[llvm] [AMDGPU] op_sel[0] must be zero for v_cvt_sr_fp8/bf8_f16 (PR #227413)

[Changpeng Fang via llvm-commits]

================
@@ -371,7 +371,7 @@ define amdgpu_ps void @test_cvt_sr_bf8_f16_byte0(half %a, i32 %sr, i32 %old, ptr
 ; GFX1250-SDAG-REAL16-NEXT:    v_nop
 ; GFX1250-SDAG-REAL16-NEXT:    global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
 ; GFX1250-SDAG-REAL16-NEXT:    v_dual_mov_b32 v5, v4 :: v_dual_mov_b32 v4, v3
-; GFX1250-SDAG-REAL16-NEXT:    v_cvt_sr_bf8_f16 v2, v0.l, v1
+; GFX1250-SDAG-REAL16-NEXT:    v_cvt_sr_bf8_f16 v2, v0, v1
----------------
changpeng wrote:

> Looks like regression. It still needs .l suffix.

What if hi16 is dead (or not defined) for v0?  I think there was a mov that is optimized away based on the pattern

https://github.com/llvm/llvm-project/pull/227413