nir: introduce bindgen_return

with vtn_bindgen2, we'll want return values without derefs. this needs some
special handholding.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Mary Guillemard <mary.guillemard@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33099>

src/compiler/nir/nir.h

@@ -5581,10 +5581,10 @@ bool nir_split_struct_vars(nir_shader *shader, nir_variable_mode modes);
 bool nir_lower_returns_impl(nir_function_impl *impl);
 bool nir_lower_returns(nir_shader *shader);
 
-void nir_inline_function_impl(struct nir_builder *b,
-                              const nir_function_impl *impl,
-                              nir_def **params,
-                              struct hash_table *shader_var_remap);
+nir_def *nir_inline_function_impl(struct nir_builder *b,
+                                  const nir_function_impl *impl,
+                                  nir_def **params,
+                                  struct hash_table *shader_var_remap);
 bool nir_inline_functions(nir_shader *shader);
 void nir_cleanup_functions(nir_shader *shader);
 bool nir_link_shader_functions(nir_shader *shader,

src/compiler/nir/nir_functions.c

@@ -96,7 +96,7 @@ fixup_cast_deref_mode(nir_deref_instr *deref)
    }
 }
 
-void
+nir_def *
 nir_inline_function_impl(struct nir_builder *b,
                          const nir_function_impl *impl,
                          nir_def **params,
@@ -106,6 +106,11 @@ nir_inline_function_impl(struct nir_builder *b,
 
    exec_list_append(&b->impl->locals, &copy->locals);
 
+   /* Normally, NIR function returns are deref based, but for bindgen, we use a
+    * sideband here to let us lower derefs ahead-of-time.
+    */
+   nir_def *ret_val = NULL;
+
    nir_foreach_block(block, copy) {
       nir_foreach_instr_safe(instr, block) {
          switch (instr->type) {
@@ -152,13 +157,16 @@ nir_inline_function_impl(struct nir_builder *b,
          }
 
          case nir_instr_type_intrinsic: {
-            nir_intrinsic_instr *load = nir_instr_as_intrinsic(instr);
-            if (load->intrinsic != nir_intrinsic_load_param)
-               break;
-
-            unsigned param_idx = nir_intrinsic_param_idx(load);
-            assert(param_idx < impl->function->num_params);
-            nir_def_replace(&load->def, params[param_idx]);
+            nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+            if (intr->intrinsic == nir_intrinsic_bindgen_return) {
+               assert(ret_val == NULL && "expected exactly one return write");
+               ret_val = intr->src[0].ssa;
+               nir_instr_remove(&intr->instr);
+            } else if (intr->intrinsic == nir_intrinsic_load_param) {
+               unsigned param_idx = nir_intrinsic_param_idx(intr);
+               assert(param_idx < impl->function->num_params);
+               nir_def_replace(&intr->def, params[param_idx]);
+            }
             break;
          }
 
@@ -191,6 +199,8 @@ nir_inline_function_impl(struct nir_builder *b,
       nir_cf_reinsert(&body, nir_before_instr(&nop->instr));
       b->cursor = nir_instr_remove(&nop->instr);
    }
+
+   return ret_val;
 }
 
 static bool inline_function_impl(nir_function_impl *impl, struct set *inlined);

src/compiler/nir/nir_intrinsics.py

@@ -340,6 +340,10 @@ intrinsic("convert_alu_types", dest_comp=0, src_comp=[0],
 
 intrinsic("load_param", dest_comp=0, indices=[PARAM_IDX], flags=[CAN_ELIMINATE])
 
+# Store a scalar value to be used as a return value. Usually store_deref is used
+# for this, but vtn_bindgen needs to lower derefs.
+intrinsic("bindgen_return", src_comp=[0])
+
 intrinsic("load_deref", dest_comp=0, src_comp=[-1],
           indices=[ACCESS], flags=[CAN_ELIMINATE])
 intrinsic("store_deref", src_comp=[-1, 0], indices=[WRITE_MASK, ACCESS])