davemgreen wrote:

> Should not introduce new target intrinsics for strictfp. Target intrinsics do 
> not work with strictfp at all today. These have no chains and will be freely 
> reordered. It's also not reasonable to duplicate every floating point math 
> intrinsic, which is the point of the proposed change to start using operand 
> bundles for strictfp

For MVE the list is not that long, and we need some way of preventing the 
intrinsics from scalarizing into multiple ops. So we either convert fadd into 
vadd, which does not obey the rounding modes so is not correct (not to mention 
denormals - those have always been bad in llvm). So we introduce intrinsics and 
at least not have them scalarize. I believe we will need them independent of 
what happens with operand bundles, and are stuck not being able to enable the 
feature in the short term if not.

https://github.com/llvm/llvm-project/pull/169795
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to