Another problem with "Recommit r265547, and r265610,r265639,r265657"

[Mikael Holmén via llvm-commits]

Hi Wei,

On 05/17/2016 11:16 PM, Quentin Colombet wrote:
>
>> On May 12, 2016, at 5:40 AM, Mikael Holmén <mikael.holmen at ericsson.com> wrote:
>>
>> Hi Wei and Quentin,
>>
>> On 05/10/2016 07:55 PM, Wei Mi wrote:
>> [...]
>>> Mikael, thanks for getting the dump. The problem is that in Phoenix
>>> Post Coalescer, the following instruction becomes dead and it is
>>> marked as dead, but it is never deleted until regsplit in regalloc
>>> finds it dead and calls eliminateDeadDef for it.
>>>
>>> %vreg51<def,dead,tied1> = addh_nimm5_a32_aNhDst_formatDstN32
>>> %vreg51<tied0>, 4, pred:0, pred:%noreg, pred:0, %CCReg<imp-def,dead>,
>>> %cuc<imp-use>; aNh_0_7:%vreg51
>>
>> Yes, this instruction is introduced by our software pipeline pass running post coalescer. In the epilogue produced by the software pipeliner there may be dead code.
>>
>>> Maybe regalloc should allow IR containing dead instruction as its
>>> input, since itself leaves some dead instructions in the IR for the
>>> sake of better rematerialization.  Quentin, what do you think? If that
>>> is the case, the patch D19486 maybe needed.
>
> As long as the IR is valid then, yes, we should add the required bits to support it.

Did you get anywhere with this?

As I think I said in some earlier mail, we've worked around the last 
problem that we're aware of with

+  // To avoid situations where Def is tied and the same Def/Use 
registers gets
+  //  new reg ids and we end up with a use of an undefined register
+  if (MI->isRegTiedToUseOperand(0)) {
+    DEBUG(dbgs() << "Can't delete, def is tied to operand: " << Idx << 
'\t' << *MI);
+    return;
+  }
+

in LiveRangeEdit::eliminateDeadDef but it sounded like you had some 
better solution, D19486?

Thanks,
Mikael

>
> Cheers,
> Q.
>
>>
>> Sounds good to me but I don't know if that breaks any general assumption.
>>
>> Quentin?
>
>>
>> Thanks,
>> Mikael
>>
>>>
>>> Wei.
>>>
>
>