static void
lower_load_const_instr_scalar(nir_load_const_instr *lower)
{
if (lower->def.num_components == 1)
return;
nir_builder b;
nir_builder_init(&b, nir_cf_node_get_function(&lower->instr.block->cf_node));
b.cursor = nir_before_instr(&lower->instr);
/* Emit the individual loads. */
nir_ssa_def *loads[4];
for (unsigned i = 0; i < lower->def.num_components; i++) {
nir_load_const_instr *load_comp = nir_load_const_instr_create(b.shader, 1);
load_comp->value.u[0] = lower->value.u[i];
nir_builder_instr_insert(&b, &load_comp->instr);
loads[i] = &load_comp->def;
}
/* Batch things back together into a vector. */
nir_ssa_def *vec = nir_vec(&b, loads, lower->def.num_components);
/* Replace the old load with a reference to our reconstructed vector. */
nir_ssa_def_rewrite_uses(&lower->def, nir_src_for_ssa(vec));
nir_instr_remove(&lower->instr);
}
static bool
constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
{
nir_const_value src[4];
if (!instr->dest.dest.is_ssa)
return false;
for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
if (!instr->src[i].src.is_ssa)
return false;
nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;
if (src_instr->type != nir_instr_type_load_const)
return false;
nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr);
for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
j++) {
src[i].u[j] = load_const->value.u[instr->src[i].swizzle[j]];
}
/* We shouldn't have any source modifiers in the optimization loop. */
assert(!instr->src[i].abs && !instr->src[i].negate);
}
/* We shouldn't have any saturate modifiers in the optimization loop. */
assert(!instr->dest.saturate);
nir_const_value dest =
nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components,
src);
nir_load_const_instr *new_instr =
nir_load_const_instr_create(mem_ctx,
instr->dest.dest.ssa.num_components);
new_instr->value = dest;
nir_instr_insert_before(&instr->instr, &new_instr->instr);
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(&new_instr->def),
mem_ctx);
nir_instr_remove(&instr->instr);
ralloc_free(instr);
return true;
}
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);
}
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 bool
constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
{
nir_const_value src[NIR_MAX_VEC_COMPONENTS];
if (!instr->dest.dest.is_ssa)
return false;
/* In the case that any outputs/inputs have unsized types, then we need to
* guess the bit-size. In this case, the validator ensures that all
* bit-sizes match so we can just take the bit-size from first
* output/input with an unsized type. If all the outputs/inputs are sized
* then we don't need to guess the bit-size at all because the code we
* generate for constant opcodes in this case already knows the sizes of
* the types involved and does not need the provided bit-size for anything
* (although it still requires to receive a valid bit-size).
*/
unsigned bit_size = 0;
if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type))
bit_size = instr->dest.dest.ssa.bit_size;
for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
if (!instr->src[i].src.is_ssa)
return false;
if (bit_size == 0 &&
!nir_alu_type_get_type_size(nir_op_infos[instr->op].input_sizes[i])) {
bit_size = instr->src[i].src.ssa->bit_size;
}
nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;
if (src_instr->type != nir_instr_type_load_const)
return false;
nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr);
for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
j++) {
switch(load_const->def.bit_size) {
case 64:
src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]];
break;
case 32:
src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]];
break;
case 16:
src[i].u16[j] = load_const->value.u16[instr->src[i].swizzle[j]];
break;
case 8:
src[i].u8[j] = load_const->value.u8[instr->src[i].swizzle[j]];
break;
default:
unreachable("Invalid bit size");
}
}
/* We shouldn't have any source modifiers in the optimization loop. */
assert(!instr->src[i].abs && !instr->src[i].negate);
}
if (bit_size == 0)
bit_size = 32;
/* We shouldn't have any saturate modifiers in the optimization loop. */
assert(!instr->dest.saturate);
nir_const_value dest =
nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components,
bit_size, src);
nir_load_const_instr *new_instr =
nir_load_const_instr_create(mem_ctx,
instr->dest.dest.ssa.num_components,
instr->dest.dest.ssa.bit_size);
new_instr->value = dest;
nir_instr_insert_before(&instr->instr, &new_instr->instr);
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
nir_src_for_ssa(&new_instr->def));
nir_instr_remove(&instr->instr);
ralloc_free(instr);
return true;
}
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 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;
}
static nir_alu_src
construct_value(const nir_search_value *value, nir_alu_type type,
unsigned num_components, struct match_state *state,
nir_instr *instr, void *mem_ctx)
{
switch (value->type) {
case nir_search_value_expression: {
const nir_search_expression *expr = nir_search_value_as_expression(value);
if (nir_op_infos[expr->opcode].output_size != 0)
num_components = nir_op_infos[expr->opcode].output_size;
nir_alu_instr *alu = nir_alu_instr_create(mem_ctx, expr->opcode);
nir_ssa_dest_init(&alu->instr, &alu->dest.dest, num_components, NULL);
alu->dest.write_mask = (1 << num_components) - 1;
alu->dest.saturate = false;
for (unsigned i = 0; i < nir_op_infos[expr->opcode].num_inputs; i++) {
/* If the source is an explicitly sized source, then we need to reset
* the number of components to match.
*/
if (nir_op_infos[alu->op].input_sizes[i] != 0)
num_components = nir_op_infos[alu->op].input_sizes[i];
alu->src[i] = construct_value(expr->srcs[i],
nir_op_infos[alu->op].input_types[i],
num_components,
state, instr, mem_ctx);
}
nir_instr_insert_before(instr, &alu->instr);
nir_alu_src val;
val.src = nir_src_for_ssa(&alu->dest.dest.ssa);
val.negate = false;
val.abs = false,
memcpy(val.swizzle, identity_swizzle, sizeof val.swizzle);
return val;
}
case nir_search_value_variable: {
const nir_search_variable *var = nir_search_value_as_variable(value);
assert(state->variables_seen & (1 << var->variable));
nir_alu_src val = { NIR_SRC_INIT };
nir_alu_src_copy(&val, &state->variables[var->variable], mem_ctx);
assert(!var->is_constant);
return val;
}
case nir_search_value_constant: {
const nir_search_constant *c = nir_search_value_as_constant(value);
nir_load_const_instr *load = nir_load_const_instr_create(mem_ctx, 1);
switch (type) {
case nir_type_float:
load->def.name = ralloc_asprintf(mem_ctx, "%f", c->data.f);
load->value.f[0] = c->data.f;
break;
case nir_type_int:
load->def.name = ralloc_asprintf(mem_ctx, "%d", c->data.i);
load->value.i[0] = c->data.i;
break;
case nir_type_unsigned:
case nir_type_bool:
load->value.u[0] = c->data.u;
break;
default:
unreachable("Invalid alu source type");
}
nir_instr_insert_before(instr, &load->instr);
nir_alu_src val;
val.src = nir_src_for_ssa(&load->def);
val.negate = false;
val.abs = false,
memset(val.swizzle, 0, sizeof val.swizzle);
return val;
}
default:
unreachable("Invalid search value type");
}
}
