Android: bake GCC flags for perf into profiling=1 in GYP_DEFINES

build/common.gypi

 # Copyright (c) 2012 The Chromium Authors. All rights reserved.
 # Use of this source code is governed by a BSD-style license that can be
 # found in the LICENSE file.
 
 # IMPORTANT:
 # Please don't directly include this file if you are building via gyp_chromium,
 # since gyp_chromium is automatically forcing its inclusion.
 {
   # Variables expected to be overriden on the GYP command line (-D) or by
   # ~/.gyp/include.gypi.
@@ skipping 3429 matching lines @@
                         '-Wl,-z,relro',
                         '-Wl,-z,now',
                         '-fuse-ld=gold',
                     ],
                     'conditions': [
                       ['arm_thumb==1', {
                         'cflags': [ '-mthumb-interwork' ],
                       }],
                       ['profiling==1', {
                         'cflags': [
-                          '-marm', # Probably reduntant, but recommend by "perf" docs.
-                          '-mapcs-frame', # Seems required by -fno-omit-frame-pointer.
+                          # Thumb code with frame pointer makes chrome crash
+                          # early.
+                          '-marm',
+                          '-mapcs-frame', # Required by -fno-omit-frame-pointer.
+                          '-fno-optimize-sibling-calls',

[Sami, 2014/05/15 13:34:32]

Looks like this one is also set by profiling_full_stack_frames=1. What's the
advantage of having it here too?

[pasko, 2014/05/15 14:34:53]

I noticed too. The profiling=1 is supposed to be the default way to use perf.
Without -fno-optimize-sibling-calls I found perf reports on Android to be often
misleading. On the other hand, profiling_full_stack_frames=1 disallows inlining
(sort of), which I never needed, so not enabling it to be closer to profiling=0
performance.

profiling=1 *and* profiling_full_stack_frames=1 together may be useful in some
cases (such as for memory profiling?), so let's have the two modes.

+                        ],
+                        'cflags!': [
+                          '-fomit-frame-pointer',

[Sami, 2014/05/15 13:34:32]

BTW, is this just a safe guard or did we end up removing frame pointers in some
cases?

[pasko, 2014/05/15 14:34:53]

We add the -fomit-frame-pointer flag on Android at line 3184 in this file.

                         ],
                       }],
                       ['clang==1', {
                         'cflags!': [
                           # Clang does not support the following options.
                           '-mthumb-interwork',
                           '-finline-limit=64',
                           '-fno-tree-sra',
                           '-fuse-ld=gold',
                           '-Wno-psabi',
@@ skipping 1759 matching lines @@
     # settings in target dicts.  SYMROOT is a special case, because many other
     # Xcode variables depend on it, including variables such as
     # PROJECT_DERIVED_FILE_DIR.  When a source group corresponding to something
     # like PROJECT_DERIVED_FILE_DIR is added to a project, in order for the
     # files to appear (when present) in the UI as actual files and not red
     # red "missing file" proxies, the correct path to PROJECT_DERIVED_FILE_DIR,
     # and therefore SYMROOT, needs to be set at the project level.
     'SYMROOT': '<(DEPTH)/xcodebuild',
   },
 }