Gyp flag to disable inlining.

Gyp flag to disable inlining.

To turn off all inlining, rebuild after running the command

$ GYP_DEFINES=skia_disable_inlining=1 ./gyp_skia

Provides an initial value in one GM to avoid an uninitialized variable warning. 

BUG=skia:2126
Committed: http://code.google.com/p/skia/source/detail?r=13437

[tomhudson, 2014-02-13 17:39:31]

PTAL

[tomhudson, 2014-02-13 17:44:17]

Without, profile of bench_record looks like:

+  10.68%  bench_record  libc-2.15.so                [.] __memcpy_ssse3_back
+   6.28%  bench_record  bench_record                [.]
SkPicturePlayback::draw(SkCanvas&, SkDrawPictureCallback*)
+   5.27%  bench_record  bench_record                [.]
SkPictureRecord::drawRect(SkRect const&, SkPaint const&)
+   5.12%  bench_record  bench_record                [.]
SkPictureRecord::drawPosTextH(void const*, unsigned long, float const*, float,
SkPaint const&)
+   4.85%  bench_record  bench_record                [.]
SkPaint::flatten(SkWriteBuffer&) const
+   3.91%  bench_record  bench_record                [.]
SkPictureRecord::drawPosTextHImpl(void const*, unsigned long, float const*,
float, SkPaint const&, SkFlatData const*)
+   2.37%  bench_record  bench_record                [.] SkPtrSet::add(void*)


With, profile is more fine-grained and has more leaf nodes:

+   4.25%  bench_record  bench_record                [.]
SkWriter32::reserve(unsigned long)
+   4.10%  bench_record  libc-2.15.so                [.] __memcpy_ssse3_back
+   3.98%  bench_record  bench_record                [.]
SkTDynamicHash<SkFlatData, SkFlatData, &(SkFlatData::Identity(SkFlatData
const&)), &(SkFlatData::Hash(SkFlatData const&)
+   2.97%  bench_record  bench_record                [.]
SkPicturePlayback::draw(SkCanvas&, SkDrawPictureCallback*)
+   2.81%  bench_record  bench_record                [.] asint(void const*)
+   2.33%  bench_record  bench_record                [.]
SkPaint::flatten(SkWriteBuffer&) const
+   1.95%  bench_record  bench_record                [.]
SkFlatData::operator==(SkFlatData const&) const
+   1.63%  bench_record  bench_record                [.] write_scalar(unsigned
int*, float)


... but, as noted in http://skbug.com/2126, avoids pollution with *::validate()
and other non-release overheads.

[scroggo, 2014-02-13 17:46:18]

lgtm

https://codereview.chromium.org/163363004/diff/1/gyp/common_conditions.gypi
File gyp/common_conditions.gypi (right):

https://codereview.chromium.org/163363004/diff/1/gyp/common_conditions.gypi#n...
gyp/common_conditions.gypi:145: [ 'skia_disable_inlining', {
Why does this depend on being on a linux-ish OS?

[tomhudson, 2014-02-13 17:48:24]

On 2014/02/13 17:46:18, scroggo wrote:
> Why does this depend on being on a linux-ish OS?

I've only tested the flags on GCC with a Linux target.
Presumably they'll work on Mac OSX, but I only have easy access to Linux /
Android while in London.

[mtklein, 2014-02-13 17:55:39]

LGTM

Just keep in mind this tips the bias in the other direction.  There are things
which are cheap when inlined and disproportionately expensive when not, e.g.
trivial getters and setters.

https://codereview.chromium.org/163363004/diff/1/gyp/common_conditions.gypi
File gyp/common_conditions.gypi (right):

https://codereview.chromium.org/163363004/diff/1/gyp/common_conditions.gypi#n...
gyp/common_conditions.gypi:145: [ 'skia_disable_inlining', {
Maybe tack on # Handy for profiling. ?

[tomhudson, 2014-02-13 18:01:46]

The CQ bit was checked by tomhudson@google.com

[commit-bot: I haz the power, 2014-02-13 18:01:57]

CQ is trying da patch. Follow status at
https://skia-tree-status.appspot.com/cq/tomhudson@chromium.org/163363004/70001

[commit-bot: I haz the power, 2014-02-13 18:36:39]

Change committed as 13437