diff --git a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
index b8e064fb94b..02545073f64 100644
--- a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
+++ b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
@@ -218,7 +218,7 @@ static void build_streamout_vertex(struct si_shader_context *ctx, LLVMValueRef *
                                    LLVMValueRef offset_vtx, LLVMValueRef vertexptr)
 {
    struct si_shader_info *info = &ctx->shader->selector->info;
-   struct pipe_stream_output_info *so = &ctx->shader->selector->so;
+   struct pipe_stream_output_info *so = &ctx->so;
    LLVMBuilderRef builder = ctx->ac.builder;
    LLVMValueRef offset[4] = {};
    LLVMValueRef tmp;
@@ -274,7 +274,7 @@ struct ngg_streamout {
 static void build_streamout(struct si_shader_context *ctx, struct ngg_streamout *nggso)
 {
    struct si_shader_info *info = &ctx->shader->selector->info;
-   struct pipe_stream_output_info *so = &ctx->shader->selector->so;
+   struct pipe_stream_output_info *so = &ctx->so;
    LLVMBuilderRef builder = ctx->ac.builder;
    LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
    LLVMValueRef tid = gfx10_get_thread_id_in_tg(ctx);
@@ -1418,7 +1418,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
 
    LLVMValueRef vertex_ptr = NULL;
 
-   if (sel->so.num_outputs || gfx10_ngg_writes_user_edgeflags(ctx->shader))
+   if (ctx->so.num_outputs || gfx10_ngg_writes_user_edgeflags(ctx->shader))
       vertex_ptr = ngg_nogs_vertex_ptr(ctx, gfx10_get_thread_id_in_tg(ctx));
 
    for (unsigned i = 0; i < info->num_outputs; i++) {
@@ -1430,7 +1430,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
          /* TODO: we may store more outputs than streamout needs,
           * but streamout performance isn't that important.
           */
-         if (sel->so.num_outputs) {
+         if (ctx->so.num_outputs) {
             tmp = ac_build_gep0(&ctx->ac, vertex_ptr, LLVMConstInt(ctx->ac.i32, 4 * i + j, false));
             tmp2 = LLVMBuildLoad(builder, addrs[4 * i + j], "");
             tmp2 = ac_to_integer(&ctx->ac, tmp2);
@@ -1452,7 +1452,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
    }
 
    bool unterminated_es_if_block =
-      !sel->so.num_outputs && !gfx10_ngg_writes_user_edgeflags(ctx->shader) &&
+      !ctx->so.num_outputs && !gfx10_ngg_writes_user_edgeflags(ctx->shader) &&
       !ctx->screen->use_ngg_streamout && /* no query buffer */
       (ctx->stage != MESA_SHADER_VERTEX || !ctx->shader->key.ge.mono.u.vs_export_prim_id);
 
@@ -1478,7 +1478,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
    /* Streamout */
    LLVMValueRef emitted_prims = NULL;
 
-   if (sel->so.num_outputs) {
+   if (ctx->so.num_outputs) {
       assert(!unterminated_es_if_block);
 
       struct ngg_streamout nggso = {};
@@ -1498,7 +1498,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
       assert(!unterminated_es_if_block);
 
       /* Streamout already inserted the barrier, so don't insert it again. */
-      if (!sel->so.num_outputs)
+      if (!ctx->so.num_outputs)
          ac_build_s_barrier(&ctx->ac);
 
       ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);
@@ -1522,7 +1522,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
       assert(!unterminated_es_if_block);
 
       /* Streamout and edge flags use LDS. Make it idle, so that we can reuse it. */
-      if (sel->so.num_outputs || gfx10_ngg_writes_user_edgeflags(ctx->shader))
+      if (ctx->so.num_outputs || gfx10_ngg_writes_user_edgeflags(ctx->shader))
          ac_build_s_barrier(&ctx->ac);
 
       ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);
@@ -1550,7 +1550,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
       tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
       ac_build_ifcc(&ctx->ac, tmp, 5030);
       tmp = LLVMBuildICmp(builder, LLVMIntULE, ac_get_thread_id(&ctx->ac),
-                          sel->so.num_outputs ? ctx->ac.i32_1 : ctx->ac.i32_0, "");
+                          ctx->so.num_outputs ? ctx->ac.i32_1 : ctx->ac.i32_0, "");
       ac_build_ifcc(&ctx->ac, tmp, 5031);
       {
          LLVMValueRef args[] = {
@@ -1561,7 +1561,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
             ctx->ac.i32_0,                        /* cachepolicy */
          };
 
-         if (sel->so.num_outputs) {
+         if (ctx->so.num_outputs) {
             args[0] = ac_build_writelane(&ctx->ac, args[0], emitted_prims, ctx->ac.i32_1);
             args[2] = ac_build_writelane(&ctx->ac, args[2], LLVMConstInt(ctx->ac.i32, 24, false),
                                          ctx->ac.i32_1);
@@ -1896,7 +1896,7 @@ void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
    LLVMValueRef num_emit_threads = ngg_get_prim_cnt(ctx);
 
    /* Streamout */
-   if (sel->so.num_outputs) {
+   if (ctx->so.num_outputs) {
       struct ngg_streamout nggso = {};
 
       nggso.num_vertices = LLVMConstInt(ctx->ac.i32, verts_per_prim, false);
@@ -1927,17 +1927,17 @@ void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
       tmp = si_unpack_param(ctx, ctx->vs_state_bits, 6, 1);
       tmp = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
       ac_build_ifcc(&ctx->ac, tmp, 5109); /* if (STREAMOUT_QUERY_ENABLED) */
-      unsigned num_query_comps = sel->so.num_outputs ? 8 : 4;
+      unsigned num_query_comps = ctx->so.num_outputs ? 8 : 4;
       tmp = LLVMBuildICmp(builder, LLVMIntULT, tid,
                           LLVMConstInt(ctx->ac.i32, num_query_comps, false), "");
       ac_build_ifcc(&ctx->ac, tmp, 5110);
       {
          LLVMValueRef offset;
          tmp = tid;
-         if (sel->so.num_outputs)
+         if (ctx->so.num_outputs)
             tmp = LLVMBuildAnd(builder, tmp, LLVMConstInt(ctx->ac.i32, 3, false), "");
          offset = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->ac.i32, 32, false), "");
-         if (sel->so.num_outputs) {
+         if (ctx->so.num_outputs) {
             tmp = LLVMBuildLShr(builder, tid, LLVMConstInt(ctx->ac.i32, 2, false), "");
             tmp = LLVMBuildNUWMul(builder, tmp, LLVMConstInt(ctx->ac.i32, 8, false), "");
             offset = LLVMBuildAdd(builder, offset, tmp, "");
@@ -1967,7 +1967,7 @@ void gfx10_ngg_gs_emit_epilogue(struct si_shader_context *ctx)
       LLVMValueRef prim_enable = LLVMBuildAnd(builder, live, is_emit, "");
 
       /* Wait for streamout to finish before we kill primitives. */
-      if (sel->so.num_outputs)
+      if (ctx->so.num_outputs)
          ac_build_s_barrier(&ctx->ac);
 
       ac_build_ifcc(&ctx->ac, prim_enable, 0);
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 3344774523c..c60909833fa 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -409,7 +409,7 @@ void si_init_shader_args(struct si_shader_context *ctx, bool ngg_cull_shader)
       ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
 
       if (ctx->shader->is_gs_copy_shader) {
-         declare_streamout_params(ctx, &shader->selector->so);
+         declare_streamout_params(ctx, &ctx->so);
          /* VGPRs */
          declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
          break;
@@ -425,7 +425,7 @@ void si_init_shader_args(struct si_shader_context *ctx, bool ngg_cull_shader)
       } else if (shader->key.ge.as_ls) {
          /* no extra parameters */
       } else {
-         declare_streamout_params(ctx, &shader->selector->so);
+         declare_streamout_params(ctx, &ctx->so);
       }
 
       /* VGPRs */
@@ -626,7 +626,7 @@ void si_init_shader_args(struct si_shader_context *ctx, bool ngg_cull_shader)
          ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
          ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.es2gs_offset);
       } else {
-         declare_streamout_params(ctx, &shader->selector->so);
+         declare_streamout_params(ctx, &ctx->so);
          ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.tess_offchip_offset);
       }
 
@@ -1501,12 +1501,14 @@ bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compi
    bool free_nir;
    struct nir_shader *nir = si_get_nir_shader(sel, &shader->key, &free_nir);
 
+   struct pipe_stream_output_info so = sel->so;
+
    /* Dump NIR before doing NIR->LLVM conversion in case the
     * conversion fails. */
    if (si_can_dump_shader(sscreen, sel->info.stage) &&
        !(sscreen->debug_flags & DBG(NO_NIR))) {
       nir_print_shader(nir, stderr);
-      si_dump_streamout(&sel->so);
+      si_dump_streamout(&so);
    }
 
    /* Initialize vs_output_ps_input_cntl to default. */
@@ -1523,7 +1525,7 @@ 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, debug, nir, free_nir))
+   if (!si_llvm_compile_shader(sscreen, compiler, shader, &so, debug, nir, free_nir))
       return false;
 
    /* The GS copy shader is compiled next. */
diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 103d78a9c52..a9747c15b7f 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -41,6 +41,7 @@ struct si_shader_context {
    struct si_shader *shader;
    struct si_shader_selector *next_shader_sel;
    struct si_screen *screen;
+   struct pipe_stream_output_info so;
 
    gl_shader_stage stage;
 
@@ -226,8 +227,9 @@ void si_build_wrapper_function(struct si_shader_context *ctx, LLVMValueRef *part
 bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shader,
                            struct nir_shader *nir, bool free_nir, bool ngg_cull_shader);
 bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
-                            struct si_shader *shader, struct util_debug_callback *debug,
-                            struct nir_shader *nir, bool free_nir);
+                            struct si_shader *shader, const struct pipe_stream_output_info *so,
+                            struct util_debug_callback *debug, struct nir_shader *nir,
+                            bool free_nir);
 
 /* si_shader_llvm_gs.c */
 LLVMValueRef si_is_es_thread(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 e8546d1748d..101d2fb116e 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm.c
@@ -933,7 +933,7 @@ bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shad
       /* This is really only needed when streamout and / or vertex
        * compaction is enabled.
        */
-      if (!ctx->gs_ngg_scratch && (sel->so.num_outputs || shader->key.ge.opt.ngg_culling)) {
+      if (!ctx->gs_ngg_scratch && (ctx->so.num_outputs || shader->key.ge.opt.ngg_culling)) {
          LLVMTypeRef asi32 = LLVMArrayType(ctx->ac.i32, gfx10_ngg_get_scratch_dw_size(shader));
          ctx->gs_ngg_scratch =
             LLVMAddGlobalInAddressSpace(ctx->ac.module, asi32, "ngg_scratch", AC_ADDR_SPACE_LDS);
@@ -1087,13 +1087,15 @@ static void si_optimize_vs_outputs(struct si_shader_context *ctx)
 }
 
 bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
-                            struct si_shader *shader, struct util_debug_callback *debug,
-                            struct nir_shader *nir, bool free_nir)
+                            struct si_shader *shader, const struct pipe_stream_output_info *so,
+                            struct util_debug_callback *debug, struct nir_shader *nir,
+                            bool free_nir)
 {
    struct si_shader_selector *sel = shader->selector;
    struct si_shader_context ctx;
 
    si_llvm_context_init(&ctx, sscreen, compiler, shader->wave_size);
+   ctx.so = *so;
 
    LLVMValueRef ngg_cull_main_fn = NULL;
    if (sel->info.stage <= MESA_SHADER_TESS_EVAL && shader->key.ge.opt.ngg_culling) {
diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
index ec77d2fc65d..9cfca4dc5dc 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
@@ -446,9 +446,11 @@ struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
    si_llvm_context_init(&ctx, sscreen, compiler, shader->wave_size);
    ctx.shader = shader;
    ctx.stage = MESA_SHADER_VERTEX;
+   ctx.so = gs_selector->so;
 
    builder = ctx.ac.builder;
 
+   /* Build the main function. */
    si_llvm_create_main_func(&ctx, false);
 
    LLVMValueRef buf_ptr = ac_get_arg(&ctx.ac, ctx.internal_bindings);
@@ -461,7 +463,7 @@ struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
    /* Fetch the vertex stream ID.*/
    LLVMValueRef stream_id;
 
-   if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs)
+   if (!sscreen->use_ngg_streamout && ctx.so.num_outputs)
       stream_id = si_unpack_param(&ctx, ctx.args.streamout_config, 24, 2);
    else
       stream_id = ctx.ac.i32_0;
@@ -485,7 +487,7 @@ struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
       if (!gsinfo->num_stream_output_components[stream])
          continue;
 
-      if (stream > 0 && !gs_selector->so.num_outputs)
+      if (stream > 0 && !ctx.so.num_outputs)
          continue;
 
       bb = LLVMInsertBasicBlockInContext(ctx.ac.context, end_bb, "out");
@@ -513,7 +515,7 @@ struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
       }
 
       /* Streamout and exports. */
-      if (!sscreen->use_ngg_streamout && gs_selector->so.num_outputs) {
+      if (!sscreen->use_ngg_streamout && ctx.so.num_outputs) {
          si_llvm_emit_streamout(&ctx, outputs, gsinfo->num_outputs, stream);
       }
 
diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
index 06b049652d3..d0c699cf924 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
@@ -315,8 +315,7 @@ void si_llvm_streamout_store_output(struct si_shader_context *ctx, LLVMValueRef
 void si_llvm_emit_streamout(struct si_shader_context *ctx, struct si_shader_output_values *outputs,
                             unsigned noutput, unsigned stream)
 {
-   struct si_shader_selector *sel = ctx->shader->selector;
-   struct pipe_stream_output_info *so = &sel->so;
+   struct pipe_stream_output_info *so = &ctx->so;
    LLVMBuilderRef builder = ctx->ac.builder;
    int i;
 
@@ -775,7 +774,7 @@ void si_llvm_emit_vs_epilogue(struct ac_shader_abi *abi)
       }
    }
 
-   if (!ctx->screen->use_ngg_streamout && ctx->shader->selector->so.num_outputs)
+   if (!ctx->screen->use_ngg_streamout && ctx->so.num_outputs)
       si_llvm_emit_streamout(ctx, outputs, i, 0);
 
    /* Export PrimitiveID. */