[llvm-commits] [llvm] r45815 - /llvm/trunk/lib/CodeGen/LiveIntervalAnalysis.cpp

[Evan Cheng]

On Jan 9, 2008, at 8:44 PM, Owen Anderson wrote:

> You really need to address these to Evan.  I asked him what to do  
> for LiveIntervalAnalysis, and this is what he told me :-)
>
> --Owen
>
> On Jan 9, 2008, at 10:08 PM, Chris Lattner wrote:
>
>>> URL: http://llvm.org/viewvc/llvm-project?rev=45815&view=rev
>>> Log:
>>> Don't use LiveVariables::VarInfo::DefInst.
>>
>> Cool, question though:
>>
>>> +++ llvm/trunk/lib/CodeGen/LiveIntervalAnalysis.cpp Wed Jan  9
>>> 21:12:54 2008
>>> @@ -309,7 +309,14 @@
>>>      // are actually two values in the live interval.  Because of
>>> this we
>>>      // need to take the LiveRegion that defines this register and
>>> split it
>>>      // into two values.
>>> -      unsigned DefIndex =
>>> getDefIndex(getInstructionIndex(vi.DefInst));
>>
>> In the old code, when it introduces multiple definitions of a vreg,
>> did it clear DefInst?  If so, this code can only be triggered for
>> vregs with a single def.

It does change DefInst. This code is triggered when a two-address  
instruction redefines a vreg so it can only be triggered for vregs  
with multiple def's.
>>
>>>
>>> +      MachineRegisterInfo& MRI = mbb->getParent()->getRegInfo();
>>
>> I think that MRI should become and ivar in liveintervalanalysis.

Yes, there is an ivar mri_.

>>
>>>
>>> +      unsigned lowIndex = ~0U;
>>> +      for (MachineRegisterInfo::def_iterator DI =
>>> MRI.def_begin(interval.reg),
>>> +          DE = MRI.def_end(); DI != DE; ++DI)
>>> +        if (getInstructionIndex(&*DI) < lowIndex)
>>> +          lowIndex = getInstructionIndex(&*DI);
>>> +
>>> +      unsigned DefIndex = getDefIndex(lowIndex);
>>
>> It would be really nice if you could drop the loop.  It's unclear why
>> it's right anyway, the numbering is not necessarily contiguous across
>> code...

Yes, I think it' can probably be simplified. But it's probably  
correct. This is before coalescing:
v1024 = v1025
v1024 = v1024 op ?

So there must be two def's of v1024 in the same BB and we want the  
first one.

Evan

>>
>> -Chris
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits