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

[David Blaikie]

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. 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.

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.

- David
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20141019/198ad73c/attachment.html>