[PATCH] D113439: [RISCV] Add IR intrinsics for reading/write vxrm.

[Eli Friedman via Phabricator via llvm-commits]

efriedma added a comment.

In D113439#3183639 <https://reviews.llvm.org/D113439#3183639>, @craig.topper wrote:

> Maybe I should also mention that my immediate goal was to enable Halide to implement halving_add on RISCV vectors. They were going to use these intrinsics and existing RVV intrinsics that are marked as having side effects.

This is part of why we want the compiler to control the rounding/saturation state, not the user.  You don't want halving_add to have side-effects. :)

> I had wondered if it would make sense to have generic IR intrinsics for halving add. I could lower that to a vxrm change plus the vaadd instruction.

You mean target-independent?  This is orthogonal to the question of what the riscv intrinsics should look like.  But it would make sense; it's a common operation in vector instruction sets.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D113439/new/

https://reviews.llvm.org/D113439