tc-build: Introduce '--muilticall'

Signed-off-by: Nathan Chancellor <nathan@kernel.org>

Author: nathanchance

build-llvm.py

@@ -236,6 +236,14 @@
                     '''),
                     type=str,
                     choices=['thin', 'full'])
+parser.add_argument('-m',
+                    '--multicall',
+                    help=textwrap.dedent('''\
+                    Build LLVM as a multicall binary via the LLVM_TOOL_LLVM_DRIVER_BUILD CMake option.
+                    This results in a much smaller installation on disk.
+
+                    '''),
+                    action='store_true')
 parser.add_argument('-n',
                     '--no-update',
                     help=textwrap.dedent('''\
@@ -536,6 +544,8 @@
 if args.vendor_string:
     common_cmake_defines['CLANG_VENDOR'] = args.vendor_string
     common_cmake_defines['LLD_VENDOR'] = args.vendor_string
+if args.multicall:
+    common_cmake_defines['LLVM_TOOL_LLVM_DRIVER_BUILD'] = 'ON'
 if args.defines:
     defines = dict(define.split('=', 1) for define in args.defines)
     common_cmake_defines.update(defines)

tc_build/kernel.py

@@ -38,8 +38,8 @@ def __init__(self, arch):
 
     def build(self):
         bin_folder = Path(self.toolchain_prefix, 'bin')
-        if self.bolt_instrumentation:
-            self.make_variables['CC'] = Path(bin_folder, 'clang.inst')
+        if self.bolt_instrumentation and (clang_inst := Path(bin_folder, 'clang.inst')).exists():
+            self.make_variables['CC'] = clang_inst
         # The user may have configured clang without the host target, in which
         # case we need to use GCC for compiling the host utilities.
         if self.can_use_clang_as_hostcc():

tc_build/llvm.py

@@ -74,26 +74,27 @@ def bolt_clang(self):
 
         tc_build.utils.print_header(f"Performing BOLT with {mode}")
 
-        # clang-#: original binary
-        # clang.bolt: BOLT optimized binary
+        # real_binary: clang-# or llvm (if multicall is enabled)
+        # bolt_binary: clang.bolt or llvm.bolt
         # .bolt will become original binary after optimization
-        clang = Path(self.folders.build, 'bin/clang').resolve()
-        clang_bolt = clang.with_name('clang.bolt')
+        real_binary = Path(self.folders.build, 'bin/clang').resolve()
+        binary_prefix = 'llvm' if self.multicall_is_enabled() else 'clang'
+        bolt_binary = real_binary.with_name(f"{binary_prefix}.bolt")
 
-        bolt_profile = Path(self.folders.build, 'clang.fdata')
+        bolt_profile = Path(self.folders.build, f"{binary_prefix}.fdata")
 
         if mode == 'instrumentation':
-            # clang.inst: instrumented binary, will be removed after generating profiles
-            clang_inst = clang.with_name('clang.inst')
+            # clang.inst / llvm.inst: instrumented binary, will be removed after generating profiles
+            inst_binary = real_binary.with_name(f"{binary_prefix}.inst")
 
             clang_inst_cmd = [
                 self.tools.llvm_bolt,
                 '--instrument',
                 f"--instrumentation-file={bolt_profile}",
                 '--instrumentation-file-append-pid',
                 '-o',
-                clang_inst,
-                clang,
+                inst_binary,
+                real_binary,
             ]
             # When running an instrumented binary on certain platforms (namely
             # Apple Silicon), there may be hangs due to instrumentation in
@@ -106,6 +107,13 @@ def bolt_clang(self):
 
             self.bolt_builder.bolt_instrumentation = True
 
+            # To avoid messing with all existing symlinks, move llvm to
+            # llvm.orig and llvm.inst to llvm
+            if binary_prefix == 'llvm':
+                orig_binary = real_binary.with_name('llvm.orig')
+                real_binary.replace(orig_binary)
+                inst_binary.replace(real_binary)
+
         if mode == 'sampling':
             self.bolt_builder.bolt_sampling_output = Path(self.folders.build, 'perf.data')
 
@@ -115,6 +123,11 @@ def bolt_clang(self):
         # With instrumentation, we need to combine the profiles we generated,
         # as they are separated by PID
         if mode == 'instrumentation':
+            # Undo shuffle from above
+            if binary_prefix == 'llvm':
+                real_binary.replace(inst_binary)
+                orig_binary.replace(real_binary)
+
             fdata_files = bolt_profile.parent.glob(f"{bolt_profile.name}.*.fdata")
 
             # merge-fdata will print one line for each .fdata it merges.
@@ -139,7 +152,7 @@ def bolt_clang(self):
                 self.bolt_builder.bolt_sampling_output,
                 '-o',
                 bolt_profile,
-                clang,
+                real_binary,
             ]
             self.run_cmd(perf2bolt_cmd)
             self.bolt_builder.bolt_sampling_output.unlink()
@@ -166,18 +179,18 @@ def bolt_clang(self):
             '--dyno-stats',
             f"--icf={icf_val}",
             '-o',
-            clang_bolt,
+            bolt_binary,
             f"--reorder-blocks={'cache+' if use_cache_plus else 'ext-tsp'}",
             f"--reorder-functions={reorder_funcs_val}",
             '--split-all-cold',
             f"--split-functions{'=3' if use_sf_val else ''}",
             '--use-gnu-stack',
-            clang,
+            real_binary,
         ]
         self.run_cmd(clang_opt_cmd)
-        clang_bolt.replace(clang)
+        bolt_binary.replace(real_binary)
         if mode == 'instrumentation':
-            clang_inst.unlink()
+            inst_binary.unlink()
 
     def build(self):
         if not self.folders.build:
@@ -388,6 +401,9 @@ def host_target(self):
     def host_target_is_enabled(self):
         return 'all' in self.targets or self.host_target() in self.targets
 
+    def multicall_is_enabled(self):
+        return self.cmake_defines.get('LLVM_TOOL_LLVM_DRIVER_BUILD', 'OFF') == 'ON'
+
     def project_is_enabled(self, project):
         return 'all' in self.projects or project in self.projects