[AMDGPU] Improve codegen for s_barrier_init (#111866)

Author: jayfoad

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -10031,9 +10031,7 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
         // If reference to barrier id is not an inline constant then it must be
         // referenced with M0[4:0]. Perform an OR with the member count to
         // include it in M0.
-        M0Val = SDValue(DAG.getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32,
-                                           Op.getOperand(2), M0Val),
-                        0);
+        M0Val = DAG.getNode(ISD::OR, DL, MVT::i32, Op.getOperand(2), M0Val);
       }
       Ops.push_back(copyToM0(DAG, Chain, DL, M0Val).getValue(0));
     } else if (IsInlinableBarID) {

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.barrier.wait.ll

@@ -737,11 +737,9 @@ define void @test5_s_barrier_init_m0(i32 %arg1 ,i32 %arg2) {
 ; GFX12-SDAG-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-SDAG-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-SDAG-NEXT:    s_wait_kmcnt 0x0
-; GFX12-SDAG-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX12-SDAG-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-SDAG-NEXT:    v_or_b32_e32 v0, v0, v1
 ; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX12-SDAG-NEXT:    s_mov_b32 m0, s0
 ; GFX12-SDAG-NEXT:    s_barrier_init m0
 ; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe