diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index d2cca025bc4..99b2910a470 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1898,14 +1898,22 @@ struct nir_shader *si_get_nir_shader(struct si_shader *shader, struct si_shader_
 
    bool opt_offsets = si_lower_io_to_mem(shader, nir, tcs_vgpr_only_inputs);
 
-   /* Assign param export indices. */
-   if (is_last_vgt_stage)
+   if (is_last_vgt_stage) {
+      /* Assign param export indices. */
       si_assign_param_offsets(nir, shader);
 
-   /* Only lower last VGT NGG shader stage. */
-   if (sel->stage <= MESA_SHADER_GEOMETRY && key->ge.as_ngg && !key->ge.as_es) {
-      si_lower_ngg(shader, nir);
-      opt_offsets = true;
+      if (key->ge.as_ngg) {
+         /* Lower last VGT NGG shader stage. */
+         si_lower_ngg(shader, nir);
+         opt_offsets = true;
+      } else if (sel->stage == MESA_SHADER_VERTEX || sel->stage == MESA_SHADER_TESS_EVAL) {
+         /* Lower last VGT none-NGG VS/TES shader stage. */
+         int primitive_id_location =
+            shader->key.ge.mono.u.vs_export_prim_id ? sel->info.num_outputs : -1;
+
+         NIR_PASS_V(nir, ac_nir_lower_legacy_vs, primitive_id_location,
+                    key->ge.opt.remove_streamout);
+      }
    }
 
    NIR_PASS(progress2, nir, si_nir_lower_abi, shader, args);
@@ -1952,6 +1960,7 @@ void si_update_shader_binary_info(struct si_shader *shader, nir_shader *nir)
 bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
                        struct si_shader *shader, struct util_debug_callback *debug)
 {
+   bool ret = true;
    struct si_shader_selector *sel = shader->selector;
 
    struct si_shader_args args;
@@ -1960,11 +1969,6 @@ bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compi
    bool free_nir;
    struct nir_shader *nir = si_get_nir_shader(shader, &args, &free_nir, 0);
 
-   struct pipe_stream_output_info so = {};
-   /* NGG streamout has been lowered to buffer store in nir. */
-   if (!sscreen->use_ngg_streamout && si_shader_uses_streamout(shader))
-      nir_gather_stream_output_info(nir, &so);
-
    /* Dump NIR before doing NIR->LLVM conversion in case the
     * conversion fails. */
    if (si_can_dump_shader(sscreen, sel->stage) &&
@@ -2016,17 +2020,24 @@ bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compi
     * with PS and NGG VS), but monolithic shaders should be compiled
     * by LLVM due to more complicated compilation.
     */
-   if (!si_llvm_compile_shader(sscreen, compiler, shader, &args, &so, debug, nir, free_nir))
-      return false;
+   if (!si_llvm_compile_shader(sscreen, compiler, shader, &args, debug, nir)) {
+      ret = false;
+      goto out;
+   }
 
    shader->config.float_mode = float_mode;
 
    /* The GS copy shader is compiled next. */
    if (sel->stage == MESA_SHADER_GEOMETRY && !shader->key.ge.as_ngg) {
+      struct pipe_stream_output_info so = {};
+      if (si_shader_uses_streamout(shader))
+         nir_gather_stream_output_info(nir, &so);
+
       shader->gs_copy_shader = si_generate_gs_copy_shader(sscreen, compiler, sel, &so, debug);
       if (!shader->gs_copy_shader) {
          fprintf(stderr, "radeonsi: can't create GS copy shader\n");
-         return false;
+         ret = false;
+         goto out;
       }
    }
 
@@ -2111,7 +2122,12 @@ bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compi
 
    si_calculate_max_simd_waves(shader);
    si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
-   return true;
+
+out:
+   if (free_nir)
+      ralloc_free(nir);
+
+   return ret;
 }
 
 /**
diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 2d9ba9b6ef7..ede39887c29 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -229,9 +229,7 @@ bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shad
                            struct nir_shader *nir, bool free_nir);
 bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
                             struct si_shader *shader, struct si_shader_args *args,
-                            const struct pipe_stream_output_info *so,
-                            struct util_debug_callback *debug, struct nir_shader *nir,
-                            bool free_nir);
+                            struct util_debug_callback *debug, struct nir_shader *nir);
 
 /* si_shader_llvm_gs.c */
 LLVMValueRef si_is_es_thread(struct si_shader_context *ctx);
@@ -269,7 +267,6 @@ void si_llvm_emit_streamout(struct si_shader_context *ctx, struct si_shader_outp
                             unsigned noutput, unsigned stream);
 void si_llvm_build_vs_exports(struct si_shader_context *ctx,
                               struct si_shader_output_values *outputs, unsigned noutput);
-void si_llvm_vs_build_end(struct si_shader_context *ctx);
 void si_llvm_build_vs_prolog(struct si_shader_context *ctx, union si_shader_part_key *key);
 void si_llvm_init_vs_callbacks(struct si_shader_context *ctx);
 
diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm.c b/src/gallium/drivers/radeonsi/si_shader_llvm.c
index b6a978e73b6..3c1d344f6a4 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm.c
@@ -1020,8 +1020,8 @@ bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shad
    ctx->abi.disable_aniso_single_level = true;
 
    unsigned num_outputs = info->num_outputs;
-   /* need extra output to hold primitive id added by nir ngg lower */
-   if (ctx->stage <= MESA_SHADER_GEOMETRY && shader->key.ge.as_ngg &&
+   /* need extra output to hold primitive id added by nir lower */
+   if (ctx->stage <= MESA_SHADER_GEOMETRY &&
        ctx->shader->key.ge.mono.u.vs_export_prim_id)
       num_outputs++;
 
@@ -1048,8 +1048,6 @@ bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shad
          si_llvm_ls_build_end(ctx);
       else if (shader->key.ge.as_es)
          si_llvm_es_build_end(ctx);
-      else if (!shader->key.ge.as_ngg)
-         si_llvm_vs_build_end(ctx);
       break;
 
    case MESA_SHADER_TESS_CTRL:
@@ -1059,8 +1057,6 @@ bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shad
    case MESA_SHADER_TESS_EVAL:
       if (ctx->shader->key.ge.as_es)
          si_llvm_es_build_end(ctx);
-      else if (!ctx->shader->key.ge.as_ngg)
-         si_llvm_vs_build_end(ctx);
       break;
 
    case MESA_SHADER_GEOMETRY:
@@ -1099,9 +1095,7 @@ static bool si_should_optimize_less(struct ac_llvm_compiler *compiler,
 
 bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
                             struct si_shader *shader, struct si_shader_args *args,
-                            const struct pipe_stream_output_info *so,
-                            struct util_debug_callback *debug, struct nir_shader *nir,
-                            bool free_nir)
+                            struct util_debug_callback *debug, struct nir_shader *nir)
 {
    struct si_shader_selector *sel = shader->selector;
    struct si_shader_context ctx;
@@ -1112,10 +1106,9 @@ bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *
       exports_color_null = si_shader_uses_discard(shader) || sscreen->info.gfx_level < GFX10;
 
    si_llvm_context_init(&ctx, sscreen, compiler, shader->wave_size, exports_color_null, exports_mrtz);
-   ctx.so = *so;
    ctx.args = args;
 
-   if (!si_llvm_translate_nir(&ctx, shader, nir, free_nir)) {
+   if (!si_llvm_translate_nir(&ctx, shader, nir, false)) {
       si_llvm_dispose(&ctx);
       return false;
    }
@@ -1168,6 +1161,8 @@ bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *
          shader_ls.is_monolithic = true;
 
          si_init_shader_args(&shader_ls, ctx.args);
+
+         bool free_nir;
          nir = si_get_nir_shader(&shader_ls, ctx.args, &free_nir, sel->info.tcs_vgpr_only_inputs);
          si_update_shader_binary_info(shader, nir);
 
@@ -1240,6 +1235,8 @@ bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *
          shader_es.is_monolithic = true;
 
          si_init_shader_args(&shader_es, ctx.args);
+
+         bool free_nir;
          nir = si_get_nir_shader(&shader_es, ctx.args, &free_nir, 0);
          si_update_shader_binary_info(shader, nir);
 
diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
index 734a15504b8..1aadd00b360 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
@@ -724,44 +724,6 @@ void si_llvm_build_vs_exports(struct si_shader_context *ctx,
       ac_build_export(&ctx->ac, &param_exports[i]);
 }
 
-void si_llvm_vs_build_end(struct si_shader_context *ctx)
-{
-   struct si_shader_info *info = &ctx->shader->selector->info;
-   struct si_shader_output_values *outputs = NULL;
-   LLVMValueRef *addrs = ctx->abi.outputs;
-   int i, j;
-
-   assert(!ctx->shader->is_gs_copy_shader);
-   assert(info->num_outputs <= AC_LLVM_MAX_OUTPUTS);
-
-   outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
-
-   for (i = 0; i < info->num_outputs; i++) {
-      outputs[i].semantic = info->output_semantic[i];
-
-      for (j = 0; j < 4; j++) {
-         outputs[i].values[j] = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addrs[4 * i + j], "");
-         outputs[i].vertex_streams = info->output_streams[i];
-      }
-   }
-
-   if (!ctx->screen->use_ngg_streamout && ctx->so.num_outputs)
-      si_llvm_emit_streamout(ctx, outputs, i, 0);
-
-   /* Export PrimitiveID. */
-   if (ctx->shader->key.ge.mono.u.vs_export_prim_id) {
-      outputs[i].semantic = VARYING_SLOT_PRIMITIVE_ID;
-      outputs[i].vertex_streams = 0;
-      outputs[i].values[0] = ac_to_float(&ctx->ac, si_get_primitive_id(ctx, 0));
-      for (j = 1; j < 4; j++)
-         outputs[i].values[j] = LLVMConstReal(ctx->ac.f32, 0);
-      i++;
-   }
-
-   si_llvm_build_vs_exports(ctx, outputs, i);
-   FREE(outputs);
-}
-
 /**
  * Build the vertex shader prolog function.
  *