[llvm] r204435 - [Constant Hoisting] Make the constant materialization cost operand dependent

[Eric Christopher]

>> I remember you fixed a fallthrough in a followup patch but I can't
>> recall if you fixed up this one or the one here:
>
> This was an intended fall-though, but that got replaced in r204739.

OK.

>
>>
>>>   case Intrinsic::experimental_stackmap:
>>> +    if (Idx < 2)
>>> +      return TCC_Free;
>>>   case Intrinsic::experimental_patchpoint_void:
>>
>> I think it was the first and if so, did you mean the fallthrough here?
>
> This was the think-o that got fixed in r204738. It wasn't meant to fall-through.

Thanks.

>>
>> Why did the actual testcase change here?
>
> The test started to fail with the operand aware cost model. Constant hoisting would start hoisting the first operand of the icmp and
> the CHECKs wouldn't match the new output. I updated the CHECKs and manually constant folded some instructions away.
>

Interesting. Is it because you were optimizing more so that the
testcase was no longer interesting? What wasn't matching out of
curiosity?

-eric