llvm dev - Feb 2016 - Possible soundness issue with available_externally (split from "RFC: Add guard intrinsics")

On Wed, Feb 24, 2016 at 10:25 PM Sanjoy Das <sanjoy at
playingwithpointers.com>
wrote:
>
> Hal Finkel wrote:
>
>  > But it is not all optimizations that are the problem. Rather, it
>  > seems like a select few (e.g. things involving collapsing allowed
>  > non-determinism in atomics), and losing those optimizations seems
>  > better than generally losing function-attribute deduction.
>
> If we go by the langref, then optimizations that fold undef are also
> problematic (though most C/C++ programs resulting in such IR would
> have UB in practice).
>
I'm also concerned about undef. It isn't clear that we can reasonably
ensure that all things involving undef suitably stem from UB in the face of
speculation to preclude all "impossible" results coming from
interposition.

I think trying to enumerate the function-local optimizations which are
unsafe is *much* more challenging than identifying the interprocedural
optimizations which are unsafe, which is essentially what disabling
function attribute deduction is doing.

>
> I think there are non-(non-deterministic) problematic cases too.  The
> following won't happen today since `readnone` does not imply
> `safe_to_speculate`, but if we add a `safe_to_speculate` property some
> day:
>
>    int foo(bool C) available_externally {
>      if (C)
>        ((int *) null)++; // UB
>      ret 42;
>    }
>
>    void bar() {
>      if (<some cond>)
>       foo(true);
>    }
>
> Now, normally you can just delete the `if (C)` branch in foo, and it
> would become just a `ret 42`, and would look like it is speculatable
> above the `<some cond>` check.  But if you then link with an -O0
> version, you'll have introduced UB if `<some cond>` is always
false at
> runtime.
>
> Today this won't happen since we don't speculate `readnone
nounwind`
> functions, but could become a problem in the future.
>
All of this makes me think we're just seeing the tip of the iceberg of ways
in which local optimizations can combine with attributes to break this. =/
I really think the problem is as Richard describes the interprocedural
movement of information across a comdat-like boundary where replacements
may occur with differently transformed code.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160225/c2f8883d/attachment.html>

----- Original Message -----
> From: "Chandler Carruth" <chandlerc at google.com>
> To: "Sanjoy Das" <sanjoy at playingwithpointers.com>,
"Hal Finkel" <hfinkel at anl.gov>
> Cc: "llvm-dev" <llvm-dev at lists.llvm.org>, "Philip
Reames" <listmail at philipreames.com>, "Duncan P. N. Exon
Smith"
> <dexonsmith at apple.com>
> Sent: Thursday, February 25, 2016 12:40:53 AM
> Subject: Re: [llvm-dev] Possible soundness issue with available_externally
(split from "RFC: Add guard intrinsics")
> 
> 
> 
> 
> On Wed, Feb 24, 2016 at 10:25 PM Sanjoy Das <
> sanjoy at playingwithpointers.com > wrote:
> 
> 
> 
> Hal Finkel wrote:
> 
> > But it is not all optimizations that are the problem. Rather, it
> > seems like a select few (e.g. things involving collapsing allowed
> > non-determinism in atomics), and losing those optimizations seems
> > better than generally losing function-attribute deduction.
> 
> If we go by the langref, then optimizations that fold undef are also
> problematic (though most C/C++ programs resulting in such IR would
> have UB in practice).
> 
> 
> 
> I'm also concerned about undef. It isn't clear that we can
reasonably
> ensure that all things involving undef suitably stem from UB in the
> face of speculation to preclude all "impossible" results coming
from
> interposition.
> 
> 
> I think trying to enumerate the function-local optimizations which
> are unsafe is *much* more challenging than identifying the
> interprocedural optimizations which are unsafe, which is essentially
> what disabling function attribute deduction is doing.
> 
It might be much more challenging, but let's try. This is not an issue we
need to fix by the end of the month, and the potential optimization regressions
are significant. Our deductions of readonly/readnone/nocapture/etc. are really
important for enabling other optimizations. Given that all of our C++
constructors, inline functions, etc. end up in comdat sections, this is really
important.
> 
> I think there are non-(non-deterministic) problematic cases too. The
> following won't happen today since `readnone` does not imply
> `safe_to_speculate`, but if we add a `safe_to_speculate` property
> some
> day:
> 
> int foo(bool C) available_externally {
> if (C)
> ((int *) null)++; // UB
> ret 42;
> }
> 
> void bar() {
> if (<some cond>)
> foo(true);
> }
> 
> Now, normally you can just delete the `if (C)` branch in foo, and it
> would become just a `ret 42`, and would look like it is speculatable
> above the `<some cond>` check. But if you then link with an -O0
> version, you'll have introduced UB if `<some cond>` is always
false
> at
> runtime.
> 
> Today this won't happen since we don't speculate `readnone
nounwind`
> functions, but could become a problem in the future.
> 
> All of this makes me think we're just seeing the tip of the iceberg
> of ways in which local optimizations can combine with attributes to
> break this. =/ I really think the problem is as Richard describes
> the interprocedural movement of information across a comdat-like
> boundary where replacements may occur with differently transformed
> code.
Yes, but it is not *all* information that is relevant here.

 -Hal

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory

On Wed, Feb 24, 2016 at 11:18 PM Hal Finkel <hfinkel at anl.gov> wrote:
> ----- Original Message -----
> > From: "Chandler Carruth" <chandlerc at google.com>
> > To: "Sanjoy Das" <sanjoy at playingwithpointers.com>,
"Hal Finkel" <
> hfinkel at anl.gov>
> > Cc: "llvm-dev" <llvm-dev at lists.llvm.org>,
"Philip Reames" <
> listmail at philipreames.com>, "Duncan P. N. Exon Smith"
> > <dexonsmith at apple.com>
> > Sent: Thursday, February 25, 2016 12:40:53 AM
> > Subject: Re: [llvm-dev] Possible soundness issue with
> available_externally (split from "RFC: Add guard intrinsics")
> >
> >
> >
> >
> > On Wed, Feb 24, 2016 at 10:25 PM Sanjoy Das <
> > sanjoy at playingwithpointers.com > wrote:
> >
> >
> >
> > Hal Finkel wrote:
> >
> > > But it is not all optimizations that are the problem. Rather, it
> > > seems like a select few (e.g. things involving collapsing allowed
> > > non-determinism in atomics), and losing those optimizations seems
> > > better than generally losing function-attribute deduction.
> >
> > If we go by the langref, then optimizations that fold undef are also
> > problematic (though most C/C++ programs resulting in such IR would
> > have UB in practice).
> >
> >
> >
> > I'm also concerned about undef. It isn't clear that we can
reasonably
> > ensure that all things involving undef suitably stem from UB in the
> > face of speculation to preclude all "impossible" results
coming from
> > interposition.
> >
> >
> > I think trying to enumerate the function-local optimizations which
> > are unsafe is *much* more challenging than identifying the
> > interprocedural optimizations which are unsafe, which is essentially
> > what disabling function attribute deduction is doing.
> >
>
> It might be much more challenging, but let's try. This is not an issue
we
> need to fix by the end of the month, and the potential optimization
> regressions are significant. Our deductions of
> readonly/readnone/nocapture/etc. are really important for enabling other
> optimizations. Given that all of our C++ constructors, inline functions,
> etc. end up in comdat sections, this is really important.
>
Well, no one has really commented on my ideas to mitigate the issue.. I'm
actually much less concerned in the face of those ideas.

But given that we can't even formulate a good statement of what transforms
would necessarily have to be forbidden, I think we should get to a sound
state first, and then work on more aggressive models to use.

I think that coming up with something that is really sound is going to be
incredibly hard here because I think it will be tantamount to removing the
concept of undef entirely and the entire optimizer's reliance on it. And
most of the obvious ideas to replace it I think will have the same
challenges as undef.

We're also, meanwhile, going to have to dramatically curtail our
investigation of more exciting IPO opportunities because all of them bust
be carefully examined to not trigger these issues more directly.
> All of this makes me think we're just seeing the tip of the iceberg
> > of ways in which local optimizations can combine with attributes to
> > break this. =/ I really think the problem is as Richard describes
> > the interprocedural movement of information across a comdat-like
> > boundary where replacements may occur with differently transformed
> > code.
>
> Yes, but it is not *all* information that is relevant here.
>
True. It is any information which is not necessarily the same in all
(possibly transformed or untransformed) variants of the function.

My only confident summary of that is: "the information we can deduce prior
to any transformations". It sounds like you think we can effectively
preclude any such transformations on functions with ODR. I think that this
will end up being extraordinarily limiting if we can do it at all. It
certainly isn't the optimizer we have today, and would require changes all
over...

>  -Hal
>
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160225/42c4de60/attachment.html>

On Wed, Feb 24, 2016 at 11:18 PM, Hal Finkel <hfinkel at anl.gov> wrote:
> It might be much more challenging, but let's try. This is not an
> issue we need to fix by the end of the month, and the potential
> optimization regressions are significant. Our deductions of
> readonly/readnone/nocapture/etc. are really important for enabling
> other optimizations. Given that all of our C++ constructors, inline
> functions, etc. end up in comdat sections, this is really important.
The only optimizations I can think of that are okay are algebraic
simplifications that don't exploit no-overflow, inbounds or exact
flags and CFG simplifications like loop unswitch and loop unroll (that
don't actually change the trace that we'll see at runtime).

-- Sanjoy