| Message ID | 20171115213428.22559-13-samitolvanen@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Wed, Nov 15, 2017 at 1:34 PM, Sami Tolvanen <samitolvanen@google.com> wrote: > This change adds the configuration option CONFIG_LTO_CLANG, and > build system support for clang's Link Time Optimization (LTO). In > preparation for LTO support for other compilers, potentially common > parts of the changes are gated behind CONFIG_LTO instead. > > With -flto, instead of object files, clang produces LLVM bitcode, > which is compiled into a native object at link time, allowing the > final binary to be optimized globally. For more details, see: > > https://llvm.org/docs/LinkTimeOptimization.html > > While the kernel normally uses GNU ld for linking, LLVM supports LTO > only with lld or GNU gold linkers. This patch set assumes gold will > be used with the LLVMgold plug-in to perform the LTO link step. Due > to potential incompatibilities with GNU ld, this change also adds > LDFINAL_vmlinux for using a different linker for the vmlinux_link > step, and defaults to using GNU ld. > > Assuming LLVMgold.so is in LD_LIBRARY_PATH and CONFIG_LTO_CLANG has > been selected, an LTO kernel can be built simply by running make > CC=clang. Recommended versions are >= 5.0 for clang, and >= 2.27 for > binutils. > > Signed-off-by: Sami Tolvanen <samitolvanen@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> I wonder if this patch could be split into a few pieces. e.g. linker script changes, modpost changes, core build changes, and Kconfig? The changes area already pretty confined to specific files, so maybe this isn't really needed, but it might make review easier. -Kees > --- > .gitignore | 2 ++ > Makefile | 54 +++++++++++++++++++++++++++- > arch/Kconfig | 32 +++++++++++++++++ > scripts/Makefile.build | 66 +++++++++++++++++++++++++++++++++- > scripts/Makefile.modpost | 63 ++++++++++++++++++++++++++++----- > scripts/link-vmlinux.sh | 92 ++++++++++++++++++++++++++++++++++++++++++------ > 6 files changed, 288 insertions(+), 21 deletions(-) > > diff --git a/.gitignore b/.gitignore > index 6c119eab5d46..ac236e2bb9b1 100644 > --- a/.gitignore > +++ b/.gitignore > @@ -11,6 +11,7 @@ > # > .* > *.a > +*.a.* > *.bin > *.bz2 > *.c.[012]*.* > @@ -28,6 +29,7 @@ > *.lzma > *.lzo > *.mod.c > +*.modversions > *.o > *.o.* > *.order > diff --git a/Makefile b/Makefile > index f976af9525bf..8141b4c8f1bf 100644 > --- a/Makefile > +++ b/Makefile > @@ -350,6 +350,7 @@ include scripts/Kbuild.include > # Make variables (CC, etc...) > AS = $(CROSS_COMPILE)as > LD = $(CROSS_COMPILE)ld > +LDGOLD = $(CROSS_COMPILE)ld.gold > CC = $(CROSS_COMPILE)gcc > CPP = $(CC) -E > AR = $(CROSS_COMPILE)ar > @@ -623,6 +624,15 @@ endif > # Defaults to vmlinux, but the arch makefile usually adds further targets > all: vmlinux > > +# Make toolchain changes before including arch/$(SRCARCH)/Makefile to ensure > +# cc-/ld-* macros return correct values. > +ifdef CONFIG_LTO_CLANG > +# use GNU gold with LLVMgold for LTO linking, and LD for vmlinux_link > +LDFINAL_vmlinux := $(LD) > +LD := $(LDGOLD) > +LDFLAGS_GOLD += -plugin LLVMgold.so > +endif > + > # The arch Makefile can set ARCH_{CPP,A,C}FLAGS to override the default > # values of the respective KBUILD_* variables > ARCH_CPPFLAGS := > @@ -789,6 +799,32 @@ KBUILD_CFLAGS += $(call cc-option,-ffunction-sections,) > KBUILD_CFLAGS += $(call cc-option,-fdata-sections,) > endif > > +ifdef CONFIG_LTO_CLANG > +lto-clang-flags := -flto -fvisibility=hidden > + > +# allow disabling only clang LTO where needed > +DISABLE_LTO_CLANG := -fno-lto > +export DISABLE_LTO_CLANG > + > +ifdef CONFIG_MODVERSIONS > +# llvm-dis is used instead of objdump to process LLVM IR files > +LLVM_DIS := llvm-dis > +export LLVM_DIS > +endif > +endif > + > +ifdef CONFIG_LTO > +lto-flags := $(lto-clang-flags) > +KBUILD_CFLAGS += $(lto-flags) > + > +DISABLE_LTO := $(DISABLE_LTO_CLANG) > +export DISABLE_LTO > + > +# LDFINAL_vmlinux and LDFLAGS_FINAL_vmlinux can be set to override > +# the linker and flags for vmlinux_link. > +export LDFINAL_vmlinux LDFLAGS_FINAL_vmlinux > +endif > + > # arch Makefile may override CC so keep this after arch Makefile is included > NOSTDINC_FLAGS += -nostdinc -isystem $(shell $(CC) -print-file-name=include) > CHECKFLAGS += $(NOSTDINC_FLAGS) > @@ -1090,6 +1126,19 @@ prepare-objtool: $(objtool_target) > # CC_STACKPROTECTOR_STRONG! Why did it build with _REGULAR?!") > PHONY += prepare-compiler-check > prepare-compiler-check: FORCE > +# Make sure we're using clang with LTO_CLANG > +ifdef CONFIG_LTO_CLANG > + ifneq ($(cc-name),clang) > + @echo Cannot use CONFIG_LTO_CLANG without CC=clang >&2 && exit 1 > + endif > +endif > +# Make sure compiler supports LTO flags > +ifdef lto-flags > + ifeq ($(call cc-option, $(lto-flags)),) > + @echo Cannot use CONFIG_LTO: $(lto-flags) not supported by compiler \ > + >&2 && exit 1 > + endif > +endif > # Make sure compiler supports requested stack protector flag. > ifdef stackp-name > ifeq ($(call cc-option, $(stackp-flag)),) > @@ -1564,7 +1613,10 @@ clean: $(clean-dirs) > -o -name modules.builtin -o -name '.tmp_*.o.*' \ > -o -name '*.c.[012]*.*' \ > -o -name '*.ll' \ > - -o -name '*.gcno' \) -type f -print | xargs rm -f > + -o -name '*.gcno' \ > + -o -name '*.[oa].objects' \ > + -o -name '*.o.symversions' \ > + -o -name '*.modversions' \) -type f -print | xargs rm -f > > # Generate tags for editors > # --------------------------------------------------------------------------- > diff --git a/arch/Kconfig b/arch/Kconfig > index 400b9e1b2f27..bb5296ecebdd 100644 > --- a/arch/Kconfig > +++ b/arch/Kconfig > @@ -585,6 +585,7 @@ config CC_STACKPROTECTOR_STRONG > endchoice > > config THIN_ARCHIVES > + depends on !LTO_CLANG > def_bool y > help > Select this if the architecture wants to use thin archives > @@ -605,6 +606,37 @@ config LD_DEAD_CODE_DATA_ELIMINATION > sections (e.g., '.text.init'). Typically '.' in section names > is used to distinguish them from label names / C identifiers. > > +config LTO > + bool > + > +config ARCH_SUPPORTS_LTO_CLANG > + bool > + help > + An architecture should select this option it supports: > + - compiling with clang, > + - compiling inline assembly with clang's integrated assembler, > + - and linking with either lld or GNU gold w/ LLVMgold. > + > +config LTO_CLANG > + bool "Use clang Link Time Optimization (LTO)" > + depends on ARCH_SUPPORTS_LTO_CLANG > + depends on !FTRACE_MCOUNT_RECORD > + select LTO > + select LD_DEAD_CODE_DATA_ELIMINATION > + help > + This option enables clang's Link Time Optimization (LTO), which allows > + the compiler to optimize the kernel globally at link time. If you > + enable this option, the compiler generates LLVM IR instead of object > + files, and the actual compilation from IR occurs at the LTO link step, > + which may take several minutes. > + > + If you select this option, you must compile the kernel with clang > + (make CC=clang) and have the LVMgold plug-in in LD_LIBRARY_PATH. > + > + Using clang >= 5.0 and GNU gold from binutils >= 2.27 is recommended > + for this option. > + > + > config HAVE_ARCH_WITHIN_STACK_FRAMES > bool > help > diff --git a/scripts/Makefile.build b/scripts/Makefile.build > index e63af4e19382..e8bf5c440612 100644 > --- a/scripts/Makefile.build > +++ b/scripts/Makefile.build > @@ -210,6 +210,23 @@ else > > cmd_cc_o_c = $(CC) $(c_flags) -c -o $(@D)/.tmp_$(@F) $< > > +ifdef CONFIG_LTO_CLANG > +# Generate .o.symversions files for each .o with exported symbols, and link these > +# to the kernel and/or modules at the end. > +cmd_modversions_c = \ > + if echo '$(c_flags)' | grep -q -- '$(DISABLE_LTO_CLANG)'; then \ > + if $(OBJDUMP) -h $(@D)/.tmp_$(@F) | grep -q __ksymtab; then \ > + $(call cmd_gensymtypes_c,$(KBUILD_SYMTYPES),$(@:.o=.symtypes)) \ > + > $(@D)/$(@F).symversions; \ > + fi; \ > + else \ > + if $(LLVM_DIS) -o=- $(@D)/.tmp_$(@F) | grep -q __ksymtab; then \ > + $(call cmd_gensymtypes_c,$(KBUILD_SYMTYPES),$(@:.o=.symtypes)) \ > + > $(@D)/$(@F).symversions; \ > + fi; \ > + fi; \ > + mv -f $(@D)/.tmp_$(@F) $@; > +else > cmd_modversions_c = \ > if $(OBJDUMP) -h $(@D)/.tmp_$(@F) | grep -q __ksymtab; then \ > $(call cmd_gensymtypes_c,$(KBUILD_SYMTYPES),$(@:.o=.symtypes)) \ > @@ -222,6 +239,7 @@ cmd_modversions_c = \ > mv -f $(@D)/.tmp_$(@F) $@; \ > fi; > endif > +endif > > ifdef CONFIG_FTRACE_MCOUNT_RECORD > ifdef BUILD_C_RECORDMCOUNT > @@ -443,6 +461,13 @@ $(obj)/%-asn1.c $(obj)/%-asn1.h: $(src)/%.asn1 $(objtree)/scripts/asn1_compiler > # To build objects in subdirs, we need to descend into the directories > $(sort $(subdir-obj-y)): $(subdir-ym) ; > > +ifdef CONFIG_LTO_CLANG > +# If LTO is enabled, we remove all intermediate linking steps and instead > +# collect a list of all objects to be linked at the end. > +cc_lto_objects = $(foreach o,$(1),\ > + [ -f ${o}.objects ] && cat ${o}.objects || echo ${o};) > +endif > + > # > # Rule to compile a set of .o files into one .o file > # > @@ -458,6 +483,16 @@ else > quiet_cmd_link_o_target = LD $@ > endif > > +ifdef CONFIG_LTO_CLANG > +builtin-cmds = $(call cc_lto_objects,$(filter $(obj-y), $^)) > + > +quiet_cmd_update_builtin = GEN $@ > +cmd_update_builtin = (cat /dev/null; $(builtin-cmds)) > $@.objects && \ > + cat /dev/null > $@ > + > +$(builtin-target): $(obj-y) FORCE > + $(call if_changed,update_builtin) > +else > # If the list of objects to link is empty, just create an empty built-in.o > cmd_link_o_target = $(if $(strip $(obj-y)),\ > $(cmd_make_builtin) $@ $(filter $(obj-y), $^) \ > @@ -466,6 +501,7 @@ cmd_link_o_target = $(if $(strip $(obj-y)),\ > > $(builtin-target): $(obj-y) FORCE > $(call if_changed,link_o_target) > +endif > > targets += $(builtin-target) > endif # builtin-target > @@ -487,6 +523,16 @@ $(modorder-target): $(subdir-ym) FORCE > # Rule to compile a set of .o files into one .a file > # > ifdef lib-target > +ifdef CONFIG_LTO_CLANG > +lib-target-cmds = $(call cc_lto_objects,$(lib-y)) > + > +quiet_cmd_update_lib_target = GEN $@ > +cmd_update_lib_target = (cat /dev/null; $(lib-target-cmds)) > $@.objects && \ > + cat /dev/null > $@ > + > +$(lib-target): $(lib-y) FORCE > + $(call if_changed,update_lib_target) > +else > quiet_cmd_link_l_target = AR $@ > > ifdef CONFIG_THIN_ARCHIVES > @@ -497,6 +543,7 @@ endif > > $(lib-target): $(lib-y) FORCE > $(call if_changed,link_l_target) > +endif > > targets += $(lib-target) > > @@ -552,14 +599,31 @@ endif > quiet_cmd_link_multi-m = LD [M] $@ > cmd_link_multi-m = $(cmd_link_multi-link) > > +ifdef CONFIG_LTO_CLANG > +multi-deps-cmds = $(call cc_lto_objects,$(link_multi_deps)) > + > +quiet_cmd_update_multi_deps = GEN $@ > +cmd_update_multi_deps = (cat /dev/null; $(multi-deps-cmds)) > $@.objects && \ > + cat /dev/null > $@ > + > +$(multi-used-y): FORCE > + $(call if_changed,update_multi_deps) > + > +$(multi-used-m): FORCE > + $(call if_changed,update_multi_deps) > + @{ echo $(@:.o=.ko); echo $(link_multi_deps); \ > + $(cmd_undef_syms); } > $(MODVERDIR)/$(@F:.o=.mod) > +else > $(multi-used-y): FORCE > $(call if_changed,link_multi-y) > -$(call multi_depend, $(multi-used-y), .o, -objs -y) > > $(multi-used-m): FORCE > $(call if_changed,link_multi-m) > @{ echo $(@:.o=.ko); echo $(link_multi_deps); \ > $(cmd_undef_syms); } > $(MODVERDIR)/$(@F:.o=.mod) > +endif > + > +$(call multi_depend, $(multi-used-y), .o, -objs -y) > $(call multi_depend, $(multi-used-m), .o, -objs -y -m) > > targets += $(multi-used-y) $(multi-used-m) > diff --git a/scripts/Makefile.modpost b/scripts/Makefile.modpost > index 991db7d6e4df..cb1c040a006c 100644 > --- a/scripts/Makefile.modpost > +++ b/scripts/Makefile.modpost > @@ -83,12 +83,46 @@ modpost = scripts/mod/modpost \ > > MODPOST_OPT=$(subst -i,-n,$(filter -i,$(MAKEFLAGS))) > > +# If CONFIG_LTO_CLANG is enabled, .o files are either LLVM IR, or empty, so we > +# need to link them into actual objects before passing them to modpost > +modpost-ext = $(if $(CONFIG_LTO_CLANG),.lto,) > + > +ifdef CONFIG_LTO_CLANG > +quiet_cmd_cc_lto_modversions = GEN [M] $@ > +cmd_cc_lto_modversions = \ > + rm -f $(@); \ > + if [ -f $(@:.modversions=.o).objects ]; then \ > + for i in `cat $(@:.modversions=.o).objects`; do \ > + [ -s $$i.symversions ] && \ > + cat $$i.symversions >> $(@); \ > + done; \ > + else \ > + [ -s $(@:.modversions=.o).symversions ] && \ > + cat $(@:.modversions=.o).symversions >> $(@); \ > + fi > + > +$(modules:.ko=.modversions): FORCE > + $(call if_changed,cc_lto_modversions) > + > +quiet_cmd_cc_lto_link_modules = LD [M] $@ > +cmd_cc_lto_link_modules = \ > + $(LD) $(ld_flags) -r -o $(@) \ > + $(shell [ -s $(@:$(modpost-ext).o=.modversions) ] && \ > + echo -T $(@:$(modpost-ext).o=.modversions)) \ > + $(shell [ -f $(@:$(modpost-ext).o=.o).objects ] && \ > + cat $(@:$(modpost-ext).o=.o).objects || \ > + echo $(@:$(modpost-ext).o=.o)) > + > +$(modules:.ko=$(modpost-ext).o): %$(modpost-ext).o: %.o %.modversions FORCE > + $(call if_changed,cc_lto_link_modules) > +endif > + > # We can go over command line length here, so be careful. > quiet_cmd_modpost = MODPOST $(words $(filter-out vmlinux FORCE, $^)) modules > - cmd_modpost = $(MODLISTCMD) | sed 's/\.ko$$/.o/' | $(modpost) $(MODPOST_OPT) -s -T - > + cmd_modpost = $(MODLISTCMD) | sed 's/\.ko$$/$(modpost-ext)\.o/' | $(modpost) $(MODPOST_OPT) -s -T - > > PHONY += __modpost > -__modpost: $(modules:.ko=.o) FORCE > +__modpost: $(modules:.ko=$(modpost-ext).o) FORCE > $(call cmd,modpost) $(wildcard vmlinux) > > quiet_cmd_kernel-mod = MODPOST $@ > @@ -98,8 +132,7 @@ vmlinux.o: FORCE > $(call cmd,kernel-mod) > > # Declare generated files as targets for modpost > -$(modules:.ko=.mod.c): __modpost ; > - > +$(modules:.ko=$(modpost-ext).mod.c): __modpost ; > > # Step 5), compile all *.mod.c files > > @@ -110,23 +143,37 @@ quiet_cmd_cc_o_c = CC $@ > cmd_cc_o_c = $(CC) $(c_flags) $(KBUILD_CFLAGS_MODULE) $(CFLAGS_MODULE) \ > -c -o $@ $< > > -$(modules:.ko=.mod.o): %.mod.o: %.mod.c FORCE > +$(modules:.ko=.mod.o): %.mod.o: %$(modpost-ext).mod.c FORCE > $(call if_changed_dep,cc_o_c) > > -targets += $(modules:.ko=.mod.o) > +targets += $(modules:.ko=$(modpost-ext).mod.o) > > ARCH_POSTLINK := $(wildcard $(srctree)/arch/$(SRCARCH)/Makefile.postlink) > > # Step 6), final link of the modules with optional arch pass after final link > quiet_cmd_ld_ko_o = LD [M] $@ > + > +ifdef CONFIG_LTO_CLANG > +lto_ko_objects = $(foreach o,$(1:$(modpost-ext).o=.o), \ > + $(shell [ -f $(o).objects ] && \ > + cat $(o).objects || echo $(o))) > + > + cmd_ld_ko_o = \ > + $(LD) -r $(LDFLAGS) \ > + $(KBUILD_LDFLAGS_MODULE) $(LDFLAGS_MODULE) \ > + $(shell [ -s $(@:.ko=.modversions) ] && \ > + echo -T $(@:.ko=.modversions)) \ > + -o $@ $(call lto_ko_objects, $(filter-out FORCE,$^)) > +else > cmd_ld_ko_o = \ > $(LD) -r $(LDFLAGS) \ > $(KBUILD_LDFLAGS_MODULE) $(LDFLAGS_MODULE) \ > -o $@ $(filter-out FORCE,$^) ; \ > $(if $(ARCH_POSTLINK), $(MAKE) -f $(ARCH_POSTLINK) $@, true) > +endif > > -$(modules): %.ko :%.o %.mod.o FORCE > - +$(call if_changed,ld_ko_o) > +$(modules): %.ko: %$(modpost-ext).o %.mod.o FORCE > + $(call if_changed,ld_ko_o) > > targets += $(modules) > > diff --git a/scripts/link-vmlinux.sh b/scripts/link-vmlinux.sh > index e6818b8e7141..fccc123b691f 100755 > --- a/scripts/link-vmlinux.sh > +++ b/scripts/link-vmlinux.sh > @@ -64,6 +64,53 @@ archive_builtin() > fi > } > > +# If CONFIG_LTO_CLANG is selected, the compiler produces LLVM IR files instead > +# of ELF object files. This function expands individual IR files from a list of > +# objects that would have otherwise been linked already. > +expand() > +{ > + if [ -z "${CONFIG_LTO_CLANG}" ]; then > + echo $* > + fi > + > + local objs > + > + for o in $*; do > + if [ -f ${o}.objects ]; then > + objs="${objs} $(xargs < ${o}.objects)" > + else > + objs="${objs} ${o}" > + fi > + done > + > + echo "${objs}" > +} > + > +# If CONFIG_LTO_CLANG is selected, collect generated symbol versions into > +# .tmp_symversions > +modversions() > +{ > + if [ -z "${CONFIG_LTO_CLANG}" ]; then > + return > + fi > + > + if [ -z "${CONFIG_MODVERSIONS}" ]; then > + return > + fi > + > + rm -f .tmp_symversions > + > + for o in $(expand ${KBUILD_VMLINUX_INIT}) \ > + $(expand ${KBUILD_VMLINUX_MAIN}) \ > + $(expand ${KBUILD_VMLINUX_LIBS}); do > + if [ -f ${o}.symversions ]; then > + cat ${o}.symversions >> .tmp_symversions > + fi > + done > + > + echo "-T .tmp_symversions" > +} > + > # Link of vmlinux.o used for section mismatch analysis > # ${1} output file > modpost_link() > @@ -78,13 +125,22 @@ modpost_link() > ${KBUILD_VMLINUX_LIBS} \ > --end-group" > else > - objects="${KBUILD_VMLINUX_INIT} \ > + objects="$(expand ${KBUILD_VMLINUX_INIT}) \ > --start-group \ > - ${KBUILD_VMLINUX_MAIN} \ > - ${KBUILD_VMLINUX_LIBS} \ > + $(expand ${KBUILD_VMLINUX_MAIN}) \ > + $(expand ${KBUILD_VMLINUX_LIBS}) \ > --end-group" > fi > - ${LD} ${LDFLAGS} -r -o ${1} ${objects} > + > + if [ -n "${CONFIG_LTO_CLANG}" ]; then > + # This might take a while, so indicate that we're doing > + # an LTO link > + info LTO vmlinux.o > + else > + info LD vmlinux.o > + fi > + > + ${LD} ${LDFLAGS} -r -o ${1} $(modversions) ${objects} > } > > # Link of vmlinux > @@ -96,6 +152,14 @@ vmlinux_link() > local objects > > if [ "${SRCARCH}" != "um" ]; then > + local ld=${LD} > + local ldflags="${LDFLAGS} ${LDFLAGS_vmlinux}" > + > + if [ -n "${LDFINAL_vmlinux}" ]; then > + ld=${LDFINAL_vmlinux} > + ldflags="${LDFLAGS_FINAL_vmlinux} ${LDFLAGS_vmlinux}" > + fi > + > if [ -n "${CONFIG_THIN_ARCHIVES}" ]; then > objects="--whole-archive \ > built-in.o \ > @@ -105,16 +169,15 @@ vmlinux_link() > --end-group \ > ${1}" > else > - objects="${KBUILD_VMLINUX_INIT} \ > + objects="$(expand ${KBUILD_VMLINUX_INIT}) \ > --start-group \ > - ${KBUILD_VMLINUX_MAIN} \ > - ${KBUILD_VMLINUX_LIBS} \ > + $(expand ${KBUILD_VMLINUX_MAIN}) \ > + $(expand ${KBUILD_VMLINUX_LIBS}) \ > --end-group \ > ${1}" > fi > > - ${LD} ${LDFLAGS} ${LDFLAGS_vmlinux} -o ${2} \ > - -T ${lds} ${objects} > + ${ld} ${ldflags} -o ${2} -T ${lds} ${objects} > else > if [ -n "${CONFIG_THIN_ARCHIVES}" ]; then > objects="-Wl,--whole-archive \ > @@ -141,7 +204,6 @@ vmlinux_link() > fi > } > > - > # Create ${2} .o file with all symbols from the ${1} object file > kallsyms() > { > @@ -192,6 +254,7 @@ cleanup() > rm -f .tmp_System.map > rm -f .tmp_kallsyms* > rm -f .tmp_version > + rm -f .tmp_symversions > rm -f .tmp_vmlinux* > rm -f built-in.o > rm -f System.map > @@ -253,12 +316,19 @@ ${MAKE} -f "${srctree}/scripts/Makefile.build" obj=init GCC_PLUGINS_CFLAGS="${GC > archive_builtin > > #link vmlinux.o > -info LD vmlinux.o > modpost_link vmlinux.o > > # modpost vmlinux.o to check for section mismatches > ${MAKE} -f "${srctree}/scripts/Makefile.modpost" vmlinux.o > > +if [ -n "${CONFIG_LTO_CLANG}" ]; then > + # Re-use vmlinux.o, so we can avoid the slow LTO link step in > + # vmlinux_link > + KBUILD_VMLINUX_INIT= > + KBUILD_VMLINUX_MAIN=vmlinux.o > + KBUILD_VMLINUX_LIBS= > +fi > + > kallsymso="" > kallsyms_vmlinux="" > if [ -n "${CONFIG_KALLSYMS}" ]; then > -- > 2.15.0.448.gf294e3d99a-goog >
On Wed, 15 Nov 2017 13:34:22 -0800 Sami Tolvanen <samitolvanen@google.com> wrote: > This change adds the configuration option CONFIG_LTO_CLANG, and > build system support for clang's Link Time Optimization (LTO). In > preparation for LTO support for other compilers, potentially common > parts of the changes are gated behind CONFIG_LTO instead. > > With -flto, instead of object files, clang produces LLVM bitcode, > which is compiled into a native object at link time, allowing the > final binary to be optimized globally. For more details, see: > > https://llvm.org/docs/LinkTimeOptimization.html > > While the kernel normally uses GNU ld for linking, LLVM supports LTO > only with lld or GNU gold linkers. This patch set assumes gold will > be used with the LLVMgold plug-in to perform the LTO link step. Due > to potential incompatibilities with GNU ld, this change also adds > LDFINAL_vmlinux for using a different linker for the vmlinux_link > step, and defaults to using GNU ld. > > Assuming LLVMgold.so is in LD_LIBRARY_PATH and CONFIG_LTO_CLANG has > been selected, an LTO kernel can be built simply by running make > CC=clang. Recommended versions are >= 5.0 for clang, and >= 2.27 for > binutils. Do you have any kind of numbers for this, out of curiosity? Binary size, performance, build time? Also > @@ -585,6 +585,7 @@ config CC_STACKPROTECTOR_STRONG > endchoice > > config THIN_ARCHIVES > + depends on !LTO_CLANG > def_bool y > help > Select this if the architecture wants to use thin archives Why is this needed? It would have been nice to get rid of the !THIN_ARCHIVES option if you can make the patches work with the thin archives paths. Thanks, Nick -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Sat, Nov 18, 2017 at 01:21:39PM +1000, Nicholas Piggin wrote: > Do you have any kind of numbers for this, out of curiosity? Binary > size, performance, build time? I don't have performance numbers to share. Are there any specific benchmarks you'd be interested in seeing? Build time typically increases with LTO and in my experience, binary size tends to increase by ~10-15% as well. > Why is this needed? It would have been nice to get rid of the > !THIN_ARCHIVES option if you can make the patches work with the thin > archives paths. I believe LLVMgold doesn't know how to deal with an archive of LLVM IR files, but I can certainly use thin archives as an index and extract the path names for linking. I'll look into it. Sami -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Mon, 20 Nov 2017 12:21:52 -0800 Sami Tolvanen <samitolvanen@google.com> wrote: > On Sat, Nov 18, 2017 at 01:21:39PM +1000, Nicholas Piggin wrote: > > Do you have any kind of numbers for this, out of curiosity? Binary > > size, performance, build time? > > I don't have performance numbers to share. Are there any specific > benchmarks you'd be interested in seeing? Build time typically > increases with LTO and in my experience, binary size tends to increase > by ~10-15% as well. By deduction, then you must see some performance improvement? :) I just wonder are you doing this because there is some worthwhile performance gain? Or to enable more testing and development of LTO? Any clues for why a user would want to enable it. > > > Why is this needed? It would have been nice to get rid of the > > !THIN_ARCHIVES option if you can make the patches work with the thin > > archives paths. > > I believe LLVMgold doesn't know how to deal with an archive of LLVM IR > files, but I can certainly use thin archives as an index and extract > the path names for linking. I'll look into it. Thanks, if you could. Possibly file a request with LLVMgold too, it seems to be that toolchain support for archives is quite strong, so it will be good to keep pushing for that. Thanks, Nick -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Nov 21, 2017 at 11:01:52AM +1000, Nicholas Piggin wrote: > I just wonder are you doing this because there is some worthwhile > performance gain? Or to enable more testing and development of LTO? > Any clues for why a user would want to enable it. I'm primarily interested in CFI, which with clang requires LTO; not for the optimizations, but for source visibility. We do expect to see performance improvements with LTO though, especially if combined with PGO. > Thanks, if you could. Possibly file a request with LLVMgold too, it > seems to be that toolchain support for archives is quite strong, so it > will be good to keep pushing for that. It turns out LLVMgold is fine with mixed IR/object archives, but we need to use llvm-ar to generate symbol tables for them, and there are some compatibility issues with objdump that I had to work around. I'll send v3 for review once I receive some feedback for the clang/gold patches we need first. Sami -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/.gitignore b/.gitignore index 6c119eab5d46..ac236e2bb9b1 100644 --- a/.gitignore +++ b/.gitignore @@ -11,6 +11,7 @@ # .* *.a +*.a.* *.bin *.bz2 *.c.[012]*.* @@ -28,6 +29,7 @@ *.lzma *.lzo *.mod.c +*.modversions *.o *.o.* *.order diff --git a/Makefile b/Makefile index f976af9525bf..8141b4c8f1bf 100644 --- a/Makefile +++ b/Makefile @@ -350,6 +350,7 @@ include scripts/Kbuild.include # Make variables (CC, etc...) AS = $(CROSS_COMPILE)as LD = $(CROSS_COMPILE)ld +LDGOLD = $(CROSS_COMPILE)ld.gold CC = $(CROSS_COMPILE)gcc CPP = $(CC) -E AR = $(CROSS_COMPILE)ar @@ -623,6 +624,15 @@ endif # Defaults to vmlinux, but the arch makefile usually adds further targets all: vmlinux +# Make toolchain changes before including arch/$(SRCARCH)/Makefile to ensure +# cc-/ld-* macros return correct values. +ifdef CONFIG_LTO_CLANG +# use GNU gold with LLVMgold for LTO linking, and LD for vmlinux_link +LDFINAL_vmlinux := $(LD) +LD := $(LDGOLD) +LDFLAGS_GOLD += -plugin LLVMgold.so +endif + # The arch Makefile can set ARCH_{CPP,A,C}FLAGS to override the default # values of the respective KBUILD_* variables ARCH_CPPFLAGS := @@ -789,6 +799,32 @@ KBUILD_CFLAGS += $(call cc-option,-ffunction-sections,) KBUILD_CFLAGS += $(call cc-option,-fdata-sections,) endif +ifdef CONFIG_LTO_CLANG +lto-clang-flags := -flto -fvisibility=hidden + +# allow disabling only clang LTO where needed +DISABLE_LTO_CLANG := -fno-lto +export DISABLE_LTO_CLANG + +ifdef CONFIG_MODVERSIONS +# llvm-dis is used instead of objdump to process LLVM IR files +LLVM_DIS := llvm-dis +export LLVM_DIS +endif +endif + +ifdef CONFIG_LTO +lto-flags := $(lto-clang-flags) +KBUILD_CFLAGS += $(lto-flags) + +DISABLE_LTO := $(DISABLE_LTO_CLANG) +export DISABLE_LTO + +# LDFINAL_vmlinux and LDFLAGS_FINAL_vmlinux can be set to override +# the linker and flags for vmlinux_link. +export LDFINAL_vmlinux LDFLAGS_FINAL_vmlinux +endif + # arch Makefile may override CC so keep this after arch Makefile is included NOSTDINC_FLAGS += -nostdinc -isystem $(shell $(CC) -print-file-name=include) CHECKFLAGS += $(NOSTDINC_FLAGS) @@ -1090,6 +1126,19 @@ prepare-objtool: $(objtool_target) # CC_STACKPROTECTOR_STRONG! Why did it build with _REGULAR?!") PHONY += prepare-compiler-check prepare-compiler-check: FORCE +# Make sure we're using clang with LTO_CLANG +ifdef CONFIG_LTO_CLANG + ifneq ($(cc-name),clang) + @echo Cannot use CONFIG_LTO_CLANG without CC=clang >&2 && exit 1 + endif +endif +# Make sure compiler supports LTO flags +ifdef lto-flags + ifeq ($(call cc-option, $(lto-flags)),) + @echo Cannot use CONFIG_LTO: $(lto-flags) not supported by compiler \ + >&2 && exit 1 + endif +endif # Make sure compiler supports requested stack protector flag. ifdef stackp-name ifeq ($(call cc-option, $(stackp-flag)),) @@ -1564,7 +1613,10 @@ clean: $(clean-dirs) -o -name modules.builtin -o -name '.tmp_*.o.*' \ -o -name '*.c.[012]*.*' \ -o -name '*.ll' \ - -o -name '*.gcno' \) -type f -print | xargs rm -f + -o -name '*.gcno' \ + -o -name '*.[oa].objects' \ + -o -name '*.o.symversions' \ + -o -name '*.modversions' \) -type f -print | xargs rm -f # Generate tags for editors # --------------------------------------------------------------------------- diff --git a/arch/Kconfig b/arch/Kconfig index 400b9e1b2f27..bb5296ecebdd 100644 --- a/arch/Kconfig +++ b/arch/Kconfig @@ -585,6 +585,7 @@ config CC_STACKPROTECTOR_STRONG endchoice config THIN_ARCHIVES + depends on !LTO_CLANG def_bool y help Select this if the architecture wants to use thin archives @@ -605,6 +606,37 @@ config LD_DEAD_CODE_DATA_ELIMINATION sections (e.g., '.text.init'). Typically '.' in section names is used to distinguish them from label names / C identifiers. +config LTO + bool + +config ARCH_SUPPORTS_LTO_CLANG + bool + help + An architecture should select this option it supports: + - compiling with clang, + - compiling inline assembly with clang's integrated assembler, + - and linking with either lld or GNU gold w/ LLVMgold. + +config LTO_CLANG + bool "Use clang Link Time Optimization (LTO)" + depends on ARCH_SUPPORTS_LTO_CLANG + depends on !FTRACE_MCOUNT_RECORD + select LTO + select LD_DEAD_CODE_DATA_ELIMINATION + help + This option enables clang's Link Time Optimization (LTO), which allows + the compiler to optimize the kernel globally at link time. If you + enable this option, the compiler generates LLVM IR instead of object + files, and the actual compilation from IR occurs at the LTO link step, + which may take several minutes. + + If you select this option, you must compile the kernel with clang + (make CC=clang) and have the LVMgold plug-in in LD_LIBRARY_PATH. + + Using clang >= 5.0 and GNU gold from binutils >= 2.27 is recommended + for this option. + + config HAVE_ARCH_WITHIN_STACK_FRAMES bool help diff --git a/scripts/Makefile.build b/scripts/Makefile.build index e63af4e19382..e8bf5c440612 100644 --- a/scripts/Makefile.build +++ b/scripts/Makefile.build @@ -210,6 +210,23 @@ else cmd_cc_o_c = $(CC) $(c_flags) -c -o $(@D)/.tmp_$(@F) $< +ifdef CONFIG_LTO_CLANG +# Generate .o.symversions files for each .o with exported symbols, and link these +# to the kernel and/or modules at the end. +cmd_modversions_c = \ + if echo '$(c_flags)' | grep -q -- '$(DISABLE_LTO_CLANG)'; then \ + if $(OBJDUMP) -h $(@D)/.tmp_$(@F) | grep -q __ksymtab; then \ + $(call cmd_gensymtypes_c,$(KBUILD_SYMTYPES),$(@:.o=.symtypes)) \ + > $(@D)/$(@F).symversions; \ + fi; \ + else \ + if $(LLVM_DIS) -o=- $(@D)/.tmp_$(@F) | grep -q __ksymtab; then \ + $(call cmd_gensymtypes_c,$(KBUILD_SYMTYPES),$(@:.o=.symtypes)) \ + > $(@D)/$(@F).symversions; \ + fi; \ + fi; \ + mv -f $(@D)/.tmp_$(@F) $@; +else cmd_modversions_c = \ if $(OBJDUMP) -h $(@D)/.tmp_$(@F) | grep -q __ksymtab; then \ $(call cmd_gensymtypes_c,$(KBUILD_SYMTYPES),$(@:.o=.symtypes)) \ @@ -222,6 +239,7 @@ cmd_modversions_c = \ mv -f $(@D)/.tmp_$(@F) $@; \ fi; endif +endif ifdef CONFIG_FTRACE_MCOUNT_RECORD ifdef BUILD_C_RECORDMCOUNT @@ -443,6 +461,13 @@ $(obj)/%-asn1.c $(obj)/%-asn1.h: $(src)/%.asn1 $(objtree)/scripts/asn1_compiler # To build objects in subdirs, we need to descend into the directories $(sort $(subdir-obj-y)): $(subdir-ym) ; +ifdef CONFIG_LTO_CLANG +# If LTO is enabled, we remove all intermediate linking steps and instead +# collect a list of all objects to be linked at the end. +cc_lto_objects = $(foreach o,$(1),\ + [ -f ${o}.objects ] && cat ${o}.objects || echo ${o};) +endif + # # Rule to compile a set of .o files into one .o file # @@ -458,6 +483,16 @@ else quiet_cmd_link_o_target = LD $@ endif +ifdef CONFIG_LTO_CLANG +builtin-cmds = $(call cc_lto_objects,$(filter $(obj-y), $^)) + +quiet_cmd_update_builtin = GEN $@ +cmd_update_builtin = (cat /dev/null; $(builtin-cmds)) > $@.objects && \ + cat /dev/null > $@ + +$(builtin-target): $(obj-y) FORCE + $(call if_changed,update_builtin) +else # If the list of objects to link is empty, just create an empty built-in.o cmd_link_o_target = $(if $(strip $(obj-y)),\ $(cmd_make_builtin) $@ $(filter $(obj-y), $^) \ @@ -466,6 +501,7 @@ cmd_link_o_target = $(if $(strip $(obj-y)),\ $(builtin-target): $(obj-y) FORCE $(call if_changed,link_o_target) +endif targets += $(builtin-target) endif # builtin-target @@ -487,6 +523,16 @@ $(modorder-target): $(subdir-ym) FORCE # Rule to compile a set of .o files into one .a file # ifdef lib-target +ifdef CONFIG_LTO_CLANG +lib-target-cmds = $(call cc_lto_objects,$(lib-y)) + +quiet_cmd_update_lib_target = GEN $@ +cmd_update_lib_target = (cat /dev/null; $(lib-target-cmds)) > $@.objects && \ + cat /dev/null > $@ + +$(lib-target): $(lib-y) FORCE + $(call if_changed,update_lib_target) +else quiet_cmd_link_l_target = AR $@ ifdef CONFIG_THIN_ARCHIVES @@ -497,6 +543,7 @@ endif $(lib-target): $(lib-y) FORCE $(call if_changed,link_l_target) +endif targets += $(lib-target) @@ -552,14 +599,31 @@ endif quiet_cmd_link_multi-m = LD [M] $@ cmd_link_multi-m = $(cmd_link_multi-link) +ifdef CONFIG_LTO_CLANG +multi-deps-cmds = $(call cc_lto_objects,$(link_multi_deps)) + +quiet_cmd_update_multi_deps = GEN $@ +cmd_update_multi_deps = (cat /dev/null; $(multi-deps-cmds)) > $@.objects && \ + cat /dev/null > $@ + +$(multi-used-y): FORCE + $(call if_changed,update_multi_deps) + +$(multi-used-m): FORCE + $(call if_changed,update_multi_deps) + @{ echo $(@:.o=.ko); echo $(link_multi_deps); \ + $(cmd_undef_syms); } > $(MODVERDIR)/$(@F:.o=.mod) +else $(multi-used-y): FORCE $(call if_changed,link_multi-y) -$(call multi_depend, $(multi-used-y), .o, -objs -y) $(multi-used-m): FORCE $(call if_changed,link_multi-m) @{ echo $(@:.o=.ko); echo $(link_multi_deps); \ $(cmd_undef_syms); } > $(MODVERDIR)/$(@F:.o=.mod) +endif + +$(call multi_depend, $(multi-used-y), .o, -objs -y) $(call multi_depend, $(multi-used-m), .o, -objs -y -m) targets += $(multi-used-y) $(multi-used-m) diff --git a/scripts/Makefile.modpost b/scripts/Makefile.modpost index 991db7d6e4df..cb1c040a006c 100644 --- a/scripts/Makefile.modpost +++ b/scripts/Makefile.modpost @@ -83,12 +83,46 @@ modpost = scripts/mod/modpost \ MODPOST_OPT=$(subst -i,-n,$(filter -i,$(MAKEFLAGS))) +# If CONFIG_LTO_CLANG is enabled, .o files are either LLVM IR, or empty, so we +# need to link them into actual objects before passing them to modpost +modpost-ext = $(if $(CONFIG_LTO_CLANG),.lto,) + +ifdef CONFIG_LTO_CLANG +quiet_cmd_cc_lto_modversions = GEN [M] $@ +cmd_cc_lto_modversions = \ + rm -f $(@); \ + if [ -f $(@:.modversions=.o).objects ]; then \ + for i in `cat $(@:.modversions=.o).objects`; do \ + [ -s $$i.symversions ] && \ + cat $$i.symversions >> $(@); \ + done; \ + else \ + [ -s $(@:.modversions=.o).symversions ] && \ + cat $(@:.modversions=.o).symversions >> $(@); \ + fi + +$(modules:.ko=.modversions): FORCE + $(call if_changed,cc_lto_modversions) + +quiet_cmd_cc_lto_link_modules = LD [M] $@ +cmd_cc_lto_link_modules = \ + $(LD) $(ld_flags) -r -o $(@) \ + $(shell [ -s $(@:$(modpost-ext).o=.modversions) ] && \ + echo -T $(@:$(modpost-ext).o=.modversions)) \ + $(shell [ -f $(@:$(modpost-ext).o=.o).objects ] && \ + cat $(@:$(modpost-ext).o=.o).objects || \ + echo $(@:$(modpost-ext).o=.o)) + +$(modules:.ko=$(modpost-ext).o): %$(modpost-ext).o: %.o %.modversions FORCE + $(call if_changed,cc_lto_link_modules) +endif + # We can go over command line length here, so be careful. quiet_cmd_modpost = MODPOST $(words $(filter-out vmlinux FORCE, $^)) modules - cmd_modpost = $(MODLISTCMD) | sed 's/\.ko$$/.o/' | $(modpost) $(MODPOST_OPT) -s -T - + cmd_modpost = $(MODLISTCMD) | sed 's/\.ko$$/$(modpost-ext)\.o/' | $(modpost) $(MODPOST_OPT) -s -T - PHONY += __modpost -__modpost: $(modules:.ko=.o) FORCE +__modpost: $(modules:.ko=$(modpost-ext).o) FORCE $(call cmd,modpost) $(wildcard vmlinux) quiet_cmd_kernel-mod = MODPOST $@ @@ -98,8 +132,7 @@ vmlinux.o: FORCE $(call cmd,kernel-mod) # Declare generated files as targets for modpost -$(modules:.ko=.mod.c): __modpost ; - +$(modules:.ko=$(modpost-ext).mod.c): __modpost ; # Step 5), compile all *.mod.c files @@ -110,23 +143,37 @@ quiet_cmd_cc_o_c = CC $@ cmd_cc_o_c = $(CC) $(c_flags) $(KBUILD_CFLAGS_MODULE) $(CFLAGS_MODULE) \ -c -o $@ $< -$(modules:.ko=.mod.o): %.mod.o: %.mod.c FORCE +$(modules:.ko=.mod.o): %.mod.o: %$(modpost-ext).mod.c FORCE $(call if_changed_dep,cc_o_c) -targets += $(modules:.ko=.mod.o) +targets += $(modules:.ko=$(modpost-ext).mod.o) ARCH_POSTLINK := $(wildcard $(srctree)/arch/$(SRCARCH)/Makefile.postlink) # Step 6), final link of the modules with optional arch pass after final link quiet_cmd_ld_ko_o = LD [M] $@ + +ifdef CONFIG_LTO_CLANG +lto_ko_objects = $(foreach o,$(1:$(modpost-ext).o=.o), \ + $(shell [ -f $(o).objects ] && \ + cat $(o).objects || echo $(o))) + + cmd_ld_ko_o = \ + $(LD) -r $(LDFLAGS) \ + $(KBUILD_LDFLAGS_MODULE) $(LDFLAGS_MODULE) \ + $(shell [ -s $(@:.ko=.modversions) ] && \ + echo -T $(@:.ko=.modversions)) \ + -o $@ $(call lto_ko_objects, $(filter-out FORCE,$^)) +else cmd_ld_ko_o = \ $(LD) -r $(LDFLAGS) \ $(KBUILD_LDFLAGS_MODULE) $(LDFLAGS_MODULE) \ -o $@ $(filter-out FORCE,$^) ; \ $(if $(ARCH_POSTLINK), $(MAKE) -f $(ARCH_POSTLINK) $@, true) +endif -$(modules): %.ko :%.o %.mod.o FORCE - +$(call if_changed,ld_ko_o) +$(modules): %.ko: %$(modpost-ext).o %.mod.o FORCE + $(call if_changed,ld_ko_o) targets += $(modules) diff --git a/scripts/link-vmlinux.sh b/scripts/link-vmlinux.sh index e6818b8e7141..fccc123b691f 100755 --- a/scripts/link-vmlinux.sh +++ b/scripts/link-vmlinux.sh @@ -64,6 +64,53 @@ archive_builtin() fi } +# If CONFIG_LTO_CLANG is selected, the compiler produces LLVM IR files instead +# of ELF object files. This function expands individual IR files from a list of +# objects that would have otherwise been linked already. +expand() +{ + if [ -z "${CONFIG_LTO_CLANG}" ]; then + echo $* + fi + + local objs + + for o in $*; do + if [ -f ${o}.objects ]; then + objs="${objs} $(xargs < ${o}.objects)" + else + objs="${objs} ${o}" + fi + done + + echo "${objs}" +} + +# If CONFIG_LTO_CLANG is selected, collect generated symbol versions into +# .tmp_symversions +modversions() +{ + if [ -z "${CONFIG_LTO_CLANG}" ]; then + return + fi + + if [ -z "${CONFIG_MODVERSIONS}" ]; then + return + fi + + rm -f .tmp_symversions + + for o in $(expand ${KBUILD_VMLINUX_INIT}) \ + $(expand ${KBUILD_VMLINUX_MAIN}) \ + $(expand ${KBUILD_VMLINUX_LIBS}); do + if [ -f ${o}.symversions ]; then + cat ${o}.symversions >> .tmp_symversions + fi + done + + echo "-T .tmp_symversions" +} + # Link of vmlinux.o used for section mismatch analysis # ${1} output file modpost_link() @@ -78,13 +125,22 @@ modpost_link() ${KBUILD_VMLINUX_LIBS} \ --end-group" else - objects="${KBUILD_VMLINUX_INIT} \ + objects="$(expand ${KBUILD_VMLINUX_INIT}) \ --start-group \ - ${KBUILD_VMLINUX_MAIN} \ - ${KBUILD_VMLINUX_LIBS} \ + $(expand ${KBUILD_VMLINUX_MAIN}) \ + $(expand ${KBUILD_VMLINUX_LIBS}) \ --end-group" fi - ${LD} ${LDFLAGS} -r -o ${1} ${objects} + + if [ -n "${CONFIG_LTO_CLANG}" ]; then + # This might take a while, so indicate that we're doing + # an LTO link + info LTO vmlinux.o + else + info LD vmlinux.o + fi + + ${LD} ${LDFLAGS} -r -o ${1} $(modversions) ${objects} } # Link of vmlinux @@ -96,6 +152,14 @@ vmlinux_link() local objects if [ "${SRCARCH}" != "um" ]; then + local ld=${LD} + local ldflags="${LDFLAGS} ${LDFLAGS_vmlinux}" + + if [ -n "${LDFINAL_vmlinux}" ]; then + ld=${LDFINAL_vmlinux} + ldflags="${LDFLAGS_FINAL_vmlinux} ${LDFLAGS_vmlinux}" + fi + if [ -n "${CONFIG_THIN_ARCHIVES}" ]; then objects="--whole-archive \ built-in.o \ @@ -105,16 +169,15 @@ vmlinux_link() --end-group \ ${1}" else - objects="${KBUILD_VMLINUX_INIT} \ + objects="$(expand ${KBUILD_VMLINUX_INIT}) \ --start-group \ - ${KBUILD_VMLINUX_MAIN} \ - ${KBUILD_VMLINUX_LIBS} \ + $(expand ${KBUILD_VMLINUX_MAIN}) \ + $(expand ${KBUILD_VMLINUX_LIBS}) \ --end-group \ ${1}" fi - ${LD} ${LDFLAGS} ${LDFLAGS_vmlinux} -o ${2} \ - -T ${lds} ${objects} + ${ld} ${ldflags} -o ${2} -T ${lds} ${objects} else if [ -n "${CONFIG_THIN_ARCHIVES}" ]; then objects="-Wl,--whole-archive \ @@ -141,7 +204,6 @@ vmlinux_link() fi } - # Create ${2} .o file with all symbols from the ${1} object file kallsyms() { @@ -192,6 +254,7 @@ cleanup() rm -f .tmp_System.map rm -f .tmp_kallsyms* rm -f .tmp_version + rm -f .tmp_symversions rm -f .tmp_vmlinux* rm -f built-in.o rm -f System.map @@ -253,12 +316,19 @@ ${MAKE} -f "${srctree}/scripts/Makefile.build" obj=init GCC_PLUGINS_CFLAGS="${GC archive_builtin #link vmlinux.o -info LD vmlinux.o modpost_link vmlinux.o # modpost vmlinux.o to check for section mismatches ${MAKE} -f "${srctree}/scripts/Makefile.modpost" vmlinux.o +if [ -n "${CONFIG_LTO_CLANG}" ]; then + # Re-use vmlinux.o, so we can avoid the slow LTO link step in + # vmlinux_link + KBUILD_VMLINUX_INIT= + KBUILD_VMLINUX_MAIN=vmlinux.o + KBUILD_VMLINUX_LIBS= +fi + kallsymso="" kallsyms_vmlinux="" if [ -n "${CONFIG_KALLSYMS}" ]; then
This change adds the configuration option CONFIG_LTO_CLANG, and build system support for clang's Link Time Optimization (LTO). In preparation for LTO support for other compilers, potentially common parts of the changes are gated behind CONFIG_LTO instead. With -flto, instead of object files, clang produces LLVM bitcode, which is compiled into a native object at link time, allowing the final binary to be optimized globally. For more details, see: https://llvm.org/docs/LinkTimeOptimization.html While the kernel normally uses GNU ld for linking, LLVM supports LTO only with lld or GNU gold linkers. This patch set assumes gold will be used with the LLVMgold plug-in to perform the LTO link step. Due to potential incompatibilities with GNU ld, this change also adds LDFINAL_vmlinux for using a different linker for the vmlinux_link step, and defaults to using GNU ld. Assuming LLVMgold.so is in LD_LIBRARY_PATH and CONFIG_LTO_CLANG has been selected, an LTO kernel can be built simply by running make CC=clang. Recommended versions are >= 5.0 for clang, and >= 2.27 for binutils. Signed-off-by: Sami Tolvanen <samitolvanen@google.com> --- .gitignore | 2 ++ Makefile | 54 +++++++++++++++++++++++++++- arch/Kconfig | 32 +++++++++++++++++ scripts/Makefile.build | 66 +++++++++++++++++++++++++++++++++- scripts/Makefile.modpost | 63 ++++++++++++++++++++++++++++----- scripts/link-vmlinux.sh | 92 ++++++++++++++++++++++++++++++++++++++++++------ 6 files changed, 288 insertions(+), 21 deletions(-)