[llvm-commits] CVS: llvm/lib/Transforms/Scalar/InstructionCombining.cpp

[Chris Lattner]

On Mar 19, 2007, at 2:21 PM, Reid Spencer wrote:
> On Mon, 2007-03-19 at 14:16 -0700, Chris Lattner wrote:
>>> Implement isOneBitSet in terms of APInt::countPopulation.
>>
>>> @@ -3474,8 +3474,7 @@
>>>  // isOneBitSet - Return true if there is exactly one bit set in
>>> the specified
>>>  // constant.
>>>  static bool isOneBitSet(const ConstantInt *CI) {
>>> -  uint64_t V = CI->getZExtValue();
>>> -  return V && (V & (V-1)) == 0;
>>> +  return CI->getValue().countPopulation() == 1;
>>>  }
>>
>> Are you sure this is a good idea?  countPopulation is *much* slower
>> than a couple of and's and a subtract.
>
> Its the temporary construction of APInts that makes the performance of
> the existing algorithm poor. This will construct 3 temporaries, each
> potentially with a malloc. Using countPopulation is constant time (I
> agree, not super fast, but consistent) and much easier to read in the
> code.

Optimizing for the "big" apint case isn't interesting, please  
optimize for the small case.

-Chris