[PATCH] D153547: [CostModel] Use min/max intrinsics for vecreduce.min/max costs
Dave Green via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Thu Jun 22 22:35:48 PDT 2023
dmgreen added a comment.
In D153547#4442566 <https://reviews.llvm.org/D153547#4442566>, @anna wrote:
> Not to hijack this patch, but I was experimenting a similar patch for Loop vectorizer when trip count of the loop is low (when we vectorize it, we get regressions since we do not consider out-of-loop reduction cost in LoopVectorizer). The plan is to add a TTIHook, I'll put the change up for review.
Sounds good. Low trip count loops are often unrolled prior to vectorization (or they are large so the overheads are quite small in comparison). The loop vectorizer hasn't in the past modelled many of the loop invariant overheads such as reductions and setup splats or constants. In here and D153548 <https://reviews.llvm.org/D153548> my main aim was just to get some basic AArch64 cost modelling for the intrinsics.
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D153547/new/
https://reviews.llvm.org/D153547
More information about the llvm-commits
mailing list