[PATCH] D114087: [SLP]Improve registering and merging of compatible shuffles.
Simon Pilgrim via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Tue Nov 30 08:30:52 PST 2021
RKSimon added inline comments.
================
Comment at: llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6807
+ return false;
+ // Check if the second instruction is more defined than the first one.
+ NewMask.assign(SI2->getShuffleMask().begin(), SI2->getShuffleMask().end());
----------------
Add a comment about keep track of the terminating (trailing?) undefmaskelems
================
Comment at: llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6815
+ else
+ LastUndefsCnt = 0;
+ if (NewMask[I] != UndefMaskElem && SM1[I] != UndefMaskElem &&
----------------
ABataev wrote:
> RKSimon wrote:
> > I don't understand - why do you reset LastUndefsCnt here?
> `LastUndefsCnt` is the number of terminating undefmaskelems, that's why I reset it each time we ran into non-undef mask elem.
Got it - thanks.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D114087/new/
https://reviews.llvm.org/D114087
More information about the llvm-commits
mailing list