builtins.cc#newcode132
src/builtins.cc:132: int arglength, Object** argobject, Isolate* isolate) { \
On 2013/04/10 14:19:20, ulan wrote:
> Nit: args_length, args_object would be better.

Done.

https://codereview.chromium.org/13818012/diff/3001/src/builtins.cc#newcode148
src/builtins.cc:148: name##ArgumentsType args(arglength, argobject);          \
This is helpful. I haven't done any performance testing of this CL. I know the
NaCl versions will be incredibly slow.

On 2013/04/10 14:19:20, ulan wrote:
> This makes builtin calls more expensive. I am running benchmarks to see the
> impact.

[Brad Chen, 2013-04-10 18:36:53]

NOTE: I am seeing a "TIMEOUT" test failure for arm.check
cctest/testapi/Threading1. I see this with or without this CL, so I'm assuming
it has nothing to do with my changes.

I'm also seeing some problems with the pre-submit checks that run with
foo.check. That might actually be my problem, although I don't yet understand
it.

On 2013/04/10 16:53:28, Brad Chen wrote:
> Thanks for the comments. I'll upload a minor update after a little bit of
local
> testing.
> 
> https://codereview.chromium.org/13818012/diff/3001/src/arguments.h
> File src/arguments.h (right):
> 
> https://codereview.chromium.org/13818012/diff/3001/src/arguments.h#newcode122
> src/arguments.h:122: Type Name(int argslength, Object** argsobject, Isolate*
> isolate) {  \
> On 2013/04/10 14:19:20, ulan wrote:
> > Nit: args_length, args_object would be better.
> 
> Done.
> 
> https://codereview.chromium.org/13818012/diff/3001/src/builtins.cc
> File src/builtins.cc (right):
> 
> https://codereview.chromium.org/13818012/diff/3001/src/builtins.cc#newcode132
> src/builtins.cc:132: int arglength, Object** argobject, Isolate* isolate) { \
> On 2013/04/10 14:19:20, ulan wrote:
> > Nit: args_length, args_object would be better.
> 
> Done.
> 
> https://codereview.chromium.org/13818012/diff/3001/src/builtins.cc#newcode148
> src/builtins.cc:148: name##ArgumentsType args(arglength, argobject);         
\
> This is helpful. I haven't done any performance testing of this CL. I know the
> NaCl versions will be incredibly slow.
> 
> On 2013/04/10 14:19:20, ulan wrote:
> > This makes builtin calls more expensive. I am running benchmarks to see the
> > impact.

[ulan, 2013-04-11 12:02:34]

LGTM. I didn't get any presubmit errors. I will land after Jakob's review.

On my machine this CL seems to be performance neutral on Octane and Sunspider:

                               benchmark:    score |  tot-oct |      % 
===================================================+==========+========
                                Richards:  14844.1 |  14835.8 |    0.1 
                               DeltaBlue:  19079.4 |  19257.1 |   -0.9 
                                  Crypto:  18723.1 |  18725.9 |   -0.0 
                                RayTrace:  19431.7 |  19538.4 |   -0.5 
                             EarleyBoyer:  30608.9 |  30630.1 |   -0.1 
                                  RegExp:   4286.6 |   4283.0 |    0.1 
                                   Splay:   5090.7 |   5093.9 |   -0.1 
                            NavierStokes:  19155.3 |  19140.6 |    0.1 
                                   PdfJS:  14752.0 |  14593.4 |    1.1 
                                Mandreel:  15092.0 |  15175.7 |   -0.6 
                                 Gameboy:  20865.4 |  20902.9 |   -0.2 
                                CodeLoad:  16330.3 |  15902.9 |    2.7 
                                   Box2D:  14565.2 |  14523.4 |    0.3 
                                  Octane:  14683.1 |  14663.1 |    0.1 
---------------------------------------------------+----------+--------

                               benchmark:    score |      tot |      % 
===================================================+==========+========
                                 3d-cube:     14.0 |     13.9 |   -0.7 
                                3d-morph:      9.1 |      9.0 |   -1.1 
                             3d-raytrace:     14.1 |     14.0 |   -0.7 
                     access-binary-trees:      2.2 |      2.2 |        
                         access-fannkuch:      7.5 |      7.6 |    1.3 
                            access-nbody:      5.5 |      5.6 |    1.8 
                           access-nsieve:      3.7 |      3.8 |    2.7 
                bitops-3bit-bits-in-byte:      0.6 |      0.7 |   16.7 
                     bitops-bits-in-byte:      4.3 |      4.4 |    2.3 
                      bitops-bitwise-and:      3.0 |      3.1 |    3.3 
                      bitops-nsieve-bits:      4.1 |      4.1 |        
                   controlflow-recursive:      2.0 |      2.0 |        
                              crypto-aes:      9.6 |      9.4 |   -2.1 
                              crypto-md5:      4.5 |      4.5 |        
                             crypto-sha1:      5.3 |      5.2 |   -1.9 
                       date-format-tofte:     11.5 |     11.1 |   -3.5 
                       date-format-xparb:     10.1 |     10.0 |   -1.0 
                             math-cordic:      2.4 |      2.8 |   16.7 
                       math-partial-sums:      9.8 |      9.6 |   -2.0 
                      math-spectral-norm:      3.0 |      3.0 |        
                              regexp-dna:      7.8 |      7.7 |   -1.3 
                           string-base64:      5.9 |      5.9 |        
                            string-fasta:      8.0 |      8.0 |        
                         string-tagcloud:     23.6 |     24.0 |    1.7 
                      string-unpack-code:     24.3 |     24.7 |    1.6 
                   string-validate-input:      9.8 |     10.0 |    2.0 
                          SunSpider-Once:    205.7 |    206.3 |    0.3 
---------------------------------------------------+----------+--------

[Jakob Kummerow, 2013-04-11 12:31:26]

lgtm

[ulan, 2013-04-11 12:40:53]

Committed patchset #3 manually as r14230 (presubmit successful).