[PATCH] D20763: [DAG] Improve legalization of INSERT_SUBVECTOR to avoid spilling

[Simon Pilgrim via llvm-commits]

RKSimon added a comment.

> Unfortunately, aside from the SAD test, this doesn't really seem to get hit, and I'm not sure how to construct a test-case for the high lanes. One alternative is to make this safer by only inserting directly when Idx is 0 (this tested, and probably the common case). Let me know if you think that would be better.


Only using this for ((IdxVal == 0) && (IdxVal + SubElems <= VecElems / 2)) is probably for the best (along with a suitable TODO comment). Alternatively investigate AVX512F tests that use (illegal) 64i8 / 32i6 types to try and trigger the insertion?


http://reviews.llvm.org/D20763