[PATCH] D135951: [X86][1/2] SUPPORT RAO-INT

[Phoebe Wang via Phabricator via llvm-commits]

pengfei added inline comments.


================
Comment at: llvm/lib/Target/X86/X86.td:259
+                                     "Support RAO-INT instructions",
+                                     [FeatureSSE2]>;
 def FeatureINVPCID : SubtargetFeature<"invpcid", "HasINVPCID", "true",
----------------
craig.topper wrote:
> pengfei wrote:
> > craig.topper wrote:
> > > Why do these require SSE2?
> > We need mfence instructions for strong orders. The mfence feature relies on SSE2.
> > I see your concern, we may need split these features from SSE2. Filed an issue https://github.com/llvm/llvm-project/issues/58388
> The lowering code for atomics could check SSE2 in addition to the RAO-INT feature. My primary concern was that the RAO-INT feature itself shouldn't require it.
> 
> Are there going to be intrinsics for RAO-INT?
I don't think there will be a target supports RAO-INT but no SSE2. All I can think out is for Kernel build that disable vector registers intentionally.

> Are there going to be intrinsics for RAO-INT?

This is a good question. In fact we wander whether providing intrinsics or not. The current preference is not. We don't want to add new intrinsics arbitrarily, there're burdens to maintenance and it's always easy to adding per to actual requests than removing afterwards.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135951/new/

https://reviews.llvm.org/D135951