static void
validate_deref_chain(nir_deref *deref, validate_state *state)
{
assert(deref->child == NULL || ralloc_parent(deref->child) == deref);
nir_deref *parent = NULL;
while (deref != NULL) {
switch (deref->deref_type) {
case nir_deref_type_array:
assert(deref->type == glsl_get_array_element(parent->type));
if (nir_deref_as_array(deref)->deref_array_type ==
nir_deref_array_type_indirect)
validate_src(&nir_deref_as_array(deref)->indirect, state);
break;
case nir_deref_type_struct:
assert(deref->type ==
glsl_get_struct_field(parent->type,
nir_deref_as_struct(deref)->index));
break;
case nir_deref_type_var:
break;
default:
assert(!"Invalid deref type");
break;
}
parent = deref;
deref = deref->child;
}
}
static void
validate_deref_chain(nir_deref *deref, nir_variable_mode mode,
validate_state *state)
{
validate_assert(state, deref->child == NULL || ralloc_parent(deref->child) == deref);
nir_deref *parent = NULL;
while (deref != NULL) {
switch (deref->deref_type) {
case nir_deref_type_array:
if (mode == nir_var_shared) {
/* Shared variables have a bit more relaxed rules because we need
* to be able to handle array derefs on vectors. Fortunately,
* nir_lower_io handles these just fine.
*/
validate_assert(state, glsl_type_is_array(parent->type) ||
glsl_type_is_matrix(parent->type) ||
glsl_type_is_vector(parent->type));
} else {
/* Most of NIR cannot handle array derefs on vectors */
validate_assert(state, glsl_type_is_array(parent->type) ||
glsl_type_is_matrix(parent->type));
}
validate_assert(state, deref->type == glsl_get_array_element(parent->type));
if (nir_deref_as_array(deref)->deref_array_type ==
nir_deref_array_type_indirect)
validate_src(&nir_deref_as_array(deref)->indirect, state, 32, 1);
break;
case nir_deref_type_struct:
assume(parent); /* cannot happen: deref change starts w/ nir_deref_var */
validate_assert(state, deref->type ==
glsl_get_struct_field(parent->type,
nir_deref_as_struct(deref)->index));
break;
case nir_deref_type_var:
break;
default:
validate_assert(state, !"Invalid deref type");
break;
}
parent = deref;
deref = deref->child;
}
}
static bool
constant_fold_deref(nir_instr *instr, nir_deref_var *deref)
{
bool progress = false;
for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) {
if (tail->deref_type != nir_deref_type_array)
continue;
nir_deref_array *arr = nir_deref_as_array(tail);
if (arr->deref_array_type == nir_deref_array_type_indirect &&
arr->indirect.is_ssa &&
arr->indirect.ssa->parent_instr->type == nir_instr_type_load_const) {
nir_load_const_instr *indirect =
nir_instr_as_load_const(arr->indirect.ssa->parent_instr);
arr->base_offset += indirect->value.u[0];
/* Clear out the source */
nir_instr_rewrite_src(instr, &arr->indirect, nir_src_for_ssa(NULL));
arr->deref_array_type = nir_deref_array_type_direct;
progress = true;
}
}
return progress;
}
static bool
deref_has_indirect(nir_deref_var *deref)
{
for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) {
if (tail->deref_type == nir_deref_type_array) {
nir_deref_array *arr = nir_deref_as_array(tail);
if (arr->deref_array_type == nir_deref_array_type_indirect)
return true;
}
}
return false;
}
/* Calculate the sampler index based on array indicies and also
* calculate the base uniform location for struct members.
*/
static void
calc_sampler_offsets(nir_deref *tail, nir_tex_instr *instr,
unsigned *array_elements, nir_ssa_def **indirect,
nir_builder *b, unsigned *location)
{
if (tail->child == NULL)
return;
switch (tail->child->deref_type) {
case nir_deref_type_array: {
nir_deref_array *deref_array = nir_deref_as_array(tail->child);
assert(deref_array->deref_array_type != nir_deref_array_type_wildcard);
calc_sampler_offsets(tail->child, instr, array_elements,
indirect, b, location);
instr->sampler_index += deref_array->base_offset * *array_elements;
if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
nir_ssa_def *mul =
nir_imul(b, nir_imm_int(b, *array_elements),
nir_ssa_for_src(b, deref_array->indirect, 1));
nir_instr_rewrite_src(&instr->instr, &deref_array->indirect,
NIR_SRC_INIT);
if (*indirect) {
*indirect = nir_iadd(b, *indirect, mul);
} else {
*indirect = mul;
}
}
*array_elements *= glsl_get_length(tail->type);
break;
}
case nir_deref_type_struct: {
nir_deref_struct *deref_struct = nir_deref_as_struct(tail->child);
*location += glsl_get_record_location_offset(tail->type, deref_struct->index);
calc_sampler_offsets(tail->child, instr, array_elements,
indirect, b, location);
break;
}
default:
unreachable("Invalid deref type");
break;
}
}
/* Walks down the deref chain and returns the next deref in the chain whose
* child is a wildcard. In other words, given the chain a[1].foo[*].bar,
* this function will return the deref to foo. Calling it a second time
* with the [*].bar, it will return NULL.
*/
static nir_deref *
deref_next_wildcard_parent(nir_deref *deref)
{
for (nir_deref *tail = deref; tail->child; tail = tail->child) {
if (tail->child->deref_type != nir_deref_type_array)
continue;
nir_deref_array *arr = nir_deref_as_array(tail->child);
if (arr->deref_array_type == nir_deref_array_type_wildcard)
return tail;
}
return NULL;
}
static void
emit_load_store(nir_builder *b, nir_intrinsic_instr *orig_instr,
nir_deref_var *deref, nir_deref *tail,
nir_ssa_def **dest, nir_ssa_def *src)
{
for (; tail->child; tail = tail->child) {
if (tail->child->deref_type != nir_deref_type_array)
continue;
nir_deref_array *arr = nir_deref_as_array(tail->child);
if (arr->deref_array_type != nir_deref_array_type_indirect)
continue;
int length = glsl_get_length(tail->type);
emit_indirect_load_store(b, orig_instr, deref, tail, -arr->base_offset,
length - arr->base_offset, dest, src);
return;
}
assert(tail && tail->child == NULL);
/* We reached the end of the deref chain. Emit the instruction */
if (src == NULL) {
/* This is a load instruction */
nir_intrinsic_instr *load =
nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_var);
load->num_components = orig_instr->num_components;
load->variables[0] =
nir_deref_as_var(nir_copy_deref(load, &deref->deref));
unsigned bit_size = orig_instr->dest.ssa.bit_size;
nir_ssa_dest_init(&load->instr, &load->dest,
load->num_components, bit_size, NULL);
nir_builder_instr_insert(b, &load->instr);
*dest = &load->dest.ssa;
} else {
/* This is a store instruction */
nir_intrinsic_instr *store =
nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_var);
store->num_components = orig_instr->num_components;
nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(orig_instr));
store->variables[0] =
nir_deref_as_var(nir_copy_deref(store, &deref->deref));
store->src[0] = nir_src_for_ssa(src);
nir_builder_instr_insert(b, &store->instr);
}
}
/* \sa deref_may_be_aliased */
static bool
deref_may_be_aliased_node(struct deref_node *node, nir_deref *deref,
struct lower_variables_state *state)
{
if (deref->child == NULL) {
return false;
} else {
switch (deref->child->deref_type) {
case nir_deref_type_array: {
nir_deref_array *arr = nir_deref_as_array(deref->child);
if (arr->deref_array_type == nir_deref_array_type_indirect)
return true;
/* If there is an indirect at this level, we're aliased. */
if (node->indirect)
return true;
assert(arr->deref_array_type == nir_deref_array_type_direct);
if (node->children[arr->base_offset] &&
deref_may_be_aliased_node(node->children[arr->base_offset],
deref->child, state))
return true;
if (node->wildcard &&
deref_may_be_aliased_node(node->wildcard, deref->child, state))
return true;
return false;
}
case nir_deref_type_struct: {
nir_deref_struct *str = nir_deref_as_struct(deref->child);
if (node->children[str->index]) {
return deref_may_be_aliased_node(node->children[str->index],
deref->child, state);
} else {
return false;
}
}
default:
unreachable("Invalid nir_deref child type");
}
}
}
struct vtn_ssa_value *
vtn_local_load(struct vtn_builder *b, nir_deref_var *src)
{
nir_deref *src_tail = get_deref_tail(src);
struct vtn_ssa_value *val = vtn_create_ssa_value(b, src_tail->type);
_vtn_local_load_store(b, true, src, src_tail, val);
if (src_tail->child) {
nir_deref_array *vec_deref = nir_deref_as_array(src_tail->child);
assert(vec_deref->deref.child == NULL);
val->type = vec_deref->deref.type;
if (vec_deref->deref_array_type == nir_deref_array_type_direct)
val->def = vtn_vector_extract(b, val->def, vec_deref->base_offset);
else
val->def = vtn_vector_extract_dynamic(b, val->def,
vec_deref->indirect.ssa);
}
return val;
}
void
vtn_local_store(struct vtn_builder *b, struct vtn_ssa_value *src,
nir_deref_var *dest)
{
nir_deref *dest_tail = get_deref_tail(dest);
if (dest_tail->child) {
struct vtn_ssa_value *val = vtn_create_ssa_value(b, dest_tail->type);
_vtn_local_load_store(b, true, dest, dest_tail, val);
nir_deref_array *deref = nir_deref_as_array(dest_tail->child);
assert(deref->deref.child == NULL);
if (deref->deref_array_type == nir_deref_array_type_direct)
val->def = vtn_vector_insert(b, val->def, src->def,
deref->base_offset);
else
val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
deref->indirect.ssa);
_vtn_local_load_store(b, false, dest, dest_tail, val);
} else {
_vtn_local_load_store(b, false, dest, dest_tail, src);
}
}
/* \sa foreach_deref_node_match */
static bool
foreach_deref_node_worker(struct deref_node *node, nir_deref *deref,
bool (* cb)(struct deref_node *node,
struct lower_variables_state *state),
struct lower_variables_state *state)
{
if (deref->child == NULL) {
return cb(node, state);
} else {
switch (deref->child->deref_type) {
case nir_deref_type_array: {
nir_deref_array *arr = nir_deref_as_array(deref->child);
assert(arr->deref_array_type == nir_deref_array_type_direct);
if (node->children[arr->base_offset] &&
!foreach_deref_node_worker(node->children[arr->base_offset],
deref->child, cb, state))
return false;
if (node->wildcard &&
!foreach_deref_node_worker(node->wildcard,
deref->child, cb, state))
return false;
return true;
}
case nir_deref_type_struct: {
nir_deref_struct *str = nir_deref_as_struct(deref->child);
return foreach_deref_node_worker(node->children[str->index],
deref->child, cb, state);
}
default:
unreachable("Invalid deref child type");
}
}
}
/* 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
lower_instr(nir_intrinsic_instr *instr,
lower_atomic_state *state)
{
nir_intrinsic_op op;
switch (instr->intrinsic) {
case nir_intrinsic_atomic_counter_read_var:
op = nir_intrinsic_atomic_counter_read;
break;
case nir_intrinsic_atomic_counter_inc_var:
op = nir_intrinsic_atomic_counter_inc;
break;
case nir_intrinsic_atomic_counter_dec_var:
op = nir_intrinsic_atomic_counter_dec;
break;
default:
return;
}
if (instr->variables[0]->var->data.mode != nir_var_uniform &&
instr->variables[0]->var->data.mode != nir_var_shader_storage)
return; /* atomics passed as function arguments can't be lowered */
void *mem_ctx = ralloc_parent(instr);
unsigned uniform_loc = instr->variables[0]->var->data.location;
nir_intrinsic_instr *new_instr = nir_intrinsic_instr_create(mem_ctx, op);
new_instr->const_index[0] =
state->shader_program->UniformStorage[uniform_loc].opaque[state->shader->stage].index;
nir_load_const_instr *offset_const = nir_load_const_instr_create(mem_ctx, 1);
offset_const->value.u[0] = instr->variables[0]->var->data.atomic.offset;
nir_instr_insert_before(&instr->instr, &offset_const->instr);
nir_ssa_def *offset_def = &offset_const->def;
nir_deref *tail = &instr->variables[0]->deref;
while (tail->child != NULL) {
assert(tail->child->deref_type == nir_deref_type_array);
nir_deref_array *deref_array = nir_deref_as_array(tail->child);
tail = tail->child;
unsigned child_array_elements = tail->child != NULL ?
glsl_get_aoa_size(tail->type) : 1;
offset_const->value.u[0] += deref_array->base_offset *
child_array_elements * ATOMIC_COUNTER_SIZE;
if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
nir_load_const_instr *atomic_counter_size =
nir_load_const_instr_create(mem_ctx, 1);
atomic_counter_size->value.u[0] = child_array_elements * ATOMIC_COUNTER_SIZE;
nir_instr_insert_before(&instr->instr, &atomic_counter_size->instr);
nir_alu_instr *mul = nir_alu_instr_create(mem_ctx, nir_op_imul);
nir_ssa_dest_init(&mul->instr, &mul->dest.dest, 1, NULL);
mul->dest.write_mask = 0x1;
nir_src_copy(&mul->src[0].src, &deref_array->indirect, mul);
mul->src[1].src.is_ssa = true;
mul->src[1].src.ssa = &atomic_counter_size->def;
nir_instr_insert_before(&instr->instr, &mul->instr);
nir_alu_instr *add = nir_alu_instr_create(mem_ctx, nir_op_iadd);
nir_ssa_dest_init(&add->instr, &add->dest.dest, 1, NULL);
add->dest.write_mask = 0x1;
add->src[0].src.is_ssa = true;
add->src[0].src.ssa = &mul->dest.dest.ssa;
add->src[1].src.is_ssa = true;
add->src[1].src.ssa = offset_def;
nir_instr_insert_before(&instr->instr, &add->instr);
offset_def = &add->dest.dest.ssa;
}
}
new_instr->src[0].is_ssa = true;
new_instr->src[0].ssa = offset_def;
if (instr->dest.is_ssa) {
nir_ssa_dest_init(&new_instr->instr, &new_instr->dest,
instr->dest.ssa.num_components, NULL);
nir_ssa_def_rewrite_uses(&instr->dest.ssa,
nir_src_for_ssa(&new_instr->dest.ssa));
} else {
nir_dest_copy(&new_instr->dest, &instr->dest, mem_ctx);
}
nir_instr_insert_before(&instr->instr, &new_instr->instr);
nir_instr_remove(&instr->instr);
}
/* Gets the deref_node for the given deref chain and creates it if it
* doesn't yet exist. If the deref is fully-qualified and direct and
* state->add_to_direct_deref_nodes is true, it will be added to the hash
* table of of fully-qualified direct derefs.
*/
static struct deref_node *
get_deref_node(nir_deref_var *deref, struct lower_variables_state *state)
{
bool is_direct = true;
/* Start at the base of the chain. */
struct deref_node *node = get_deref_node_for_var(deref->var, state);
assert(deref->deref.type == node->type);
for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) {
switch (tail->deref_type) {
case nir_deref_type_struct: {
nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
assert(deref_struct->index < glsl_get_length(node->type));
if (node->children[deref_struct->index] == NULL)
node->children[deref_struct->index] =
deref_node_create(node, tail->type, state->dead_ctx);
node = node->children[deref_struct->index];
break;
}
case nir_deref_type_array: {
nir_deref_array *arr = nir_deref_as_array(tail);
switch (arr->deref_array_type) {
case nir_deref_array_type_direct:
/* This is possible if a loop unrolls and generates an
* out-of-bounds offset. We need to handle this at least
* somewhat gracefully.
*/
if (arr->base_offset >= glsl_get_length(node->type))
return NULL;
if (node->children[arr->base_offset] == NULL)
node->children[arr->base_offset] =
deref_node_create(node, tail->type, state->dead_ctx);
node = node->children[arr->base_offset];
break;
case nir_deref_array_type_indirect:
if (node->indirect == NULL)
node->indirect = deref_node_create(node, tail->type,
state->dead_ctx);
node = node->indirect;
is_direct = false;
break;
case nir_deref_array_type_wildcard:
if (node->wildcard == NULL)
node->wildcard = deref_node_create(node, tail->type,
state->dead_ctx);
node = node->wildcard;
is_direct = false;
break;
default:
unreachable("Invalid array deref type");
}
break;
}
default:
unreachable("Invalid deref type");
}
}
assert(node);
/* Only insert if it isn't already in the list. */
if (is_direct && state->add_to_direct_deref_nodes &&
node->direct_derefs_link.next == NULL) {
node->deref = deref;
assert(deref->var != NULL);
exec_list_push_tail(&state->direct_deref_nodes,
&node->direct_derefs_link);
}
return node;
}
static nir_src
get_deref_reg_src(nir_deref_var *deref, nir_instr *instr,
struct locals_to_regs_state *state)
{
nir_src src;
src.is_ssa = false;
src.reg.reg = get_reg_for_deref(deref, state);
src.reg.base_offset = 0;
src.reg.indirect = NULL;
/* It is possible for a user to create a shader that has an array with a
* single element and then proceed to access it indirectly. Indirectly
* accessing a non-array register is not allowed in NIR. In order to
* handle this case we just convert it to a direct reference.
*/
if (src.reg.reg->num_array_elems == 0)
return src;
nir_deref *tail = &deref->deref;
while (tail->child != NULL) {
const struct glsl_type *parent_type = tail->type;
tail = tail->child;
if (tail->deref_type != nir_deref_type_array)
continue;
nir_deref_array *deref_array = nir_deref_as_array(tail);
src.reg.base_offset *= glsl_get_length(parent_type);
src.reg.base_offset += deref_array->base_offset;
if (src.reg.indirect) {
nir_load_const_instr *load_const =
nir_load_const_instr_create(state->shader, 1, 32);
load_const->value.u32[0] = glsl_get_length(parent_type);
nir_instr_insert_before(instr, &load_const->instr);
nir_alu_instr *mul = nir_alu_instr_create(state->shader, nir_op_imul);
mul->src[0].src = *src.reg.indirect;
mul->src[1].src.is_ssa = true;
mul->src[1].src.ssa = &load_const->def;
mul->dest.write_mask = 1;
nir_ssa_dest_init(&mul->instr, &mul->dest.dest, 1, 32, NULL);
nir_instr_insert_before(instr, &mul->instr);
src.reg.indirect->is_ssa = true;
src.reg.indirect->ssa = &mul->dest.dest.ssa;
}
if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
if (src.reg.indirect == NULL) {
src.reg.indirect = ralloc(state->shader, nir_src);
nir_src_copy(src.reg.indirect, &deref_array->indirect,
state->shader);
} else {
nir_alu_instr *add = nir_alu_instr_create(state->shader,
nir_op_iadd);
add->src[0].src = *src.reg.indirect;
nir_src_copy(&add->src[1].src, &deref_array->indirect, add);
add->dest.write_mask = 1;
nir_ssa_dest_init(&add->instr, &add->dest.dest, 1, 32, NULL);
nir_instr_insert_before(instr, &add->instr);
src.reg.indirect->is_ssa = true;
src.reg.indirect->ssa = &add->dest.dest.ssa;
}
}
}
return src;
}
static void
lower_instr(nir_intrinsic_instr *instr, nir_function_impl *impl)
{
nir_intrinsic_op op;
switch (instr->intrinsic) {
case nir_intrinsic_atomic_counter_read_var:
op = nir_intrinsic_atomic_counter_read;
break;
case nir_intrinsic_atomic_counter_inc_var:
op = nir_intrinsic_atomic_counter_inc;
break;
case nir_intrinsic_atomic_counter_dec_var:
op = nir_intrinsic_atomic_counter_dec;
break;
default:
return;
}
if (instr->variables[0]->var->data.mode != nir_var_uniform)
return; /* atomics passed as function arguments can't be lowered */
void *mem_ctx = ralloc_parent(instr);
nir_intrinsic_instr *new_instr = nir_intrinsic_instr_create(mem_ctx, op);
new_instr->const_index[0] =
(int) instr->variables[0]->var->data.atomic.buffer_index;
nir_load_const_instr *offset_const = nir_load_const_instr_create(mem_ctx, 1);
offset_const->value.u[0] = instr->variables[0]->var->data.atomic.offset;
nir_instr_insert_before(&instr->instr, &offset_const->instr);
nir_ssa_def *offset_def = &offset_const->def;
if (instr->variables[0]->deref.child != NULL) {
assert(instr->variables[0]->deref.child->deref_type ==
nir_deref_type_array);
nir_deref_array *deref_array =
nir_deref_as_array(instr->variables[0]->deref.child);
assert(deref_array->deref.child == NULL);
offset_const->value.u[0] +=
deref_array->base_offset * ATOMIC_COUNTER_SIZE;
if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
nir_load_const_instr *atomic_counter_size =
nir_load_const_instr_create(mem_ctx, 1);
atomic_counter_size->value.u[0] = ATOMIC_COUNTER_SIZE;
nir_instr_insert_before(&instr->instr, &atomic_counter_size->instr);
nir_alu_instr *mul = nir_alu_instr_create(mem_ctx, nir_op_imul);
nir_ssa_dest_init(&mul->instr, &mul->dest.dest, 1, NULL);
mul->dest.write_mask = 0x1;
nir_src_copy(&mul->src[0].src, &deref_array->indirect, mem_ctx);
mul->src[1].src.is_ssa = true;
mul->src[1].src.ssa = &atomic_counter_size->def;
nir_instr_insert_before(&instr->instr, &mul->instr);
nir_alu_instr *add = nir_alu_instr_create(mem_ctx, nir_op_iadd);
nir_ssa_dest_init(&add->instr, &add->dest.dest, 1, NULL);
add->dest.write_mask = 0x1;
add->src[0].src.is_ssa = true;
add->src[0].src.ssa = &mul->dest.dest.ssa;
add->src[1].src.is_ssa = true;
add->src[1].src.ssa = &offset_const->def;
nir_instr_insert_before(&instr->instr, &add->instr);
offset_def = &add->dest.dest.ssa;
}
}
new_instr->src[0].is_ssa = true;
new_instr->src[0].ssa = offset_def;;
if (instr->dest.is_ssa) {
nir_ssa_dest_init(&new_instr->instr, &new_instr->dest,
instr->dest.ssa.num_components, NULL);
nir_ssa_def_rewrite_uses(&instr->dest.ssa,
nir_src_for_ssa(&new_instr->dest.ssa),
mem_ctx);
} else {
nir_dest_copy(&new_instr->dest, &instr->dest, mem_ctx);
}
nir_instr_insert_before(&instr->instr, &new_instr->instr);
nir_instr_remove(&instr->instr);
}
static void
lower_sampler(nir_tex_instr *instr, struct gl_shader_program *shader_program,
const struct gl_program *prog, void *mem_ctx)
{
if (instr->sampler == NULL)
return;
/* Get the name and the offset */
instr->sampler_index = 0;
bool has_indirect = false;
char *name = ralloc_strdup(mem_ctx, instr->sampler->var->name);
for (nir_deref *deref = &instr->sampler->deref;
deref->child; deref = deref->child) {
switch (deref->child->deref_type) {
case nir_deref_type_array: {
nir_deref_array *deref_array = nir_deref_as_array(deref->child);
/* XXX: We're assuming here that the indirect is the last array
* thing we have. This should be ok for now as we don't support
* arrays_of_arrays yet.
*/
assert(!has_indirect);
instr->sampler_index *= glsl_get_length(deref->type);
switch (deref_array->deref_array_type) {
case nir_deref_array_type_direct:
instr->sampler_index += deref_array->base_offset;
if (deref_array->deref.child)
ralloc_asprintf_append(&name, "[%u]", deref_array->base_offset);
break;
case nir_deref_array_type_indirect: {
assert(!has_indirect);
instr->src = reralloc(mem_ctx, instr->src, nir_tex_src,
instr->num_srcs + 1);
memset(&instr->src[instr->num_srcs], 0, sizeof *instr->src);
instr->src[instr->num_srcs].src_type = nir_tex_src_sampler_offset;
instr->num_srcs++;
nir_instr_rewrite_src(&instr->instr,
&instr->src[instr->num_srcs - 1].src,
deref_array->indirect);
instr->sampler_array_size = glsl_get_length(deref->type);
nir_src empty;
memset(&empty, 0, sizeof empty);
nir_instr_rewrite_src(&instr->instr, &deref_array->indirect, empty);
if (deref_array->deref.child)
ralloc_strcat(&name, "[0]");
break;
}
case nir_deref_array_type_wildcard:
unreachable("Cannot copy samplers");
default:
unreachable("Invalid deref array type");
}
break;
}
case nir_deref_type_struct: {
nir_deref_struct *deref_struct = nir_deref_as_struct(deref->child);
const char *field = glsl_get_struct_elem_name(deref->type,
deref_struct->index);
ralloc_asprintf_append(&name, ".%s", field);
break;
}
default:
unreachable("Invalid deref type");
break;
}
}
instr->sampler_index += get_sampler_index(shader_program, name, prog);
instr->sampler = NULL;
}
static void
emit_indirect_load_store(nir_builder *b, nir_intrinsic_instr *orig_instr,
nir_deref_var *deref, nir_deref *arr_parent,
int start, int end,
nir_ssa_def **dest, nir_ssa_def *src)
{
assert(arr_parent->child &&
arr_parent->child->deref_type == nir_deref_type_array);
nir_deref_array *arr = nir_deref_as_array(arr_parent->child);
assert(arr->deref_array_type == nir_deref_array_type_indirect);
assert(arr->indirect.is_ssa);
assert(start < end);
if (start == end - 1) {
/* Base case. Just emit the load/store op */
nir_deref_array direct = *arr;
direct.deref_array_type = nir_deref_array_type_direct;
direct.base_offset += start;
direct.indirect = NIR_SRC_INIT;
arr_parent->child = &direct.deref;
emit_load_store(b, orig_instr, deref, &arr->deref, dest, src);
arr_parent->child = &arr->deref;
} else {
int mid = start + (end - start) / 2;
nir_ssa_def *then_dest, *else_dest;
nir_if *if_stmt = nir_if_create(b->shader);
if_stmt->condition = nir_src_for_ssa(nir_ilt(b, arr->indirect.ssa,
nir_imm_int(b, mid)));
nir_cf_node_insert(b->cursor, &if_stmt->cf_node);
b->cursor = nir_after_cf_list(&if_stmt->then_list);
emit_indirect_load_store(b, orig_instr, deref, arr_parent,
start, mid, &then_dest, src);
b->cursor = nir_after_cf_list(&if_stmt->else_list);
emit_indirect_load_store(b, orig_instr, deref, arr_parent,
mid, end, &else_dest, src);
b->cursor = nir_after_cf_node(&if_stmt->cf_node);
if (src == NULL) {
/* We're a load. We need to insert a phi node */
nir_phi_instr *phi = nir_phi_instr_create(b->shader);
unsigned bit_size = then_dest->bit_size;
nir_ssa_dest_init(&phi->instr, &phi->dest,
then_dest->num_components, bit_size, NULL);
nir_phi_src *src0 = ralloc(phi, nir_phi_src);
src0->pred = nir_cf_node_as_block(nir_if_last_then_node(if_stmt));
src0->src = nir_src_for_ssa(then_dest);
exec_list_push_tail(&phi->srcs, &src0->node);
nir_phi_src *src1 = ralloc(phi, nir_phi_src);
src1->pred = nir_cf_node_as_block(nir_if_last_else_node(if_stmt));
src1->src = nir_src_for_ssa(else_dest);
exec_list_push_tail(&phi->srcs, &src1->node);
nir_builder_instr_insert(b, &phi->instr);
*dest = &phi->dest.ssa;
}
}
}
static unsigned
get_io_offset(nir_deref_var *deref, nir_instr *instr, nir_src *indirect,
struct lower_io_state *state)
{
bool found_indirect = false;
unsigned base_offset = 0;
nir_deref *tail = &deref->deref;
while (tail->child != NULL) {
const struct glsl_type *parent_type = tail->type;
tail = tail->child;
if (tail->deref_type == nir_deref_type_array) {
nir_deref_array *deref_array = nir_deref_as_array(tail);
unsigned size = type_size(tail->type);
base_offset += size * deref_array->base_offset;
if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
nir_load_const_instr *load_const =
nir_load_const_instr_create(state->mem_ctx, 1);
load_const->value.u[0] = size;
nir_instr_insert_before(instr, &load_const->instr);
nir_alu_instr *mul = nir_alu_instr_create(state->mem_ctx,
nir_op_imul);
mul->src[0].src.is_ssa = true;
mul->src[0].src.ssa = &load_const->def;
nir_src_copy(&mul->src[1].src, &deref_array->indirect,
state->mem_ctx);
mul->dest.write_mask = 1;
nir_ssa_dest_init(&mul->instr, &mul->dest.dest, 1, NULL);
nir_instr_insert_before(instr, &mul->instr);
if (found_indirect) {
nir_alu_instr *add = nir_alu_instr_create(state->mem_ctx,
nir_op_iadd);
add->src[0].src = *indirect;
add->src[1].src.is_ssa = true;
add->src[1].src.ssa = &mul->dest.dest.ssa;
add->dest.write_mask = 1;
nir_ssa_dest_init(&add->instr, &add->dest.dest, 1, NULL);
nir_instr_insert_before(instr, &add->instr);
indirect->is_ssa = true;
indirect->ssa = &add->dest.dest.ssa;
} else {
indirect->is_ssa = true;
indirect->ssa = &mul->dest.dest.ssa;
found_indirect = true;
}
}
} else if (tail->deref_type == nir_deref_type_struct) {
nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
for (unsigned i = 0; i < deref_struct->index; i++)
base_offset += type_size(glsl_get_struct_field(parent_type, i));
}
}
return base_offset;
}
