*/
/*
- * Implements a pass that lowers output variables to a temporary plus an
- * output variable with a single copy at each exit point of the shader.
- * This way the output variable is only ever written.
+ * Implements a pass that lowers output and/or input variables to a
+ * temporary plus an output variable with a single copy at each exit
+ * point of the shader and/or an input variable with a single copy
+ * at the entrance point of the shader. This way the output variable
+ * is only ever written once and/or input is only read once, and there
+ * are no indirect outut/input accesses.
*/
#include "nir.h"
+#include "nir_builder.h"
struct lower_io_state {
nir_shader *shader;
+ nir_function_impl *entrypoint;
struct exec_list old_outputs;
+ struct exec_list old_inputs;
};
static void
-emit_output_copies(nir_cursor cursor, struct lower_io_state *state)
+emit_copies(nir_builder *b, struct exec_list *dest_vars,
+ struct exec_list *src_vars)
{
- assert(exec_list_length(&state->shader->outputs) ==
- exec_list_length(&state->old_outputs));
+ assert(exec_list_length(dest_vars) == exec_list_length(src_vars));
+
+ foreach_two_lists(dest_node, dest_vars, src_node, src_vars) {
+ nir_variable *dest = exec_node_data(nir_variable, dest_node, node);
+ nir_variable *src = exec_node_data(nir_variable, src_node, node);
+
+ /* No need to copy the contents of a non-fb_fetch_output output variable
+ * to the temporary allocated for it, since its initial value is
+ * undefined.
+ */
+ if (src->data.mode == nir_var_shader_out &&
+ !src->data.fb_fetch_output)
+ continue;
- foreach_two_lists(out_node, &state->shader->outputs,
- temp_node, &state->old_outputs) {
- nir_variable *output = exec_node_data(nir_variable, out_node, node);
- nir_variable *temp = exec_node_data(nir_variable, temp_node, node);
+ /* Can't copy the contents of the temporary back to a read-only
+ * interface variable. The value of the temporary won't have been
+ * modified by the shader anyway.
+ */
+ if (dest->data.read_only)
+ continue;
- nir_intrinsic_instr *copy =
- nir_intrinsic_instr_create(state->shader, nir_intrinsic_copy_var);
- copy->variables[0] = nir_deref_var_create(copy, output);
- copy->variables[1] = nir_deref_var_create(copy, temp);
+ nir_copy_var(b, dest, src);
+ }
+}
- nir_instr_insert(cursor, ©->instr);
+static void
+emit_output_copies_impl(struct lower_io_state *state, nir_function_impl *impl)
+{
+ nir_builder b;
+ nir_builder_init(&b, impl);
+
+ if (state->shader->info.stage == MESA_SHADER_GEOMETRY) {
+ /* For geometry shaders, we have to emit the output copies right
+ * before each EmitVertex call.
+ */
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ if (intrin->intrinsic == nir_intrinsic_emit_vertex) {
+ b.cursor = nir_before_instr(&intrin->instr);
+ emit_copies(&b, &state->shader->outputs, &state->old_outputs);
+ }
+ }
+ }
+ } else if (impl == state->entrypoint) {
+ b.cursor = nir_before_block(nir_start_block(impl));
+ emit_copies(&b, &state->old_outputs, &state->shader->outputs);
+
+ /* For all other shader types, we need to do the copies right before
+ * the jumps to the end block.
+ */
+ struct set_entry *block_entry;
+ set_foreach(impl->end_block->predecessors, block_entry) {
+ struct nir_block *block = (void *)block_entry->key;
+ b.cursor = nir_after_block_before_jump(block);
+ emit_copies(&b, &state->shader->outputs, &state->old_outputs);
+ }
}
}
+static void
+emit_input_copies_impl(struct lower_io_state *state, nir_function_impl *impl)
+{
+ if (impl == state->entrypoint) {
+ nir_builder b;
+ nir_builder_init(&b, impl);
+ b.cursor = nir_before_block(nir_start_block(impl));
+ emit_copies(&b, &state->old_inputs, &state->shader->inputs);
+ }
+}
+
+static nir_variable *
+create_shadow_temp(struct lower_io_state *state, nir_variable *var)
+{
+ nir_variable *nvar = ralloc(state->shader, nir_variable);
+ memcpy(nvar, var, sizeof *nvar);
+
+ /* The original is now the temporary */
+ nir_variable *temp = var;
+
+ /* Reparent the name to the new variable */
+ ralloc_steal(nvar, nvar->name);
+
+ assert(nvar->constant_initializer == NULL);
+
+ /* Give the original a new name with @<mode>-temp appended */
+ const char *mode = (temp->data.mode == nir_var_shader_in) ? "in" : "out";
+ temp->name = ralloc_asprintf(var, "%s@%s-temp", mode, nvar->name);
+ temp->data.mode = nir_var_global;
+ temp->data.read_only = false;
+ temp->data.fb_fetch_output = false;
+ temp->data.compact = false;
+
+ return nvar;
+}
+
void
-nir_lower_io_to_temporaries(nir_shader *shader, nir_function *entrypoint)
+nir_lower_io_to_temporaries(nir_shader *shader, nir_function_impl *entrypoint,
+ bool outputs, bool inputs)
{
struct lower_io_state state;
- if (shader->stage == MESA_SHADER_TESS_CTRL)
+ if (shader->info.stage == MESA_SHADER_TESS_CTRL)
return;
state.shader = shader;
- exec_list_move_nodes_to(&shader->outputs, &state.old_outputs);
+ state.entrypoint = entrypoint;
+
+ if (inputs)
+ exec_list_move_nodes_to(&shader->inputs, &state.old_inputs);
+ else
+ exec_list_make_empty(&state.old_inputs);
+
+ if (outputs)
+ exec_list_move_nodes_to(&shader->outputs, &state.old_outputs);
+ else
+ exec_list_make_empty(&state.old_outputs);
/* Walk over all of the outputs turn each output into a temporary and
* make a new variable for the actual output.
*/
nir_foreach_variable(var, &state.old_outputs) {
- nir_variable *output = ralloc(shader, nir_variable);
- memcpy(output, var, sizeof *output);
-
- /* The orignal is now the temporary */
- nir_variable *temp = var;
-
- /* Reparent the name to the new variable */
- ralloc_steal(output, output->name);
-
- /* Reparent the constant initializer (if any) */
- ralloc_steal(output, output->constant_initializer);
-
- /* Give the output a new name with @out-temp appended */
- temp->name = ralloc_asprintf(var, "%s@out-temp", output->name);
- temp->data.mode = nir_var_global;
- temp->constant_initializer = NULL;
-
+ nir_variable *output = create_shadow_temp(&state, var);
exec_list_push_tail(&shader->outputs, &output->node);
}
+ /* and same for inputs: */
+ nir_foreach_variable(var, &state.old_inputs) {
+ nir_variable *input = create_shadow_temp(&state, var);
+ exec_list_push_tail(&shader->inputs, &input->node);
+ }
+
nir_foreach_function(function, shader) {
if (function->impl == NULL)
continue;
- if (shader->stage == MESA_SHADER_GEOMETRY) {
- /* For geometry shaders, we have to emit the output copies right
- * before each EmitVertex call.
- */
- nir_foreach_block(block, function->impl) {
- nir_foreach_instr(instr, block) {
- if (instr->type != nir_instr_type_intrinsic)
- continue;
-
- nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
- if (intrin->intrinsic == nir_intrinsic_emit_vertex) {
- emit_output_copies(nir_before_instr(&intrin->instr),
- &state);
- }
- }
- }
- } else if (function == entrypoint) {
- /* For all other shader types, we need to do the copies right before
- * the jumps to the end block.
- */
- struct set_entry *block_entry;
- set_foreach(function->impl->end_block->predecessors, block_entry) {
- struct nir_block *block = (void *)block_entry->key;
- emit_output_copies(nir_after_block_before_jump(block), &state);
- }
- }
+ if (inputs)
+ emit_input_copies_impl(&state, function->impl);
+
+ if (outputs)
+ emit_output_copies_impl(&state, function->impl);
nir_metadata_preserve(function->impl, nir_metadata_block_index |
nir_metadata_dominance);
}
+ exec_list_append(&shader->globals, &state.old_inputs);
exec_list_append(&shader->globals, &state.old_outputs);
+
+ nir_fixup_deref_modes(shader);
}