On Wed, 3 Jun 2020 at 11:15, Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote: > > On Wed, Jun 3, 2020 at 6:02 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote: > > > > On Wed, 3 Jun 2020 at 10:59, Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote: > > > > > > On Wed, Jun 3, 2020 at 5:45 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote: > > > > > > > > On Wed, 3 Jun 2020 at 10:36, Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote: > > > > > > > > > > On Wed, Jun 3, 2020 at 3:45 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote: > > > > > > > > > > > > On Wed, 3 Jun 2020 at 07:34, Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote: > > > > > > > > > > > > > > Documentation/kbuild/makefiles.rst says: > > > > > > > > > > > > > > Use of lib-y is normally restricted to `lib/` and `arch/*/lib`. > > > > > > > > > > > > > > I want to disallow lib-y outside of them. > > > > > > > > > > > > > > > > > > > Why? > > > > > > > > > > > > > > > Because I plan to remove lib-y entirely at some point. > > > > > > > > > > lib-y is not so useful to shrink the image size because: > > > > > > > > > > - An object in lib.a can be omitted only when no symbol > > > > > in that object is referenced. This rarely happens. > > > > > > > > > > - lib-y objects are often exported by nature > > > > > because lib-y is a collection of utility functions. > > > > > Even if no in-tree user, we always need to keep them > > > > > because EXPORT_SYMBOL() is the interface to modules. > > > > > > > > > > > > > > > When I worked on commit 7273ad2b08f8ac9563579d16a3cf528857b26f49, > > > > > I made some research. > > > > > > > > > > The benefit of lib-y is just 362 byte for x86_64_defconfig. > > > > > ( Before: 26578002, After: 26578364) > > > > > > > > > > My hope is lib-y will be replaced by dead-code elimination or > > > > > ultimately by LTO. > > > > > > > > > > drivers/firmware/efi/libstub/Makefile > > > > > is the only Makefile that breaks the rule: > > > > > "Use of lib-y is normally restricted to `lib/` and `arch/*/lib`" > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Add a custom rule to build lib.a, which is linked to the decompressor > > > > > > > for ARCH=x86, ARCH=arm. > > > > > > > > > > > > > > For ARCH=arm64, use obj-y to link objects to vmlinux in the ordinary > > > > > > > way. > > > > > > > > > > > > > > > > > > > The code works perfectly fine as is, and I don't see what is > > > > > > fundamentally wrong with using static libraries outside of lib/ and > > > > > > arch/*/lib. > > > > > > > > > > The intended usage of lib-y is to hook lib.a > > > > > to scripts/vmlinux.sh via KBUILD_VMLINUX_LIBS. > > > > > > > > > > This Makefile is just what you found to work. > > > > > > > > > > > > > > > > > > > > > > Also, I would like this code to still be incorporated as a static > > > > > > library into arm64 as well, so that only pieces that are actually > > > > > > needed are incorporated into the final image. > > > > > > > > > > No. > > > > > It is not working like that because you set > > > > > lib.a to core-y. > > > > > > > > > > All objects in core-y are always linked to vmlinux. > > > > > > > > > > > > > The lib.a file is passed to the linker as a static library, so it will > > > > only grab what it needs. > > > > > > > > For instance, if you build arm64 from mainline today, the > > > > efi_relocate_kernel will not be in the final image, even though it is > > > > built as part of libstub > > > > > > > > > I built today's mainline kernel > > > (d6f9469a03d832dcd17041ed67774ffb5f3e73b3). > > > > > > > > > I see it in vmlinux. > > > > > > > > > $ make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- defconfig > > > $ make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- -j24 > > > ... > > > $ aarch64-linux-gnu-nm -n vmlinux | grep efi_relocate_kernel > > > ffff8000114afb90 t __efistub_efi_relocate_kernel > > > > > > > That is strange. I tested this before, and it worked. > > > > Did anything change recently in the way the linker is invoked? > > > Nothing recently. > > This is obvious result because > drivers/firmware/efi/libstub/lib.a > is passed after ----whole-archive flag. > > > I guess the following commit is it, > but it is already 3 years ago. > Right. So that means there is no point at all in using static libraries, no? Or does --whole-archive still allow duplicate definitions like static libraries do usually?
