diff --git a/src/broadcom/compiler/qpu_schedule.c b/src/broadcom/compiler/qpu_schedule.c
index af864b86e64..3a818d64d96 100644
--- a/src/broadcom/compiler/qpu_schedule.c
+++ b/src/broadcom/compiler/qpu_schedule.c
@@ -83,6 +83,7 @@ struct schedule_state {
         struct schedule_node *last_vpm;
         struct schedule_node *last_unif;
         struct schedule_node *last_rtop;
+        struct schedule_node *last_unifa;
         enum direction dir;
         /* Estimated cycle when the current instruction would start. */
         uint32_t time;
@@ -228,6 +229,11 @@ process_waddr_deps(struct schedule_state *state, struct schedule_node *n,
                         add_write_dep(state, &state->last_tmu_write, n);
                         break;
 
+                case V3D_QPU_WADDR_UNIFA:
+                        if (state->devinfo->ver >= 40)
+                                add_write_dep(state, &state->last_unifa, n);
+                        break;
+
                 case V3D_QPU_WADDR_NOP:
                         break;
 
@@ -400,6 +406,10 @@ calculate_deps(struct schedule_state *state, struct schedule_node *n)
         if (vir_has_uniform(qinst))
                 add_write_dep(state, &state->last_unif, n);
 
+        /* Both unifa and ldunifa must preserve ordering */
+        if (inst->sig.ldunifa || inst->sig.ldunifarf)
+                add_write_dep(state, &state->last_unifa, n);
+
         if (v3d_qpu_reads_flags(inst))
                 add_read_dep(state, state->last_sf, n);
         if (v3d_qpu_writes_flags(inst))