diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 537e0646e99..a970d2c1aeb 100644
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -12598,6 +12598,11 @@ select_rt_prolog(Program* program, ac_shader_config* config,
    assert(in_stack_base == out_launch_size_z);
    assert(in_local_ids[0] == out_launch_ids[0]);
 
+   /* <gfx9 reads in_scratch_offset at the end of the prolog to write out the scratch_offset
+    * arg. Make sure no other outputs have overwritten it by then.
+    */
+   assert(in_scratch_offset.reg() >= out_args->num_sgprs_used);
+
    /* load raygen sbt */
    bld.smem(aco_opcode::s_load_dwordx2, Definition(tmp_raygen_sbt, s2), Operand(in_sbt_desc, s2),
             Operand::c32(0u));
@@ -12649,15 +12654,6 @@ select_rt_prolog(Program* program, ac_shader_config* config,
    bld.vop3(aco_opcode::v_mad_u32_u24, Definition(out_launch_ids[0], v1), Operand(in_wg_id_x, s1),
             Operand::c32(8), Operand(in_local_ids[0], v1));
 
-   if (options->gfx_level < GFX9) {
-      /* write scratch/ring offsets to outputs, if needed */
-      bld.sop1(aco_opcode::s_mov_b32,
-               Definition(get_arg_reg(out_args, out_args->scratch_offset), s1),
-               Operand(in_scratch_offset, s1));
-      bld.sop1(aco_opcode::s_mov_b64, Definition(get_arg_reg(out_args, out_args->ring_offsets), s2),
-               Operand(tmp_ring_offsets, s2));
-   }
-
    /* calculate shader record ptr: SBT + RADV_RT_HANDLE_SIZE */
    if (options->gfx_level < GFX9) {
       bld.vop2_e64(aco_opcode::v_add_co_u32, Definition(out_record_ptr, v1), Definition(vcc, s2),
@@ -12699,6 +12695,15 @@ select_rt_prolog(Program* program, ac_shader_config* config,
    bld.vop2(aco_opcode::v_cndmask_b32, Definition(out_launch_ids[1], v1), Operand::zero(),
             Operand(out_launch_ids[1], v1), Operand(vcc, bld.lm));
 
+   if (options->gfx_level < GFX9) {
+      /* write scratch/ring offsets to outputs, if needed */
+      bld.sop1(aco_opcode::s_mov_b32,
+               Definition(get_arg_reg(out_args, out_args->scratch_offset), s1),
+               Operand(in_scratch_offset, s1));
+      bld.sop1(aco_opcode::s_mov_b64, Definition(get_arg_reg(out_args, out_args->ring_offsets), s2),
+               Operand(tmp_ring_offsets, s2));
+   }
+
    /* jump to raygen */
    bld.sop1(aco_opcode::s_setpc_b64, Operand(out_uniform_shader_addr, s2));