llvm-avx2.patch

diff --git a/tools/clang/scripts/build.py b/tools/clang/scripts/build.py
index bb0f6e16b4649..89d3fb5299e9f 100755
--- a/tools/clang/scripts/build.py
+++ b/tools/clang/scripts/build.py
@@ -72,6 +72,8 @@ BUG_REPORT_URL = ('https://crbug.com in the Tools>LLVM component,'
 LIBXML2_VERSION = 'libxml2-v2.9.12'
 ZSTD_VERSION = 'zstd-1.5.5'
 
+MIMALLOC_URL = 'https://github.com/microsoft/mimalloc/archive/refs/tags/v3.1.4.tar.gz'
+
 win_sdk_dir = None
 def GetWinSDKDir():
   """Get the location of the current SDK."""
@@ -164,7 +166,7 @@ def CheckoutGitRepo(name, git_url, commit, dir):
         ['git', 'diff-index', '--exit-code', 'HEAD'], fail_hard=False)
         and RunCommand(['git', 'fetch'], fail_hard=False)
         and RunCommand(['git', 'checkout', commit], fail_hard=False)
-        and RunCommand(['git', 'clean', '-f'], fail_hard=False)):
+        and RunCommand(['git', 'clean', '-ffd'], fail_hard=False)):
       return
 
     # If we can't use the current repo, delete it.
@@ -349,6 +351,13 @@ def BuildLibXml2():
           '-DCMAKE_INSTALL_LIBDIR=lib',
           '-DCMAKE_MSVC_RUNTIME_LIBRARY=MultiThreaded',  # /MT to match LLVM.
           '-DBUILD_SHARED_LIBS=OFF',
+          '-DCMAKE_C_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang'),
+          '-DCMAKE_CXX_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang++'),
+          '-DCMAKE_C_FLAGS_RELEASE=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -w ',
+          '-DCMAKE_CXX_FLAGS_RELEASE=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -w ',
+          '-DCMAKE_EXE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -w ',
+          '-DCMAKE_SHARED_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -w ',
+          '-DCMAKE_MODULE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -w ',
           '-DLIBXML2_WITH_C14N=OFF',
           '-DLIBXML2_WITH_CATALOG=OFF',
           '-DLIBXML2_WITH_DEBUG=OFF',
@@ -408,6 +417,52 @@ def BuildLibXml2():
 
   return extra_cmake_flags, extra_cflags
 
+class LibMimallocDirs:
+  """
+  The set of directories where mimalloc is located.
+
+  Includes the diractories where the source is unpacked, where it is built,
+  and installed.
+  """
+  def __init__(self):
+    self.unzip_dir = LLVM_BUILD_TOOLS_DIR
+    self.src_dir = os.path.join(self.unzip_dir, 'mimalloc-3.1.4')
+    self.build_dir = os.path.join(self.src_dir, 'build')
+
+def BuildLibMimalloc():
+  """Download and build mimalloc"""
+
+  dirs = LibMimallocDirs()
+  if os.path.exists(dirs.src_dir):
+    RmTree(dirs.src_dir)
+  DownloadAndUnpack(MIMALLOC_URL, dirs.unzip_dir)
+  os.mkdir(dirs.build_dir)
+  os.chdir(dirs.build_dir)
+
+  RunCommand(
+    [
+        'cmake',
+        '-GNinja',
+        '-DCMAKE_BUILD_TYPE=Release',
+        '-DCMAKE_C_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang'),
+        '-DCMAKE_CXX_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang++'),
+        '-DCMAKE_C_FLAGS_RELEASE=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -w ',
+        '-DCMAKE_CXX_FLAGS_RELEASE=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -w ',
+        '-DCMAKE_EXE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -w ',
+        '-DCMAKE_SHARED_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -w ',
+        '-DCMAKE_MODULE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -w ',
+        '-DMI_OVERRIDE=ON',
+        '-DMI_SHOW_ERRORS=OFF',
+        '..',
+    ],
+    setenv=True)
+  RunCommand(['ninja'], setenv=True)
+
+  MIMALLOC_LIB = os.path.abspath(os.path.join(LLVM_DIR, '..', 'llvm-build-tools/mimalloc-3.1.4')) + '/build/mimalloc.o'
+
+  mimalloc_ldflags = [ '' + MIMALLOC_LIB ]
+
+  return mimalloc_ldflags
 
 class ZStdDirs:
   """
@@ -457,6 +512,13 @@ def BuildZStd():
           '-DCMAKE_INSTALL_LIBDIR=lib',
           '-DCMAKE_MSVC_RUNTIME_LIBRARY=MultiThreaded',  # /MT to match LLVM.
           '-DZSTD_BUILD_SHARED=OFF',
+          '-DCMAKE_C_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang'),
+          '-DCMAKE_CXX_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang++'),
+          '-DCMAKE_C_FLAGS_RELEASE=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -w ',
+          '-DCMAKE_CXX_FLAGS_RELEASE=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -w ',
+          '-DCMAKE_EXE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -w ',
+          '-DCMAKE_SHARED_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -w ',
+          '-DCMAKE_MODULE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -w ',
           '../build/cmake',
       ],
       setenv=True)
@@ -703,21 +765,9 @@ def main():
                       action='store_true',
                       help='use the cmake from PATH instead of downloading '
                       'and using prebuilt cmake binaries')
-  parser.add_argument('--tf-path',
-                      help='path to python tensorflow pip package. '
-                      'Used for embedding an MLGO model')
-  parser.add_argument(
-      '--with-ml-inliner-model',
-      help='path to MLGO inliner model to embed. Setting to '
-      '\'default\', will download an official model which was '
-      'trained for Chrome on Android',
-      default='default' if sys.platform.startswith('linux') else '')
   parser.add_argument('--with-android', type=gn_arg, nargs='?', const=True,
                       help='build the Android ASan runtime (linux only)',
                       default=sys.platform.startswith('linux'))
-  parser.add_argument('--pic',
-                      action='store_true',
-                      help='Uses PIC when building LLVM')
   parser.add_argument('--with-fuchsia',
                       type=gn_arg,
                       nargs='?',
@@ -734,10 +784,6 @@ def main():
   parser.add_argument('--with-ccache',
                       action='store_true',
                       help='Use ccache to build the stage 1 compiler')
-  parser.add_argument('--without-zstd',
-                      dest='with_zstd',
-                      action='store_false',
-                      help='Disable zstd in the build')
 
   args = parser.parse_args()
 
@@ -748,8 +794,14 @@ def main():
     global RELEASE_VERSION
     RELEASE_VERSION = '22'
 
-  if (args.pgo or args.thinlto) and not args.bootstrap:
-    print('--pgo/--thinlto requires --bootstrap')
+  if not sys.platform.startswith('linux'):
+    print('Customized build script only works with Linux')
+    return 1
+  if not args.bootstrap:
+    print('Customized build script requires --bootstrap')
+    return 1
+  if args.skip_checkout:
+    print('Customized build script does not support --skip-checkout')
     return 1
   if args.with_android and not os.path.exists(ANDROID_NDK_DIR):
     print('Android NDK not found at ' + ANDROID_NDK_DIR)
@@ -774,10 +826,6 @@ def main():
     print('for general Fuchsia build instructions.')
     return 1
 
-  if args.with_ml_inliner_model and not sys.platform.startswith('linux'):
-    print('--with-ml-inliner-model only supports linux hosts')
-    return 1
-
   # Don't buffer stdout, so that print statements are immediately flushed.
   # LLVM tests print output without newlines, so with buffering they won't be
   # immediately printed.
@@ -838,12 +886,12 @@ def main():
   # LLVM_ENABLE_LLD).
   cc, cxx, lld = None, None, None
 
-  cflags = []
-  cxxflags = []
-  ldflags = []
+  cflags = ['-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -pthread -w']
+  cxxflags = ['-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -pthread -w']
+  ldflags = ['-fuse-ld=lld  -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -pthread -w']
 
   targets = 'AArch64;ARM;LoongArch;Mips;PowerPC;RISCV;SystemZ;WebAssembly;X86'
-  projects = 'clang;lld'
+  projects = 'clang;lld;polly'
   if not args.no_tools:
     projects += ';clang-tools-extra'
   if args.bolt:
@@ -856,7 +904,6 @@ def main():
     runtimes = 'compiler-rt'
 
   pic_default = sys.platform == 'win32'
-  pic_mode = 'ON' if args.pic or pic_default else 'OFF'
 
   base_cmake_args = [
       '-GNinja',
@@ -865,8 +912,8 @@ def main():
       f'-DLLVM_ENABLE_PROJECTS={projects}',
       f'-DLLVM_ENABLE_RUNTIMES={runtimes}',
       f'-DLLVM_TARGETS_TO_BUILD={targets}',
-      f'-DLLVM_ENABLE_PIC={pic_mode}',
-      '-DLLVM_ENABLE_TERMINFO=OFF',
+      f'-DLLVM_ENABLE_PIC=ON',
+      # '-DLLVM_ENABLE_TERMINFO=OFF',
       '-DLLVM_ENABLE_Z3_SOLVER=OFF',
       '-DCLANG_PLUGIN_SUPPORT=OFF',
       '-DCLANG_ENABLE_STATIC_ANALYZER=OFF',
@@ -890,7 +937,8 @@ def main():
       # `bootstrap` tool:
       # https://github.com/rust-lang/rust/blob/021861aea8de20c76c7411eb8ada7e8235e3d9b5/src/bootstrap/src/core/build_steps/llvm.rs#L348
       '-DLLVM_INSTALL_UTILS=ON',
-      '-DLLVM_ENABLE_ZSTD=%s' % ('ON' if args.with_zstd else 'OFF'),
+      '-DLLVM_ENABLE_ZSTD=ON',
+      '-DLLVM_ENABLE_BINDINGS=OFF'
   ]
 
   if sys.platform == 'darwin':
@@ -953,6 +1001,10 @@ def main():
     base_cmake_args.append('-DLLVM_WINSYSROOT="%s"' %
                            os.path.dirname(os.path.dirname(GetWinSDKDir())))
 
+  #Build mimalloc
+  libmimalloc_ldflags = BuildLibMimalloc()
+  ldflags += libmimalloc_ldflags
+
   # Statically link libxml2 to make lld-link not require mt.exe on Windows,
   # and to make sure lld-link output on other platforms is identical to
   # lld-link on Windows (for cross-builds).
@@ -962,13 +1014,12 @@ def main():
   cflags += libxml_cflags
   cxxflags += libxml_cflags
 
-  if args.with_zstd:
-    # Statically link zstd to make lld support zstd compression for debug info.
-    with timer.time('zstd build'):
-      zstd_cmake_args, zstd_cflags = BuildZStd()
-    base_cmake_args += zstd_cmake_args
-    cflags += zstd_cflags
-    cxxflags += zstd_cflags
+  # Statically link zstd to make lld support zstd compression for debug info.
+  with timer.time('zstd build'):
+    zstd_cmake_args, zstd_cflags = BuildZStd()
+  base_cmake_args += zstd_cmake_args
+  cflags += zstd_cflags
+  cxxflags += zstd_cflags
 
   lit_excludes = []
   if sys.platform.startswith('linux'):
@@ -1027,7 +1078,7 @@ def main():
       bootstrap_targets += ';ARM;AArch64'
     bootstrap_args = base_cmake_args + ccache_cmake_args + [
         '-DLLVM_TARGETS_TO_BUILD=' + bootstrap_targets,
-        '-DLLVM_ENABLE_PROJECTS=clang;lld',
+        '-DLLVM_ENABLE_PROJECTS=clang;lld;polly',
         '-DLLVM_ENABLE_RUNTIMES=' + ';'.join(runtimes),
         '-DCMAKE_INSTALL_PREFIX=' + LLVM_BOOTSTRAP_INSTALL_DIR,
         '-DCMAKE_C_FLAGS=' + ' '.join(cflags),
@@ -1085,6 +1136,9 @@ def main():
 
     print('Bootstrap compiler installed.')
 
+  cflags += ['-mllvm -polly']
+  cxxflags += ['-mllvm -polly']
+
   if args.pgo:
     print('Building instrumented compiler')
     if os.path.exists(LLVM_INSTRUMENTED_DIR):
@@ -1098,9 +1152,9 @@ def main():
     cxxflags += ['-Wno-backend-plugin']
 
     instrument_args = base_cmake_args + [
-        '-DLLVM_ENABLE_PROJECTS=clang',
-        '-DCMAKE_C_FLAGS=' + ' '.join(cflags),
-        '-DCMAKE_CXX_FLAGS=' + ' '.join(cxxflags),
+        '-DLLVM_ENABLE_PROJECTS=clang;polly',
+        '-DCMAKE_C_FLAGS= -mllvm -vp-counters-per-site=3 ' + ' '.join(cflags),
+        '-DCMAKE_CXX_FLAGS= -mllvm -vp-counters-per-site=3 ' + ' '.join(cxxflags),
         '-DCMAKE_EXE_LINKER_FLAGS=' + ' '.join(ldflags),
         '-DCMAKE_SHARED_LINKER_FLAGS=' + ' '.join(ldflags),
         '-DCMAKE_MODULE_LINKER_FLAGS=' + ' '.join(ldflags),
@@ -1146,8 +1200,11 @@ def main():
       with open(training_source, 'wb') as f:
         DownloadUrl(CDS_URL + '/' + training_source, f)
       train_cmd = [os.path.join(LLVM_INSTRUMENTED_DIR, 'bin', 'clang++'),
-                  '-target', 'x86_64-unknown-unknown', '-O2', '-g', '-std=c++14',
-                   '-fno-exceptions', '-fno-rtti', '-w', '-c', training_source]
+                 '-target', 'x86_64-unknown-unknown', '-O3', '-g', '-std=c++20',
+                 '-march=x86-64-v3', '-ffp-contract=fast', '-fmerge-all-constants',
+                 '-fdata-sections', '-ffunction-sections', '-fno-exceptions',
+                 '-fno-rtti', '-mllvm', '-polly', '-mllvm', '-polly-process-unprofitable',
+                 '-Rpass-analysis=polly', '-w', '-c', training_source]
       if sys.platform == 'darwin':
         train_cmd.extend(['-isysroot', isysroot])
       RunCommand(train_cmd, setenv=True)
@@ -1198,13 +1255,18 @@ def main():
   if cxx is not None: base_cmake_args.append('-DCMAKE_CXX_COMPILER=' + cxx)
   if lld is not None: base_cmake_args.append('-DCMAKE_LINKER=' + lld)
   final_install_dir = args.install_dir if args.install_dir else LLVM_BUILD_DIR
+  tf_path = subprocess.check_output(
+    ['vpython3', os.path.join(THIS_DIR, 'get_tensorflow.py')],
+    universal_newlines=True).rstrip()
+  base_cmake_args += [
+    '-DTENSORFLOW_AOT_PATH=%s' % tf_path,
+  ]
   cmake_args = base_cmake_args + [
-      '-DCMAKE_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),
       '-DCMAKE_INSTALL_PREFIX=' + final_install_dir,
+      '-DLLVM_INLINER_MODEL_CURRENT_URL=https://commondatastorage.googleapis.com/chromium-browser-clang/tools/mlgo_model3.tgz',
+      '-DLLVM_INLINER_MODEL_PATH=download',
+      '-DLLVM_RAEVICT_MODEL_CURRENT_URL=https://github.com/google/ml-compiler-opt/releases/download/regalloc-evict-v1.1/model.zip',
+      '-DLLVM_RAEVICT_MODEL_PATH=download',
   ]
   if not args.no_tools:
     cmake_args.extend([
@@ -1213,10 +1275,36 @@ def main():
         os.path.join(CHROMIUM_DIR, 'tools', 'clang'),
         '-DCHROMIUM_TOOLS=%s' % ';'.join(chrome_tools)
     ])
-  if args.pgo:
+  if args.pgo and not args.thinlto:
+    cmake_args.append('-DLLVM_PROFDATA_FILE=' + LLVM_PROFDATA_FILE)
+    cmake_args.append('-DCMAKE_C_FLAGS=-mllvm -enable-ext-tsp-block-placement ' + ' '.join(cflags))
+    cmake_args.append('-DCMAKE_CXX_FLAGS=-mllvm -enable-ext-tsp-block-placement ' + ' '.join(cxxflags))
+    cmake_args.append('-DCMAKE_EXE_LINKER_FLAGS=-Wl,-z,keep-text-section-prefix ' + ' '.join(ldflags))
+    cmake_args.append('-DCMAKE_SHARED_LINKER_FLAGS=-Wl,-z,keep-text-section-prefix ' + ' '.join(ldflags))
+    cmake_args.append('-DCMAKE_MODULE_LINKER_FLAGS=-Wl,-z,keep-text-section-prefix ' + ' '.join(ldflags))
+  if args.thinlto and not args.pgo:
+    cmake_args.append('-DLLVM_ENABLE_LTO=Thin')
+    cmake_args.append('-DLLVM_PARALLEL_LINK_JOBS=16')
+    cmake_args.append('-DCMAKE_C_FLAGS=-flto=thin -fwhole-program-vtables -fsplit-lto-unit ' + ' '.join(cflags))
+    cmake_args.append('-DCMAKE_CXX_FLAGS=-flto=thin -fwhole-program-vtables -fsplit-lto-unit ' + ' '.join(cxxflags))
+    cmake_args.append('-DCMAKE_EXE_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,-mllvm,-import-instr-limit=5 -Wl,-mllvm,-polly ' + ' '.join(ldflags))
+    cmake_args.append('-DCMAKE_SHARED_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,-mllvm,-import-instr-limit=5 -Wl,-mllvm,-polly ' + ' '.join(ldflags))
+    cmake_args.append('-DCMAKE_MODULE_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,-mllvm,-import-instr-limit=5 -Wl,-mllvm,-polly ' + ' '.join(ldflags))
+  if args.thinlto and args.pgo:
     cmake_args.append('-DLLVM_PROFDATA_FILE=' + LLVM_PROFDATA_FILE)
-  if args.thinlto:
     cmake_args.append('-DLLVM_ENABLE_LTO=Thin')
+    cmake_args.append('-DLLVM_PARALLEL_LINK_JOBS=16')
+    cmake_args.append('-DCMAKE_C_FLAGS=-flto=thin -fwhole-program-vtables -fsplit-lto-unit ' + ' '.join(cflags))
+    cmake_args.append('-DCMAKE_CXX_FLAGS=-flto=thin -fwhole-program-vtables -fsplit-lto-unit ' + ' '.join(cxxflags))
+    cmake_args.append('-DCMAKE_EXE_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,-mllvm,-import-instr-limit=5 -Wl,-mllvm,-enable-ext-tsp-block-placement -Wl,-z,keep-text-section-prefix -Wl,-mllvm,-polly ' + ' '.join(ldflags))
+    cmake_args.append('-DCMAKE_SHARED_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,-mllvm,-import-instr-limit=5 -Wl,-mllvm,-enable-ext-tsp-block-placement -Wl,-z,keep-text-section-prefix -Wl,-mllvm,-polly ' + ' '.join(ldflags))
+    cmake_args.append('-DCMAKE_MODULE_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,-mllvm,-import-instr-limit=5 -Wl,-mllvm,-enable-ext-tsp-block-placement -Wl,-z,keep-text-section-prefix -Wl,-mllvm,-polly ' + ' '.join(ldflags))
+  else:
+    cmake_args.append('-DCMAKE_C_FLAGS= ' + ' '.join(cflags))
+    cmake_args.append('-DCMAKE_CXX_FLAGS= ' + ' '.join(cxxflags))
+    cmake_args.append('-DCMAKE_EXE_LINKER_FLAGS= ' + ' '.join(ldflags))
+    cmake_args.append('-DCMAKE_SHARED_LINKER_FLAGS= ' + ' '.join(ldflags))
+    cmake_args.append('-DCMAKE_MODULE_LINKER_FLAGS= ' + ' '.join(ldflags))
   if sys.platform == 'win32':
     cmake_args.append('-DLLVM_ENABLE_ZLIB=FORCE_ON')
 
@@ -1258,6 +1346,11 @@ def main():
             # TODO(crbug.com/40242553): pass proper flags to i386 tests so they
             # compile correctly
             'LLVM_INCLUDE_TESTS=OFF',
+            'CMAKE_C_FLAGS=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -mllvm -polly -w ',
+            'CMAKE_CXX_FLAGS=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -mllvm -polly -w ',
+            'CMAKE_EXE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -mllvm -polly -w ',
+            'CMAKE_SHARED_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -mllvm -polly -w ',
+            'CMAKE_MODULE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -mllvm -polly -w '
         ],
         "profile":
         True,
@@ -1267,6 +1360,12 @@ def main():
     runtimes_triples_args['x86_64-unknown-linux-gnu'] = {
         "args": [
             'CMAKE_SYSROOT=%s' % sysroot_amd64,
+            'CMAKE_C_FLAGS=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -mllvm -polly -w ',
+            'CMAKE_CXX_FLAGS=-O3 -march=x86-64-v3 -fdata-sections -ffunction-sections -mllvm -polly -w ',
+            'CMAKE_EXE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -mllvm -polly -w ',
+            'CMAKE_SHARED_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -mllvm -polly -w ',
+            'CMAKE_MODULE_LINKER_FLAGS=-fuse-ld=lld -Wl,-O2 -Wl,--gc-sections -Wl,--icf=all -mllvm -polly -w ',
+
         ],
         "profile": True,
         "sanitizers": True,
@@ -1450,32 +1549,6 @@ def main():
           "profile": build_profile
       }
 
-  # Embed MLGO inliner model. If tf_path is not specified, a vpython3 env
-  # will be created which contains the necessary source files for compilation.
-  # MLGO is only officially supported on linux. This condition is checked at
-  # the top of main()
-  if args.with_ml_inliner_model:
-    if args.with_ml_inliner_model == 'default':
-      model_path = ('https://commondatastorage.googleapis.com/'
-                    'chromium-browser-clang/tools/mlgo_model3.tgz')
-    else:
-      model_path = args.with_ml_inliner_model
-    if not args.tf_path:
-      with timer.time('get tensorflow'):
-        tf_path = subprocess.check_output(
-            ['vpython3', os.path.join(THIS_DIR, 'get_tensorflow.py')],
-            universal_newlines=True).rstrip()
-    else:
-      tf_path = args.tf_path
-    print('Embedding MLGO inliner model at %s using Tensorflow at %s' %
-          (model_path, tf_path))
-    cmake_args += [
-        '-DLLVM_INLINER_MODEL_PATH=%s' % model_path,
-        '-DTENSORFLOW_AOT_PATH=%s' % tf_path,
-        # Disable Regalloc model generation since it is unused
-        '-DLLVM_RAEVICT_MODEL_PATH=none'
-    ]
-
   # Convert FOO=BAR CMake flags per triple into
   # -DBUILTINS_$triple_FOO=BAR/-DRUNTIMES_$triple_FOO=BAR and build up
   # -DLLVM_BUILTIN_TARGETS/-DLLVM_RUNTIME_TARGETS.
@@ -1534,7 +1607,7 @@ def main():
     with timer.time('bolt instrument'):
       RunCommand([
           'bin/llvm-bolt', 'bin/clang', '-o', 'bin/clang-bolt.inst',
-          '-instrument', '--instrumentation-file-append-pid',
+          '-lite=false', '-instrument', '--instrumentation-file-append-pid',
           '--instrumentation-file=' +
           os.path.join(bolt_profiles_dir, 'prof.fdata')
       ])
@@ -1547,14 +1620,18 @@ def main():
     # Train by building a part of Clang.
     os.mkdir('bolt-training')
     os.chdir('bolt-training')
+    if args.pgo:
+      cflags += ['-mllvm -enable-ext-tsp-block-placement']
+      cxxflags += ['-mllvm -enable-ext-tsp-block-placement']
+      ldflags += ['-Wl,-z,keep-text-section-prefix']
     bolt_train_cmake_args = base_cmake_args + [
         '-DLLVM_TARGETS_TO_BUILD=X86',
         '-DLLVM_ENABLE_PROJECTS=clang',
-        '-DCMAKE_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),
+        '-DCMAKE_C_FLAGS=-g -fmerge-all-constants -Rpass-analysis=polly ' + ' '.join(cflags),
+        '-DCMAKE_CXX_FLAGS=-g -fmerge-all-constants -Rpass-analysis=polly ' + ' '.join(cxxflags),
+        '-DCMAKE_EXE_LINKER_FLAGS= ' + ' '.join(ldflags),
+        '-DCMAKE_SHARED_LINKER_FLAGS= ' + ' '.join(ldflags),
+        '-DCMAKE_MODULE_LINKER_FLAGS= ' + ' '.join(ldflags),
         '-DCMAKE_C_COMPILER=' +
         os.path.join(LLVM_BUILD_DIR, 'bin/clang-bolt.inst'),
         '-DCMAKE_CXX_COMPILER=' +
@@ -1563,6 +1640,8 @@ def main():
         os.path.join(LLVM_BUILD_DIR, 'bin/clang-bolt.inst'),
         '-DCMAKE_ASM_COMPILER_ID=Clang',
     ]
+    if args.pgo:
+      bolt_train_cmake_args.append('-DLLVM_PROFDATA_FILE=' + LLVM_PROFDATA_FILE)
     with timer.time('bolt training cmake'):
       RunCommand(['cmake'] + bolt_train_cmake_args +
                  [os.path.join(LLVM_DIR, 'llvm')])
@@ -1582,9 +1661,15 @@ def main():
       ])
       RunCommand([
           'bin/llvm-bolt', 'bin/clang', '-o', 'bin/clang-bolt.opt', '-data',
-          'merged.fdata', '-reorder-blocks=ext-tsp', '-reorder-functions=hfsort+',
-          '-split-functions', '-split-all-cold', '-split-eh', '-dyno-stats',
-          '-icf=1', '-use-gnu-stack', '-use-old-text'
+          'merged.fdata', '-reorder-blocks=ext-tsp', '-reorder-functions=cdsort',
+          '-split-functions', '-split-strategy=cdsplit',
+          '-split-all-cold', '-split-eh', '-dyno-stats',
+          '-icf=all', '-use-gnu-stack', '-peepholes=all', '-elim-link-veneers',
+          '-group-stubs', '-align-blocks', '-sctc-mode=heuristic',
+          '-jump-tables=aggressive', '-simplify-rodata-loads',
+          '-tail-duplication=cache', '-indirect-call-promotion=all', '-icp-eliminate-loads',
+          '-hot-data', '-x86-strip-redundant-address-size', '-reorder-functions-use-hot-size',
+          '-reorder-data-algo=funcs', '-lite=false',
       ])
 
     # Overwrite clang, preserving its timestamp so ninja doesn't rebuild it.
@@ -1593,8 +1678,7 @@ def main():
 
   VerifyVersionOfBuiltClangMatchesVERSION()
   VerifyZlibSupport()
-  if args.with_zstd:
-    VerifyZStdSupport()
+  VerifyZStdSupport()
 
   # Run tests.
   if (chrome_tools and (args.run_tests or args.llvm_force_head_revision)):
@@ -1612,6 +1696,8 @@ def main():
 
   WriteStampFile(PACKAGE_VERSION, STAMP_FILE)
   WriteStampFile(PACKAGE_VERSION, FORCE_HEAD_REVISION_FILE)
+  RunCommand(['ln', '--symbolic', '--relative', '-f', 'bin/clang', 'bin/clang++'])
+  RunCommand(['ln', '--symbolic', '--relative', '-f', 'bin/clang', 'bin/clang-cl'])
 
   print('Clang build was successful.')