[llvm-dev] RFC: Stop using redundant PHI node entries for multi-edge predecessors

[Sanjoy Das via llvm-dev]

Hi,

On Mon, May 1, 2017 at 8:47 AM, Daniel Berlin via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
>> Today, the IR requires that if you have multiple edges from A to B
>> (typically with a switch) any phi nodes in B must have an equal number of
>> entries for A, but that all of them must have the same value.
>
>> This seems rather annoying....
>> 1) It creates multiple uses of values in A for no apparently good reason
>> 2) It makes updating PHI nodes using sets of predecessors incredibly hard
>> 3) There is no correspondence between the edges and the PHI node entries
>> other than number, making most of the code that does handle this very ad-hoc
>> "count off N repetitions of the same thing" style code, which is brittle and
>> hard to test effectively.
>
> Having written this code a few times, i'd agree.
>
>>
>> 4) I suspect bugs are quite common given that PHINode::removeIncomingValue
>> accepts a basic block but only removes the first incoming value, leaving the
>> others in place.
>
>> I can't see any serious use case for this symmetry either. These aren't
>> uses (and can't be), and there doesn't seem to be any lost functionality if
>> we only have a single
>> entry.http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>> It also seems likely to open up more efficient in-memory representations
>> if it is possible at some point to build maps of these things.
>>
>> Thoughts?
>
>
>
> So would this lead to a case where PHI->getNumOperands() !=
> std::distance(pred_begin(phiblock), pred_end(phiblock))
>
> If so, that would seem problematic to handle as well.
> :(

Also, IIUC, today deleting an edge from A to B only requires
removeIncomingValue(A) on B's PHI nodes, but after this change you'll
have to check if you're deleting the last edge or not.

Can (2), (3) and (4) be fixed by changing the API instead of the
deeper IR change you're proposing?

-- Sanjoy