You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2022/05/02 15:10:31 UTC

[GitHub] [tvm] Lunderberg commented on pull request #11181: [LLVM] Fix a possible tbaa issue

Lunderberg commented on PR #11181:
URL: https://github.com/apache/tvm/pull/11181#issuecomment-1115016980

   > After a re look-through, IIUC, I think the "index" for alias info should keep the same element unit for all accesses, so we may not change it either the last_index is corrected or not.
   
   That makes sense to me, so long as all access uses the vectorized data type or the scalar data type, but not both.
   
   > However, before any changes, I find another concern on aliased buffers. since in the new convention, two buffer object (maybe of different datatype) with the same buffer var alias to each other, it may not be safe to use current buffer's dtype as index unit for alias info.
   
   Good point, and thank you for the examples.  I agree with the conclusion that this can impact any aliased buffer that has a different element type, either differing by number of lanes or scalar datatype.  There's a few cases I can think of where that could occur in practice, such as `StorageRewrite` merging multiple buffers of different types into a single allocation.
   
   What are the restrictions on the element type presented to the tbaa annotations?  If we write the alias information using a byte-based indexing, I think that would solve both the vector size and the dtype size issues.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org