[llvm] [AMDGPU][GlobalISel] Legalize odd bf16 vectors (PR #226707)

[Kazu Hirata via llvm-commits]

================
@@ -131,6 +131,57 @@ define amdgpu_ps <2 x bfloat> @fmax_v2bf16_vl(<2 x bfloat> %a) {
   ret <2 x bfloat> %result
 }
 
+define amdgpu_ps <3 x bfloat> @fmin_v3bf16_vv(<3 x bfloat> %a, <3 x bfloat> %b) {
+; GFX9-LABEL: fmin_v3bf16_vv:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_pk_min_num_bf16 v0, v0, v2
+; GFX9-NEXT:    v_pk_min_num_bf16 v1, v1, v3
+; GFX9-NEXT:    ; return to shader part epilog
+;
+; GFX12-LABEL: fmin_v3bf16_vv:
+; GFX12:       ; %bb.0:
+; GFX12-NEXT:    v_pk_min_num_bf16 v0, v0, v2
+; GFX12-NEXT:    v_pk_min_num_bf16 v1, v1, v3
+; GFX12-NEXT:    ; return to shader part epilog
----------------
kazutakahirata wrote:

Did you mean to add corresponding `RUN: ... | FileCheck` lines at the beginning of this file?  Or should these lines be removed since this file only tests `GFX1250`?

https://github.com/llvm/llvm-project/pull/226707