[llvm-commits] [llvm] r137250 - in /llvm/trunk: lib/CodeGen/LiveDebugVariables.cpp test/CodeGen/ARM/debug-info-sreg2.ll test/CodeGen/X86/dbg-value-isel.ll

[Jakob Stoklund Olesen]

On Aug 10, 2011, at 2:50 PM, Devang Patel wrote:

> 
> On Aug 10, 2011, at 2:47 PM, Jakob Stoklund Olesen wrote:
> 
>> 
>> On Aug 10, 2011, at 2:42 PM, Devang Patel wrote:
>> 
>>> 
>>> On Aug 10, 2011, at 2:33 PM, Jakob Stoklund Olesen wrote:
>>> 
>>>> 
>>>> On Aug 10, 2011, at 2:25 PM, Devang Patel wrote:
>>>> 
>>>>> void UserValue::extendDef(SlotIndex Idx, unsigned LocNo,
>>>>>                       LiveInterval *LI, const VNInfo *VNI,
>>>>>                       SmallVectorImpl<SlotIndex> *Kills,
>>>>> -                          LiveIntervals &LIS, MachineDominatorTree &MDT) {
>>>>> +                          LiveIntervals &LIS, MachineDominatorTree &MDT,
>>>>> +                          LexicalScopes &LS) {
>>>>> SmallVector<SlotIndex, 16> Todo;
>>>>> Todo.push_back(Idx);
>>>>> -
>>>>> +  SmallPtrSet<const MachineBasicBlock *, 4> LBlocks;
>>>>> +  LS.getMachineBasicBlocks(dl, LBlocks);
>>>> 
>>>> Hi Devang,
>>>> 
>>>> I think the DebugLoc in a UserValue is bogus and should be removed.
>>>> 
>>>> Shouldn't the lexical scope be derived from the MDNode instead?
>>> 
>>> DebugLoc has inline info which is required to properly determine lexical scopes in case of inlining. DebugLoc is more compressed form, and it provides utilities to convert to/from MDNode.
>> 
>> But UserValue's DebugLoc comes from the first DBG_VALUE instruction to mention that MDNode. That's not very robust.
> 
>> I thought we were going to stop setting DebugLocs on DBG_VALUE instructions? You moved the inline location into the MDNode, right?
> 
> oh.. that MDNode, in other words DIVariable,

Right ;-)

> Yes, that's more robust now. I'll tackle it next.

Thanks. I want to eliminate UserValue::dl entirely.

/jakob