static bool
split_var_copies_block(nir_block *block, void *void_state)
{
struct split_var_copies_state *state = void_state;
nir_foreach_instr_safe(block, instr) {
if (instr->type != nir_instr_type_intrinsic)
continue;
nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr);
if (intrinsic->intrinsic != nir_intrinsic_copy_var)
continue;
nir_deref *dest_head = &intrinsic->variables[0]->deref;
nir_deref *src_head = &intrinsic->variables[1]->deref;
nir_deref *dest_tail = nir_deref_tail(dest_head);
nir_deref *src_tail = nir_deref_tail(src_head);
switch (glsl_get_base_type(src_tail->type)) {
case GLSL_TYPE_ARRAY:
case GLSL_TYPE_STRUCT:
split_var_copy_instr(intrinsic, dest_head, src_head,
dest_tail, src_tail, state);
nir_instr_remove(&intrinsic->instr);
ralloc_steal(state->dead_ctx, instr);
break;
case GLSL_TYPE_FLOAT:
case GLSL_TYPE_INT:
case GLSL_TYPE_UINT:
case GLSL_TYPE_BOOL:
if (glsl_type_is_matrix(src_tail->type)) {
split_var_copy_instr(intrinsic, dest_head, src_head,
dest_tail, src_tail, state);
nir_instr_remove(&intrinsic->instr);
ralloc_steal(state->dead_ctx, instr);
}
break;
default:
unreachable("Invalid type");
break;
}
}
return true;
}
/* This function recursively walks the given deref chain and replaces the
* given copy instruction with an equivalent sequence load/store
* operations.
*
* @copy_instr The copy instruction to replace; new instructions will be
* inserted before this one
*
* @dest_head The head of the destination variable deref chain
*
* @src_head The head of the source variable deref chain
*
* @dest_tail The current tail of the destination variable deref chain;
* this is used for recursion and external callers of this
* function should call it with tail == head
*
* @src_tail The current tail of the source variable deref chain;
* this is used for recursion and external callers of this
* function should call it with tail == head
*
* @state The current variable lowering state
*/
static void
emit_copy_load_store(nir_intrinsic_instr *copy_instr,
nir_deref_var *dest_head, nir_deref_var *src_head,
nir_deref *dest_tail, nir_deref *src_tail, void *mem_ctx)
{
/* Find the next pair of wildcards */
nir_deref *src_arr_parent = deref_next_wildcard_parent(src_tail);
nir_deref *dest_arr_parent = deref_next_wildcard_parent(dest_tail);
if (src_arr_parent || dest_arr_parent) {
/* Wildcards had better come in matched pairs */
assert(dest_arr_parent && dest_arr_parent);
nir_deref_array *src_arr = nir_deref_as_array(src_arr_parent->child);
nir_deref_array *dest_arr = nir_deref_as_array(dest_arr_parent->child);
unsigned length = glsl_get_length(src_arr_parent->type);
/* The wildcards should represent the same number of elements */
assert(length == glsl_get_length(dest_arr_parent->type));
assert(length > 0);
/* Walk over all of the elements that this wildcard refers to and
* call emit_copy_load_store on each one of them */
src_arr->deref_array_type = nir_deref_array_type_direct;
dest_arr->deref_array_type = nir_deref_array_type_direct;
for (unsigned i = 0; i < length; i++) {
src_arr->base_offset = i;
dest_arr->base_offset = i;
emit_copy_load_store(copy_instr, dest_head, src_head,
&dest_arr->deref, &src_arr->deref, mem_ctx);
}
src_arr->deref_array_type = nir_deref_array_type_wildcard;
dest_arr->deref_array_type = nir_deref_array_type_wildcard;
} else {
/* In this case, we have no wildcards anymore, so all we have to do
* is just emit the load and store operations. */
src_tail = nir_deref_tail(src_tail);
dest_tail = nir_deref_tail(dest_tail);
assert(src_tail->type == dest_tail->type);
unsigned num_components = glsl_get_vector_elements(src_tail->type);
nir_intrinsic_instr *load =
nir_intrinsic_instr_create(mem_ctx, nir_intrinsic_load_var);
load->num_components = num_components;
load->variables[0] = nir_deref_as_var(nir_copy_deref(load, &src_head->deref));
nir_ssa_dest_init(&load->instr, &load->dest, num_components, NULL);
nir_instr_insert_before(©_instr->instr, &load->instr);
nir_intrinsic_instr *store =
nir_intrinsic_instr_create(mem_ctx, nir_intrinsic_store_var);
store->num_components = num_components;
store->const_index[0] = (1 << num_components) - 1;
store->variables[0] = nir_deref_as_var(nir_copy_deref(store, &dest_head->deref));
store->src[0].is_ssa = true;
store->src[0].ssa = &load->dest.ssa;
nir_instr_insert_before(©_instr->instr, &store->instr);
}
}
static void
validate_intrinsic_instr(nir_intrinsic_instr *instr, validate_state *state)
{
unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
for (unsigned i = 0; i < num_srcs; i++) {
unsigned components_read =
nir_intrinsic_infos[instr->intrinsic].src_components[i];
if (components_read == 0)
components_read = instr->num_components;
validate_assert(state, components_read > 0);
if (instr->src[i].is_ssa) {
validate_assert(state, components_read <= instr->src[i].ssa->num_components);
} else if (!instr->src[i].reg.reg->is_packed) {
validate_assert(state, components_read <= instr->src[i].reg.reg->num_components);
}
validate_src(&instr->src[i], state);
}
unsigned num_vars = nir_intrinsic_infos[instr->intrinsic].num_variables;
for (unsigned i = 0; i < num_vars; i++) {
validate_deref_var(instr, instr->variables[i], state);
}
if (nir_intrinsic_infos[instr->intrinsic].has_dest) {
unsigned components_written =
nir_intrinsic_infos[instr->intrinsic].dest_components;
if (components_written == 0)
components_written = instr->num_components;
validate_assert(state, components_written > 0);
if (instr->dest.is_ssa) {
validate_assert(state, components_written <= instr->dest.ssa.num_components);
} else if (!instr->dest.reg.reg->is_packed) {
validate_assert(state, components_written <= instr->dest.reg.reg->num_components);
}
validate_dest(&instr->dest, state);
}
switch (instr->intrinsic) {
case nir_intrinsic_load_var: {
const struct glsl_type *type =
nir_deref_tail(&instr->variables[0]->deref)->type;
validate_assert(state, glsl_type_is_vector_or_scalar(type) ||
(instr->variables[0]->var->data.mode == nir_var_uniform &&
glsl_get_base_type(type) == GLSL_TYPE_SUBROUTINE));
validate_assert(state, instr->num_components == glsl_get_vector_elements(type));
break;
}
case nir_intrinsic_store_var: {
const struct glsl_type *type =
nir_deref_tail(&instr->variables[0]->deref)->type;
validate_assert(state, glsl_type_is_vector_or_scalar(type) ||
(instr->variables[0]->var->data.mode == nir_var_uniform &&
glsl_get_base_type(type) == GLSL_TYPE_SUBROUTINE));
validate_assert(state, instr->num_components == glsl_get_vector_elements(type));
validate_assert(state, instr->variables[0]->var->data.mode != nir_var_shader_in &&
instr->variables[0]->var->data.mode != nir_var_uniform &&
instr->variables[0]->var->data.mode != nir_var_shader_storage);
validate_assert(state, (nir_intrinsic_write_mask(instr) & ~((1 << instr->num_components) - 1)) == 0);
break;
}
case nir_intrinsic_copy_var:
validate_assert(state, nir_deref_tail(&instr->variables[0]->deref)->type ==
nir_deref_tail(&instr->variables[1]->deref)->type);
validate_assert(state, instr->variables[0]->var->data.mode != nir_var_shader_in &&
instr->variables[0]->var->data.mode != nir_var_uniform &&
instr->variables[0]->var->data.mode != nir_var_shader_storage);
break;
default:
break;
}
}