[llvm] [AMDGPU] Add custom lowering of llvm.convert.to.arbitrary.fp for FP8 (PR #213813)

[Arseniy Obolenskiy via llvm-commits]

================
@@ -0,0 +1,284 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc < %s -mtriple=amdgpu9.50-amd-amdhsa | FileCheck -check-prefix=GFX950 %s
+; RUN: llc < %s -mtriple=amdgpu12.00-amd-amdhsa | FileCheck -check-prefix=GFX1200 %s
+; RUN: llc < %s -mtriple=amdgpu12.50-amd-amdhsa -mattr=+real-true16 | FileCheck -check-prefixes=GFX1250,GFX1250-TRUE16 %s
+; RUN: llc < %s -mtriple=amdgpu12.50-amd-amdhsa -mattr=-real-true16 | FileCheck -check-prefixes=GFX1250,GFX1250-FAKE16 %s
+; RUN: llc < %s -mtriple=amdgpu13.10-amd-amdhsa | FileCheck -check-prefix=GFX1310 %s
+
+; Element counts the vector legalizer neither splits nor widens evenly, so the
+; conversion is unrolled. Each lane still selects a single packed HW
+; instruction rather than a software expansion, so only the OCP subtargets are
+; run here.
+
----------------
aobolensk wrote:

Shall we tests conversion from fp16 as well? 

https://github.com/llvm/llvm-project/pull/213813