[PATCH] Add iterator_adaptor to iterate over SDNode operand SDNode's

[Pete Cooper]

> On Jun 25, 2015, at 5:49 PM, David Blaikie <dblaikie at gmail.com> wrote:
> 
> 
> 
> On Thu, Jun 25, 2015 at 5:36 PM, Pete Cooper <peter_cooper at apple.com <mailto:peter_cooper at apple.com>> wrote:
> 
>> On Jun 25, 2015, at 5:21 PM, David Blaikie <dblaikie at gmail.com <mailto:dblaikie at gmail.com>> wrote:
>> 
>> 
>> 
>> On Thu, Jun 25, 2015 at 5:11 PM, Pete Cooper <peter_cooper at apple.com <mailto:peter_cooper at apple.com>> wrote:
>> 
>>> On Jun 25, 2015, at 2:40 PM, David Blaikie <dblaikie at gmail.com <mailto:dblaikie at gmail.com>> wrote:
>>> 
>>> 
>>> 
>>> On Thu, Jun 25, 2015 at 2:30 PM, Pete Cooper <peter_cooper at apple.com <mailto:peter_cooper at apple.com>> wrote:
>>> Hi David
>>> 
>>> Currently the only SDNode iterator over operands does so with SDUse*.  Users frequently then call getNode() on the operand.
>>> 
>>> This patch adds an iterator to SDNode which returns the SDNode of the operand.  This allows more patterns to be converted to foreach.  It is based on value_op_iterator which I found in User.h.
>>> 
>>> For now i’ve only used it in a single place, but I found a bunch more in DAGCombiner for example which should be applicable.  I would convert those in a later commit assuming you are ok with this solution.
>>> 
>>> This immediately raises red flags:
>>> 
>>>   SDNode *operator*() const { return I->getNode(); }
>>>   SDNode *operator->() const { return operator*(); }
>>> 
>>> op* should return a T& and op-> should return T*
>> I’d forgotten about that.
>>> 
>>> If these SDNode*s can never be null, then perhaps this should be:
>> I wasn’t actually sure if they could be.  My initial reaction was that null operands wouldn’t make sense, but it turns out we never checked.  So here’s a patch which does actually ensure that the SDNode's referenced as operands are never null.  It passes make check.  I can put it on another email for review if you prefer I don’t add it here.
>> 
>> It might be worth a separate thread, or at least a drive-by by someone who deals with this part of the code. I don't really understand the necessity/merits/drawbacks of the 'SDUse::reset()" member function you've introduced.
> No problem.  Thanks for taking a look.  I’ve just sent out an email to llvm-commits and asked Hal for review as I know he’s done lots of SD work.
>>  
>> 
>> 
>>> 
>>>   SDNode &operator*
>>>   SDNode *operator->
>>> 
>>> ? (because I assume you don't have an SDNode* lvalue to return a reference to) I assume the adapter helper can implement one of those in terms of the other so you only have to implement one of them? I forget how the adapter utility works.
>> I think it makes sense to do this.  This will unfortunately be one of the few SDNode & in the entire codebase though, which makes it stand out.  SDNode really does seem to always be a pointer.  I’ll fix up the patch to do this soon.
>> 
>> Yeah, there are a few types (read: Lots) like this in LLVM. I personally don't mind being more referential in spite of that, but I can understand others might feel less comfortable with that.
>> 
>> If you wanted to preserve the pointer-ness, you'd have op* return SDNode * (this would be a bit incorrect, it should really be SDNode * const &, and you can do that by having an SDNode* member in the iterator that you init and return a ref to... technically that's the more correct option - I'm not entirely sure where that matters) and no op->.
> I like this solution.  I tried SDNode * const & earlier but of course getNode() is a temporary so this doesn’t work.  I’ll get a patch together which makes this change.  Technically I guess that means we don’t need the nonnull SDNode patch, but I don’t see any harm in it anyway.
> 
> Personally I'd go the reference route - the more references the clearer the documentation of nullness contracts (& tools like ubsan null reference checking can catch bugs sooner/closer to the point of the problem rather than later on). But I realize that's a bit jarring, stylistically, so I certainly wouldn't insist on it.
I tried making the op* return an SDNode& then propagated that through to the dyn_cast in AArch64ISelLowering.cpp.  Unfortunately after that things got messy.

I could have done this

for (const SDNode &Elt : N->op_values()) {
    if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(&Elt)) {

but it seemed a shame to have to add the & in the dyn_cast.  So I tried to actually make a version of dyn_cast on references to non-pointers, and return an optional if the cast fails.  Unfortunately that wasn’t much better.

But it did lead me to an interesting piece of code in Casting.h.  We have a struct called simplify_type which can be used to automatically do things like convert to other classes.  This is implemented to go from SDValue to SDNode* which is really what I want.

So, i’ve totally changed the iterator now to return SDValue& and SDValue*.  Thats really what the SDUse’s used as operands were storing anyway.  Then the dyn_cast can be left unchanged.

What do you think?

Cheers,
Pete


> 
> Mightn't hurt to think a bit about the tradeoff of using temporary storage or just cheating and returning a pointer by value from op*... I mean it'll probably work for the range-for. I'm not sure where it'd really break - if the value_type of the iterator is "const SDNode" then it might be hard-to-impossible to for any algorithm to really have a problem with this.
> 
>  
> 
> Thanks,
> Pete
>>>  
>>> 
>>> BTW, been trying to work out if there would ever be a good solution for an iterator combined with isa<> or dyn_cast<>.  If you look at the code this patch touches in AArch64ISelLowering, it is immediately followed by a dyn_cast.  I’d really like to find a clean way to fold that it to the foreach loop, i.e.,
>>> 
>>> for (auto *C : dyn_cast<ConstantSDNode>(N->op_nodes()))
>>> 
>>> just a thought, but thats unrelated to this patch for now.
>>> 
>>> Yep, though probably more in the form of a filtered range, I suspect:
>>> 
>>>   for (auto &C : filtered_transform(N->op_nodes(), [](SDUse *U) { return U->getNode(); }))
>>> 
>>> It'd be a bit tricky to deal with the value type of this range's iterators - chances are the predicate should return an Optional<T&> (Hmm, don't think our Optional template supports ref parameters yet anyway) or T* (not sure if we could generalize it so it could cope with Optional<T>, maybe - so we could support generators where the values are not already/permanently in-memory) and then the value_type is T.
>> Interesting.  I hadn’t though to use Optional.  I might try to implement something like this if i get time.
>> 
>> Cheers,
>> Pete
>>>  
>>> 
>>> Cheers,
>>> Pete

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150626/da123299/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: value_iterator.patch
Type: application/octet-stream
Size: 2073 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150626/da123299/attachment.obj>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150626/da123299/attachment-0001.html>