asb wrote:

We discussed again in the sync-up call to try to move this forwards and per a 
discussion with @topperc and @preames were wondering if relying on -mattr and a 
corresponding feature for the profile might be the better approach after all.
* There's not a lot of instances of this, but I do see at least some ARM tests 
using -mattr=+armv8.2-a or +armv8.3-a. So there's some precedent there.
  * Looking again, x86 does indeed use -mcpu in llc for x86_64-v3 etc. So that 
could be an argument for following suit, but for some reason it still feels 
like defining new generic CPUs isn't very consistent with how we normally 
specific extensions in RISC-V. So -march for clang and -mattr for llc feels 
consistent to me.
* Philip raised the point that if we're making this change primarily to ease 
usage of llc, it's unfortunate if we're adding something that is exposed to 
clang that we wouldn't necessarily want otherwise (in this case, the generic 
cpu for the profile).

That said, it was a brief discussion and we may have missed something or Philip 
or Craig may have more thoughts afterwards.

Do you see any disadvantage to going back to adding a feature for the profiles?

https://github.com/llvm/llvm-project/pull/84877
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to