[PATCH] D22547: [LSV] Don't move stores across may-load instrs, and loosen restrictions on moving loads.

[Alina Sbirlea via llvm-commits]

asbirlea added inline comments.

================
Comment at: test/Transforms/LoadStoreVectorizer/NVPTX/merge-across-side-effects.ll:2
@@ -1,2 +1,3 @@
 ; RUN: opt -mtriple=nvptx64-nvidia-cuda -load-store-vectorizer -S -o - %s | FileCheck %s
 
+; Check that the load/store vectorizer is willing to move loads/stores across
----------------
jlebar wrote:
> asbirlea wrote:
> > Could you double check the validation for these tests is done in getVectorizablePrefix, and there isn't an earlier exit on the data layout (just before the calls to getVectorizablePrefix)? Possibly add a data layout to avoid that, in case the defaults change?
> > Could you double check the validation for these tests is done in getVectorizablePrefix
> 
> Sorry, not sure what you mean.
> 
> > there isn't an earlier exit on the data layout
> 
> I think we should be good here because if the pass doesn't modify any of the functions, the test will also fail.
Ok, you're right, on early exit it wouldn't vectorize any of these.


https://reviews.llvm.org/D22547