On Thu, Jul 18, 2019 at 1:46 AM Segher Boessenkool <seg...@kernel.crashing.org> wrote: > > On Thu, Jul 18, 2019 at 12:19:36AM +0900, Masahiro Yamada wrote: > > On Wed, Jul 17, 2019 at 11:38 PM Segher Boessenkool > > <seg...@kernel.crashing.org> wrote: > > > > > > On Tue, Jul 16, 2019 at 10:15:47PM +1000, Michael Ellerman wrote: > > > > Segher Boessenkool <seg...@kernel.crashing.org> writes: > > > > And it's definitely calling ar with no flags, eg: > > > > > > > > rm -f init/built-in.a; powerpc-linux-ar rcSTPD init/built-in.a > > > > init/main.o init/version.o init/do_mounts.o init/do_mounts_rd.o > > > > init/do_mounts_initrd.o init/do_mounts_md.o init/initramfs.o > > > > init/init_task.o > > > > > > This uses thin archives. Those will work fine. > > > > > > The failing case was empty files IIRC, stuff created from no inputs. > > > > Actually, empty files are created everywhere. > > > cmd_ar_builtin = rm -f $@; $(AR) rcSTP$(KBUILD_ARFLAGS) $@ > > $(real-prereqs) > > You use thin archives. > > Does every config use thin archives always nowadays?
Kbuild always uses thin archives as far as vmlinux is concerned. But, there are some other call-sites. masahiro@pug:~/ref/linux$ git grep '$(AR)' -- :^Documentation :^tools arch/powerpc/boot/Makefile: BOOTAR := $(AR) arch/unicore32/lib/Makefile: $(Q)$(AR) p $(GNU_LIBC_A) $(notdir $@) > $@ arch/unicore32/lib/Makefile: $(Q)$(AR) p $(GNU_LIBGCC_A) $(notdir $@) > $@ lib/raid6/test/Makefile: $(AR) cq $@ $^ scripts/Kbuild.include:ar-option = $(call try-run, $(AR) rc$(1) "$$TMP",$(1),$(2)) scripts/Makefile.build: cmd_ar_builtin = rm -f $@; $(AR) rcSTP$(KBUILD_ARFLAGS) $@ $(real-prereqs) scripts/Makefile.lib: cmd_ar = rm -f $@; $(AR) rcsTP$(KBUILD_ARFLAGS) $@ $(real-prereqs) Probably, you are interested in arch/powerpc/boot/Makefile. This does not seem a thin archive. > > BTW, your commit 8995ac8702737147115e1c75879a1a2d75627b9e > > dates back to 2008. > > > > At that time, thin archive was not used. > > Yes, I know. This isn't about built-in.[oa], it is about *other* > archives we at least *used to* create. If we *know* we do not anymore, > then this workaround can of course be removed (and good riddance). If it is not about built-in.[oa], which archive are you talking about? Can you pin-point the one? masahiro@pug:~/ref/linux$ git log --oneline -1 8995ac870273 (HEAD) [POWERPC] Specify GNUTARGET on $(AR) invocations masahiro@pug:~/ref/linux$ git grep '(AR)' Documentation/kbuild/makefiles.txt: per-directory options to $(LD) and $(AR). arch/powerpc/Makefile:CROSS32AR := GNUTARGET=elf32-powerpc $(AR) arch/powerpc/Makefile:override AR := GNUTARGET=elf$(SZ)-powerpc $(AR) drivers/md/raid6test/Makefile: $(AR) cq $@ $^ scripts/Makefile.build: rm -f $@; $(AR) rcs $@) scripts/Makefile.build:cmd_link_l_target = rm -f $@; $(AR) $(EXTRA_ARFLAGS) rcs $@ $(lib-y) masahiro@pug:~/ref/linux$ git grep '(CROSS32AR)' arch/powerpc/boot/Makefile: cmd_bootar = $(CROSS32AR) -cr $@.$$$$ $(filter-out FORCE,$^); mv $@.$$$$ $@ > If ar creates an archive file (a real one, not a thin archive), and it > has no input files, it uses its default object format as destination > format, if it isn't told to use something else. And that doesn't work, > it needs to use some format compatible with what that archive later is > linked with. I compile-tested v4.10, which was before the thin-archive migration, but I did not see any problem for building ppc32/64. Whether or not it is a thin archive, an empty archive is always 8-byte file. masahiro@pug:~/ref/linux$ cat kernel/livepatch/built-in.o !<arch> Is there a room for caring about the under-lying architecture? -- Best Regards Masahiro Yamada