Hi Josh and Peter,
An in-review LLVM patch [1] introduces an optimization around switch
cases, marking the default case unreachable when all known values are
covered. This ended up introducing a boot regression when building a
kernel with ThinLTO, which Sami reported initially [2] and I reported it
upstream, generating some discussion starting from [3].
There is an objtool warning about the function that the author of the
patch tracked down to be problematic:
vmlinux.o: warning: objtool: cfg80211_edmg_chandef_valid()+0x169: can't
find jump dest instruction at .text.cfg80211_edmg_chandef_valid+0x17b
The LLVM developers are under the impression that this is an issue with
objtool; specifically quoting Eli Friedman:
"The backend can, in general, create basic blocks that don't contain any
instructions, and don't fall through to another block. A jump table
entry can refer to such a block. I guess certain tools could be confused
by this.
If that's the issue, it should be possible to work around it using
'-mllvm -trap-unreachable'."
I can confirm that adding '-mllvm -trap-unreachable' to
KBUILD_{C,LD}FLAGS does resolve the issue but I am curious if that is an
adequate solution (as Nick has frowned upon the presence of '-mllvm'
flags) or if there is something that could be done on the objtool side.
I do have a small reproducer on Phabricator [4] and I can provide any
binary files that would be helpful for seeing what is going on here.
[1] https://reviews.llvm.org/D106056
[2] https://github.com/ClangBuiltLinux/linux/issues/1440
[3] https://reviews.llvm.org/D106056#2953104
[4] https://reviews.llvm.org/D106056#2961136
Cheers,
Nathan
On 8/24/2021 2:05 PM, Peter Zijlstra wrote:
> On Tue, Aug 24, 2021 at 01:08:58PM -0700, Nathan Chancellor wrote:
>
>> The LLVM developers are under the impression that this is an issue with
>> objtool; specifically quoting Eli Friedman:
>>
>> "The backend can, in general, create basic blocks that don't contain any
>> instructions, and don't fall through to another block. A jump table entry
>> can refer to such a block. I guess certain tools could be confused by this.
>>
>> If that's the issue, it should be possible to work around it using '-mllvm
>> -trap-unreachable'."
>
> So jump-tables are a weak point; ARM64 was having worse problems than
> x86 there, they can't even locate them.
>
> As to having a jump-table entry to an empty block and not falling
> through; how are we supposed to know?
Fair enough. It does make me wonder why LLVM does that.
> Emitting them is a waste of space, so I'd say it's a compiler bug :-))
Isn't it always? :)
Turns out Nick brought up an issue very similar to this (unreachable
conditions with switches) on LLVM's issue tracker
(https://bugs.llvm.org/show_bug.cgi?id=50080) with the same workaround
suggestion ('-mllvm -trap-unreachable') and there was no follow up after
that so maybe that is one thing to look into once Nick is back online.
> It's been brought up before; but perhaps we should look at an 'informal'
> ABI for jump-tables ?
Not a bad idea, especially if this has come up before.
Cheers,
Nathan
On Tue, Aug 24, 2021 at 02:19:07PM -0700, Nathan Chancellor wrote:
> On 8/24/2021 2:05 PM, Peter Zijlstra wrote:
> > On Tue, Aug 24, 2021 at 01:08:58PM -0700, Nathan Chancellor wrote:
> >
> > > The LLVM developers are under the impression that this is an issue with
> > > objtool; specifically quoting Eli Friedman:
> > >
> > > "The backend can, in general, create basic blocks that don't contain any
> > > instructions, and don't fall through to another block. A jump table entry
> > > can refer to such a block. I guess certain tools could be confused by this.
> > >
> > > If that's the issue, it should be possible to work around it using '-mllvm
> > > -trap-unreachable'."
> >
> > So jump-tables are a weak point; ARM64 was having worse problems than
> > x86 there, they can't even locate them.
> >
> > As to having a jump-table entry to an empty block and not falling
> > through; how are we supposed to know?
>
> Fair enough. It does make me wonder why LLVM does that.
>
> > Emitting them is a waste of space, so I'd say it's a compiler bug :-))
>
> Isn't it always? :)
>
> Turns out Nick brought up an issue very similar to this (unreachable
> conditions with switches) on LLVM's issue tracker
> (https://bugs.llvm.org/show_bug.cgi?id=50080) with the same workaround
> suggestion ('-mllvm -trap-unreachable') and there was no follow up after
> that so maybe that is one thing to look into once Nick is back online.
>
> > It's been brought up before; but perhaps we should look at an 'informal'
> > ABI for jump-tables ?
> Not a bad idea, especially if this has come up before.
This is definitely needed. Jump tables have always been a major thorn
in objtool's side. I think I volunteered to write up a proposal for the
linux-toolchains list but I've just been waaay too busy.
--
Josh