[lworld+fp16] RFR: 8338061: Add support for FP16 unary and ternary operations

[Jatin Bhateja]

On Tue, 20 Aug 2024 11:56:30 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

>> I have the conversion ops and isFinite/Infinite/NaN intrinsics still left to complete.
>
> I am in process of porting Float16 extension to VectorAPI into lworld+fp16, that will consolidate Float16 effort on lworld+fp16 branch and we can leverage the unified backend implementation. With VectorAPI we will also need to support Reductions for Float16 vectors :-)

> @jatin-bhateja Thanks for the effort. It will be good to test Float16 with vectorAPI and add the missing operations. I think we need to support Float16 reductions in auto-vectorization as well? I dont think those operations are added for FP16. Also about the design part, I was thinking instead of modifying the existing opcode or adding a separate secondary opcode for the operations, can we reuse the same existing FP32 nodes for FP16 as well? We can add flags for the FP32 IR nodes to check if it is a FP16 op or eventually FP8 op ? 

With new flag in existing IR we may end up adding complimentary predication checks in existing floating point instruction selection patterns.

> That way we can reuse the same Value(), Ideal(), identity() methods and can modify those functions based on the value these flags returns in case we do not want a particular optimization to be applied for FP16/FP8? 

Yes, idea is to reuse existing idealization routines, either by factoring out the meaty parts, or by inheritance. 

> In the backend, we can use these flags to emit the correct instruction with the correct SIMD arrangement ? I guess this might introduce a lot of checks (to check value of the flags) in the compiler code.. What are your thoughts about it?

I agree.
 
To me having a secondary opcode along with some code re-factoring around value / ideal / identity routines looks cleaner, but we can think over alternatives which fits best in existing infrastructure.

-------------

PR Review Comment: https://git.openjdk.org/valhalla/pull/1211#discussion_r1725446668