llvm-znver3-win.patch

diff --git a/tools/clang/scripts/build.py b/tools/clang/scripts/build.py
index 6d57def43567e..afd0e1a7864d0 100644
--- a/tools/clang/scripts/build.py
+++ b/tools/clang/scripts/build.py
@@ -69,7 +69,11 @@ ANDROID_NDK_CANARY_TOOLCHAIN_DIR = os.path.join(
     ANDROID_NDK_CANARY_DIR, ANDROID_NDK_TOOLCHAIN_RELATIVE_DIR)
 FUCHSIA_SDK_DIR = os.path.join(CHROMIUM_DIR, 'third_party', 'fuchsia-sdk',
                                'sdk')
-PINNED_CLANG_DIR = os.path.join(LLVM_BUILD_TOOLS_DIR, 'pinned-clang')
+
+if sys.platform == 'win32':
+    PINNED_CLANG_DIR = os.path.join('C:\Program Files', 'llvm').replace('\\', '/')
+else:
+    PINNED_CLANG_DIR = os.path.join(LLVM_BUILD_TOOLS_DIR, 'pinned-clang')
 
 BUG_REPORT_URL = ('https://crbug.com in the Tools>LLVM component,'
                   ' run tools/clang/scripts/process_crashreports.py'
@@ -78,6 +82,8 @@ 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/v2.1.2.tar.gz'
+
 win_sdk_dir = None
 def GetWinSDKDir():
   """Get the location of the current SDK."""
@@ -167,7 +173,7 @@ def CheckoutGitRepo(name, git_url, commit, dir):
     if (RunCommand(['git', 'diff-index', '--quiet', '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.

@@ -377,6 +383,11 @@ def BuildLibXml2():
           '-DLIBXML2_WITH_XPATH=OFF',
           '-DLIBXML2_WITH_XPTR=OFF',
           '-DLIBXML2_WITH_ZLIB=OFF',
+          '-DCMAKE_C_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang-cl.exe').replace('\\', '/'),
+          '-DCMAKE_C_FLAGS=-O3 -march=znver3 -w -D_WIN32 -D_CRT_DECLARE_NONSTDC_NAMES=1',
+          '-DCMAKE_EXE_LINKER_FLAGS=-w ',
+          '-DCMAKE_SHARED_LINKER_FLAGS=-w ',
+          '-DCMAKE_MODULE_LINKER_FLAGS=-w ',
           '..',
       ],
       setenv=True)
@@ -441,6 +452,8 @@ def BuildZStd():
   os.mkdir(dirs.build_dir)
   os.chdir(dirs.build_dir)
 
+
+
   RunCommand(
       [
           'cmake',
@@ -455,6 +468,13 @@ def BuildZStd():
           '-DCMAKE_MSVC_RUNTIME_LIBRARY=MultiThreaded',  # /MT to match LLVM.
           '-DZSTD_BUILD_SHARED=OFF',
           '../build/cmake',
+          '-DCMAKE_C_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang.exe').replace('\\', '/'),
+          '-DCMAKE_CXX_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang++.exe').replace('\\', '/'),
+          '-DCMAKE_C_FLAGS=-O3 -march=znver3 -w ',
+          '-DCMAKE_CXX_FLAGS=-O3 -march=znver3 -w ',
+          '-DCMAKE_EXE_LINKER_FLAGS=-w ',
+          '-DCMAKE_SHARED_LINKER_FLAGS=-w ',
+          '-DCMAKE_MODULE_LINKER_FLAGS=-w ',
       ],
       setenv=True)
   RunCommand(['ninja', 'install'], setenv=True)
@@ -473,6 +493,52 @@ def BuildZStd():
 
   return extra_cmake_flags, extra_cflags
 
+class LibMimallocDirs:
+  """
+  The set of directories where zstd 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-2.1.2')
+    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.exe').replace('\\', '/'),
+          '-DCMAKE_CXX_COMPILER=' + os.path.join(PINNED_CLANG_DIR, 'bin', 'clang++.exe').replace('\\', '/'),
+          '-DCMAKE_C_FLAGS=-O3 -march=znver3 -w ',
+          '-DCMAKE_CXX_FLAGS=-O3 -march=znver3 -w ',
+          '-DCMAKE_EXE_LINKER_FLAGS=-w ',
+          '-DCMAKE_SHARED_LINKER_FLAGS=-w ',
+          '-DCMAKE_MODULE_LINKER_FLAGS=-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-2.1.2')) + '/build/mimalloc-static.lib'
+
+  mimalloc_ldflags = [ '' + MIMALLOC_LIB ]
+
+  return mimalloc_ldflags
 
 def DownloadRPMalloc():
   """Download rpmalloc."""
@@ -508,7 +574,8 @@ def StartGomaAndGetGomaCCPath():
 
 def DownloadPinnedClang():
   PINNED_CLANG_VERSION = 'llvmorg-17-init-16420-g0c545a44-1'
-  DownloadAndUnpackPackage('clang', PINNED_CLANG_DIR, GetDefaultHostOs(),
+  if sys.platform != 'win32':
+    DownloadAndUnpackPackage('clang', PINNED_CLANG_DIR, GetDefaultHostOs(),
                            PINNED_CLANG_VERSION)
 
 
@@ -692,9 +759,6 @@ def main():
   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='?',
@@ -721,8 +785,11 @@ def main():
 
   global CLANG_REVISION, PACKAGE_VERSION, LLVM_BUILD_DIR
 
-  if (args.pgo or args.thinlto) and not args.bootstrap:
-    print('--pgo/--thinlto requires --bootstrap')
+  if not args.bootstrap:
+    print('Customized build script requires --bootstrap')
+    return 1
+  if args.with_goma:
+    print('Customized build script does not support --with-goma')
     return 1
   if args.with_android and not os.path.exists(ANDROID_NDK_DIR):
     print('Android NDK not found at ' + ANDROID_NDK_DIR)
@@ -804,17 +871,12 @@ def main():
   # LLVM_ENABLE_LLD).
   cc, cxx, lld = None, None, None
 
-  cflags = []
-  cxxflags = []
-  ldflags = []
+  cflags = ['-O3 -march=znver3 -w']
+  cxxflags = ['-O3 -march=znver3 -w']
+  ldflags = ['']
 
-  targets = 'AArch64;ARM;LoongArch;Mips;PowerPC;RISCV;SystemZ;WebAssembly;X86'
-  projects = 'clang;lld;clang-tools-extra'
-  if args.bolt:
-    projects += ';bolt'
-
-  pic_default = sys.platform == 'win32'
-  pic_mode = 'ON' if args.pic or pic_default else 'OFF'
+  targets = 'AArch64;ARM;X86'
+  projects = 'clang;lld;clang-tools-extra;polly'
 
   base_cmake_args = [
       '-GNinja',
@@ -823,7 +885,7 @@ def main():
       '-DLLVM_ENABLE_PROJECTS=' + projects,
       '-DLLVM_ENABLE_RUNTIMES=compiler-rt',
       '-DLLVM_TARGETS_TO_BUILD=' + targets,
-      f'-DLLVM_ENABLE_PIC={pic_mode}',
+      '-DLLVM_ENABLE_PIC=ON',
       '-DLLVM_ENABLE_TERMINFO=OFF',
       '-DLLVM_ENABLE_Z3_SOLVER=OFF',
       '-DCLANG_PLUGIN_SUPPORT=OFF',
@@ -919,9 +981,9 @@ def main():
     rpmalloc_dir = DownloadRPMalloc()
     base_cmake_args.append('-DLLVM_INTEGRATED_CRT_ALLOC=' + rpmalloc_dir)
 
-    # Set a sysroot to make the build more hermetic.
-    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
@@ -1021,8 +1083,8 @@ def main():
 
     instrument_args = base_cmake_args + [
         '-DLLVM_ENABLE_PROJECTS=clang',
-        '-DCMAKE_C_FLAGS=' + ' '.join(cflags),
-        '-DCMAKE_CXX_FLAGS=' + ' '.join(cxxflags),
+        '-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),
@@ -1039,37 +1101,40 @@ def main():
     RunCommand(['ninja', 'clang'], setenv=True)
     print('Instrumented compiler built.')
 
-    # Train by building some C++ code.
-    #
-    # pgo_training-1.ii is a preprocessed (on Linux) version of
-    # src/third_party/blink/renderer/core/layout/layout_object.cc, selected
-    # because it's a large translation unit in Blink, which is normally the
-    # slowest part of Chromium to compile. Using this, we get ~20% shorter
-    # build times for Linux, Android, and Mac, which is also what we got when
-    # training by actually building a target in Chromium. (For comparison, a
-    # C++-y "Hello World" program only resulted in 14% faster builds.)
-    # See https://crbug.com/966403#c16 for all numbers.
-    #
-    # Although the training currently only exercises Clang, it does involve LLVM
-    # internals, and so LLD also benefits when used for ThinLTO links.
-    #
-    # NOTE: Tidy uses binaries built with this profile, but doesn't seem to
-    # gain much from it. If tidy's execution time becomes a concern, it might
-    # be good to investigate that.
-    #
-    # TODO(hans): Enhance the training, perhaps by including preprocessed code
-    # from more platforms, and by doing some linking so that lld can benefit
-    # from PGO as well. Perhaps the training could be done asynchronously by
-    # dedicated buildbots that upload profiles to the cloud.
-    training_source = 'pgo_training-1.ii'
-    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]
-    if sys.platform == 'darwin':
-      train_cmd.extend(['-isysroot', isysroot])
-    RunCommand(train_cmd, setenv=True)
+    if sys.platform != 'win32':
+        # Train by building some C++ code.
+        #
+        # pgo_training-1.ii is a preprocessed (on Linux) version of
+        # src/third_party/blink/renderer/core/layout/layout_object.cc, selected
+        # because it's a large translation unit in Blink, which is normally the
+        # slowest part of Chromium to compile. Using this, we get ~20% shorter
+        # build times for Linux, Android, and Mac, which is also what we got when
+        # training by actually building a target in Chromium. (For comparison, a
+        # C++-y "Hello World" program only resulted in 14% faster builds.)
+        # See https://crbug.com/966403#c16 for all numbers.
+        #
+        # Although the training currently only exercises Clang, it does involve LLVM
+        # internals, and so LLD also benefits when used for ThinLTO links.
+        #
+        # NOTE: Tidy uses binaries built with this profile, but doesn't seem to
+        # gain much from it. If tidy's execution time becomes a concern, it might
+        # be good to investigate that.
+        #
+        # TODO(hans): Enhance the training, perhaps by including preprocessed code
+        # from more platforms, and by doing some linking so that lld can benefit
+        # from PGO as well. Perhaps the training could be done asynchronously by
+        # dedicated buildbots that upload profiles to the cloud.
+        training_source = 'pgo_training-1.ii'
+        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-pc-windows-msvc', '-O3', '-g', '-std=c++20',
+                     '-march=znver3', '-ffp-contract=fast', '-fmerge-all-constants',
+                     '-fdata-sections', '-ffunction-sections', '-fno-exceptions',
+                     '-fno-rtti', '-w', '-c', training_source]
+        if sys.platform == 'darwin':
+          train_cmd.extend(['-isysroot', isysroot])
+        RunCommand(train_cmd, setenv=True)
 
     # Merge profiles.
     profdata = os.path.join(LLVM_BOOTSTRAP_INSTALL_DIR, 'bin', 'llvm-profdata')
@@ -1116,11 +1181,6 @@ def main():
   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
   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,
   ]
   if not args.no_tools:
@@ -1130,10 +1190,51 @@ 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=-mllvm -enable-ext-tsp-block-placement ' + 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))
+    if sys.platform == 'win32':
+        cmake_args.append('-DCMAKE_EXE_LINKER_FLAGS=/OPT:REF /opt:ICF ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_SHARED_LINKER_FLAGS=/OPT:REF /opt:ICF ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_MODULE_LINKER_FLAGS=/OPT:REF /opt:ICF ' + ' '.join(ldflags))
+    else:
+        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))
+    if sys.platform == 'win32':
+        cmake_args.append('-DCMAKE_EXE_LINKER_FLAGS=/LTCG /OPT:REF /opt:ICF -mllvm:-import-instr-limit=30 ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_SHARED_LINKER_FLAGS=/LTCG /OPT:REF /opt:ICF -mllvm:-import-instr-limit=30 ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_MODULE_LINKER_FLAGS=/LTCG /OPT:REF /opt:ICF -mllvm:-import-instr-limit=30 ' + ' '.join(ldflags))
+    else:
+        cmake_args.append('-DCMAKE_EXE_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,--lto-CGO3 -Wl,-mllvm,-import-instr-limit=30 ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_SHARED_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,--lto-CGO3 -Wl,-mllvm,-import-instr-limit=30 ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_MODULE_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,--lto-CGO3 -Wl,-mllvm,-import-instr-limit=30 ' + ' '.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))
+    if sys.platform == 'win32':
+        cmake_args.append('-DCMAKE_EXE_LINKER_FLAGS=/LTCG /OPT:REF /opt:ICF -mllvm:-import-instr-limit=30 -mllvm:-enable-ext-tsp-block-placement ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_SHARED_LINKER_FLAGS=/LTCG /OPT:REF /opt:ICF -mllvm:-import-instr-limit=30 -mllvm:-enable-ext-tsp-block-placement ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_MODULE_LINKER_FLAGS=/LTCG /OPT:REF /opt:ICF -mllvm:-import-instr-limit=30 -mllvm:-enable-ext-tsp-block-placement ' + ' '.join(ldflags))
+    else:
+        cmake_args.append('-DCMAKE_EXE_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,--lto-CGO3 -Wl,-mllvm,-import-instr-limit=30 -Wl,-mllvm,-enable-ext-tsp-block-placement -Wl,-z,keep-text-section-prefix ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_SHARED_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,--lto-CGO3 -Wl,-mllvm,-import-instr-limit=30 -Wl,-mllvm,-enable-ext-tsp-block-placement -Wl,-z,keep-text-section-prefix ' + ' '.join(ldflags))
+        cmake_args.append('-DCMAKE_MODULE_LINKER_FLAGS=-flto=thin -fwhole-program-vtables -Wl,--lto-O3 -Wl,--lto-CGO3 -Wl,-mllvm,-import-instr-limit=30 -Wl,-mllvm,-enable-ext-tsp-block-placement -Wl,-z,keep-text-section-prefix ' + ' '.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')
 
@@ -1180,6 +1281,11 @@ def main():
             'CMAKE_SYSROOT=%s' % sysroot_i386,
             # TODO(https://crbug.com/1374690): pass proper flags to i386 tests so they compile correctly
             'LLVM_INCLUDE_TESTS=OFF',
+            'CMAKE_C_FLAGS=-O3 -march=znver3 -w ',
+            'CMAKE_CXX_FLAGS=-O3 -march=znver3 -w ',
+            'CMAKE_EXE_LINKER_FLAGS=-fuse-ld=lld -w ',
+            'CMAKE_SHARED_LINKER_FLAGS=-fuse-ld=lld -w ',
+            'CMAKE_MODULE_LINKER_FLAGS=-fuse-ld=lld -w ',
         ],
         "profile":
         True,
@@ -1189,6 +1295,11 @@ def main():
     runtimes_triples_args['x86_64-unknown-linux-gnu'] = {
         "args": [
             'CMAKE_SYSROOT=%s' % sysroot_amd64,
+            'CMAKE_C_FLAGS=-O3 -march=znver3 -w ',
+            'CMAKE_CXX_FLAGS=-O3 -march=znver3 -w ',
+            'CMAKE_EXE_LINKER_FLAGS=-fuse-ld=lld -w ',
+            'CMAKE_SHARED_LINKER_FLAGS=-fuse-ld=lld -w ',
+            'CMAKE_MODULE_LINKER_FLAGS=-fuse-ld=lld -w ',
         ],
         "profile": True,
         "sanitizers": True,
@@ -1227,7 +1338,9 @@ def main():
     runtimes_triples_args['i386-pc-windows-msvc'] = {
         "args": [
             'LLVM_ENABLE_PER_TARGET_RUNTIME_DIR=OFF',
-            'LLVM_WINSYSROOT="%s"' % sysroot,
+            'CMAKE_C_FLAGS=-O3 -march=znver3 -w ',
+            'CMAKE_CXX_FLAGS=-O3 -march=znver3 -w ',
+            'CMAKE_EXE_LINKER_FLAGS=/OPT:REF /OPT:ICF ',
         ],
         "profile":
         True,
@@ -1237,7 +1350,9 @@ def main():
     runtimes_triples_args['x86_64-pc-windows-msvc'] = {
         "args": [
             'LLVM_ENABLE_PER_TARGET_RUNTIME_DIR=OFF',
-            'LLVM_WINSYSROOT="%s"' % sysroot,
+            'CMAKE_C_FLAGS=-O3 -march=znver3 -w ',
+            'CMAKE_CXX_FLAGS=-O3 -march=znver3 -w ',
+            'CMAKE_EXE_LINKER_FLAGS=/OPT:REF /OPT:ICF ',
         ],
         "profile":
         True,
@@ -1247,9 +1362,11 @@ def main():
     runtimes_triples_args['aarch64-pc-windows-msvc'] = {
         "args": [
             'LLVM_ENABLE_PER_TARGET_RUNTIME_DIR=OFF',
-            'LLVM_WINSYSROOT="%s"' % sysroot,
             # Can't run tests on x86 host.
             'LLVM_INCLUDE_TESTS=OFF',
+            'CMAKE_C_FLAGS=-O3 -w ',
+            'CMAKE_CXX_FLAGS=-O3 -w ',
+            'CMAKE_EXE_LINKER_FLAGS=/OPT:REF /OPT:ICF ',
         ],
         "profile":
         True,
@@ -1308,7 +1425,6 @@ def main():
           'CMAKE_BUILD_TYPE=RelWithDebInfo',
           'CMAKE_C_FLAGS=' + ' '.join(android_cflags),
           'CMAKE_CXX_FLAGS=' + ' '.join(android_cflags),
-          'CMAKE_ASM_FLAGS=' + ' '.join(android_cflags),
           'COMPILER_RT_USE_BUILTINS_LIBRARY=ON',
           'SANITIZER_CXX_ABI=libcxxabi',
           'CMAKE_SHARED_LINKER_FLAGS=-Wl,-u__cxa_demangle',
@@ -1438,8 +1554,8 @@ def main():
     # Instrument.
     RunCommand([
         'bin/llvm-bolt', 'bin/clang', '-o', 'bin/clang-bolt.inst',
-        '-instrument', '--instrumentation-file-append-pid',
-        '--instrumentation-file=' +
+        '-lite=false','-instrument', '-instrumentation-file-append-pid',
+        '-instrumentation-file=' +
         os.path.join(bolt_profiles_dir, 'prof.fdata')
     ])
     RunCommand([
@@ -1454,8 +1570,8 @@ def main():
     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_C_FLAGS=-ffp-contract=fast -fmerge-all-constants ' + ' '.join(cflags),
+        '-DCMAKE_CXX_FLAGS=-ffp-contract=fast -fmerge-all-constants ' + ' '.join(cxxflags),
         '-DCMAKE_EXE_LINKER_FLAGS=' + ' '.join(ldflags),
         '-DCMAKE_SHARED_LINKER_FLAGS=' + ' '.join(ldflags),
         '-DCMAKE_MODULE_LINKER_FLAGS=' + ' '.join(ldflags),
@@ -1485,7 +1601,12 @@ def main():
         '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'
+        '-icf', '-use-gnu-stack', '-peepholes=all', '-elim-link-veneers',
+        '-group-stubs', '-align-blocks', '-sctc-mode=heuristic',
+        '-jump-tables=aggressive', '-simplify-rodata-loads', '-align-macro-fusion=all',
+        '-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.
@@ -1527,6 +1648,10 @@ 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.')
   return 0