[PATCH] D129013: [RISCV] Enable (scalable) vectorization by default

[Jessica Clarke via Phabricator via llvm-commits]

jrtc27 added a comment.

In D129013#3707411 <https://reviews.llvm.org/D129013#3707411>, @craig.topper wrote:

> In D129013#3707331 <https://reviews.llvm.org/D129013#3707331>, @jrtc27 wrote:
>
>> This is user error, plain and simple. If you say you have extension foo then the compiler should be free to use it. V isn't special; specifying +zba for example when your implementation lacks it has the exact same behaviour. If anything, leaving autovectorisation off would be more inconsistent.
>
> What if we had a -mcpu that included V? Would specifying that CPU when the kernel doesn't support it also be user error?

I don't see how it's any different from equivalent situations on other architectures... it's a bit confusing but can't avoid it. Would you say the same thing for -mcpu enabling F on an OS that doesn't support floating-point state and then users being sad their float-using code traps rather than uses soft-float routines? Or what about SVE on Arm? V isn't special, it's just the most likely situation you'll hit the issue at the moment.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D129013/new/

https://reviews.llvm.org/D129013