Add --enable-pgo option to build clang with PGO

tools/clang/scripts/update.py

Index: tools/clang/scripts/update.py
diff --git a/tools/clang/scripts/update.py b/tools/clang/scripts/update.py
index 3571d3159462e951e913a72eb89ca1e4cba98652..c4729a31fb8dc6e967cd6c0330dfe319577c03df 100755
--- a/tools/clang/scripts/update.py
+++ b/tools/clang/scripts/update.py
@@ -375,6 +375,69 @@ def CopyDiaDllTo(target_dir):
   CopyFile(dia_dll, target_dir)
 
 
+def GetNewCCAndCXX(install_dir):
+  if sys.platform == 'win32':
+    cc = os.path.join(install_dir, 'bin', 'clang-cl.exe')
+    cxx = os.path.join(install_dir, 'bin', 'clang-cl.exe')
+    # CMake has a hard time with backslashes in compiler paths:
+    # https://stackoverflow.com/questions/13050827
+    cc = cc.replace('\\', '/')
+    cxx = cxx.replace('\\', '/')
+  else:
+    cc = os.path.join(install_dir, 'bin', 'clang')
+    cxx = os.path.join(install_dir, 'bin', 'clang++')
+
+  return cc, cxx
+
+
+def BuildLLVMgold(args, install_dir, base_cmake_args, binutils_incdir, cc, cxx):
+  # Build LLVM gold plugin with LTO. That speeds up the linker by ~10%.
+  # We only use LTO for Linux now.
+  print 'Building LTO LLVM Gold plugin'
+  if os.path.exists(LLVM_LTO_GOLD_PLUGIN_DIR):
+    RmTree(LLVM_LTO_GOLD_PLUGIN_DIR)
+  EnsureDirExists(LLVM_LTO_GOLD_PLUGIN_DIR)
+  os.chdir(LLVM_LTO_GOLD_PLUGIN_DIR)
+
+  # Create a symlink to LLVMgold.so build in the previous step so that ar
+  # and ranlib could find it while linking LLVMgold.so with LTO.
+  EnsureDirExists(BFD_PLUGINS_DIR)
+  RunCommand(['ln', '-sf',
+              os.path.join(install_dir, 'lib', 'LLVMgold.so'),
+              os.path.join(BFD_PLUGINS_DIR, 'LLVMgold.so')])
+
+  # Link against binutils's copy of tcmalloc to speed up the linker by ~10%.
+  # In package.py we copy the .so into our package.
+  tcmalloc_ldflags = ['-L' + BINUTILS_LIB_DIR, '-ltcmalloc_minimal']
+  # Make sure that tblgen and the test suite can find tcmalloc.
+  os.environ['LD_LIBRARY_PATH'] = \
+      BINUTILS_LIB_DIR + os.pathsep + os.environ.get('LD_LIBRARY_PATH', '')
+
+  lto_cflags = ['-flto=thin']
+  lto_ldflags = ['-fuse-ld=lld']
+  if args.gcc_toolchain:
+    # Tell the bootstrap compiler to use a specific gcc prefix to search
+    # for standard library headers and shared object files.
+    lto_cflags += ['--gcc-toolchain=' + args.gcc_toolchain]
+  lto_cmake_args = base_cmake_args + [
+      '-DLLVM_BINUTILS_INCDIR=' + binutils_incdir,
+      '-DCMAKE_C_COMPILER=' + cc,
+      '-DCMAKE_CXX_COMPILER=' + cxx,
+      '-DCMAKE_C_FLAGS=' + ' '.join(lto_cflags),
+      '-DCMAKE_CXX_FLAGS=' + ' '.join(lto_cflags),
+      '-DCMAKE_EXE_LINKER_FLAGS=' + ' '.join(lto_ldflags + tcmalloc_ldflags),
+      '-DCMAKE_SHARED_LINKER_FLAGS=' + ' '.join(lto_ldflags),
+      '-DCMAKE_MODULE_LINKER_FLAGS=' + ' '.join(lto_ldflags)]
+
+  # We need to use the proper binutils which support LLVM Gold plugin.
+  lto_env = os.environ.copy()
+  lto_env['PATH'] = BINUTILS_BIN_DIR + os.pathsep + lto_env.get('PATH', '')
+
+  RmCmakeCache('.')
+  RunCommand(['cmake'] + lto_cmake_args + [LLVM_DIR], env=lto_env)
+  RunCommand(['ninja', 'LLVMgold', 'lld'], env=lto_env)
+
+
 def VeryifyVersionOfBuiltClangMatchesVERSION():
   """Checks that `clang --version` outputs VERSION.  If this fails, VERSION
   in this file is out-of-date and needs to be updated (possibly in an
@@ -462,7 +525,10 @@ def UpdateClang(args):
     # In case someone sends a tryjob that temporary adds lld to the checkout,
     # make sure it's not around on future builds.
     RmTree(LLD_DIR)
-  Checkout('compiler-rt', LLVM_REPO_URL + '/compiler-rt/trunk', COMPILER_RT_DIR)
+  if args.enable_pgo:
+    # PGO requires compiler-rt.
+    Checkout('compiler-rt', LLVM_REPO_URL + '/compiler-rt/trunk',
+             os.path.join(LLVM_DIR, 'projects', 'compiler-rt'))
   if sys.platform == 'darwin':
     # clang needs a libc++ checkout, else -stdlib=libc++ won't find includes
     # (i.e. this is needed for bootstrap builds).
@@ -504,105 +570,13 @@ def UpdateClang(args):
   if sys.platform.startswith('linux'):
     binutils_incdir = os.path.join(BINUTILS_DIR, 'include')
 
-  if args.bootstrap:
-    print 'Building bootstrap compiler'
-    EnsureDirExists(LLVM_BOOTSTRAP_DIR)
-    os.chdir(LLVM_BOOTSTRAP_DIR)
-    bootstrap_args = base_cmake_args + [
-        '-DLLVM_BINUTILS_INCDIR=' + binutils_incdir,
-        '-DLLVM_TARGETS_TO_BUILD=host',
-        '-DCMAKE_INSTALL_PREFIX=' + LLVM_BOOTSTRAP_INSTALL_DIR,
-        '-DCMAKE_C_FLAGS=' + ' '.join(cflags),
-        '-DCMAKE_CXX_FLAGS=' + ' '.join(cxxflags),
-        ]
-    if cc is not None:  bootstrap_args.append('-DCMAKE_C_COMPILER=' + cc)
-    if cxx is not None: bootstrap_args.append('-DCMAKE_CXX_COMPILER=' + cxx)
-    RmCmakeCache('.')
-    RunCommand(['cmake'] + bootstrap_args + [LLVM_DIR], msvc_arch='x64')
-    RunCommand(['ninja'], msvc_arch='x64')
-    if args.run_tests:
-      if sys.platform == 'win32':
-        CopyDiaDllTo(os.path.join(LLVM_BOOTSTRAP_DIR, 'bin'))
-      RunCommand(['ninja', 'check-all'], msvc_arch='x64')
-    RunCommand(['ninja', 'install'], msvc_arch='x64')
-    if args.gcc_toolchain:
-      # Copy that gcc's stdlibc++.so.6 to the build dir, so the bootstrap
-      # compiler can start.
-      CopyFile(libstdcpp, os.path.join(LLVM_BOOTSTRAP_INSTALL_DIR, 'lib'))
-
-    if sys.platform == 'win32':
-      cc = os.path.join(LLVM_BOOTSTRAP_INSTALL_DIR, 'bin', 'clang-cl.exe')
-      cxx = os.path.join(LLVM_BOOTSTRAP_INSTALL_DIR, 'bin', 'clang-cl.exe')
-      # CMake has a hard time with backslashes in compiler paths:
-      # https://stackoverflow.com/questions/13050827
-      cc = cc.replace('\\', '/')
-      cxx = cxx.replace('\\', '/')
-    else:
-      cc = os.path.join(LLVM_BOOTSTRAP_INSTALL_DIR, 'bin', 'clang')
-      cxx = os.path.join(LLVM_BOOTSTRAP_INSTALL_DIR, 'bin', 'clang++')
-
-    if args.gcc_toolchain:
-      # Tell the bootstrap compiler to use a specific gcc prefix to search
-      # for standard library headers and shared object files.
-      cflags = ['--gcc-toolchain=' + args.gcc_toolchain]
-      cxxflags = ['--gcc-toolchain=' + args.gcc_toolchain]
-    print 'Building final compiler'
-
-  # Build LLVM gold plugin with LTO. That speeds up the linker by ~10%.
-  # We only use LTO for Linux now.
-  if args.bootstrap and args.lto_gold_plugin:
-    print 'Building LTO LLVM Gold plugin'
-    if os.path.exists(LLVM_LTO_GOLD_PLUGIN_DIR):
-      RmTree(LLVM_LTO_GOLD_PLUGIN_DIR)
-    EnsureDirExists(LLVM_LTO_GOLD_PLUGIN_DIR)
-    os.chdir(LLVM_LTO_GOLD_PLUGIN_DIR)
-
-    # Create a symlink to LLVMgold.so build in the previous step so that ar
-    # and ranlib could find it while linking LLVMgold.so with LTO.
-    EnsureDirExists(BFD_PLUGINS_DIR)
-    RunCommand(['ln', '-sf',
-                os.path.join(LLVM_BOOTSTRAP_INSTALL_DIR, 'lib', 'LLVMgold.so'),
-                os.path.join(BFD_PLUGINS_DIR, 'LLVMgold.so')])
-
-    # Link against binutils's copy of tcmalloc to speed up the linker by ~10%.
-    # In package.py we copy the .so into our package.
-    tcmalloc_ldflags = ['-L' + BINUTILS_LIB_DIR, '-ltcmalloc_minimal']
-    # Make sure that tblgen and the test suite can find tcmalloc.
-    os.environ['LD_LIBRARY_PATH'] = \
-        BINUTILS_LIB_DIR + os.pathsep + os.environ.get('LD_LIBRARY_PATH', '')
-
-    lto_cflags = ['-flto=thin']
-    lto_ldflags = ['-fuse-ld=lld']
-    if args.gcc_toolchain:
-      # Tell the bootstrap compiler to use a specific gcc prefix to search
-      # for standard library headers and shared object files.
-      lto_cflags += ['--gcc-toolchain=' + args.gcc_toolchain]
-    lto_cmake_args = base_cmake_args + [
-        '-DLLVM_BINUTILS_INCDIR=' + binutils_incdir,
-        '-DCMAKE_C_COMPILER=' + cc,
-        '-DCMAKE_CXX_COMPILER=' + cxx,
-        '-DCMAKE_C_FLAGS=' + ' '.join(lto_cflags),
-        '-DCMAKE_CXX_FLAGS=' + ' '.join(lto_cflags),
-        '-DCMAKE_EXE_LINKER_FLAGS=' + ' '.join(lto_ldflags + tcmalloc_ldflags),
-        '-DCMAKE_SHARED_LINKER_FLAGS=' + ' '.join(lto_ldflags),
-        '-DCMAKE_MODULE_LINKER_FLAGS=' + ' '.join(lto_ldflags)]
-
-    # We need to use the proper binutils which support LLVM Gold plugin.
-    lto_env = os.environ.copy()
-    lto_env['PATH'] = BINUTILS_BIN_DIR + os.pathsep + lto_env.get('PATH', '')
-
-    RmCmakeCache('.')
-    RunCommand(['cmake'] + lto_cmake_args + [LLVM_DIR], env=lto_env)
-    RunCommand(['ninja', 'LLVMgold', 'lld'], env=lto_env)
-
-
   # LLVM uses C++11 starting in llvm 3.5. On Linux, this means libstdc++4.7+ is
   # needed, on OS X it requires libc++. clang only automatically links to libc++
   # when targeting OS X 10.9+, so add stdlib=libc++ explicitly so clang can run
   # on OS X versions as old as 10.7.
   deployment_target = ''
 
-  if sys.platform == 'darwin' and args.bootstrap:
+  if sys.platform == 'darwin' and args.enable_pgo:
     # When building on 10.9, /usr/include usually doesn't exist, and while
     # Xcode's clang automatically sets a sysroot, self-built clangs don't.
     cflags = ['-isysroot', subprocess.check_output(
@@ -610,10 +584,6 @@ def UpdateClang(args):
     cxxflags = ['-stdlib=libc++'] + cflags
     ldflags += ['-stdlib=libc++']
     deployment_target = '10.7'
-    # Running libc++ tests takes a long time. Since it was only needed for
-    # the install step above, don't build it as part of the main build.
-    # This makes running package.py over 10% faster (30 min instead of 34 min)
-    RmTree(LIBCXX_DIR)
 
   # Build clang.
 
@@ -644,17 +614,49 @@ def UpdateClang(args):
   if cc is not None:  cc_args.append('-DCMAKE_C_COMPILER=' + cc)
   if cxx is not None: cc_args.append('-DCMAKE_CXX_COMPILER=' + cxx)
   chrome_tools = list(set(['plugins', 'blink_gc_plugin'] + args.extra_tools))
-  cmake_args += base_cmake_args + [
-      '-DLLVM_ENABLE_THREADS=OFF',
+
+
+  stage2_cache = os.path.join(CHROMIUM_DIR, 'tools', 'clang', 'stage2.cmake')
+  if args.bootstrap:
+    additional_cmake_args = [
+      '-DBOOTSTRAP_CHROMIUM_TOOLS=%s' % ','.join(chrome_tools),
+      '-DBOOTSTRAP_CMAKE_EXE_LINKER_FLAGS=' + ' '.join(ldflags),
+      '-DBOOTSTRAP_CMAKE_SHARED_LINKER_FLAGS=' + ' '.join(ldflags),
+      '-DBOOTSTRAP_CMAKE_MODULE_LINKER_FLAGS=' + ' '.join(ldflags),
+      '-DSTAGE2_CACHE_FILE=%s' % stage2_cache,
+      '-C', os.path.join(LLVM_DIR, 'tools', 'clang', 'cmake', 'caches',
+                         'DistributionExample.cmake'),
+    ]
+  elif args.enable_pgo:
+    additional_cmake_args = [
+      '-DBOOTSTRAP_BOOTSTRAP_CHROMIUM_TOOLS=%s' % ','.join(chrome_tools),
+      '-DBOOTSTRAP_BOOTSTRAP_LLVM_BINUTILS_INCDIR=' + binutils_incdir,
+      '-DBOOTSTRAP_BOOTSTRAP_CMAKE_C_FLAGS=' + ' '.join(cflags),
+      '-DBOOTSTRAP_BOOTSTRAP_CMAKE_CXX_FLAGS=' + ' '.join(cxxflags),
+      '-DBOOTSTRAP_BOOTSTRAP_CMAKE_EXE_LINKER_FLAGS=' + ' '.join(ldflags),
+      '-DBOOTSTRAP_BOOTSTRAP_CMAKE_SHARED_LINKER_FLAGS=' + ' '.join(ldflags),
+      '-DBOOTSTRAP_BOOTSTRAP_CMAKE_MODULE_LINKER_FLAGS=' + ' '.join(ldflags),
+      '-DPGO_BUILD_CONFIGURATION=%s' % stage2_cache,
+      '-C', os.path.join(LLVM_DIR, 'tools', 'clang', 'cmake', 'caches',
+                         'PGO.cmake')
+    ]
+
+  cmake_args += base_cmake_args + additional_cmake_args + [
+      # Override lto setting from DistributionExample.cmake cache, if used.
+      '-DPACKAGE_VENDOR=',
+      '-DBOOTSTRAP_LLVM_ENABLE_LTO=OFF',
+      '-DCMAKE_VERBOSE_MAKEFILE=ON',
+      '-DBOOTSTRAP_CMAKE_VERBOSE_MAKEFILE=ON',
+      '-DBOOTSTRAP_BOOTSTRAP_CMAKE_VERBOSE_MAKEFILE=ON',
       '-DLLVM_BINUTILS_INCDIR=' + binutils_incdir,
+      '-DBOOTSTRAP_LLVM_BINUTILS_INCDIR=' + binutils_incdir,
       '-DCMAKE_C_FLAGS=' + ' '.join(cflags),
+      '-DBOOTSTRAP_CMAKE_C_FLAGS=' + ' '.join(cflags),
       '-DCMAKE_CXX_FLAGS=' + ' '.join(cxxflags),
-      '-DCMAKE_EXE_LINKER_FLAGS=' + ' '.join(ldflags),
-      '-DCMAKE_SHARED_LINKER_FLAGS=' + ' '.join(ldflags),
-      '-DCMAKE_MODULE_LINKER_FLAGS=' + ' '.join(ldflags),
+      '-DBOOTSTRAP_CMAKE_CXX_FLAGS=' + ' '.join(cxxflags),
       '-DCMAKE_INSTALL_PREFIX=' + LLVM_BUILD_DIR,
-      '-DCHROMIUM_TOOLS_SRC=%s' % os.path.join(CHROMIUM_DIR, 'tools', 'clang'),
-      '-DCHROMIUM_TOOLS=%s' % ';'.join(chrome_tools)]
+      '-DCHROMIUM_TOOLS=%s' % ','.join(chrome_tools),
+  ]
 
   EnsureDirExists(LLVM_BUILD_DIR)
   os.chdir(LLVM_BUILD_DIR)
@@ -662,6 +664,18 @@ def UpdateClang(args):
   RunCommand(['cmake'] + cmake_args + [LLVM_DIR],
              msvc_arch='x64', env=deployment_env)
 
+  RunCommand(['ninja', 'stage2'], env=deployment_env, msvc_arch='x64')
+
+  if args.bootstrap:
+    global LLVM_BUILD_DIR
+    LLVM_BUILD_DIR = os.path.join(
+        LLVM_BUILD_DIR, 'tools', 'clang', 'stage2-bins')
+  elif args.enable_pgo:
+    global LLVM_BUILD_DIR
+    LLVM_BUILD_DIR = os.path.join(LLVM_BUILD_DIR, 'tools', 'clang',
+        'stage2-instrumented-bins', 'tools', 'clang', 'stage2-bins')
+  os.chdir(LLVM_BUILD_DIR)
+
   if args.gcc_toolchain:
     # Copy in the right stdlibc++.so.6 so clang can start.
     if not os.path.exists(os.path.join(LLVM_BUILD_DIR, 'lib')):
@@ -670,17 +684,16 @@ def UpdateClang(args):
         [cxx] + cxxflags + ['-print-file-name=libstdc++.so.6']).rstrip()
     CopyFile(libstdcpp, os.path.join(LLVM_BUILD_DIR, 'lib'))
 
-  RunCommand(['ninja'], msvc_arch='x64')
+  RunCommand(['ninja', 'cr-install'], msvc_arch='x64')
 
-  # Copy LTO-optimized lld, if any.
-  if args.bootstrap and args.lto_gold_plugin:
+  # Build & copy LTO-optimized lld, if any.
+  if args.lto_gold_plugin:
+    cc, cxx = GetNewCCAndCXX(LLVM_BUILD_DIR)
+    BuildLLVMgold(args, LLVM_BUILD_DIR, base_cmake_args,
+                  binutils_incdir, cc, cxx)
     CopyFile(os.path.join(LLVM_LTO_GOLD_PLUGIN_DIR, 'bin', 'lld'),
              os.path.join(LLVM_BUILD_DIR, 'bin'))
 
-  if chrome_tools:
-    # If any Chromium tools were built, install those now.
-    RunCommand(['ninja', 'cr-install'], msvc_arch='x64')
-
   if sys.platform == 'darwin':
     # See http://crbug.com/256342
     RunCommand(['strip', '-x', os.path.join(LLVM_BUILD_DIR, 'bin', 'clang')])
@@ -694,6 +707,7 @@ def UpdateClang(args):
   # TODO(hans): Remove once the regular build above produces this.
   # On Mac and Linux, this is used to get the regular 64-bit run-time.
   # Do a clobbered build due to cmake changes.
+  Checkout('compiler-rt', LLVM_REPO_URL + '/compiler-rt/trunk', COMPILER_RT_DIR)
   if os.path.isdir(COMPILER_RT_BUILD_DIR):
     RmTree(COMPILER_RT_BUILD_DIR)
   os.makedirs(COMPILER_RT_BUILD_DIR)
@@ -868,14 +882,25 @@ def main():
                       help='don\'t build Android ASan runtime (linux only)',
                       dest='with_android',
                       default=sys.platform.startswith('linux'))
+  parser.add_argument('--enable-pgo', action='store_true',
+                      help='build clang in 2 stages using PGO')
   args = parser.parse_args()
 
-  if args.lto_gold_plugin and not args.bootstrap:
-    print '--lto-gold-plugin requires --bootstrap'
+  if args.lto_gold_plugin and not args.enable_pgo:
+    print '--lto-gold-plugin requires --enable-pgo'
     return 1
   if args.lto_gold_plugin and not sys.platform.startswith('linux'):
     print '--lto-gold-plugin is only effective on Linux. Ignoring the option.'
     args.lto_gold_plugin = False
+  if args.enable_pgo:
+    if args.bootstrap:
+      print '--bootstrap makes no sense with --enable-pgo specified'
+      return 1
+    if not args.force_local_build:
+      print '--enable-pgo makes no sense without --force-local-build'
+      return 1
+    if sys.platform == 'win32':
+      print '--enable-pgo not supported on Windows'
 
   if args.if_needed:
     # TODO(thakis): Can probably remove this and --if-needed altogether.