[llvm-dev] Full restrict support - status update

[Jeroen Dobbelaere via llvm-dev]

Hi Johannes et al,

> -----Original Message-----
> From: Doerfert, Johannes <jdoerfert at anl.gov>
[..]
> On 11/06, Jeroen Dobbelaere wrote:
> > >From: Alexey Zhikhartsev
> > [..]
> > > We would love to see your patches merged as soon as possible, so I was
> wondering: do you think the lack of bitcode support will prevent that from
> happening?
> >
> > Yes, I think that the lack of bitcode support will prevent it.
> >
> > During the Developers meeting, I also talked with Hal and Johannes.
> > They had some extra remarks:
> > - (1) the restrict implementation deserves a separate document. (I am
> working on that one)
> > - (2) they don't like the naming of 'noalias_sidechannel'.
> > - (3) they also have some other mechanisms in mind to add the 'sidechannel'
> to the load/store instructions
> >        (and maybe to function calls, intrinsics; currently that is handled through
> llvm.noalias.arg.guard)
> >
> > For (2) and (3), I am waiting for a proposal from them ;)
> 
> I would like to see the restrict support be merged but, as Jeroen
> mentions above, I feel there are two design choices we have to
> overthink. Here are short descriptions to get some feedback from the
> community:
> 
> (A) Naming and restriction
> 
> The name "sidechannel" is unfortunate, it has various negative
> connotations, e.g., the release notes that read:
>  "LLVM 10.0 now has sidechannel support for your restrict pointer"
> will raise a lot of follow up questions.
> 
> What I think we actually do, and what we should call it, is "provenance"
> tracking.
> 
> Now beyond the pure renaming of "sidechannel" into "provenance" (or sth.
> similar) I want us to decouple provenance tracking from the noalias
> logic. Noalias/restrict is one use case in which (pointer) provenance
> information is useful but not the only one. If we add some mechanism to
> track provenance, let's make it generic and reusable. Note that the
> basic ideas are not much different to what the noalias RFC proposed.
> The major difference would be that we have provenance information and if
> that originates in an `llvm.restrict.decl` call we can use it for
> (no)alias queries.

"provenance" might indeed be a good name.

There is a big difference between a restrict declaration, and a restrict usage: 
- the declaration intrinsic (llvm.noalias.decl) is used to track in the cfg the location
   where the restrict variable was declared. This is important to handle code motion,
   merging, duplication in a correct way (inlining, loop unrolling, ...)
- the restrict usage intrinsics (llvm.noalias and llvm.side.noalias) are used to indicate
   that from that point on, restrict (noalias) properties are introduced for that pointer.
  They can exist without an associated 'llvm.noalias.decl' (when the declaration is outside
   the function.)
Given that, I assume that you mean 'llvm.provenance.noalias' (~ llvm.side.noalias) instead
of 'llvm.restrict.decl'.

> 
> 
> 
> (B) Using operand bundles
> 
> Right now, loads and stores are treated differently and given a new
> operand. Then there are intrinsics to decode other kinds of information.
> As an alternative, we could allow operand bundles on all instructions
> and use them to tie information to an instruction. The "sidechannel"
> operand of a load would then look something like:
>   load i32* %p [ "ptr_provenance"(%p_decl) ]
> and for a store we could have
>   store i32** %p.addr, i32* %p [ "ptr_provenance"(%p_decl) ]
> 
> The benefit is that we do not change the operand count (which causes a
> lot of noise) but we still have to make sure ptr/value uses are not
> confused with operand bundle uses. We can attach the information to more
> than load/store instructions, also to remove the need for some of the
> intrinsics.

To me, operand bundles sound to be more or less equivalent to the current
solution. It  might also make the 'instruction cloning' easier, if we can omit the
'ptr_provenance' there. The change of the number of operands caused some
noise, but it is the changes in the amount of 'uses' of a pointer that refer to the
same instruction that caused the most problems. Especially when that instruction
was going to be erased. Operand bundles will still need those code changes.
(like in parts of D68516 and D68518)

As the 'Call' instruction already supports operand bundles, it could eliminate the need
for the 'llvm.noalias.arg.guard' intrinsic, which combines the normal pointer with the
side channel (aka provenance). But, after inlining, we still need to put that information
somewhere. Or it should be propagated during inlining.
Care must be taken not to lose that information when the 'call' is changed by optimizations
as, after inlining, that might result in wrong alias analysis conclusions.

Are you thinking of "operand bundles" support for just LoadInst/StoreInst, or for all
instructions ?

Greetings,

Jeroen Dobbelaere