[llvm-commits] CVS: llvm/lib/CodeGen/LiveIntervalAnalysis.cpp RegAllocLinearScan.cpp VirtRegMap.cpp VirtRegMap.h

[Evan Cheng]

On Mar 30, 2007, at 3:09 PM, Chris Lattner wrote:

>
> On Mar 30, 2007, at 3:03 PM, Evan Cheng wrote:
>
>>>>              unsigned reg = rep(mop.getReg());
>>>>              mii->getOperand(i).setReg(reg);
>>>>
>>>> +            // If the definition instruction is re- 
>>>> materializable, its spill
>>>> +            // weight is zero.
>>>>              LiveInterval &RegInt = getInterval(reg);
>>>> -            RegInt.weight +=
>>>> -              (mop.isUse() + mop.isDef()) * pow(10.0F, (int) 
>>>> loopDepth);
>>>> +            if (!RegInt.remat) {
>>>> +              RegInt.weight +=
>>>> +                (mop.isUse() + mop.isDef()) * pow(10.0F, (int) 
>>>> loopDepth);
>>>> +            }
>>>
>>> Should it really be zero?  It should definitely be lower that a  
>>> normal value, but zero seems low.  The loop depth should still be  
>>> taken into consideration.
>>
>> Funny you should mention this. I am re-evaluating this right now.  
>> My thinking is loads (load from constantpool or stack slots) would  
>> have the normal weight. The rest will have half the weight.
>
> Why not just make everything half weight?  What would be lost?

To avoid a common problem. A value live-in to a function and is  
loaded from stack slot and used in a loop. You do not want to  
rematerialize this and end up issuing the load in the loop.

This is tricky to get right. I am not sure simple heuristics can  
cover all the cases.

Evan

>
> -Chris