On Mon, May 26, 2025 at 08:27:19PM +0200, Peter Zijlstra wrote:
> On Fri, May 09, 2025 at 01:17:16PM -0700, Josh Poimboeuf wrote:
> > +static void scan_for_insn(struct section *sec, unsigned long offset,
> > + unsigned long *insn_off, unsigned int *insn_len)
> > +{
> > + unsigned long o = 0;
> > + struct insn insn;
> > +
> > + while (1) {
> > +
> > + insn_decode(&insn, sec->data->d_buf + o, sec_size(sec) - o,
> > + INSN_MODE_64);
> > +
> > + if (o + insn.length > offset) {
> > + *insn_off = o;
> > + *insn_len = insn.length;
> > + return;
> > + }
> > +
> > + o += insn.length;
> > + }
> > +}
> > +
> > +u64 arch_adjusted_addend(struct reloc *reloc)
> > +{
> > + unsigned int type = reloc_type(reloc);
> > + s64 addend = reloc_addend(reloc);
> > + unsigned long insn_off;
> > + unsigned int insn_len;
> > +
> > + if (type == R_X86_64_PLT32)
> > + return addend + 4;
> > +
> > + if (type != R_X86_64_PC32 || !is_text_sec(reloc->sec->base))
> > + return addend;
> > +
> > + scan_for_insn(reloc->sec->base, reloc_offset(reloc),
> > + &insn_off, &insn_len);
> > +
> > + return addend + insn_off + insn_len - reloc_offset(reloc);
> > +}
>
> This looks like a rather expensive proposition; it will have to decode
> the section nr_reloc times.
>
> Does it not make more sense to fully decode the section like 'normal' ?
Yeah, I'm not crazy about it either, but it at least keeps the pain
nicely localized to x86, and avoids pulling in struct instruction,
struct objtool_file, etc.
Also this typically doesn't need to be all that fast as this is only
done for changed functions, and only for a subset of relocations (those
which might be references to non-bundled data in a text section).
To give a general idea, in one of my tests, for a patch with 22
functions, it only calls scan_for_insn() 41 times.
--
Josh
