[PATCH] Add iterator for PHINode value/BB pair

Duncan P. N. Exon Smith dexonsmith at apple.com
Mon Aug 3 11:07:36 PDT 2015


> On 2015-Aug-03, at 10:54, David Blaikie <dblaikie at gmail.com> wrote:
> 
> 
> 
> On Mon, Aug 3, 2015 at 10:38 AM, Pete Cooper <peter_cooper at apple.com> wrote:
> Sorry, just getting back to this
> 
> No worries
>  
> since i’m starting to see other places where it could be useful.
>> On Jul 23, 2015, at 1:07 PM, David Blaikie <dblaikie at gmail.com> wrote:
>> 
>> 
>> 
>> On Thu, Jul 23, 2015 at 12:29 PM, Pete Cooper <peter_cooper at apple.com> wrote:
>> Sorry this took so long.  Had to learn a whole bunch about varargs templates.
>> 
>> This is a first attempt at an LLVM zip iterator.  I needed to add methods to STLExtras.h which do ++tuple<> and *tuple<>.  Originally I put those under different names and didn’t override the operators, but then I didn’t see any harm in overriding them.  I can rename them if anyone has a preference for that.
>> 
>> Yeah, those do make me twitch a bit - I'm not sure if there's a more coherent argument against them or not.
> So i can totally understand the ‘twitch’.  Is it a bad twitch, or just something that takes a while to accept as how i’ve implemented it?
> 
> Uncertain twitch - a "this feels a bit not-good, but I'm not sure if it's unacceptably so, or just a weird thing I could get used to". I wouldn't mind roping in Mr. Smith or other C++ expert opinions (if you want to Dig Doug out of the Swift Mines, etc).
> 
> If we can't find some more/better opinions it doesn't seem like it'd be too costly to just avoid the quirk and implement a custom iterator that wraps all the other iterators and has the operator overloads, etc. 

I have the same twitch... it seems dangerous somehow to add API to
std::tuple.  I think adding our own class (that wraps a `std::tuple<...>`)
would feel safer, I guess because it's more explicit.  Kind of on the
fence myself though.

>  I’d like to try get this to a state where we can consider committing it, so want to make sure everyone is happy with the solution.
>> 
>> Oh, I see, this allows you to make the tuple itself an iterator. Yeah, also makes me twitch, but again - not sure if that's problematic or not. (I'm just imagining two libraries trying to pull the same game - and colliding in some painful way)
>> 
>> (regardless of the name, wonder if the ++ and * could be generalized over each other (presumably the latter, which is more powerful (since it has a return value) than the former - but I don't know if tuple<void, void, ...> is valid... ;)))
> Yeah, i’ll try generalizing these over one another.  They also have different solutions for how to walk the tuple which isn’t ideal.  Better to just have a single way to apply an operator across a whole tuple.
>>  
>> The zip iterator itself is almost identical to the edge_iterator I had before, just that it is now templated on its iterator type.
>> 
>> There seem to be some naming issues here... wait, no, I see it - the tuple is the iterator (retcon'd comments in above once I realized this)
>>  
>>  I’ve provided a subclass in zip_input_iterator for users who don’t need all the parameters and just want to provide a varargs of iterator types.  All of this is of course up for discussion, this is just one way I could see to implement it.
>> 
>> Probably could keep it in one class rather than zip_iterator and zip_input_iterator - just always pass the collection of iterator types and then deduce the tag from the lowest common denominator of the iterator tags of the specified iterators (not sure if there's already a trait/etc for joining those different things together and getting the least powerful iterator tag as a result). Save us having a taxonomy of zip_*_iterators.
> Nice idea to derive a common trait here.  I’ll give that a try.  That will make zip_iterator just take a T… as its only template parameter and remove the input_iterator.
>>  
>> I also allowed for zip_iterator_traits so that users could provide their own type to return from operator*.  This is so that users don’t have to remember which iterator is first, second, get<0>, etc.
>> 
>> Not sure if it's in 11 or 14, but there is type based access into tuple, get<Value*>, get<BasicBlock*>, etc - perhaps that's sufficient?
> Looks like its std::get<T>(tuple) and is in C++11, with more variety in C++14.  I think thats sufficient for many of the cases...
>>  
>>  Personally I think its much more readable than to have .Value, .Block, etc.  This is optional though, as there’s a default implementation to provide a dereferenced tuple type from operator*.
>> 
>> Interesting extension point, certainly. Could be problematic if there's the same pairing in two places with different meaning (int, int is size/length in one place, and something else elsewhere) and doesn't match up with different iterators (which might be weird to users - they change container type and then their code breaks because the trait is based on the iterator type, not the value type)
> Yeah, i think this is interesting, but you’re right to raise some questions here.  I think we would have to be careful where and when to expose this.  I’ll drop this extension for any initial commit, but perhaps if we need it later we can revisit it.
>>  
>> BTW, no idea how much of this is MSVC safe.  I did find a user of 'auto fn()->decltype()’ in PDBTypes.h, so i’m not the first user of that, but this will need someone with MSVC to ok the final implementation.
>> 
>> I also need to write a unit test for all the functionality this enables.  Duncan mentioned std::copy requiring operator* to return a reference.
>> 
>> Interesting - I wonder how much/how it requires that? Because tuple<T&, U&> should be assignable (and assign to the reference members), etc, should work... even if the tuple is returned by value.
> Yeah, I think this is fine.  I just need to write a bunch of tests to double check, but as you’ve said, its a <T&, U&> tuple which is assignable.
>>  
>>  I’m still working through whether thats working or whether we only really have foreach support right now.
>> 
>> Duncan also pointed out in person, perhaps this is overkill as the edge_iterator I had before was easier to understand than what we have here.  Perhaps we’d prefer to hold off on a zip_iterator until we are sure we have more users for it.
>> 
>> There are a fair few places where we iterate over two things in parallel and they're pretty awkward right now - could find a smattering of those and clean them up with the use of this tool (with a helper function to allow building these on-site within the range-for). I'm not sure of the easiest way to find places that could use this cleanup though... probably looking for for loops that either increment two things, or that use integers (because they're indexing into two containers).
> The case i’ve now found is in AArchISelLowering where we do:
> 
> for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
>     CCValAssign &VA = ArgLocs[i];
> 
>     if (Ins[i].Flags.isByVal()) {
> 
> So we need to zip (ArgLocs, Ins).
> 
> Given that this is done in pretty much every backend, I think there are going to be enough cases to justify adding zip_iterator as i’m sure we’ll find more over time (including the one Daniel B has)
> 
> Will start trying to improve the patch based on all the feedback.
> 
> Cheers,
> Pete
> 
>>  
>>  If thats the case, then i’m happy for this to just be a demonstration of how it could be done, but not necessarily something to commit now.
>> 
>> Fair, too.
>> 
>> - Dave
>>  
>> 
>> Cheers,
>> Pete
>> 
>> 
>> 
>>> On Jul 22, 2015, at 5:10 PM, David Blaikie <dblaikie at gmail.com> wrote:
>>> 
>>> (oops, dropped the list by accident)
>>> ---------- Forwarded message ----------
>>> From: David Blaikie <dblaikie at gmail.com>
>>> Date: Wed, Jul 22, 2015 at 5:10 PM
>>> Subject: Re: [PATCH] Add iterator for PHINode value/BB pair
>>> To: "Duncan P. N. Exon Smith" <dexonsmith at apple.com>
>>> 
>>> 
>>> 
>>> 
>>> On Wed, Jul 22, 2015 at 5:05 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
>>> 
>>> > On 2015-Jul-22, at 16:57, David Blaikie <dblaikie at gmail.com> wrote:
>>> >
>>> >
>>> >
>>> > On Wed, Jul 22, 2015 at 4:50 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
>>> >
>>> > > On 2015-Jul-22, at 15:07, David Blaikie <dblaikie at gmail.com> wrote:
>>> > >
>>> > >
>>> > >
>>> > > On Wed, Jul 22, 2015 at 12:11 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
>>> > >
>>> > > > On 2015-Jul-21, at 21:20, Pete Cooper <peter_cooper at apple.com> wrote:
>>> > > >
>>> > > > Thanks for all the feedback.  This is a patch which addresses all of it.
>>> > > >
>>> > > > <phinode.diff>
>>> > > >
>>> > >
>>> > > > +
>>> > > > +    const PHINodeEdge operator*() const {
>>> > >
>>> > > No reason for the const in `const PHINodeEdge` here.
>>> > >
>>> > > to support operator-> you have to return a pointer, which means you need the PHINodeEdge storage inside the iterator to point to (& then you can just return a const ref from op*)
>>> >
>>> > Do we need operator->()?
>>> >
>>> > Seems poor form not to provide it (someone'll trip over it pretty quickly, I'd imagine)
>>> >
>>> > If so, we can return a proxy object:
>>> >
>>> >     struct PHINodeEdgeArrowProxy {
>>> >       const PHINodeEdge RefProxy;
>>> >     public:
>>> >       PHINodeEdgeArrowProxy(PHINodeEdge Edge) : RefProxy(Edge) {}
>>> >       const PHINodeEdge *operator->() const { return &RefProxy; }
>>> >     };
>>> >
>>> >     PHINodeEdgeArrowProxy operator->() const { return operator*(); }
>>> >
>>> > Then we avoid bloating the iterator, and only make the copy when we
>>> > actually need it.
>>> >
>>> > Non-conforming in terms of the iterator traits, I would imagine
>>> 
>>> AFAICT, iterators only require that `i->m` has the same semantics as
>>> `(*i).m`.  The return type isn't specified.
>>> 
>>> Fair enough - can't quite find the wording on what constitutes the iterators value type (as is mentioned in the iterator traits) or how it might relate, but the basic definition is as you've mentioned.
>>>  
>>> > - but I take it that's the N1550 stuff you're talking about below? It makes these sort of proxy solutions valid?
>>> >
>>> 
>>> IIRC, proxy solutions are always valid for InputIterator and for
>>> OutputIterator; it's ForwardIterator that prevents `operator*()`
>>> from returning a proxy.
>>> 
>>> *nod* I don't mind violating that too much, if that's the preference.
>>>  
>>> (This makes `std::vector<bool>::iterator` invalid.)
>>> 
>>> Yep. Fails in so many ways.
>>>  
>>> 
>>> >
>>> > >
>>> > >
>>> > > > +      return { *Values, *Blocks };
>>> > > > +    }
>>> > > > +  };
>>> > > > +
>>> > >
>>> > > Otherwise LGTM.  Might want to pass it through clang-format; I noticed
>>> > > some minor whitespace oddities.
>>> > >
>>> > > > *snip*
>>> > > >> On Jul 21, 2015, at 8:21 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
>>> > > >>
>>> > > >>
>>> > > >>
>>> > > >> If you drop the requirements from `forward_iterator` to
>>> > > >> `input_iterator`, then you're allowed to return a `PHINodeEdge` by-value
>>> > > >> here instead of by-reference (unfortunately this makes it illegal to use
>>> > > >> a bunch of STL algorithms; STL iterator traits are completely broken
>>> > > >> IMO).
>>> > > > I’m fine with this.  David, Chandler, please let me know how you feel about this.
>>> > > >
>>> > > > Also, i forgot to say that I considered doing a zip iterator and inheriting this from it.  This is something I think Chandler or David mentioned a few months ago.  If there’s been any progress in the C++ committee on that then i’m happy to try implement something better.  If not, then i don’t think what I have here should be difficult to change in future.
>>> > >
>>> > > I guess a generalized version would return a
>>> > > `std::tuple<Value *const &, BasicBlock *const &>` or some such.  Not
>>> > > sure how to actually make zip iterators work well without something
>>> > > like N1550 though.
>>> > >
>>> > > What's N1550 offer to make zip work?
>>> >
>>> > s/work/& well/
>>> >
>>> > I think bloated iterators are bad, but without bloating them (so you
>>> > can return a T&), you can't call a zip_iterator a ForwardIterator,
>>> > which means you can't use it in a bunch of algorithms that you might
>>> > want to (such as the destination for `std::copy()`).
>>> >
>>> > N1550 let's you correctly identify the type of traversal the
>>> > zip_iterator can do, without requiring a T& return from operator*().
>>> >
>>> > > I wouldn't mind a slightly half-hearted version that works for basic/common cases...
>>> 
>>> 
>>> 
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>> 
>> 
>> 
> 
> 





More information about the llvm-commits mailing list