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

[Janek van Oirschot via llvm-commits]

================
@@ -16,10 +16,12 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
 ; CHECK-NEXT:    s_bitcmp1_b32 s0, 8
 ; CHECK-NEXT:    s_cselect_b64 s[2:3], -1, 0
 ; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[2:3]
-; CHECK-NEXT:    s_xor_b64 s[20:21], s[2:3], -1
 ; CHECK-NEXT:    v_cmp_ne_u32_e64 s[0:1], 1, v0
-; CHECK-NEXT:    s_and_b64 s[2:3], exec, s[2:3]
 ; CHECK-NEXT:    v_mov_b32_e32 v0, 0x9037ab78
+; CHECK-NEXT:    v_accvgpr_write_b32 a3, v1
+; CHECK-NEXT:    s_xor_b64 s[20:21], s[2:3], -1
+; CHECK-NEXT:    s_and_b64 s[2:3], exec, s[2:3]
+; CHECK-NEXT:    v_accvgpr_write_b32 a2, v0
----------------
JanekvO wrote:

Seems to be still present so I'll continue looking into this

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