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);
}
}
static void
handle_glsl450_interpolation(struct vtn_builder *b, enum GLSLstd450 opcode,
const uint32_t *w, unsigned count)
{
const struct glsl_type *dest_type =
vtn_value(b, w[1], vtn_value_type_type)->type->type;
struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
val->ssa = vtn_create_ssa_value(b, dest_type);
nir_intrinsic_op op;
switch (opcode) {
case GLSLstd450InterpolateAtCentroid:
op = nir_intrinsic_interp_var_at_centroid;
break;
case GLSLstd450InterpolateAtSample:
op = nir_intrinsic_interp_var_at_sample;
break;
case GLSLstd450InterpolateAtOffset:
op = nir_intrinsic_interp_var_at_offset;
break;
default:
unreachable("Invalid opcode");
}
nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->nb.shader, op);
nir_deref_var *deref = vtn_nir_deref(b, w[5]);
intrin->variables[0] =
nir_deref_as_var(nir_copy_deref(intrin, &deref->deref));
switch (opcode) {
case GLSLstd450InterpolateAtCentroid:
break;
case GLSLstd450InterpolateAtSample:
case GLSLstd450InterpolateAtOffset:
intrin->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
break;
default:
unreachable("Invalid opcode");
}
intrin->num_components = glsl_get_vector_elements(dest_type);
nir_ssa_dest_init(&intrin->instr, &intrin->dest,
glsl_get_vector_elements(dest_type),
glsl_get_bit_size(dest_type), NULL);
val->ssa->def = &intrin->dest.ssa;
nir_builder_instr_insert(&b->nb, &intrin->instr);
}
/* 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);
}
}
/* 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");
}
}
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;
}