compilers: Add support for using multiple threads with lto

Both Clang and GCC support using multiple threads for preforming link
time optimizaions, and they can now be configured using the
`-Db_lto_threads` option.

Fixes #7820

docs/markdown/Builtin-options.md

@@ -120,29 +120,32 @@ The following options are available. Note that they may not be
 available on all platforms or with all compilers:
 
 | Option        | Default value  | Possible values                                                  | Description                                                                   |
-| ----------- | ------------- | ---------------         | ----------- |
+|---------------|----------------|------------------------------------------------------------------|-------------------------------------------------------------------------------|
 | b_asneeded    | true           | true, false                                                      | Use -Wl,--as-needed when linking                                              |
 | b_bitcode     | false          | true, false                                                      | Embed Apple bitcode, see below                                                |
 | b_colorout    | always         | auto, always, never                                              | Use colored output                                                            |
 | b_coverage    | false          | true, false                                                      | Enable coverage tracking                                                      |
 | b_lundef      | true           | true, false                                                      | Don't allow undefined symbols when linking                                    |
 | b_lto         | false          | true, false                                                      | Use link time optimization                                                    |
+| b_lto_threads | 0              | Any integer*                                                     | Use multiple threads for lto. *(Added in 0.57.0)*                             |
 | b_ndebug      | false          | true, false, if-release                                          | Disable asserts                                                               |
 | b_pch         | true           | true, false                                                      | Use precompiled headers                                                       |
 | b_pgo         | off            | off, generate, use                                               | Use profile guided optimization                                               |
 | b_sanitize    | none           | see below                                                        | Code sanitizer to use                                                         |
 | b_staticpic   | true           | true, false                                                      | Build static libraries as position independent                                |
-| b_pie       | false         | true, false             | Build position-independent executables (since 0.49.0)|
-| b_vscrt     | from_buildtype| none, md, mdd, mt, mtd, from_buildtype, static_from_buildtype | VS runtime library to use (since 0.48.0) (static_from_buildtype since 0.56.0) |
+| b_pie         | false          | true, false                                                      | Build position-independent executables (since 0.49.0)                         |
+| b_vscrt       | from_buildtype | none, md, mdd, mt, mtd, from_buildtype, static_from_buildtype    | VS runtime library to use (since 0.48.0) (static_from_buildtype since 0.56.0) |
 
 The value of `b_sanitize` can be one of: `none`, `address`, `thread`,
 `undefined`, `memory`, `address,undefined`.
 
-<a name="b_vscrt-from_buildtype"></a> The default value of `b_vscrt`
-is `from_buildtype`. The following table is used internally to pick
-the CRT compiler arguments for `from_buildtype` or
-`static_from_buildtype` *(since 0.56)* based on the value of the
-`buildtype` option:
+* < 0 means disable, == 0 means automatic selection, > 0 sets a specific number to use
+
+<a name="b_vscrt-from_buildtype"></a>
+The default value of `b_vscrt` is `from_buildtype`. The following table is
+used internally to pick the CRT compiler arguments for `from_buildtype` or
+`static_from_buildtype` *(since 0.56)* based on the value of the `buildtype`
+option:
 
 | buildtype      | from_buildtype | static_from_buildtype |
 | --------       | -------------- | --------------------- |

docs/markdown/snippets/lto_threads.md

@@ -0,0 +1,7 @@
+## Knob to control LTO thread
+
+Both the gnu linker and lld support using threads for speading up LTO, meson
+now provides a knob for this: `-Db_lto_threads`. Currently this is only
+supported for clang and gcc. Any positive integer is supported, `0` means
+`auto`. If the compiler or linker implemnets it's on `auto` we use that,
+otherwise the number of threads on the machine is used.

mesonbuild/compilers/compilers.py

@@ -268,6 +268,8 @@ clike_debug_args = {False: [],
 base_options: 'KeyedOptionDictType' = {
     OptionKey('b_pch'): coredata.UserBooleanOption('Use precompiled headers', True),
     OptionKey('b_lto'): coredata.UserBooleanOption('Use link time optimization', False),
+    OptionKey('b_lto'): coredata.UserBooleanOption('Use link time optimization', False),
+    OptionKey('b_lto_threads'): coredata.UserIntegerOption('Use multiple threads for Link Time Optimization', (None, None,0)),
     OptionKey('b_sanitize'): coredata.UserComboOption('Code sanitizer to use',
                                                       ['none', 'address', 'thread', 'undefined', 'memory', 'address,undefined'],
                                                       'none'),
@@ -300,11 +302,25 @@ def option_enabled(boptions: T.Set[OptionKey], options: 'KeyedOptionDictType',
     except KeyError:
         return False
 
+
+def get_option_value(options: 'KeyedOptionDictType', opt: OptionKey, fallback: '_T') -> '_T':
+    """Get the value of an option, or the fallback value."""
+    try:
+        v: '_T' = options[opt].value
+    except KeyError:
+        return fallback
+
+    assert isinstance(v, type(fallback)), f'Should have {type(fallback)!r} but was {type(v)!r}'
+    # Mypy doesn't understand that the above assert ensures that v is type _T
+    return v
+
+
 def get_base_compile_args(options: 'KeyedOptionDictType', compiler: 'Compiler') -> T.List[str]:
     args = []  # type T.List[str]
     try:
         if options[OptionKey('b_lto')].value:
-            args.extend(compiler.get_lto_compile_args())
+            args.extend(compiler.get_lto_compile_args(
+                threads=get_option_value(options, OptionKey('b_lto_threads'), 0)))
     except KeyError:
         pass
     try:
@@ -926,7 +942,7 @@ class Compiler(metaclass=abc.ABCMeta):
             ret.append(arg)
         return ret
 
-    def get_lto_compile_args(self) -> T.List[str]:
+    def get_lto_compile_args(self, *, threads: int = 0) -> T.List[str]:
         return []
 
     def get_lto_link_args(self) -> T.List[str]:

mesonbuild/compilers/mixins/clang.py

@@ -49,7 +49,8 @@ class ClangCompiler(GnuLikeCompiler):
         super().__init__()
         self.id = 'clang'
         self.defines = defines or {}
-        self.base_options.add(OptionKey('b_colorout'))
+        self.base_options.update({OptionKey('b_colorout'), OptionKey('b_lto_threads')})
+
         # TODO: this really should be part of the linker base_options, but
         # linkers don't have base_options.
         if isinstance(self.linker, AppleDynamicLinker):
@@ -135,3 +136,10 @@ class ClangCompiler(GnuLikeCompiler):
 
     def get_coverage_link_args(self) -> T.List[str]:
         return ['--coverage']
+
+    def get_lto_compile_args(self, *, threads: int = 0) -> T.List[str]:
+        args = super().get_lto_compile_args(threads=threads)
+        # In clang -flto=0 means auto
+        if threads >= 0:
+            args.append(f'-flto-jobs={threads}')
+        return args

mesonbuild/compilers/mixins/gnu.py

@@ -17,6 +17,7 @@
 import abc
 import functools
 import os
+import multiprocessing
 import pathlib
 import re
 import subprocess
@@ -281,7 +282,9 @@ class GnuLikeCompiler(Compiler, metaclass=abc.ABCMeta):
                 return self._split_fetch_real_dirs(line.split('=', 1)[1])
         return []
 
-    def get_lto_compile_args(self) -> T.List[str]:
+    def get_lto_compile_args(self, *, threads: int = 0) -> T.List[str]:
+        # This provides a base for many compilers, GCC and Clang override this
+        # for their specific arguments
         return ['-flto']
 
     def sanitizer_compile_args(self, value: str) -> T.List[str]:
@@ -330,7 +333,7 @@ class GnuCompiler(GnuLikeCompiler):
         super().__init__()
         self.id = 'gcc'
         self.defines = defines or {}
-        self.base_options.add(OptionKey('b_colorout'))
+        self.base_options.update({OptionKey('b_colorout'), OptionKey('b_lto_threads')})
 
     def get_colorout_args(self, colortype: str) -> T.List[str]:
         if mesonlib.version_compare(self.version, '>=4.9.0'):
@@ -383,3 +386,13 @@ class GnuCompiler(GnuLikeCompiler):
 
     def get_prelink_args(self, prelink_name: str, obj_list: T.List[str]) -> T.List[str]:
         return ['-r', '-o', prelink_name] + obj_list
+
+    def get_lto_compile_args(self, *, threads: int = 0) -> T.List[str]:
+        if threads == 0:
+            if mesonlib.version_compare(self.version, '>= 10.0'):
+                return ['-flto=auto']
+            # This matches clang's behavior of using the number of cpus
+            return [f'-flto={multiprocessing.cpu_count()}']
+        elif threads > 0:
+            return [f'-flto={threads}']
+        return super().get_lto_compile_args(threads=threads)

run_unittests.py

@@ -2843,6 +2843,27 @@ class AllPlatformTests(BasePlatformTests):
         self.build()
         self.run_tests()
 
+    @skip_if_not_base_option('b_lto_threads')
+    def test_lto_threads(self):
+        testdir = os.path.join(self.common_test_dir, '6 linkshared')
+
+        env = get_fake_env(testdir, self.builddir, self.prefix)
+        cc = env.detect_c_compiler(MachineChoice.HOST)
+        if cc.get_id() == 'clang' and is_windows():
+            raise unittest.SkipTest('LTO not (yet) supported by windows clang')
+
+        self.init(testdir, extra_args=['-Db_lto=true', '-Db_lto_threads=8'])
+        self.build()
+        self.run_tests()
+
+        expected = set(cc.get_lto_compile_args(threads=8))
+        targets = self.introspect('--targets')
+        # This assumes all of the targets support lto
+        for t in targets:
+            for s in t['target_sources']:
+                for e in expected:
+                    self.assertIn(e, s['parameters'])
+
     def test_dist_git(self):
         if not shutil.which('git'):
             raise unittest.SkipTest('Git not found')