[llvm] [X86] Add missing vNbf16 handling in X86CallingConv.td file (PR #127102)

[Mikołaj Piróg via llvm-commits]

mikolaj-pirog wrote:

> please can you regenerate the breaking tests?



> > please can you regenerate the breaking tests?
> 
> Take a look at the tests, we cannot simply remove the handling in X86ISelLoweringCall. Unlike FP16, BF16 vectors are legal only when AVX512BF16/AVXNECONVERT are ready. Though we may make it legal for SSE2, we may just leave the code as is.

I have reverted the removal of the handling, it caused tests to fail (some crashing the compiler along the way). So the idea behind this piece of code is to pretend that the bf16 is fp16 for the call lowering purpose? `Unlike FP16, BF16 vectors are legal only when AVX512BF16/AVXNECONVERT are ready` I don't quite follow, what do you mean? The SSE2 tests work for bfloat x n types (fminimum-fmaximum.ll). 

Regarding the testing of this change, there already is a test for it, `bfloat-calling-conv-no-sse2.ll`. The problem with testing this, is that there is no way to reliably force execution of the `fast-isel` path; it can fail and fallback.

https://github.com/llvm/llvm-project/pull/127102