[llvm-commits] [llvm] r69631 - /llvm/trunk/lib/Transforms/Scalar/LoopUnroll.cpp

[Dale Johannesen]

On Apr 20, 2009, at 9:00 PMPDT, Chris Lattner wrote:

> On Apr 20, 2009, at 3:32 PM, Evan Cheng wrote:
>>> Well... That's not entirely true. Each GEP can also become a several
>>> instructions. Is there a test case to demonstrate the merit of this
>>> change?
>>>
>>
>> How about changing it to what InlineCost.cpp does?
>
> Better yet, change both to use if GEP->hasAllConstantIndices()  :)

I explained why that won't work.   What we actually need to do is  
figure out which GEP's are going to become instructions and which  
aren't, which is of course target-dependent.  I'm hoping to avoid that:)