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);
}
/**
* 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;
}
