[llvm-commits] [PATCH] Make Inits FoldingSetNodes

[Chris Lattner]

On Jul 14, 2011, at 2:47 PM, David A. Greene wrote:

> Chris Lattner <clattner at apple.com> writes:
> 
>> On Jul 13, 2011, at 3:33 PM, David Greene wrote:
>>    @@ -530,6 +629,23 @@ const Init *OpInit::resolveListElementReference(Record &R, const 
>>    +const UnOpInit *UnOpInit::get(UnaryOp opc, const Init *lhs, RecTy *Type) {
>>    +  FoldingSetNodeID ID;
>>    +  ID.AddInteger(initUnOp);
>>    +  ID.AddInteger(opc);
>>    +  ID.AddString(Type->getAsString());
>>    +  ID.AddPointer(lhs);
>> 
>> DenseMap makes sense here.
>> 
>>    +const BinOpInit *BinOpInit::get(BinaryOp opc, const Init *lhs,
>>    +                                   const Init *rhs, RecTy *Type) {
>> 
>> and here.
>> 
>>    +const TernOpInit *TernOpInit::get(TernaryOp opc, const Init *lhs,
>>    +                                  const Init *mhs, const Init *rhs,
>>    +                                  RecTy *Type) {
>> 
>> and here.
> 
> I am going through and uniquing Inits in their own unique ways.
> 
> Question on these three.  DenseMap of what?  There's an opcode, multiple
> operands and a type.  What's the Key, a struct of these items?  Is a
> DenseMap much more efficient than a FoldingSetNode when string
> comparison/hash is involved?  Wouldn't hashing a struct of these and
> putting them in a DenseMap basically be equivalent to a FoldingSet?

FoldingSet is only efficient for very large and heterogenous objects.  Its constant factor is really very very high, so densemap is almost always a win for something with a few elements.

-Chris