[PATCH] D100663: [LV] Add undef incoming value to loop-exit phis for the middle-block.

[Philip Reames via llvm-commits]

On 4/20/21 2:10 PM, Florian Hahn wrote:
>
>> On Apr 20, 2021, at 00:28, Philip Reames <listmail at philipreames.com> wrote:
>>
>>
>> On 4/19/21 3:20 PM, Florian Hahn wrote:
>>>> On 19 Apr 2021, at 17:57, Philip Reames via Phabricator <reviews at reviews.llvm.org> wrote:
>>>> I think what we need here is an actual "unknown until later" value.  I'd be tempted to add an intrinsic for that, except I'm not quite sure if we need such a thing in a constant expression anywhere.   This was one of the things I'd been hoping to chat through with you offline when you had a moment.  :)
>>> I think it would be great to discuss this offline.
>>> For me either Wednesday or Friday between 10am and 1pm California time would be ideal or Thursday before 10 am California time. Would any of those times work for you?
>> Let's do 10am Pacific on Wednesday.  I think you have my number right?  Or would you rather I send you a zoom invite?
>
> Sounds good to me! It would be great if you could send me a link.

Brief summary of our call.

We discussed ideas for an @llvm.unknown() intrinsic for use inside 
passes to allow queries over partially constructed IR. With existing 
attributes and metadata, we can restrict the unknown semantics 
appropriately.

We both thought this was potentially an interesting direction to 
explore, but that it felt like possible overkill for the current 
vectorizer problem.  We're going to move forward with one of the two 
workarounds discussed so far, and return the general topic when/if we 
have a second use case.

Philip