Re: [PATCH] x86/speculation, objtool: Use absolute relocations for annotations

From: Fangrui Song
Date: Thu Sep 21 2023 - 16:27:17 EST


On Thu, Sep 21, 2023 at 12:22 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, Sep 21, 2023 at 10:36:27AM -0700, Fangrui Song wrote:
>
> > Well, only if the clever assembler doesn't support 32-bit absolute
> > relocation for a 64-bit architecture.
> > I don't know such an architecture. In addition, as long as the
> > architecture intends to support DWARF32, it has to support 32-bit
> > absolute relocations for a 64-bit architecture.
>
> Ooh... my bad. For some reason I thought that absolute meant native word
> size. But you already mentioned R_X86_64_32 (and I failed to check) and
> that is indeed an absolute (S+A) relocation of 32bit (dword) size.
>
> And apparently we also have R_X64_64_16 and R_X86_64_8, which would even
> allow something like:
>
> #define OBJTOOL_ANNOTATE(type) \
> "999:\n\t" \
> ".pushsection .discard.objtool_annotate\n\t" \
> ".byte 999b\n\t" \
> ".byte " __stringify(type) "\n\t" \
> ".popsection\n\t"
>
> And since we only read the relocation and don't care for the actual
> result that would actually work just fine.
>
> Anyway, thanks for bearing with me.
>
> Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>

Thanks! I assume that the patch can be picked up without me doing anything :)

Yes, x86 also supports 1-byte and 2-byte absolute relocation types,
which are missing from some other architectures.
If we want to optimize for RELA, we can use R_*_NONE by utilizing
`.reloc ., BFD_RELOC_NONE, .text.foo` (as mentioned in the commit
message).
(binutils 2.26 https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=740bdc67c057ee8012327420848eb134e1db4211;
I added the support to most LLVM targets in 2021).

However, this is just 4 byte per entry difference compared with the
large sizeof(Elf{32,64}_Rel), so not worth the trouble:)

--
宋方睿