[llvm] [AMDGPU] Elide bitcast fold i64 imm to build_vector (PR #154115)

[Janek van Oirschot via llvm-commits]

================
@@ -93,8 +93,10 @@ define amdgpu_kernel void @sint_to_fp_i1_f64(ptr addrspace(1) %out, i32 %in) {
 ; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX942-NEXT:    s_cmp_eq_u32 s2, 0
 ; GFX942-NEXT:    s_cselect_b32 s2, 0xbff00000, 0
-; GFX942-NEXT:    v_mov_b32_e32 v1, s2
-; GFX942-NEXT:    global_store_dwordx2 v0, v[0:1], s[0:1]
+; GFX942-NEXT:    s_cselect_b32 s3, 0, 0
+; GFX942-NEXT:    v_mov_b32_e32 v2, s3
+; GFX942-NEXT:    v_mov_b32_e32 v3, s2
+; GFX942-NEXT:    global_store_dwordx2 v0, v[2:3], s[0:1]
----------------
JanekvO wrote:

Not sure of this myself, looking

https://github.com/llvm/llvm-project/pull/154115