[PATCH] Prevent MDNode's RAUW from introducing a reference to a void Value.

[Duncan P. N. Exon Smith]

> On 2014-Oct-20, at 10:48, David Blaikie <dblaikie at gmail.com> wrote:
> 
> 
> 
> On Mon, Oct 20, 2014 at 10:42 AM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
> 
> > On Oct 20, 2014, at 10:40 AM, David Blaikie <dblaikie at gmail.com> wrote:
> >
> >
> >
> > On Mon, Oct 20, 2014 at 10:39 AM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
> >
> > > On Oct 20, 2014, at 10:34 AM, David Blaikie <dblaikie at gmail.com> wrote:
> > >
> > >
> > >
> > > On Mon, Oct 20, 2014 at 9:55 AM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
> > >
> > > > On Oct 20, 2014, at 9:36 AM, David Blaikie <dblaikie at gmail.com> wrote:
> > > >
> > > >
> > > >
> > > >> On Mon, Oct 20, 2014 at 8:45 AM, Frédéric Riss <friss at apple.com> wrote:
> > > >>
> > > >>> On Oct 19, 2014, at 6:43 PM, David Blaikie <dblaikie at gmail.com> wrote:
> > > >>>
> > > >>>
> > > >>>
> > > >>> On Sat, Oct 18, 2014 at 11:06 AM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
> > > >>> > On 2014 Oct 17, at 10:08, David Blaikie <dblaikie at gmail.com> wrote:
> > > >>> >
> > > >>> > I'm floating the idea that the right fix might be to drop the code, yes. We do this will null pointer dereferences and other cases of UB already (& yes, in LTO this means you might get some surprises - it's rather the nature of UB and optimizations).
> > > >>> >
> > > >>> >> Coming back to the patch in itself, I also don’t see how it could be the wrong thing to do. The TrackingVH in the MDNode is meant to handle the case of the Value going away. RAUWing with a void value is just another case of a vanishing value.
> > > >>> >
> > > >>> > I don't think it's really the same though, or at least I don't think it's clean that it needs to be the same. It's a pretty weird way for a value to go away & I'd consider not accepting it as a valid way for a value to go away without, possibly, a stronger example of where this is a reasonable way for a value to go away.
> > > >>>
> > > >>> I guess you're suggesting that optimizations shouldn't be allowed to
> > > >>> RAUW a non-void instruction to a void one.  Assuming all the IR `Use`s
> > > >>> have been updated,
> > > >>>
> > > >>> What uses of a void value can/could exist? I assume I can't have an Instruction who's operand is a void value...
> > > >>>
> > > >>> this RAUW seems legitimate -- it's the only way to
> > > >>> update all the ValueHandles to the instruction.
> > > >>> I think you'd need strong justification to remove this feature.
> > > >>>
> > > >>> I suppose the narrower claim would be: If an Instruction has (even debug/metadata) uses, you can't replace all its uses with a void value.
> > > >>
> > > >> You are aware of that, but I’ll mention it anyway. This sentence makes it look like having a debug metadata node referencing a value could change the outcome of an optimization. That would be very bad.
> > > >>
> > > > Yep (though I'm not sure how bad it is - we don't actually have any continuous integration tests to ensure we hold this invariant, but we do try to avoid breaking it and fix it when we discover it).
> > > >
> > > > I wasn't suggesting that - I was merely suggesting changing algorithms to ensure this doesn't happen ever (even when debug info is in use).
> > > >
> > > >
> > > >>> I don't think it's too unreasonable to ask for a single example of a reasonable use of this feature if we're going to consciously support it.
> > > >>
> > > >> I tried hard to come up with non-UB code that triggers this code path but couldn’t find one that gets us exactly in that situation. Using the DeadArgumentElimination pass to remove the unused return value gets us close, but it special cases a replacement by a void value by setting all references to a null value.
> > > >
> > > > Sounds plausible - though I'm not sure which uses (other than debug info/metadata) of the return value exist and why the comment says "they will get removed later on" - I wonder what the comment is alluding to.
> > > >
> > > >
> > > >>> Though I appreciate orthogonality of features and the ability to RAUW any value, even one with no uses, if its compatible with its users (a value with no uses is trivially compatible with all its users). But once this requires explicit support, I'm not sure how orthogonal it is.
> > > >>
> > > >> Would you prefer a patch that introduces an assert in the MDNode RAUW callback and that special cases the introduction of a void call in InstCombine (In the same way it is done in DAE for example) ?
> > > >
> > > > I think that might be reasonable - I'd be inclined to remove the UB code as dead to address this, but at least replacing with null limits the scope of the weirdness to InstCombine, rather than allowing it for all uses of metadata.
> > > >
> > > > (again, other people might disagree with me & think what I'm suggesting is silly/wrong - I'm not the owner/arbiter/gate-keeper here, just trying to describe my perspective on the matter - if other people who have more experience with/own metadata/IR think the original patch is reasonable & I'm being silly, they can sign off on it)
> > >
> > > I disagree with this direction.  I think ValueHandles should be updated
> > > to track the new call instruction.
> > >
> > > What use would other ValueHandle users have for a void value? Should metadata be updated to handle referencing a void value instead, if it's a meaningful/useful thing to reference?
> >
> > There are uses for ValueHandles other than as metadata operands.
> >
> > This particular `void` value is a call instruction.  If an optimization
> > (or anything else) has a side map of interesting calls (for some
> > definition of "interesting"), the map should get updated when this call
> > is replaced... not just redirected at null.  For someone interested in
> > call instructions, the return type isn't necessarily relevant.
> >
> > & why would this not be true of metadata? Its a reference to a value of interest, even if that value is void, no?
> >
> 
> It's invalid IR to have something of type `void` as an operand.
> 
> Should that be true of metadata, though? It doesn't really make sense for IR instructions, certainly, but I don't see why it would be fundamentally wrong for metadata.
>  

Personally, I don't see a problem with it, although I could be
missing something.  If you're interested in pursuing a change to
the IR semantics, you should write a proposal to llvmdev so a wider
audience sees it.

In the meantime I think Fred should commit the bugfix.