kbuild: distributed build support for Clang ThinLTO

Add distributed ThinLTO build support for the Linux kernel.
This new mode offers several advantages: (1) Increased
flexibility in handling user-specified build options.
(2) Improved user-friendliness for developers. (3) Greater
convenience for integrating with objtool and livepatch.

Note that "distributed" in this context refers to a term
that differentiates in-process ThinLTO builds by invoking
backend compilation through the linker, not necessarily
building in distributed environments.

Distributed ThinLTO is enabled via the
`CONFIG_LTO_CLANG_THIN_DIST` Kconfig option. For example:
 > make LLVM=1 defconfig
 > scripts/config -e LTO_CLANG_THIN_DIST
 > make LLVM=1 oldconfig
 > make LLVM=1 vmlinux -j <..>

The build flow proceeds in four stages:
  1. Perform FE compilation, mirroring the in-process ThinLTO mode.
  2. Thin-link the generated IR files and object files.
  3. Find all IR files and perform BE compilation, using the flags
    stored in the .*.o.cmd files.
  4. Link the BE results to generate the final vmlinux.o.

NOTE: This patch currently implements the build for the main kernel
image (vmlinux) only. Kernel module support is planned for a
subsequent patch.

Tested on the following arch: x86, arm64, loongarch, and
riscv.

The earlier implementation details can be found here:
https://discourse.llvm.org/t/rfc-distributed-thinlto-build-for-kernel/85934

Signed-off-by: Rong Xu <xur@google.com>
Co-developed-by: Masahiro Yamada <masahiroy@kernel.org>
Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
Tested-by: Piotr Gorski <piotrgorski@cachyos.org>
Tested-by: Nathan Chancellor <nathan@kernel.org>
Link: https://patch.msgid.link/20260529185347.2418373-4-xur@google.com
Signed-off-by: Nathan Chancellor <nathan@kernel.org>

.gitignore

@@ -57,6 +57,7 @@
 *.zst
 Module.symvers
 dtbs-list
+builtin.order
 modules.order
 
 #
@@ -68,6 +69,7 @@ modules.order
 /vmlinux.32
 /vmlinux.map
 /vmlinux.symvers
+/vmlinux.thinlto-index
 /vmlinux.unstripped
 /vmlinux-gdb.py
 /vmlinuz

Makefile

@@ -1071,11 +1071,16 @@ export CC_FLAGS_SCS
 endif
 
 ifdef CONFIG_LTO_CLANG
-ifdef CONFIG_LTO_CLANG_THIN
-CC_FLAGS_LTO	:= -flto=thin -fsplit-lto-unit
-KBUILD_LDFLAGS += $(call ld-option,--lto-whole-program-visibility -mllvm -always-rename-promoted-locals=false)
-else
+ifdef CONFIG_LTO_CLANG_FULL
 CC_FLAGS_LTO	:= -flto
+else
+CC_FLAGS_LTO	:= -flto=thin -fsplit-lto-unit
+
+# These LLVM options were initially added with only in-process ThinLTO
+# support, so avoid distributed ThinLTO support for now.
+ifdef CONFIG_LTO_CLANG_THIN
+KBUILD_LDFLAGS += $(call ld-option,--lto-whole-program-visibility -mllvm -always-rename-promoted-locals=false)
+endif
 endif
 CC_FLAGS_LTO	+= -fvisibility=hidden
 
@@ -1684,6 +1689,7 @@ endif # CONFIG_MODULES
 CLEAN_FILES += vmlinux.symvers modules-only.symvers \
 	       modules.builtin modules.builtin.modinfo modules.nsdeps \
 	       modules.builtin.ranges vmlinux.o.map vmlinux.unstripped \
+	       vmlinux.thinlto-index builtin.order \
 	       compile_commands.json rust/test \
 	       rust-project.json .vmlinux.objs .vmlinux.export.c \
                .builtin-dtbs-list .builtin-dtbs.S
@@ -2145,7 +2151,7 @@ clean: $(clean-dirs)
 	$(call cmd,rmfiles)
 	@find . $(RCS_FIND_IGNORE) \
 		\( -name '*.[aios]' -o -name '*.rsi' -o -name '*.ko' -o -name '.*.cmd' \
-		-o -name '*.ko.*' \
+		-o -name '*.ko.*' -o -name '*.o.thinlto.bc' \
 		-o -name '*.dtb' -o -name '*.dtbo' \
 		-o -name '*.dtb.S' -o -name '*.dtbo.S' \
 		-o -name '*.dt.yaml' -o -name 'dtbs-list' \

arch/Kconfig

@@ -858,6 +858,25 @@ config LTO_CLANG_THIN
 	    https://clang.llvm.org/docs/ThinLTO.html
 
 	  If unsure, say Y.
+
+config LTO_CLANG_THIN_DIST
+	bool "Clang ThinLTO in distributed mode (EXPERIMENTAL)"
+	depends on HAS_LTO_CLANG && ARCH_SUPPORTS_LTO_CLANG_THIN
+	select LTO_CLANG
+	help
+	  This option enables Clang's ThinLTO in distributed build mode.
+	  In this mode, the linker performs the thin-link, generating
+	  ThinLTO index files. Subsequently, the build system explicitly
+	  invokes ThinLTO backend compilation using these index files
+	  and pre-linked IR objects. The resulting native object files
+	  are with the .thinlto-native.o suffix.
+
+	  This build mode offers improved visibility into the ThinLTO
+	  process through explicit subcommand exposure. It also makes
+	  final native object files directly available, benefiting
+	  tools like objtool and kpatch. Additionally, it provides
+	  crucial granular control over back-end options, enabling
+	  module-specific compiler options, and simplifies debugging.
 endchoice
 
 config ARCH_SUPPORTS_AUTOFDO_CLANG

scripts/Makefile.lib

@@ -249,6 +249,13 @@ ifdef CONFIG_LTO_CLANG
 cmd_ld_single = $(if $(objtool-enabled)$(is-single-obj-m), ; $(LD) $(ld_flags) -r -o $(tmp-target) $@; mv $(tmp-target) $@)
 endif
 
+ifdef CONFIG_LTO_CLANG_THIN_DIST
+# Save the _c_flags, sliently.
+quiet_cmd_save_c_flags =
+      saved_c_flags = $(_c_flags) $(modkern_cflags)
+      cmd_save_c_flags = printf '\n%s\n' 'saved_c_flags_$@ := $(call escsq,$(saved_c_flags))' >> $(dot-target).cmd
+endif
+
 quiet_cmd_cc_o_c = CC $(quiet_modtag)  $@
       cmd_cc_o_c = $(CC) $(c_flags) -c -o $@ $< \
 		$(cmd_ld_single) \
@@ -256,6 +263,7 @@ quiet_cmd_cc_o_c = CC $(quiet_modtag)  $@
 
 define rule_cc_o_c
 	$(call cmd_and_fixdep,cc_o_c)
+	$(call cmd,save_c_flags)
 	$(call cmd,checksrc)
 	$(call cmd,checkdoc)
 	$(call cmd,gen_objtooldep)

scripts/Makefile.thinlto

@@ -0,0 +1,40 @@
+PHONY := __default
+__default:
+
+include include/config/auto.conf
+include $(srctree)/scripts/Kbuild.include
+include $(srctree)/scripts/Makefile.lib
+
+native-objs := $(patsubst %.o,%.thinlto-native.o,$(call read-file, vmlinux.thinlto-index))
+
+__default: $(native-objs)
+
+# Generate .thinlto-native.o (obj) from .o (bitcode) and .thinlto.bc (summary) files
+# ---------------------------------------------------------------------------
+quiet_cmd_cc_o_bc = CC $(quiet_modtag)  $@
+      be_flags = $(shell sed -n '/saved_c_flags_/s/.*:= //p' \
+                $(dir $(<)).$(notdir $(<)).cmd)
+      cmd_cc_o_bc = \
+      $(CC) $(be_flags) -x ir -fno-lto -Wno-unused-command-line-argument \
+      -fthinlto-index=$(word 2, $^) -c -o $@ $<
+
+targets += $(native-objs)
+$(native-objs): %.thinlto-native.o: %.o %.o.thinlto.bc   FORCE
+	$(call if_changed,cc_o_bc)
+
+# Add FORCE to the prerequisites of a target to force it to be always rebuilt.
+# ---------------------------------------------------------------------------
+
+PHONY += FORCE
+FORCE:
+
+# Read all saved command lines and dependencies for the $(targets) we
+# may be building above, using $(if_changed{,_dep}). As an
+# optimization, we don't need to read them if the target does not
+# exist, we will rebuild anyway in that case.
+
+existing-targets := $(wildcard $(sort $(targets)))
+
+-include $(foreach f, $(existing-targets),$(dir $(f)).$(notdir $(f)).cmd)
+
+.PHONY: $(PHONY)

scripts/Makefile.vmlinux_a

@@ -21,6 +21,41 @@ targets += built-in-fixup.a
 built-in-fixup.a: $(KBUILD_VMLINUX_OBJS) scripts/head-object-list.txt FORCE
 	$(call if_changed,ar_builtin_fixup)
 
+ifdef CONFIG_LTO_CLANG_THIN_DIST
+
+quiet_cmd_builtin.order = GEN     $@
+      cmd_builtin.order = $(AR) t $< > $@
+
+targets += builtin.order
+builtin.order: built-in-fixup.a FORCE
+	$(call if_changed,builtin.order)
+
+quiet_cmd_ld_thinlto_index = LD      $@
+      cmd_ld_thinlto_index = \
+	$(LD) $(KBUILD_LDFLAGS) -r --thinlto-index-only=$@ @$<
+
+targets += vmlinux.thinlto-index
+vmlinux.thinlto-index: builtin.order FORCE
+	$(call if_changed,ld_thinlto_index)
+
+quiet_cmd_ar_vmlinux.a = GEN     $@
+      cmd_ar_vmlinux.a =					\
+	rm -f $@;						\
+	while read -r obj; do					\
+		if grep -Fqx $${obj} $(word 2, $^); then	\
+			echo $${obj%.o}.thinlto-native.o;	\
+		else						\
+			echo $${obj};				\
+		fi;						\
+	done < $< | xargs $(AR) cDPrS --thin $@
+
+targets += vmlinux.a
+vmlinux.a: builtin.order vmlinux.thinlto-index FORCE
+	$(Q)$(MAKE) -f $(srctree)/scripts/Makefile.thinlto
+	$(call if_changed,ar_vmlinux.a)
+
+else
+
 # vmlinux.a
 # ---------------------------------------------------------------------------
 
@@ -28,6 +63,8 @@ targets += vmlinux.a
 vmlinux.a: built-in-fixup.a FORCE
 	$(call if_changed,copy)
 
+endif
+
 # Add FORCE to the prerequisites of a target to force it to be always rebuilt.
 # ---------------------------------------------------------------------------
 

scripts/mod/modpost.c

@@ -1487,13 +1487,22 @@ static void extract_crcs_for_object(const char *object, struct module *mod)
 	char cmd_file[PATH_MAX];
 	char *buf, *p;
 	const char *base;
-	int dirlen, ret;
+	int dirlen, baselen_without_suffix, ret;
 
 	base = get_basename(object);
 	dirlen = base - object;
 
-	ret = snprintf(cmd_file, sizeof(cmd_file), "%.*s.%s.cmd",
-		       dirlen, object, base);
+	baselen_without_suffix = strlen(object) - dirlen - strlen(".o");
+
+	/*
+	 * When CONFIG_LTO_CLANG_THIN_DIST=y, the ELF is *.thinlto-native.o
+	 * but the symbol CRCs are recorded in *.o.cmd file.
+	 */
+	if (strends(object, ".thinlto-native.o"))
+		baselen_without_suffix -= strlen(".thinlto-native");
+
+	ret = snprintf(cmd_file, sizeof(cmd_file), "%.*s.%.*s.o.cmd",
+		       dirlen, object, baselen_without_suffix, base);
 	if (ret >= sizeof(cmd_file)) {
 		error("%s: too long path was truncated\n", cmd_file);
 		return;