Re: [PATCH 05/24] efi/libstub: Optimize for size instead of speed

From: Ard Biesheuvel <ardb@kernel.org>
To: Arvind Sankar <nivedita@alum.mit.edu>
Cc: Andrey Ignatov <rdna@fb.com>,
	linux-efi <linux-efi@vger.kernel.org>,
	bpf@vger.kernel.org
Subject: Re: [PATCH 05/24] efi/libstub: Optimize for size instead of speed
Date: Fri, 5 Jun 2020 16:53:59 +0200	[thread overview]
Message-ID: <CAMj1kXG936NeN7+Mf42bL-7V5pRVjoNmCKmVT3EcB5EGh2y5fQ@mail.gmail.com> (raw)
In-Reply-To: <20200605133232.GA616374@rani.riverdale.lan>

On Fri, 5 Jun 2020 at 15:32, Arvind Sankar <nivedita@alum.mit.edu> wrote:
>
> On Fri, Jun 05, 2020 at 09:14:19AM -0400, Arvind Sankar wrote:
> > On Fri, Jun 05, 2020 at 08:33:22AM +0200, Ard Biesheuvel wrote:
> > > Hello Andrey,
> > >
> > > On Fri, 5 Jun 2020 at 02:31, Andrey Ignatov <rdna@fb.com> wrote:
> > > >
> > > > Arvind Sankar <nivedita@alum.mit.edu> [Wed, 1969-12-31 23:00 -0800]:
> > > > > Reclaim the bloat from the addition of printf by optimizing the stub for
> > > > > size. With gcc 9, the text size of the stub is:
> > > > >
> > > > > ARCH    before  +printf    -Os
> > > > > arm      35197    37889  34638
> > > > > arm64    34883    38159  34479
> > > > > i386     18571    21657  17025
> > > > > x86_64   25677    29328  22144
> > > > >
> > > > > Signed-off-by: Arvind Sankar <nivedita@alum.mit.edu>
> > > > > ---
> > > > >  drivers/firmware/efi/libstub/Makefile | 4 ++--
> > > > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/drivers/firmware/efi/libstub/Makefile b/drivers/firmware/efi/libstub/Makefile
> > > > > index fb34c9d14a3c..034d71663b1e 100644
> > > > > --- a/drivers/firmware/efi/libstub/Makefile
> > > > > +++ b/drivers/firmware/efi/libstub/Makefile
> > > > > @@ -7,7 +7,7 @@
> > > > >  #
> > > > >  cflags-$(CONFIG_X86_32)              := -march=i386
> > > > >  cflags-$(CONFIG_X86_64)              := -mcmodel=small
> > > > > -cflags-$(CONFIG_X86)         += -m$(BITS) -D__KERNEL__ -O2 \
> > > > > +cflags-$(CONFIG_X86)         += -m$(BITS) -D__KERNEL__ \
> > > > >                                  -fPIC -fno-strict-aliasing -mno-red-zone \
> > > > >                                  -mno-mmx -mno-sse -fshort-wchar \
> > > > >                                  -Wno-pointer-sign \
> > > > > @@ -25,7 +25,7 @@ cflags-$(CONFIG_ARM)                := $(subst $(CC_FLAGS_FTRACE),,$(KBUILD_CFLAGS)) \
> > > > >
> > > > >  cflags-$(CONFIG_EFI_GENERIC_STUB) += -I$(srctree)/scripts/dtc/libfdt
> > > > >
> > > > > -KBUILD_CFLAGS                        := $(cflags-y) -DDISABLE_BRANCH_PROFILING \
> > > > > +KBUILD_CFLAGS                        := $(cflags-y) -Os -DDISABLE_BRANCH_PROFILING \
> > > > >                                  -include $(srctree)/drivers/firmware/efi/libstub/hidden.h \
> > > > >                                  -D__NO_FORTIFY \
> > > > >                                  $(call cc-option,-ffreestanding) \
> > > >
> > > > Hi Arvind,
> > > >
> > > > This patch breaks build for me:
> > > >
> > > > $>make -j32 -s bzImage
> > > > drivers/firmware/efi/libstub/alignedmem.c: In function \x2018efi_allocate_pages_aligned\x2019:
> > > > drivers/firmware/efi/libstub/alignedmem.c:38:9: sorry, unimplemented: ms_abi attribute requires -maccumulate-outgoing-args or subtarget optimization implying it
> > > >   status = efi_bs_call(allocate_pages, EFI_ALLOCATE_MAX_ADDRESS,
> > > >          ^
> > >
> > > Which version of GCC are you using?
> >
> > gcc-4.8.5 from the config. I got a copy and can reproduce it. Just
> > adding -maccumulate-outgoing-args appears to fix it, checking some more.
> >
>
> On a simple test:
>         extern void __attribute__ (( ms_abi )) ms_abi();
>         void sysv_abi(void) { ms_abi(); }
> it only breaks with -Os -fno-asynchronous-unwind-tables, weirdly enough.

I guess the logic that decides whether -maccumulate-outgoing-args is
enabled is somewhat opaque.

Could we perhaps back out the -Os change for 4.8 and earlier?