[llvm-commits] [llvm] r89639 - in /llvm/trunk: lib/Transforms/Scalar/InstructionCombining.cpp test/Transforms/InstCombine/compare-signs.ll

[Nick Lewycky]

Chris Lattner wrote:
>
> On Nov 23, 2009, at 12:33 AM, Duncan Sands wrote:
>
>>>>> + if (KnownZeroLHS.countLeadingOnes() == BitWidth-1&&
>>>>> + KnownZeroRHS.countLeadingOnes() == BitWidth-1) {
>>>>
>>>> == ->  >= :)
>>>
>>> Nope, look again!
>>>
>>> +        APInt TypeMask(APInt::getHighBitsSet(BitWidth, BitWidth-1));
>>>
>>> Thus, it will never return a knownzero with all bits set. :)
>>
>> Ha ha, you got me there!
>
> Please add a comment.  Obviously it isn't clear what is going on here.

Sorry Chris, but I don't think a comment would do anything but confuse 
people further. A reader who wasn't thinking of the earlier version of 
the patch that behaved differently wouldn't be asking this question.

Also, the issue Duncan raised is one of possible missed optimization, 
not correctness. It addressed handling the case where the comparison is 
with a constant, but those should be getting optimized away elsewhere 
anyways.

Nick

PS. Of course, if you were to tell me to stop arguing and just add a 
damned comment, I will.