[PATCH] D101769: [LegalizeIntegerTypes] Promote EXTEND_VECTOR_INREG operands

[Fraser Cormack via Phabricator via llvm-commits]

frasercrmck added a comment.

In D101769#2733902 <https://reviews.llvm.org/D101769#2733902>, @craig.topper wrote:

> This does feel like an imbalance in the legal RISCV vector types. Is it really worth adding another legalization case rather than just adding more simple vector types or making v256i32 illegal?

Yeah it's definitely an imbalance. I've been thinking that we need to round out the legal RVV vector types to an agreed-upon max vector length, like 512 or 1024 bytes.

But then again this is a crash in LLVM where there shouldn't be one, so I was kind of assuming that we'd do both. The downside, though, is that we might not have enough test cases to prevent regressions in this legalizer case.

If I had to choose one, I'd add more MVTs up to 512B/1024B depending on what people generally see as a useful maximum. I'm willing to go with the majority.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D101769/new/

https://reviews.llvm.org/D101769