[PATCH] D38315: [ARM] Armv8.2-A FP16 code generation (part 1/2)

[Sjoerd Meijer via Phabricator via llvm-commits]

SjoerdMeijer added a comment.

I've looked into it, and yes you're right, default legalisation for bitcasts in the SOFTFP-FULLFP16 is happening
resulting slightly worse codegen than SOFTFP-FP16. 
In my previous approach and implementation, I had to custom lower Bitcasts (for a different reason though). 
In a quick experiment, I copy-pasted the code, and that does what we want. However, I need to make a few
tweaks to it. For reviewing and testing purposes, i.e. not to change too many things at the same time, I am 
suggesting this approach:

- This groundwork is the 1st of 3 patches, let's see how this passes testing first.
- Patch 2/3: improve lowering for Bitcasts. FP16 codegen is unaffected by patch 1/3, and the the custom lowering addresses an inefficiency and not a correctness issue for the new FULLFP16 codegen cases.
- Patch 3/3: fill in the remaining FP16 match rules.

Does that sound ok?


https://reviews.llvm.org/D38315