diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 8402ee4727a..7b5779ea306 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -711,16 +711,20 @@ void si_init_shader_args(struct si_shader *shader, struct si_shader_args *args)
          num_prolog_vgprs += num_color_elements;
       }
 
-      /* Outputs for the epilog. */
-      num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
-      num_returns = num_return_sgprs + util_bitcount(shader->selector->info.colors_written) * 4 +
-                    shader->selector->info.writes_z + shader->selector->info.writes_stencil +
-                    shader->selector->info.writes_samplemask + 1 /* SampleMaskIn */;
+      /* Monolithic PS emit epilog in NIR directly. */
+      if (!shader->is_monolithic) {
+         /* Outputs for the epilog. */
+         num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
+         num_returns =
+            num_return_sgprs + util_bitcount(shader->selector->info.colors_written) * 4 +
+            shader->selector->info.writes_z + shader->selector->info.writes_stencil +
+            shader->selector->info.writes_samplemask + 1 /* SampleMaskIn */;
 
-      for (i = 0; i < num_return_sgprs; i++)
-         ac_add_return(&args->ac, AC_ARG_SGPR);
-      for (; i < num_returns; i++)
-         ac_add_return(&args->ac, AC_ARG_VGPR);
+         for (i = 0; i < num_return_sgprs; i++)
+            ac_add_return(&args->ac, AC_ARG_SGPR);
+         for (; i < num_returns; i++)
+            ac_add_return(&args->ac, AC_ARG_VGPR);
+      }
       break;
 
    case MESA_SHADER_COMPUTE:
@@ -1999,6 +2003,23 @@ struct nir_shader *si_get_nir_shader(struct si_shader *shader,
       }
    } else if (is_legacy_gs) {
       NIR_PASS_V(nir, ac_nir_lower_legacy_gs, false, sel->screen->use_ngg, output_info);
+   } else if (sel->stage == MESA_SHADER_FRAGMENT && shader->is_monolithic) {
+      ac_nir_lower_ps_options options = {
+         .gfx_level = sel->screen->info.gfx_level,
+         .family = sel->screen->info.family,
+         .uses_discard = si_shader_uses_discard(shader),
+         .alpha_to_coverage_via_mrtz = key->ps.part.epilog.alpha_to_coverage_via_mrtz,
+         .dual_src_blend_swizzle = key->ps.part.epilog.dual_src_blend_swizzle,
+         .spi_shader_col_format = key->ps.part.epilog.spi_shader_col_format,
+         .color_is_int8 = key->ps.part.epilog.color_is_int8,
+         .color_is_int10 = key->ps.part.epilog.color_is_int10,
+         .clamp_color = key->ps.part.epilog.clamp_color,
+         .alpha_to_one = key->ps.part.epilog.alpha_to_one,
+         .alpha_func = key->ps.part.epilog.alpha_func,
+         .broadcast_last_cbuf = key->ps.part.epilog.last_cbuf,
+      };
+
+      NIR_PASS_V(nir, ac_nir_lower_ps, &options);
    }
 
    NIR_PASS(progress2, nir, si_nir_lower_abi, shader, args);
diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm.c b/src/gallium/drivers/radeonsi/si_shader_llvm.c
index 685e152c2d9..304d626ca0c 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm.c
@@ -1050,7 +1050,8 @@ static bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shade
       break;
 
    case MESA_SHADER_FRAGMENT:
-      si_llvm_ps_build_end(ctx);
+      if (!shader->is_monolithic)
+         si_llvm_ps_build_end(ctx);
       break;
 
    default:
diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_ps.c b/src/gallium/drivers/radeonsi/si_shader_llvm_ps.c
index 41cb49a33d4..d6e0c7e3cb1 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_ps.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_ps.c
@@ -850,30 +850,26 @@ void si_llvm_build_ps_epilog(struct si_shader_context *ctx, union si_shader_part
 
 void si_llvm_build_monolithic_ps(struct si_shader_context *ctx, struct si_shader *shader)
 {
-   struct ac_llvm_pointer parts[3];
-   unsigned num_parts = 0, main_index;
+   union si_shader_part_key prolog_key;
+   si_get_ps_prolog_key(shader, &prolog_key, false);
+
+   /* If no prolog is needed, we only have the main part, no need to build wrapper function. */
+   if (!si_need_ps_prolog(&prolog_key))
+      return;
+
    struct ac_llvm_pointer main_fn = ctx->main_fn;
+
    /* Preserve main arguments. */
    enum ac_arg_type main_arg_types[AC_MAX_ARGS];
    for (int i = 0; i < ctx->args->ac.arg_count; i++)
       main_arg_types[i] = ctx->args->ac.args[i].type;
 
+   si_llvm_build_ps_prolog(ctx, &prolog_key, false);
 
-   union si_shader_part_key prolog_key;
-   si_get_ps_prolog_key(shader, &prolog_key, false);
+   struct ac_llvm_pointer parts[2] = {
+      ctx->main_fn, /* prolog */
+      main_fn,      /* main */
+   };
 
-   if (si_need_ps_prolog(&prolog_key)) {
-      si_llvm_build_ps_prolog(ctx, &prolog_key, false);
-      parts[num_parts++] = ctx->main_fn;
-   }
-
-   main_index = num_parts;
-   parts[num_parts++] = main_fn;
-
-   union si_shader_part_key epilog_key;
-   si_get_ps_epilog_key(shader, &epilog_key);
-   si_llvm_build_ps_epilog(ctx, &epilog_key, false);
-   parts[num_parts++] = ctx->main_fn;
-
-   si_build_wrapper_function(ctx, parts, num_parts, main_index, 0, main_arg_types, false);
+   si_build_wrapper_function(ctx, parts, 2, 1, 0, main_arg_types, false);
 }