[Mesa-dev] [PATCH 009/133] nir: add a pass to lower variables for scalar backends

Jason Ekstrand jason at jlekstrand.net
Mon Dec 15 22:04:19 PST 2014


From: Connor Abbott <connor.abbott at intel.com>

---
 src/glsl/Makefile.sources                 |    1 +
 src/glsl/nir/nir.h                        |    4 +
 src/glsl/nir/nir_lower_variables_scalar.c | 1241 +++++++++++++++++++++++++++++
 3 files changed, 1246 insertions(+)
 create mode 100644 src/glsl/nir/nir_lower_variables_scalar.c

diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources
index 57ca0ea..83a99c1 100644
--- a/src/glsl/Makefile.sources
+++ b/src/glsl/Makefile.sources
@@ -18,6 +18,7 @@ NIR_FILES = \
 	$(GLSL_SRCDIR)/nir/nir.h \
 	$(GLSL_SRCDIR)/nir/nir_intrinsics.c \
 	$(GLSL_SRCDIR)/nir/nir_intrinsics.h \
+	$(GLSL_SRCDIR)/nir/nir_lower_variables_scalar.c \
 	$(GLSL_SRCDIR)/nir/nir_opcodes.c \
 	$(GLSL_SRCDIR)/nir/nir_opcodes.h \
 	$(GLSL_SRCDIR)/nir/nir_print.c \
diff --git a/src/glsl/nir/nir.h b/src/glsl/nir/nir.h
index c46fb2c..b8041bb 100644
--- a/src/glsl/nir/nir.h
+++ b/src/glsl/nir/nir.h
@@ -1242,6 +1242,10 @@ void nir_print_shader(nir_shader *shader, FILE *fp);
 
 void nir_validate_shader(nir_shader *shader);
 
+void nir_lower_variables_scalar(nir_shader *shader, bool lower_globals,
+                                bool lower_io, bool add_names,
+                                bool native_integers);
+
 #ifdef __cplusplus
 } /* extern "C" */
 #endif
diff --git a/src/glsl/nir/nir_lower_variables_scalar.c b/src/glsl/nir/nir_lower_variables_scalar.c
new file mode 100644
index 0000000..97b68d3
--- /dev/null
+++ b/src/glsl/nir/nir_lower_variables_scalar.c
@@ -0,0 +1,1241 @@
+/*
+ * Copyright © 2014 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Connor Abbott (cwabbott0 at gmail.com)
+ *
+ */
+
+/*
+ * This lowering pass converts references to variables with loads/stores to
+ * registers or inputs/outputs. We assume that structure splitting has already
+ * been run, or else structures with indirect references can't be split. We
+ * also assume that this pass will be consumed by a scalar backend, so we pack
+ * things more tightly.
+ */
+
+#include "nir.h"
+
+static unsigned
+type_size(const struct glsl_type *type)
+{
+   unsigned int size, i;
+
+   switch (glsl_get_base_type(type)) {
+   case GLSL_TYPE_UINT:
+   case GLSL_TYPE_INT:
+   case GLSL_TYPE_FLOAT:
+   case GLSL_TYPE_BOOL:
+      return glsl_get_components(type);
+   case GLSL_TYPE_ARRAY:
+      return type_size(glsl_get_array_element(type)) * glsl_get_length(type);
+   case GLSL_TYPE_STRUCT:
+      size = 0;
+      for (i = 0; i < glsl_get_length(type); i++) {
+         size += type_size(glsl_get_struct_elem_type(type, i));
+      }
+      return size;
+   case GLSL_TYPE_SAMPLER:
+      return 0;
+   case GLSL_TYPE_ATOMIC_UINT:
+      return 0;
+   case GLSL_TYPE_INTERFACE:
+      return 0;
+   case GLSL_TYPE_IMAGE:
+      return 0;
+   case GLSL_TYPE_VOID:
+   case GLSL_TYPE_ERROR:
+      unreachable("not reached");
+   }
+
+   return 0;
+}
+
+/*
+ * for inputs, outputs, and uniforms, assigns starting locations for variables
+ */
+
+static void
+assign_var_locations(struct hash_table *ht)
+{
+   unsigned location = 0;
+
+   struct hash_entry *entry;
+   hash_table_foreach(ht, entry) {
+      nir_variable *var = (nir_variable *) entry->data;
+
+      /*
+       * UBO's have their own address spaces, so don't count them towards the
+       * number of global uniforms
+       */
+      if (var->data.mode == nir_var_uniform && var->interface_type != NULL)
+         continue;
+
+      var->data.driver_location = location;
+      location += type_size(var->type);
+   }
+}
+
+static void
+assign_var_locations_shader(nir_shader *shader)
+{
+   assign_var_locations(shader->inputs);
+   assign_var_locations(shader->outputs);
+   assign_var_locations(shader->uniforms);
+}
+
+static void
+init_reg(nir_variable *var, nir_register *reg, struct hash_table *ht,
+         bool add_names)
+{
+   if (!glsl_type_is_scalar(var->type) &&
+       !glsl_type_is_vector(var->type)) {
+      reg->is_packed = true;
+      reg->num_components = 1;
+      reg->num_array_elems = type_size(var->type);
+   } else {
+      reg->num_components = glsl_get_components(var->type);
+   }
+   if (add_names)
+      reg->name = ralloc_strdup(reg, var->name);
+   _mesa_hash_table_insert(ht, _mesa_hash_pointer(var), var, reg);
+}
+
+static struct hash_table *
+init_var_ht(nir_shader *shader, bool lower_globals, bool lower_io,
+            bool add_names)
+{
+   struct hash_table *ht = _mesa_hash_table_create(NULL,
+                                                   _mesa_key_pointer_equal);
+
+   if (lower_globals) {
+      foreach_list_typed(nir_variable, var, node, &shader->globals) {
+         nir_register *reg = nir_global_reg_create(shader);
+         init_reg(var, reg, ht, add_names);
+      }
+   }
+
+   if (lower_io) {
+      struct hash_entry *entry;
+      hash_table_foreach(shader->outputs, entry) {
+         nir_variable *var = (nir_variable *) entry->data;
+         nir_register *reg = nir_global_reg_create(shader);
+         init_reg(var, reg, ht, add_names);
+      }
+   }
+
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl) {
+         nir_function_impl *impl = overload->impl;
+
+         foreach_list_typed(nir_variable, var, node, &impl->locals) {
+            nir_register *reg = nir_local_reg_create(impl);
+            init_reg(var, reg, ht, add_names);
+         }
+      }
+   }
+
+   return ht;
+}
+
+static bool
+deref_has_indirect(nir_deref_var *deref_var)
+{
+   nir_deref *deref = &deref_var->deref;
+
+   while (deref->child != NULL) {
+      deref = deref->child;
+      if (deref->deref_type == nir_deref_type_array) {
+         nir_deref_array *deref_array = nir_deref_as_array(deref);
+         if (deref_array->has_indirect)
+            return true;
+      }
+   }
+
+   return false;
+}
+
+static unsigned
+get_deref_offset(nir_deref_var *deref_var, nir_instr *instr,
+                 nir_function_impl *impl, bool native_integers,
+                 nir_src *indirect)
+{
+   void *mem_ctx = ralloc_parent(instr);
+
+   bool first_indirect = true;
+
+   unsigned base_offset = 0;
+   nir_deref *deref = &deref_var->deref;
+   while (deref->child != NULL) {
+      const struct glsl_type *parent_type = deref->type;
+      deref = deref->child;
+
+      if (deref->deref_type == nir_deref_type_array) {
+         nir_deref_array *deref_array = nir_deref_as_array(deref);
+         unsigned size = type_size(deref->type);
+
+         base_offset += size * deref_array->base_offset;
+
+         if (deref_array->has_indirect) {
+            nir_src src;
+            if (size == 1) {
+               src = deref_array->indirect;
+            } else {
+               /* temp1 = size * deref_array->indirect */
+
+               nir_register *const_reg = nir_local_reg_create(impl);
+               const_reg->num_components = 1;
+
+               nir_load_const_instr *load_const =
+                  nir_load_const_instr_create(mem_ctx);
+               load_const->dest.reg.reg = const_reg;
+               load_const->num_components = 1;
+               load_const->value.u[0] = size;
+               nir_instr_insert_before(instr, &load_const->instr);
+
+               nir_register *reg = nir_local_reg_create(impl);
+               reg->num_components = 1;
+
+               nir_op op;
+               if (native_integers)
+                  op = nir_op_imul;
+               else
+                  op = nir_op_fmul;
+               nir_alu_instr *mul_instr = nir_alu_instr_create(mem_ctx, op);
+               mul_instr->dest.write_mask = 1;
+               mul_instr->dest.dest.reg.reg = reg;
+               mul_instr->src[0].src = deref_array->indirect;
+               mul_instr->src[1].src.reg.reg = const_reg;
+               nir_instr_insert_before(instr, &mul_instr->instr);
+
+               src.is_ssa = false;
+               src.reg.reg = reg;
+               src.reg.base_offset = 0;
+               src.reg.indirect = NULL;
+            }
+
+            if (!first_indirect) {
+               /* temp2 = indirect + temp1 */
+
+               nir_register *reg = nir_local_reg_create(impl);
+               reg->num_components = 1;
+
+               nir_op op;
+               if (native_integers)
+                  op = nir_op_iadd;
+               else
+                  op = nir_op_fadd;
+               nir_alu_instr *add_instr = nir_alu_instr_create(mem_ctx, op);
+               add_instr->dest.write_mask = 1;
+               add_instr->dest.dest.reg.reg = reg;
+               add_instr->src[0].src = *indirect;
+               add_instr->src[1].src = src;
+               nir_instr_insert_before(instr, &add_instr->instr);
+
+               src.is_ssa = false;
+               src.reg.reg = reg;
+               src.reg.base_offset = 0;
+               src.reg.indirect = NULL;
+            }
+
+            /* indirect = tempX */
+            *indirect = src;
+            first_indirect = false;
+         }
+      } else {
+         nir_deref_struct *deref_struct = nir_deref_as_struct(deref);
+
+         unsigned i = 0;
+         while(strcmp(glsl_get_struct_elem_name(parent_type, i),
+                      deref_struct->elem) != 0) {
+            base_offset += type_size(glsl_get_struct_elem_type(parent_type, i));
+            i++;
+         }
+      }
+   }
+
+   return base_offset;
+}
+
+/*
+ * We cannot convert variables used in calls, so remove them from the hash
+ * table.
+ */
+
+static bool
+remove_call_vars_cb(nir_block *block, void *state)
+{
+   struct hash_table *ht = (struct hash_table *) state;
+
+   nir_foreach_instr(block, instr) {
+      if (instr->type == nir_instr_type_call) {
+         nir_call_instr *call = nir_instr_as_call(instr);
+         if (call->return_deref) {
+            struct hash_entry *entry =
+               _mesa_hash_table_search(ht,
+                                       _mesa_hash_pointer(call->return_deref->var),
+                                       call->return_deref->var);
+            if (entry)
+               _mesa_hash_table_remove(ht, entry);
+         }
+
+         for (unsigned i = 0; i < call->num_params; i++) {
+            struct hash_entry *entry =
+               _mesa_hash_table_search(ht,
+                                       _mesa_hash_pointer(call->params[i]->var),
+                                       call->params[i]->var);
+            if (entry)
+               _mesa_hash_table_remove(ht, entry);
+         }
+      }
+   }
+
+   return true;
+}
+
+static void
+remove_local_vars(nir_function_impl *impl, struct hash_table *ht)
+{
+   if (impl->return_var) {
+      struct hash_entry *entry =
+         _mesa_hash_table_search(ht, _mesa_hash_pointer(impl->return_var),
+                                 impl->return_var);
+
+      if (entry)
+         _mesa_hash_table_remove(ht, entry);
+   }
+
+   for (unsigned i = 0; i < impl->num_params; i++) {
+      struct hash_entry *entry =
+         _mesa_hash_table_search(ht,
+                                 _mesa_hash_pointer(impl->params[i]),
+                                 impl->params[i]);
+      if (entry)
+         _mesa_hash_table_remove(ht, entry);
+   }
+
+   nir_foreach_block(impl, remove_call_vars_cb, ht);
+}
+
+static void
+remove_local_vars_shader(nir_shader *shader, struct hash_table *ht)
+{
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl)
+         remove_local_vars(overload->impl, ht);
+   }
+}
+
+static nir_deref *
+get_deref_tail(nir_deref *deref)
+{
+   while (deref->child != NULL)
+      deref = deref->child;
+   return deref;
+}
+
+/* helper for reg_const_load which emits a single instruction */
+static void
+reg_const_load_single_instr(nir_reg_dest reg, nir_constant *constant,
+                            unsigned num_components, unsigned offset,
+                            nir_function_impl *impl, void *mem_ctx)
+{
+   nir_load_const_instr *instr = nir_load_const_instr_create(mem_ctx);
+   instr->num_components = num_components;
+   for (unsigned i = 0; i < num_components; i++) {
+      instr->value.u[i] = constant->value.u[i + offset];
+   }
+   instr->dest.reg = reg;
+   instr->dest.reg.base_offset += offset;
+
+   nir_instr_insert_before_cf_list(&impl->body, &instr->instr);
+}
+
+/* loads a constant value into a register */
+static void
+reg_const_load(nir_reg_dest reg, nir_constant *constant,
+               const struct glsl_type *type, nir_function_impl *impl,
+               void *mem_ctx)
+{
+   unsigned offset = 0;
+   const struct glsl_type *subtype;
+   unsigned subtype_size;
+
+   switch (glsl_get_base_type(type)) {
+      case GLSL_TYPE_FLOAT:
+      case GLSL_TYPE_INT:
+      case GLSL_TYPE_UINT:
+      case GLSL_TYPE_BOOL:
+         if (glsl_type_is_matrix(type)) {
+            for (unsigned i = 0; i < glsl_get_matrix_columns(type); i++) {
+               reg_const_load_single_instr(reg, constant,
+                                           glsl_get_vector_elements(type),
+                                           i * glsl_get_vector_elements(type),
+                                           impl, mem_ctx);
+            }
+         } else {
+            reg_const_load_single_instr(reg, constant,
+                                        glsl_get_vector_elements(type), 0,
+                                        impl, mem_ctx);
+         }
+         break;
+
+      case GLSL_TYPE_STRUCT:
+         for (unsigned i = 0; i < glsl_get_length(type); i++) {
+            const struct glsl_type *field = glsl_get_struct_elem_type(type, i);
+            nir_reg_dest new_reg = reg;
+            new_reg.base_offset += offset;
+            reg_const_load(new_reg, constant->elements[i], field, impl,
+                           mem_ctx);
+            offset += type_size(field);
+         }
+         break;
+
+      case GLSL_TYPE_ARRAY:
+         subtype = glsl_get_array_element(type);
+         subtype_size = type_size(subtype);
+         for (unsigned i = 0; i < glsl_get_length(type); i++) {
+            nir_reg_dest new_reg = reg;
+            new_reg.base_offset += subtype_size * i;
+            reg_const_load(new_reg, constant->elements[i], subtype, impl,
+                           mem_ctx);
+         }
+         break;
+
+      default:
+         assert(0);
+         break;
+   }
+}
+
+/* recursively emits a register <-> dereference block copy */
+static void
+var_reg_block_copy_impl(nir_reg_src reg, nir_deref_var *deref_head,
+                        nir_src *predicate, const struct glsl_type *type,
+                        nir_instr *after, bool var_dest, void *mem_ctx)
+{
+   unsigned offset;
+
+   switch (glsl_get_base_type(type)) {
+      case GLSL_TYPE_FLOAT:
+      case GLSL_TYPE_INT:
+      case GLSL_TYPE_UINT:
+      case GLSL_TYPE_BOOL:
+         if (glsl_type_is_matrix(type)) {
+            for (unsigned i = 0; i < glsl_get_matrix_columns(type); i++) {
+               nir_deref_array *deref_array = nir_deref_array_create(mem_ctx);
+               deref_array->base_offset = i;
+               deref_array->deref.type = glsl_get_column_type(type);
+
+               nir_deref_var *new_deref_head =
+                  nir_deref_as_var(nir_copy_deref(mem_ctx, &deref_head->deref));
+               get_deref_tail(&new_deref_head->deref)->child =
+                  &deref_array->deref;
+
+               nir_reg_src new_reg = reg;
+               new_reg.base_offset += i * glsl_get_vector_elements(type);
+
+               var_reg_block_copy_impl(new_reg, new_deref_head, predicate,
+                                       glsl_get_column_type(type), after,
+                                       var_dest, mem_ctx);
+            }
+         } else {
+            if (var_dest) {
+               nir_intrinsic_op op;
+               switch (glsl_get_vector_elements(type)) {
+                  case 1: op = nir_intrinsic_store_var_vec1; break;
+                  case 2: op = nir_intrinsic_store_var_vec2; break;
+                  case 3: op = nir_intrinsic_store_var_vec3; break;
+                  case 4: op = nir_intrinsic_store_var_vec4; break;
+                  default: assert(0); break;
+               }
+
+               nir_intrinsic_instr *store =
+                  nir_intrinsic_instr_create(mem_ctx, op);
+               store->variables[0] = deref_head;
+               store->src[0].reg.reg = reg.reg;
+               store->src[0].reg.base_offset = reg.base_offset;
+               if (reg.indirect) {
+                  store->src[0].reg.indirect = ralloc(mem_ctx, nir_src);
+                  *store->src[0].reg.indirect = *reg.indirect;
+               }
+
+               if (predicate) {
+                  store->has_predicate = true;
+                  store->predicate = nir_src_copy(*predicate, mem_ctx);
+               }
+
+               nir_instr_insert_before(after, &store->instr);
+            } else {
+               nir_intrinsic_op op;
+               switch (glsl_get_vector_elements(type)) {
+                  case 1: op = nir_intrinsic_load_var_vec1; break;
+                  case 2: op = nir_intrinsic_load_var_vec2; break;
+                  case 3: op = nir_intrinsic_load_var_vec3; break;
+                  case 4: op = nir_intrinsic_load_var_vec4; break;
+                  default: assert(0); break;
+               }
+
+               nir_intrinsic_instr *load =
+                  nir_intrinsic_instr_create(mem_ctx, op);
+               load->variables[0] = deref_head;
+               load->dest.reg.reg = reg.reg;
+               load->dest.reg.base_offset = reg.base_offset;
+               if (reg.indirect) {
+                  load->dest.reg.indirect = ralloc(mem_ctx, nir_src);
+                  *load->dest.reg.indirect = *reg.indirect;
+               }
+
+               if (predicate) {
+                  load->has_predicate = true;
+                  load->predicate = nir_src_copy(*predicate, mem_ctx);
+               }
+
+               nir_instr_insert_before(after, &load->instr);
+            }
+         }
+         break;
+
+      case GLSL_TYPE_STRUCT:
+         offset = 0;
+         for (unsigned i = 0; i < glsl_get_length(type); i++) {
+            const struct glsl_type *field_type =
+               glsl_get_struct_elem_type(type, i);
+            const char *field_name = glsl_get_struct_elem_name(type, i);
+
+            nir_deref_struct *deref_struct =
+               nir_deref_struct_create(mem_ctx, field_name);
+            deref_struct->deref.type = field_type;
+            deref_struct->elem = field_name;
+
+            nir_deref_var *new_deref_head =
+               nir_deref_as_var(nir_copy_deref(mem_ctx, &deref_head->deref));
+            get_deref_tail(&new_deref_head->deref)->child =
+               &deref_struct->deref;
+
+            nir_reg_src new_reg = reg;
+            new_reg.base_offset += offset;
+
+            var_reg_block_copy_impl(new_reg, new_deref_head, predicate,
+                                    field_type, after, var_dest, mem_ctx);
+
+            offset += type_size(field_type);
+         }
+         break;
+
+      case GLSL_TYPE_ARRAY:
+         for (unsigned i = 0; i < glsl_get_length(type);
+                  i++) {
+            const struct glsl_type *elem_type = glsl_get_array_element(type);
+
+            nir_deref_array *deref_array = nir_deref_array_create(mem_ctx);
+            deref_array->base_offset = i;
+            deref_array->deref.type = elem_type;
+
+            nir_deref_var *new_deref_head =
+               nir_deref_as_var(nir_copy_deref(mem_ctx, &deref_head->deref));
+            get_deref_tail(&new_deref_head->deref)->child =
+               &deref_array->deref;
+
+            nir_reg_src new_reg = reg;
+            new_reg.base_offset += i * type_size(elem_type);
+
+            var_reg_block_copy_impl(new_reg, new_deref_head, predicate,
+                                    elem_type, after, var_dest, mem_ctx);
+         }
+         break;
+
+      default:
+         break;
+   }
+}
+
+static nir_intrinsic_op
+get_load_op(nir_variable_mode mode, bool indirect, unsigned num_components)
+{
+   if (indirect) {
+      switch (mode) {
+         case nir_var_shader_in:
+            switch (num_components) {
+               case 1: return nir_intrinsic_load_input_vec1_indirect;
+               case 2: return nir_intrinsic_load_input_vec2_indirect;
+               case 3: return nir_intrinsic_load_input_vec3_indirect;
+               case 4: return nir_intrinsic_load_input_vec4_indirect;
+               default: assert(0); break;
+            }
+            break;
+
+         case nir_var_uniform:
+            switch (num_components) {
+               case 1: return nir_intrinsic_load_uniform_vec1_indirect;
+               case 2: return nir_intrinsic_load_uniform_vec2_indirect;
+               case 3: return nir_intrinsic_load_uniform_vec3_indirect;
+               case 4: return nir_intrinsic_load_uniform_vec4_indirect;
+               default: assert(0); break;
+            }
+            break;
+
+         default:
+            assert(0);
+            break;
+      }
+   } else {
+      switch (mode) {
+         case nir_var_shader_in:
+            switch (num_components) {
+               case 1: return nir_intrinsic_load_input_vec1;
+               case 2: return nir_intrinsic_load_input_vec2;
+               case 3: return nir_intrinsic_load_input_vec3;
+               case 4: return nir_intrinsic_load_input_vec4;
+               default: assert(0); break;
+            }
+            break;
+
+         case nir_var_uniform:
+            switch (num_components) {
+               case 1: return nir_intrinsic_load_uniform_vec1;
+               case 2: return nir_intrinsic_load_uniform_vec2;
+               case 3: return nir_intrinsic_load_uniform_vec3;
+               case 4: return nir_intrinsic_load_uniform_vec4;
+               default: assert(0); break;
+            }
+            break;
+
+         default:
+            assert(0);
+            break;
+      }
+   }
+
+   return nir_intrinsic_load_input_vec1;
+}
+
+/* emits an input -> reg block copy */
+
+static void
+reg_input_block_copy(nir_reg_dest dest, unsigned src_index, nir_src *indirect,
+                     nir_src *predicate, unsigned size,
+                     unsigned num_components, nir_variable_mode mode,
+                     nir_instr *after, void *mem_ctx)
+{
+   nir_intrinsic_op op = get_load_op(mode, indirect != NULL, num_components);
+
+   nir_intrinsic_instr *load = nir_intrinsic_instr_create(mem_ctx, op);
+   load->const_index[0] = src_index;
+   load->const_index[1] = size;
+   if (indirect)
+      load->src[0] = *indirect;
+   if (predicate) {
+      load->has_predicate = true;
+      load->predicate = nir_src_copy(*predicate, mem_ctx);
+   }
+   load->dest.reg = dest;
+   nir_instr_insert_before(after, &load->instr);
+}
+
+/* emits a variable/input -> register block copy */
+
+static void
+var_reg_block_copy(nir_deref_var *src, nir_reg_dest dest, nir_src *predicate,
+                   bool lower_io, nir_instr *after, nir_function_impl *impl,
+                   bool native_integers, void *mem_ctx)
+{
+   const struct glsl_type *src_type = get_deref_tail(&src->deref)->type;
+
+   if (lower_io && (src->var->data.mode == nir_var_shader_in ||
+       src->var->data.mode == nir_var_uniform)) {
+      unsigned size, num_components;
+      if (glsl_type_is_scalar(src_type) || glsl_type_is_vector(src_type)) {
+         num_components = glsl_get_vector_elements(src_type);
+         size = 1;
+      } else {
+         num_components = 1;
+         size = type_size(src_type);
+      }
+      bool has_indirect = deref_has_indirect(src);
+      nir_src indirect;
+      nir_src *indirect_ptr = has_indirect ? &indirect : NULL;
+      unsigned offset = get_deref_offset(src, after, impl, native_integers,
+                                         indirect_ptr);
+      offset += src->var->data.driver_location;
+
+      reg_input_block_copy(dest, offset, indirect_ptr, predicate, size,
+                           num_components, src->var->data.mode, after,
+                           mem_ctx);
+   } else {
+      nir_reg_src reg;
+      reg.reg = dest.reg;
+      reg.base_offset = dest.base_offset;
+      reg.indirect = dest.indirect;
+
+      var_reg_block_copy_impl(reg, src, predicate, src_type, after, false,
+                              mem_ctx);
+   }
+}
+
+/* emits a register -> variable copy */
+static void
+reg_var_block_copy(nir_reg_src src, nir_deref_var *dest, nir_src *predicate,
+                   nir_instr *after, void *mem_ctx)
+{
+   const struct glsl_type *dest_type = get_deref_tail(&dest->deref)->type;
+
+   var_reg_block_copy_impl(src, dest, predicate, dest_type, after, true,
+                           mem_ctx);
+}
+
+/*
+ * emits an input -> variable block copy using an intermediate register
+ */
+static void
+var_var_block_copy(nir_deref_var *src, nir_deref_var *dest, nir_src *predicate,
+                   nir_instr *after, nir_function_impl *impl,
+                   bool native_integers, void *mem_ctx)
+{
+   const struct glsl_type *type = get_deref_tail(&dest->deref)->type;
+   nir_register *reg = nir_local_reg_create(impl);
+   if (glsl_type_is_scalar(type) || glsl_type_is_vector(type)) {
+      reg->num_components = glsl_get_vector_elements(type);
+   } else {
+      reg->is_packed = true;
+      reg->num_components = 1;
+      reg->num_array_elems = type_size(type);
+   }
+
+   nir_reg_src reg_src;
+   reg_src.base_offset = 0;
+   reg_src.indirect = NULL;
+   reg_src.reg = reg;
+
+   nir_reg_dest reg_dest;
+   reg_dest.base_offset = 0;
+   reg_dest.indirect = NULL;
+   reg_dest.reg = reg;
+
+   var_reg_block_copy(src, reg_dest, predicate, true, after, impl,
+                      native_integers, mem_ctx);
+   reg_var_block_copy(reg_src, dest, predicate, after, mem_ctx);
+}
+
+/* emits a register -> register block copy */
+static void
+reg_reg_block_copy(nir_reg_dest dest, nir_reg_src src, nir_src *predicate,
+                   const struct glsl_type *type, nir_instr *after,
+                   void *mem_ctx)
+{
+   if (!dest.reg->is_packed && !src.reg->is_packed)
+      assert(dest.reg->num_components == src.reg->num_components);
+
+   unsigned size, num_components;
+   if (dest.reg->is_packed && src.reg->is_packed) {
+      size = type_size(type);
+      num_components = 1;
+   } else {
+      size = 1;
+      if (dest.reg->is_packed)
+         num_components = src.reg->num_components;
+      else
+         num_components = dest.reg->num_components;
+   }
+
+   for (unsigned i = 0; i < size; i++) {
+      nir_alu_instr *move = nir_alu_instr_create(mem_ctx, nir_op_imov);
+      move->dest.write_mask = (1 << num_components) - 1;
+
+      move->dest.dest.reg.reg = dest.reg;
+      move->dest.dest.reg.base_offset = dest.base_offset + i;
+      if (dest.indirect != NULL) {
+         move->dest.dest.reg.indirect = ralloc(mem_ctx, nir_src);
+         *move->dest.dest.reg.indirect = *dest.indirect;
+      }
+
+      if (predicate) {
+         move->has_predicate = true;
+         move->predicate = nir_src_copy(*predicate, mem_ctx);
+      }
+
+      move->src[0].src.reg = src;
+      move->src[0].src.reg.base_offset += i;
+
+      nir_instr_insert_before(after, &move->instr);
+   }
+}
+
+static nir_reg_dest
+create_dest(nir_deref_var *deref, nir_instr *instr, nir_register *reg,
+            nir_function_impl *impl, bool native_integers, void *mem_ctx)
+{
+   nir_reg_dest dest;
+   if (deref_has_indirect(deref)) {
+      dest.indirect = ralloc(mem_ctx, nir_src);
+      dest.indirect->is_ssa = false;
+      dest.base_offset = get_deref_offset(deref, instr,
+                                          impl, native_integers,
+                                          dest.indirect);
+   } else {
+      dest.base_offset = get_deref_offset(deref, instr,
+                                          impl, native_integers, NULL);
+      dest.indirect = NULL;
+   }
+   dest.reg = reg;
+
+   return dest;
+}
+
+static nir_reg_src
+create_src(nir_deref_var *deref, nir_instr *instr, nir_register *reg,
+           nir_function_impl *impl, bool native_integers, void *mem_ctx)
+{
+   nir_reg_src src;
+   if (deref_has_indirect(deref)) {
+      src.indirect = ralloc(mem_ctx, nir_src);
+      src.indirect->is_ssa = false;
+      src.base_offset = get_deref_offset(deref, instr,
+                                         impl, native_integers,
+                                         src.indirect);
+   } else {
+      src.base_offset = get_deref_offset(deref, instr,
+                                         impl, native_integers, NULL);
+      src.indirect = NULL;
+   }
+   src.reg = reg;
+
+   return src;
+}
+
+static void
+handle_var_copy(nir_intrinsic_instr *instr, nir_function_impl *impl,
+                bool native_integers, bool lower_io, struct hash_table *ht)
+{
+   void *mem_ctx = ralloc_parent(instr);
+
+   struct hash_entry *entry;
+
+   nir_variable *dest_var = instr->variables[0]->var;
+   nir_variable *src_var = instr->variables[1]->var;
+
+   const struct glsl_type *type =
+      get_deref_tail(&instr->variables[0]->deref)->type;
+
+   nir_src *predicate = instr->has_predicate ? &instr->predicate : NULL;
+
+   /*
+    * The source can be either:
+    * 1. a variable we're lowering to a register
+    * 2. an input or uniform we're lowering to loads from an index
+    * 3. a variable we can't lower yet
+    *
+    * and similarly, the destination can be either:
+    * 1. a variable we're lowering to a register
+    * 2. a variable we can't lower yet
+    *
+    * meaning that there are six cases, including the trivial one (where
+    * source and destination are #3 and #2 respectively) where we can't do
+    * anything.
+    */
+
+   entry = _mesa_hash_table_search(ht, _mesa_hash_pointer(dest_var), dest_var);
+   if (entry) {
+      nir_reg_dest dest = create_dest(instr->variables[0], &instr->instr,
+                                      (nir_register *) entry->data, impl,
+                                      native_integers, mem_ctx);
+
+      entry = _mesa_hash_table_search(ht, _mesa_hash_pointer(src_var), src_var);
+      if (entry) {
+         nir_reg_src src = create_src(instr->variables[1], &instr->instr,
+                                     (nir_register *) entry->data, impl,
+                                     native_integers, mem_ctx);
+
+         reg_reg_block_copy(dest, src, predicate, type, &instr->instr, mem_ctx);
+      } else {
+         var_reg_block_copy(instr->variables[1], dest, predicate, lower_io,
+                            &instr->instr, impl, native_integers, mem_ctx);
+      }
+   } else {
+      entry = _mesa_hash_table_search(ht, _mesa_hash_pointer(src_var), src_var);
+      if (entry) {
+         nir_reg_src src = create_src(instr->variables[1], &instr->instr,
+                                     (nir_register *) entry->data, impl,
+                                     native_integers, mem_ctx);
+
+         reg_var_block_copy(src, instr->variables[0], predicate, &instr->instr,
+                            mem_ctx);
+      } else {
+         if (!lower_io || (src_var->data.mode != nir_var_shader_in &&
+             src_var->data.mode != nir_var_uniform)) {
+            /* nothing to do here */
+            return;
+         }
+
+         var_var_block_copy(instr->variables[1], instr->variables[0], predicate,
+                            &instr->instr, impl, native_integers, mem_ctx);
+      }
+   }
+
+   nir_instr_remove(&instr->instr);
+}
+
+static void
+handle_var_load(nir_intrinsic_instr *instr, nir_function_impl *impl,
+                bool native_integers, bool lower_io, struct hash_table *ht)
+{
+   void *mem_ctx = ralloc_parent(instr);
+
+   struct hash_entry *entry =
+      _mesa_hash_table_search(ht, _mesa_hash_pointer(instr->variables[0]->var),
+                              instr->variables[0]->var);
+
+   if (entry == NULL) {
+      nir_variable *src_var = instr->variables[0]->var;
+
+      if (lower_io && (src_var->data.mode == nir_var_shader_in ||
+          src_var->data.mode == nir_var_uniform)) {
+         bool has_indirect = deref_has_indirect(instr->variables[0]);
+         unsigned num_components =
+            nir_intrinsic_infos[instr->intrinsic].dest_components;
+         nir_src indirect;
+         unsigned offset = get_deref_offset(instr->variables[0], &instr->instr,
+                                            impl, native_integers, &indirect);
+         offset += src_var->data.driver_location;
+
+         nir_intrinsic_op op = get_load_op(src_var->data.mode, has_indirect,
+                                           num_components);
+         nir_intrinsic_instr *load = nir_intrinsic_instr_create(mem_ctx, op);
+         load->dest = instr->dest;
+         load->const_index[0] = (int) offset;
+         load->const_index[1] = 1;
+         if (has_indirect)
+            load->src[0] = indirect;
+
+         if (instr->has_predicate) {
+            load->has_predicate = true;
+            load->predicate = nir_src_copy(instr->predicate, mem_ctx);
+         }
+
+         nir_instr_insert_before(&instr->instr, &load->instr);
+      } else {
+         return;
+      }
+   } else {
+      nir_register *reg = (nir_register *) entry->data;
+
+      nir_alu_instr *move = nir_alu_instr_create(mem_ctx, nir_op_imov);
+      unsigned dest_components =
+         nir_intrinsic_infos[instr->intrinsic].dest_components;
+      move->dest.dest = instr->dest;
+      move->dest.write_mask = (1 << dest_components) - 1;
+      move->src[0].src.reg = create_src(instr->variables[0], &instr->instr,
+                                        reg, impl, native_integers, mem_ctx);
+      if (instr->has_predicate) {
+         move->has_predicate = true;
+         move->predicate = nir_src_copy(instr->predicate, mem_ctx);
+      }
+      nir_instr_insert_before(&instr->instr, &move->instr);
+   }
+
+   nir_instr_remove(&instr->instr);
+}
+
+static void
+handle_var_store(nir_intrinsic_instr *instr, nir_function_impl *impl,
+                 bool native_integers, bool lower_io, struct hash_table *ht)
+{
+   void *mem_ctx = ralloc_parent(instr);
+
+   struct hash_entry *entry =
+      _mesa_hash_table_search(ht, _mesa_hash_pointer(instr->variables[0]->var),
+                              instr->variables[0]->var);
+   if (entry == NULL)
+         return;
+
+   nir_register *reg = (nir_register *) entry->data;
+
+   nir_alu_instr *move = nir_alu_instr_create(mem_ctx, nir_op_imov);
+   unsigned src_components =
+      nir_intrinsic_infos[instr->intrinsic].src_components[0];
+   move->dest.dest.reg = create_dest(instr->variables[0], &instr->instr,
+                                       reg, impl, native_integers, mem_ctx);
+   move->dest.write_mask = (1 << src_components) - 1;
+   move->src[0].src = instr->src[0];
+   if (instr->has_predicate) {
+      move->has_predicate = true;
+      move->predicate = nir_src_copy(instr->predicate, mem_ctx);
+   }
+   nir_instr_insert_before(&instr->instr, &move->instr);
+   nir_instr_remove(&instr->instr);
+}
+
+typedef struct {
+   struct hash_table *ht;
+   bool native_integers, lower_io;
+   nir_function_impl *impl;
+} rewrite_state;
+
+static bool
+rewrite_block_cb(nir_block *block, void *_state)
+{
+   rewrite_state *state = (rewrite_state *) _state;
+
+   nir_foreach_instr_safe(block, instr) {
+      if (instr->type == nir_instr_type_intrinsic) {
+         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+         switch (intrin->intrinsic) {
+            case nir_intrinsic_load_var_vec1:
+            case nir_intrinsic_load_var_vec2:
+            case nir_intrinsic_load_var_vec3:
+            case nir_intrinsic_load_var_vec4:
+               handle_var_load(intrin, state->impl, state->native_integers,
+                               state->lower_io, state->ht);
+               break;
+
+            case nir_intrinsic_store_var_vec1:
+            case nir_intrinsic_store_var_vec2:
+            case nir_intrinsic_store_var_vec3:
+            case nir_intrinsic_store_var_vec4:
+               handle_var_store(intrin, state->impl, state->native_integers,
+                                state->lower_io, state->ht);
+               break;
+
+            case nir_intrinsic_copy_var:
+               handle_var_copy(intrin, state->impl, state->native_integers,
+                               state->lower_io, state->ht);
+               break;
+
+            default:
+               break;
+         }
+      }
+   }
+
+   return true;
+}
+
+static void
+rewrite_impl(nir_function_impl *impl, struct hash_table *ht,
+             bool native_integers, bool lower_io)
+{
+   rewrite_state state;
+   state.ht = ht;
+   state.native_integers = native_integers;
+   state.lower_io = lower_io;
+   state.impl = impl;
+
+   nir_foreach_block(impl, rewrite_block_cb, &state);
+}
+
+static void
+insert_load_const_impl(nir_function_impl *impl, struct exec_list *vars,
+                       struct hash_table *ht)
+{
+   void *mem_ctx = ralloc_parent(impl);
+
+   foreach_list_typed(nir_variable, var, node, vars) {
+      if (var->constant_initializer == NULL)
+         continue;
+
+      struct hash_entry *entry =
+         _mesa_hash_table_search(ht, _mesa_hash_pointer(var), var);
+      if (entry) {
+         nir_register *reg = (nir_register *) entry->data;
+         nir_reg_dest dest;
+         dest.reg = reg;
+         dest.base_offset = 0;
+         dest.indirect = NULL;
+         reg_const_load(dest, var->constant_initializer, var->type, impl,
+                        mem_ctx);
+      }
+   }
+}
+
+static nir_intrinsic_op
+get_store_op(bool indirect, unsigned num_components)
+{
+   if (indirect) {
+      switch (num_components) {
+         case 1: return nir_intrinsic_store_output_vec1_indirect;
+         case 2: return nir_intrinsic_store_output_vec2_indirect;
+         case 3: return nir_intrinsic_store_output_vec3_indirect;
+         case 4: return nir_intrinsic_store_output_vec4_indirect;
+         default: assert(0); break;
+      }
+   } else {
+      switch (num_components) {
+         case 1: return nir_intrinsic_store_output_vec1;
+         case 2: return nir_intrinsic_store_output_vec2;
+         case 3: return nir_intrinsic_store_output_vec3;
+         case 4: return nir_intrinsic_store_output_vec4;
+         default: assert(0); break;
+      }
+   }
+
+   return nir_intrinsic_store_output_vec1;
+}
+
+/* emits a reg -> output block copy after a block */
+static void
+reg_output_block_copy_block(nir_reg_src src, unsigned dest_index,
+                            unsigned num_components, unsigned size,
+                            nir_block *block, void *mem_ctx)
+{
+   nir_intrinsic_op op = get_store_op(false, num_components);
+
+   nir_intrinsic_instr *store = nir_intrinsic_instr_create(mem_ctx, op);
+   store->const_index[0] = dest_index;
+   store->const_index[1] = (size == 0) ? 1 : size;
+   store->src[0].reg = src;
+   nir_instr_insert_after_block(block, &store->instr);
+}
+
+/* emits a reg -> output copy after an instruction */
+static void
+reg_output_block_copy_instr(nir_reg_src src, unsigned dest_index,
+                            unsigned num_components, unsigned size,
+                            nir_instr *after, void *mem_ctx)
+{
+   nir_intrinsic_op op = get_store_op(false, num_components);
+
+   nir_intrinsic_instr *store = nir_intrinsic_instr_create(mem_ctx, op);
+   store->const_index[0] = dest_index;
+   store->const_index[1] = (size == 0) ? 1 : size;
+   store->src[0].reg = src;
+   nir_instr_insert_before(after, &store->instr);
+}
+
+static nir_function_impl *
+find_main(nir_shader *shader)
+{
+   foreach_list_typed(nir_function, func, node, &shader->functions) {
+      if (strcmp(func->name, "main") == 0) {
+         assert(exec_list_length(&func->overload_list) == 1);
+         nir_function_overload *overload = nir_function_first_overload(func);
+         return overload->impl;
+      }
+   }
+
+   assert(0);
+   return NULL;
+}
+
+static void
+insert_output_reg_copies(nir_shader *shader, nir_block *block,
+                         nir_instr *after, struct hash_table *ht)
+{
+   struct hash_entry *entry;
+   hash_table_foreach(shader->outputs, entry) {
+      nir_variable *var = (nir_variable *) entry->data;
+
+      struct hash_entry *entry2;
+      entry2 = _mesa_hash_table_search(ht, _mesa_hash_pointer(var), var);
+      if (entry2) {
+         nir_register *reg = (nir_register *) entry2->data;
+         nir_reg_src src;
+         src.reg = reg;
+         src.base_offset = 0;
+         src.indirect = NULL;
+
+         if (after) {
+            reg_output_block_copy_instr(src, var->data.driver_location,
+                                        reg->num_components,
+                                        reg->num_array_elems,
+                                        after, shader);
+         } else {
+            reg_output_block_copy_block(src, var->data.driver_location,
+                                        reg->num_components,
+                                        reg->num_array_elems,
+                                        block, shader);
+         }
+      }
+   }
+}
+
+typedef struct {
+   struct hash_table *ht;
+   nir_shader *shader;
+   bool found_emit_vertex;
+} reg_output_state;
+
+static bool
+insert_output_reg_copies_emit_vertex(nir_block *block, void *_state)
+{
+   reg_output_state *state = (reg_output_state *) _state;
+
+   nir_foreach_instr(block, instr) {
+      if (instr->type == nir_instr_type_intrinsic) {
+         nir_intrinsic_instr *intrin_instr = nir_instr_as_intrinsic(instr);
+         if (intrin_instr->intrinsic == nir_intrinsic_emit_vertex) {
+            insert_output_reg_copies(state->shader, NULL, instr, state->ht);
+            state->found_emit_vertex = true;
+         }
+      }
+   }
+
+   return true;
+}
+
+static void
+insert_output_reg_copies_shader(nir_shader *shader, struct hash_table *ht)
+{
+   nir_function_impl *main_impl = find_main(shader);
+
+   reg_output_state state;
+   state.shader = shader;
+   state.ht = ht;
+   state.found_emit_vertex = false;
+   nir_foreach_block(main_impl, insert_output_reg_copies_emit_vertex, &state);
+
+   if (!state.found_emit_vertex) {
+      struct set_entry *entry;
+      set_foreach(main_impl->end_block->predecessors, entry) {
+         nir_block *block = (nir_block *) entry->key;
+         insert_output_reg_copies(shader, block, NULL, ht);
+      }
+   }
+}
+
+static void
+rewrite_shader(nir_shader *shader, struct hash_table *ht, bool native_integers,
+               bool lower_globals, bool lower_io)
+{
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl) {
+         insert_load_const_impl(overload->impl, &overload->impl->locals, ht);
+         if (lower_globals && strcmp(overload->function->name, "main") == 0)
+            insert_load_const_impl(overload->impl, &shader->globals, ht);
+         rewrite_impl(overload->impl, ht, native_integers, lower_io);
+      }
+   }
+}
+
+void
+nir_lower_variables_scalar(nir_shader *shader, bool lower_globals,
+                           bool lower_io, bool add_names, bool native_integers)
+{
+   if (lower_io)
+      assign_var_locations_shader(shader);
+   struct hash_table *ht = init_var_ht(shader, lower_globals, lower_io,
+                                       add_names);
+   remove_local_vars_shader(shader, ht);
+   rewrite_shader(shader, ht, native_integers, lower_globals, lower_io);
+   if (lower_io)
+      insert_output_reg_copies_shader(shader, ht);
+   _mesa_hash_table_destroy(ht, NULL);
+}
-- 
2.2.0



More information about the mesa-dev mailing list