[llvm] [X86] Support lowering of FMINIMUMNUM/FMAXIMUMNUM (PR #121464)

[Matt Arsenault via llvm-commits]

================
@@ -1751,26 +1751,42 @@ define <2 x half> @v_maximumnum_v2f16(<2 x half> %x, <2 x half> %y) {
 ; GFX9-LABEL: v_maximumnum_v2f16:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    v_pk_max_f16 v1, v1, v1
-; GFX9-NEXT:    v_pk_max_f16 v0, v0, v0
-; GFX9-NEXT:    v_pk_max_f16 v0, v0, v1
+; GFX9-NEXT:    v_max_f16_sdwa v2, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
----------------
arsenm wrote:

This is worse, the expansion is now fully scalarized. Previously it decomposed into the legal vector ops 

https://github.com/llvm/llvm-project/pull/121464