[cfe-commits] r75168 - in /cfe/trunk: include/clang/Basic/DiagnosticGroups.td include/clang/Basic/DiagnosticSemaKinds.td lib/Sema/SemaDeclCXX.cpp test/SemaCXX/constructor-initializer.cpp test/SemaCXX/warn-reorder-ctor-initialization.cpp

[Douglas Gregor]

On Jul 24, 2009, at 4:34 PM, Daniel Dunbar wrote:

> On Tue, Jul 21, 2009 at 1:32 PM, Fariborz  
> Jahanian<fjahanian at apple.com> wrote:
>> On Jul 21, 2009, at 1:16 PM, Daniel Dunbar wrote:
>>> On Tue, Jul 21, 2009 at 10:08 AM, Douglas  
>>> Gregor<dgregor at apple.com> wrote:
>>>> On Jul 9, 2009, at 1:00 PM, Fariborz Jahanian wrote:
>>>>>
>>>>> +    int Last = AllBaseOrMembers.size();
>>>>> +    int curIndex = 0;
>>>>> +    CXXBaseOrMemberInitializer *PrevMember = 0;
>>>>> +    for (unsigned i = 0; i < NumMemInits; i++) {
>>>>> +      CXXBaseOrMemberInitializer *Member =
>>>>> +        static_cast<CXXBaseOrMemberInitializer*>(MemInits[i]);
>>>>> +      void *MemberInCtorList;
>>>>> +      if (Member->isBaseInitializer())
>>>>> +        MemberInCtorList = Member->getBaseClass();
>>>>> +      else
>>>>> +        MemberInCtorList = Member->getMember();
>>>>> +
>>>>> +      int j;
>>>>> +      for (j = curIndex; j < Last; j++)
>>>>> +        if (MemberInCtorList == AllBaseOrMembers[j])
>>>>> +          break;
>>>>
>>>> It's too bad that this is O(N^2), but I can't think of a way to  
>>>> make
>>>> it faster without having the ability to sort the initializer list  
>>>> by
>>>> its initialization-position. Besides, the initializer list will
>>>> generally be short.
>>>
>>> Am I missing something, why not just create a DenseMap from
>>> AllBaseOrMembers[j] -> j? (Assuming the contents of that array are
>>> distinct).
>>>
>>> (A SmallDenseMap ideally, if we happened to have that ADT)
>>
>> I think Doug was making the point that a very small list of  
>> initializers may
>> not justify
>> changing the algorithm to make it faster. I don't think setting up a
>> DenseMap and populating it
>> justifies finding a uniqueness of 4 or 5 elements.
>
> Well, I can't speak for what point Doug was making, but personally
> given a choice between fast and O(N^2) and slightly slower and O(N), I
> think the correct choice in a compiler is always the latter.

I was pretty unclear :(

I, too, favor the O(N) solution even if it's slightly slower for small  
N. Otherwise, some idiot [*] is going to come up with a template  
metaprogram somewhere that exposes the O(N^2) behavior, and we'll get  
flogged for it.

> In addition, all we need is a SmallDenseMap ADT and the latter should
> be categorically better than the former, and probably less code.


A SmallDenseMap ADT would be great. Even a SmallPtrMap to complement  
SmallPtrSet would be extremely useful.

   - Doug

[*] Naturally, I have such a template metaprogram. My preferred  
implementation of C++0x's tuple has N direct base classes, where N is  
the length of the tuple. Since tuple is a common metaprogramming  
primitive, that N can get quite large.