diff --git a/src/amd/compiler/aco_insert_waitcnt.cpp b/src/amd/compiler/aco_insert_waitcnt.cpp
index 6a7b65ce31f..3a1fddc4331 100644
--- a/src/amd/compiler/aco_insert_waitcnt.cpp
+++ b/src/amd/compiler/aco_insert_waitcnt.cpp
@@ -716,6 +716,14 @@ gen(Instruction* instr, wait_ctx& ctx)
          update_counters(ctx, event_sendmsg);
       break;
    }
+   case Format::SOP1: {
+      if (instr->opcode == aco_opcode::s_sendmsg_rtn_b32 ||
+          instr->opcode == aco_opcode::s_sendmsg_rtn_b64) {
+         update_counters(ctx, event_sendmsg);
+         insert_wait_entry(ctx, instr->definitions[0], event_sendmsg);
+      }
+      break;
+   }
    default: break;
    }
 }
diff --git a/src/amd/compiler/aco_opt_value_numbering.cpp b/src/amd/compiler/aco_opt_value_numbering.cpp
index 8b2fa3591ef..db977fe486b 100644
--- a/src/amd/compiler/aco_opt_value_numbering.cpp
+++ b/src/amd/compiler/aco_opt_value_numbering.cpp
@@ -202,6 +202,12 @@ struct InstrPred {
       }
 
       switch (a->format) {
+      case Format::SOP1: {
+         if (a->opcode == aco_opcode::s_sendmsg_rtn_b32 ||
+             a->opcode == aco_opcode::s_sendmsg_rtn_b64)
+            return false;
+         return true;
+      }
       case Format::SOPK: {
          if (a->opcode == aco_opcode::s_getreg_b32)
             return false;
diff --git a/src/amd/compiler/aco_scheduler.cpp b/src/amd/compiler/aco_scheduler.cpp
index 88498ef4565..d588b8000e4 100644
--- a/src/amd/compiler/aco_scheduler.cpp
+++ b/src/amd/compiler/aco_scheduler.cpp
@@ -568,7 +568,8 @@ perform_hazard_query(hazard_query* query, Instruction* instr, bool upwards)
    /* don't move non-reorderable instructions */
    if (instr->opcode == aco_opcode::s_memtime || instr->opcode == aco_opcode::s_memrealtime ||
        instr->opcode == aco_opcode::s_setprio || instr->opcode == aco_opcode::s_getreg_b32 ||
-       instr->opcode == aco_opcode::p_init_scratch || instr->opcode == aco_opcode::p_jump_to_epilog)
+       instr->opcode == aco_opcode::p_init_scratch || instr->opcode == aco_opcode::p_jump_to_epilog ||
+       instr->opcode == aco_opcode::s_sendmsg_rtn_b32 || instr->opcode == aco_opcode::s_sendmsg_rtn_b64)
       return hazard_fail_unreorderable;
 
    memory_event_set instr_set;
@@ -644,7 +645,10 @@ schedule_SMEM(sched_ctx& ctx, Block* block, std::vector<RegisterDemand>& registe
    int16_t k = 0;
 
    /* don't move s_memtime/s_memrealtime */
-   if (current->opcode == aco_opcode::s_memtime || current->opcode == aco_opcode::s_memrealtime)
+   if (current->opcode == aco_opcode::s_memtime ||
+       current->opcode == aco_opcode::s_memrealtime ||
+       current->opcode == aco_opcode::s_sendmsg_rtn_b32 ||
+       current->opcode == aco_opcode::s_sendmsg_rtn_b64)
       return;
 
    /* first, check if we have instructions before current to move down */