[PATCH] D107541: [AArch64] Cost-model vector concatenation

[Dave Green via Phabricator via llvm-commits]

dmgreen added a comment.

In D107541#2930890 <https://reviews.llvm.org/D107541#2930890>, @SjoerdMeijer wrote:

> In D107541#2929005 <https://reviews.llvm.org/D107541#2929005>, @dmgreen wrote:
>
>> Can we also add some more tests, for things that are insert_subvector's but not just the simple concat's?
>
> Yeah, I have tried, but unsuccessfully. The shuffle cost for a `SK_InsertSubvector` is guarded by `isInsertSubvectorMask()`:
>
>   if (Shuffle->isInsertSubvectorMask(NumSubElts, SubIndex))
>     return TargetTTI->getShuffleCost(
>       TTI::SK_InsertSubvector, VecTy, Shuffle->getShuffleMask(),
>       SubIndex,
>       FixedVectorType::get(VecTy->getScalarType(), NumSubElts));
>
> And this `isInsertSubvectorMask()` seems to only return true for subverter masks that are identity masks, so I haven't been able to trigger this with other masks.

This would count too `%X = shufflevector <8 x i16> undef, <8 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 8, i32 9>`. It has an identity for the first vector, with the second inserted into it.

I think the condition can be more specific to a concat if those are the cases that are really cheap. Two 64bit vectors combined together into a 128bit vector.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107541/new/

https://reviews.llvm.org/D107541