[PATCH] D123911: [VectorCombine] Fold shuffle select pattern

[Sam Tebbs via Phabricator via llvm-commits]

samtebbs accepted this revision.
samtebbs added a comment.
This revision is now accepted and ready to land.

LGTM



================
Comment at: llvm/lib/Transforms/Vectorize/VectorCombine.cpp:1300-1302
+  // For each of the output shuffles, we try to sort the first vector elements
+  // to the beginning, and the second array elements to the end to allow us to
+  // only use half of each of the binops. We compute the ReconstructMask mask
----------------
dmgreen wrote:
> samtebbs wrote:
> > I'm not exactly sure what is meant by "we try to sort the first vector elements to the beginning, and the second array elements to the end". Does it mean sorting e.g. `shuffle <9, 4, 11, 12, 3>` to `shuffle<3, 4, 9, 11, 12>`?
> > 
> > How does that then allow us to only use half of the binops? The number of binary operations in the output seems to remain the same. This is just for my own understanding, not because I think it's wrong.
> Yeah I don't feel like that was written very well. I've tried to update it a little.
> 
> The idea is to take a shuffle of the form `shuffle A, B <0, 8, 2, 3, 12, 13, 6, 15>` and turn that into a shuffle that only uses the first 4 lanes (0,1,2,3) from A and the first 4 lanes from B (8,9,10,11). We need to recreate the original, so we create a reconstruction mask of `shuffle A', B' <0, 8, 1, 2, 9, 10, 3, 11>`. The shuffles into A and B are altered to keep the lanes valid, and the whole thing is costed to make sure the total cost of the new shuffles is lower than the originals.
> If A and B are <8 x i32>, for example, then we only need the first <4 x i32> from each, cutting the number of operations down from 2 to 1 for each of the binops. Depending on the cost of the shuffles, this can be better overall.
That clears it all up, thanks :)


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D123911/new/

https://reviews.llvm.org/D123911