lkml.org 
[lkml]   [2020]   [Apr]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 3/5] objtool: Support Clang non-section symbols in ORC generation
On Wed, Apr 01, 2020 at 02:05:51PM -0500, Josh Poimboeuf wrote:
> On Wed, Apr 01, 2020 at 08:49:53PM +0200, Peter Zijlstra wrote:
> > On Wed, Apr 01, 2020 at 01:23:27PM -0500, Josh Poimboeuf wrote:
> >
> > > @@ -105,8 +100,32 @@ static int create_orc_entry(struct elf *elf, struct section *u_sec, struct secti
> > > }
> > > memset(rela, 0, sizeof(*rela));
> > >
> > > - rela->sym = insn_sec->sym;
> > > - rela->addend = insn_off;
> > > + if (insn_sec->sym) {
> > > + rela->sym = insn_sec->sym;
> > > + rela->addend = insn_off;
> > > + } else {
> > > + /*
> > > + * The Clang assembler doesn't produce section symbols, so we
> > > + * have to reference the function symbol instead:
> > > + */
> > > + rela->sym = find_symbol_containing(insn_sec, insn_off);
> >
> > It's a good thing I made that a lot faster I suppose ;-)
>
> :-)
>
> > > + if (!rela->sym) {
> > > + /*
> > > + * Hack alert. This happens when we need to reference
> > > + * the NOP pad insn immediately after the function.
> > > + */
> > > + rela->sym = find_symbol_containing(insn_sec,
> > > + insn_off - 1);
> >
> > Urgh, when does that happen?
>
> It happens naturally in the padding between functions, since objtool
> doesn't traverse those instructions. So they have undefined entries
> like
>
> .text+68: sp:(und) bp:(und) type:call end:0
>
> I suppose those aren't technically necessary.

In fact, we could probably get substantial savings in the ORC table if
we skipped those, i.e.

.text+0: sp:sp+8 bp:(und) type:call end:0
.text+8: sp:(und) bp:(und) type:call end:0
.text+10: sp:sp+8 bp:(und) type:call end:0
.text+17: sp:sp+16 bp:(und) type:call end:0
.text+18: sp:sp+24 bp:prevsp-24 type:call end:0
.text+1c: sp:sp+32 bp:prevsp-24 type:call end:0
.text+5a: sp:sp+24 bp:prevsp-24 type:call end:0
.text+61: sp:sp+16 bp:(und) type:call end:0
.text+63: sp:sp+8 bp:(und) type:call end:0
.text+68: sp:(und) bp:(und) type:call end:0
.text+70: sp:sp+8 bp:(und) type:call end:0
.text+8c: sp:(und) bp:(und) type:call end:0
.text+90: sp:sp+8 bp:(und) type:call end:0
.text+cd: sp:(und) bp:(und) type:call end:0
.text+d0: sp:sp+8 bp:(und) type:call end:0

would be compressed to

.text+0: sp:sp+8 bp:(und) type:call end:0
.text+17: sp:sp+16 bp:(und) type:call end:0
.text+18: sp:sp+24 bp:prevsp-24 type:call end:0
.text+1c: sp:sp+32 bp:prevsp-24 type:call end:0
.text+5a: sp:sp+24 bp:prevsp-24 type:call end:0
.text+61: sp:sp+16 bp:(und) type:call end:0
.text+63: sp:sp+8 bp:(und) type:call end:0

but I can do that in a separate patch, and if it works I can remove this
hack.

--
Josh

\
 
 \ /
  Last update: 2020-04-01 21:09    [W:0.118 / U:0.240 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site