[llvm] AMDGPU: Improve codegen for intrinsic llvm.fptrunc.round (PR #104486)

[Changpeng Fang via llvm-commits]

================
@@ -3,6 +3,15 @@
 ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=CHECK,SDAG %s
 ; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1030 < %s | FileCheck -check-prefixes=CHECK,GISEL %s
 
+define amdgpu_gs half @v_fptrunc_round_f32_to_f16_tonearest(float %a) {
+; CHECK-LABEL: v_fptrunc_round_f32_to_f16_tonearest:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; CHECK-NEXT:    ; return to shader part epilog
+  %res = call half @llvm.fptrunc.round.f16.f32(float %a, metadata !"round.tonearest")
----------------
changpeng wrote:

At custom lowering the the intrinsic, we can turn it to fp_round, which can also generated the expected output for "round.tonearest". However, it seems the lowering of fp_round ignores the rounding mode operand in most cases (just assume default of round-to-nearest-even) 

https://github.com/llvm/llvm-project/pull/104486