const glsl_type *
glsl_channel_type(const glsl_type *t)
{
switch (glsl_get_base_type(t)) {
case GLSL_TYPE_ARRAY: {
const glsl_type *base = glsl_channel_type(glsl_get_array_element(t));
return glsl_array_type(base, glsl_get_length(t));
}
case GLSL_TYPE_UINT:
return glsl_uint_type();
case GLSL_TYPE_INT:
return glsl_int_type();
case GLSL_TYPE_FLOAT:
return glsl_float_type();
case GLSL_TYPE_BOOL:
return glsl_bool_type();
case GLSL_TYPE_DOUBLE:
return glsl_double_type();
case GLSL_TYPE_UINT64:
return glsl_uint64_t_type();
case GLSL_TYPE_INT64:
return glsl_int64_t_type();
case GLSL_TYPE_FLOAT16:
return glsl_float16_t_type();
case GLSL_TYPE_UINT16:
return glsl_uint16_t_type();
case GLSL_TYPE_INT16:
return glsl_int16_t_type();
default:
unreachable("Unhandled base type glsl_channel_type()");
}
}
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 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_field(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:
case GLSL_TYPE_DOUBLE:
unreachable("not reached");
}
return 0;
}
/* Crawls a chain of array derefs and rewrites the types so that the
* lengths stay the same but the terminal type is the one given by
* tail_type. This is useful for split structures.
*/
static void
rewrite_deref_types(nir_deref *deref, const struct glsl_type *type)
{
deref->type = type;
if (deref->child) {
assert(deref->child->deref_type == nir_deref_type_array);
assert(glsl_type_is_array(deref->type));
rewrite_deref_types(deref->child, glsl_get_array_element(type));
}
}
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;
}
}
/**
* Mark an entire variable as used. Caller must ensure that the variable
* represents a shader input or output.
*/
static void
mark_whole_variable(nir_shader *shader, nir_variable *var, bool is_output_read)
{
const struct glsl_type *type = var->type;
if (nir_is_per_vertex_io(var, shader->info.stage)) {
assert(glsl_type_is_array(type));
type = glsl_get_array_element(type);
}
const unsigned slots =
var->data.compact ? DIV_ROUND_UP(glsl_get_length(type), 4)
: glsl_count_attribute_slots(type, false);
set_io_mask(shader, var, 0, slots, is_output_read);
}
/* Recursively constructs deref chains to split a copy instruction into
* multiple (if needed) copy instructions with full-length deref chains.
* External callers of this function should pass the tail and head of the
* deref chains found as the source and destination of the copy instruction
* into this function.
*
* \param old_copy The copy instruction we are splitting
* \param dest_head The head of the destination deref chain we are building
* \param src_head The head of the source deref chain we are building
* \param dest_tail The tail of the destination deref chain we are building
* \param src_tail The tail of the source deref chain we are building
* \param state The current split_var_copies_state object
*/
static void
split_var_copy_instr(nir_intrinsic_instr *old_copy,
nir_deref *dest_head, nir_deref *src_head,
nir_deref *dest_tail, nir_deref *src_tail,
struct split_var_copies_state *state)
{
assert(src_tail->type == dest_tail->type);
/* Make sure these really are the tails of the deref chains */
assert(dest_tail->child == NULL);
assert(src_tail->child == NULL);
switch (glsl_get_base_type(src_tail->type)) {
case GLSL_TYPE_ARRAY: {
/* Make a wildcard dereference */
nir_deref_array *deref = nir_deref_array_create(state->dead_ctx);
deref->deref.type = glsl_get_array_element(src_tail->type);
deref->deref_array_type = nir_deref_array_type_wildcard;
/* Set the tail of both as the newly created wildcard deref. It is
* safe to use the same wildcard in both places because a) we will be
* copying it before we put it in an actual instruction and b)
* everything that will potentially add another link in the deref
* chain will also add the same thing to both chains.
*/
src_tail->child = &deref->deref;
dest_tail->child = &deref->deref;
split_var_copy_instr(old_copy, dest_head, src_head,
dest_tail->child, src_tail->child, state);
/* Set it back to the way we found it */
src_tail->child = NULL;
dest_tail->child = NULL;
break;
}
case GLSL_TYPE_STRUCT:
/* This is the only part that actually does any interesting
* splitting. For array types, we just use wildcards and resolve
* them later. For structure types, we need to emit one copy
* instruction for every structure element. Because we may have
* structs inside structs, we just recurse and let the next level
* take care of any additional structures.
*/
for (unsigned i = 0; i < glsl_get_length(src_tail->type); i++) {
nir_deref_struct *deref = nir_deref_struct_create(state->dead_ctx, i);
deref->deref.type = glsl_get_struct_field(src_tail->type, i);
/* Set the tail of both as the newly created structure deref. It
* is safe to use the same wildcard in both places because a) we
* will be copying it before we put it in an actual instruction
* and b) everything that will potentially add another link in the
* deref chain will also add the same thing to both chains.
*/
src_tail->child = &deref->deref;
dest_tail->child = &deref->deref;
split_var_copy_instr(old_copy, dest_head, src_head,
dest_tail->child, src_tail->child, state);
}
/* Set it back to the way we found it */
src_tail->child = NULL;
dest_tail->child = NULL;
break;
case GLSL_TYPE_UINT:
case GLSL_TYPE_INT:
case GLSL_TYPE_FLOAT:
case GLSL_TYPE_BOOL:
if (glsl_type_is_matrix(src_tail->type)) {
nir_deref_array *deref = nir_deref_array_create(state->dead_ctx);
deref->deref.type = glsl_get_column_type(src_tail->type);
deref->deref_array_type = nir_deref_array_type_wildcard;
/* Set the tail of both as the newly created wildcard deref. It
* is safe to use the same wildcard in both places because a) we
* will be copying it before we put it in an actual instruction
* and b) everything that will potentially add another link in the
* deref chain will also add the same thing to both chains.
*/
src_tail->child = &deref->deref;
dest_tail->child = &deref->deref;
split_var_copy_instr(old_copy, dest_head, src_head,
dest_tail->child, src_tail->child, state);
/* Set it back to the way we found it */
src_tail->child = NULL;
dest_tail->child = NULL;
} else {
/* At this point, we have fully built our deref chains and can
* actually add the new copy instruction.
*/
nir_intrinsic_instr *new_copy =
nir_intrinsic_instr_create(state->mem_ctx, nir_intrinsic_copy_var);
/* We need to make copies because a) this deref chain actually
* belongs to the copy instruction and b) the deref chains may
* have some of the same links due to the way we constructed them
*/
nir_deref *src = nir_copy_deref(new_copy, src_head);
nir_deref *dest = nir_copy_deref(new_copy, dest_head);
new_copy->variables[0] = nir_deref_as_var(dest);
new_copy->variables[1] = nir_deref_as_var(src);
/* Emit the copy instruction after the old instruction. We'll
* remove the old one later.
*/
nir_instr_insert_after(&old_copy->instr, &new_copy->instr);
state->progress = true;
}
break;
case GLSL_TYPE_SAMPLER:
case GLSL_TYPE_IMAGE:
case GLSL_TYPE_ATOMIC_UINT:
case GLSL_TYPE_INTERFACE:
default:
unreachable("Cannot copy these types");
}
}
/**
* Try to mark a portion of the given varying as used. Caller must ensure
* that the variable represents a shader input or output.
*
* If the index can't be interpreted as a constant, or some other problem
* occurs, then nothing will be marked and false will be returned.
*/
static bool
try_mask_partial_io(nir_shader *shader, nir_variable *var,
nir_deref_instr *deref, bool is_output_read)
{
const struct glsl_type *type = var->type;
if (nir_is_per_vertex_io(var, shader->info.stage)) {
assert(glsl_type_is_array(type));
type = glsl_get_array_element(type);
}
/* The code below only handles:
*
* - Indexing into matrices
* - Indexing into arrays of (arrays, matrices, vectors, or scalars)
*
* For now, we just give up if we see varying structs and arrays of structs
* here marking the entire variable as used.
*/
if (!(glsl_type_is_matrix(type) ||
(glsl_type_is_array(type) && !var->data.compact &&
(glsl_type_is_numeric(glsl_without_array(type)) ||
glsl_type_is_boolean(glsl_without_array(type)))))) {
/* If we don't know how to handle this case, give up and let the
* caller mark the whole variable as used.
*/
return false;
}
unsigned offset = get_io_offset(deref, false);
if (offset == -1)
return false;
unsigned num_elems;
unsigned elem_width = 1;
unsigned mat_cols = 1;
if (glsl_type_is_array(type)) {
num_elems = glsl_get_aoa_size(type);
if (glsl_type_is_matrix(glsl_without_array(type)))
mat_cols = glsl_get_matrix_columns(glsl_without_array(type));
} else {
num_elems = glsl_get_matrix_columns(type);
}
/* double element width for double types that takes two slots */
if (glsl_type_is_dual_slot(glsl_without_array(type)))
elem_width *= 2;
if (offset >= num_elems * elem_width * mat_cols) {
/* Constant index outside the bounds of the matrix/array. This could
* arise as a result of constant folding of a legal GLSL program.
*
* Even though the spec says that indexing outside the bounds of a
* matrix/array results in undefined behaviour, we don't want to pass
* out-of-range values to set_io_mask() (since this could result in
* slots that don't exist being marked as used), so just let the caller
* mark the whole variable as used.
*/
return false;
}
set_io_mask(shader, var, offset, elem_width, is_output_read);
return true;
}
static void
_vtn_local_load_store(struct vtn_builder *b, bool load, nir_deref_var *deref,
nir_deref *tail, struct vtn_ssa_value *inout)
{
/* The deref tail may contain a deref to select a component of a vector (in
* other words, it might not be an actual tail) so we have to save it away
* here since we overwrite it later.
*/
nir_deref *old_child = tail->child;
if (glsl_type_is_vector_or_scalar(tail->type)) {
/* Terminate the deref chain in case there is one more link to pick
* off a component of the vector.
*/
tail->child = NULL;
nir_intrinsic_op op = load ? nir_intrinsic_load_var :
nir_intrinsic_store_var;
nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
intrin->variables[0] =
nir_deref_as_var(nir_copy_deref(intrin, &deref->deref));
intrin->num_components = glsl_get_vector_elements(tail->type);
if (load) {
nir_ssa_dest_init(&intrin->instr, &intrin->dest,
intrin->num_components,
glsl_get_bit_size(glsl_get_base_type(tail->type)),
NULL);
inout->def = &intrin->dest.ssa;
} else {
nir_intrinsic_set_write_mask(intrin, (1 << intrin->num_components) - 1);
intrin->src[0] = nir_src_for_ssa(inout->def);
}
nir_builder_instr_insert(&b->nb, &intrin->instr);
} else if (glsl_get_base_type(tail->type) == GLSL_TYPE_ARRAY ||
glsl_type_is_matrix(tail->type)) {
unsigned elems = glsl_get_length(tail->type);
nir_deref_array *deref_arr = nir_deref_array_create(b);
deref_arr->deref_array_type = nir_deref_array_type_direct;
deref_arr->deref.type = glsl_get_array_element(tail->type);
tail->child = &deref_arr->deref;
for (unsigned i = 0; i < elems; i++) {
deref_arr->base_offset = i;
_vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
}
} else {
assert(glsl_get_base_type(tail->type) == GLSL_TYPE_STRUCT);
unsigned elems = glsl_get_length(tail->type);
nir_deref_struct *deref_struct = nir_deref_struct_create(b, 0);
tail->child = &deref_struct->deref;
for (unsigned i = 0; i < elems; i++) {
deref_struct->index = i;
deref_struct->deref.type = glsl_get_struct_field(tail->type, i);
_vtn_local_load_store(b, load, deref, tail->child, inout->elems[i]);
}
}
tail->child = old_child;
}
