/**
* For a given starting writemask channel and corresponding source index in
* the vec instruction, insert a MOV to the vec instruction's dest of all the
* writemask channels that get read from the same src reg.
*
* Returns the writemask of our MOV, so the parent loop calling this knows
* which ones have been processed.
*/
static unsigned
insert_mov(nir_alu_instr *vec, unsigned start_channel,
unsigned start_src_idx, nir_shader *shader)
{
unsigned src_idx = start_src_idx;
assert(src_idx < nir_op_infos[vec->op].num_inputs);
nir_alu_instr *mov = nir_alu_instr_create(shader, nir_op_imov);
nir_alu_src_copy(&mov->src[0], &vec->src[src_idx], mov);
nir_alu_dest_copy(&mov->dest, &vec->dest, mov);
mov->dest.write_mask = (1u << start_channel);
mov->src[0].swizzle[start_channel] = vec->src[src_idx].swizzle[0];
src_idx++;
for (unsigned i = start_channel + 1; i < 4; i++) {
if (!(vec->dest.write_mask & (1 << i)))
continue;
if (nir_srcs_equal(vec->src[src_idx].src, vec->src[start_src_idx].src)) {
mov->dest.write_mask |= (1 << i);
mov->src[0].swizzle[i] = vec->src[src_idx].swizzle[0];
}
src_idx++;
}
nir_instr_insert_before(&vec->instr, &mov->instr);
return mov->dest.write_mask;
}
static bool
opt_undef_alu(nir_alu_instr *instr)
{
if (instr->op != nir_op_bcsel && instr->op != nir_op_fcsel)
return false;
assert(instr->dest.dest.is_ssa);
for (int i = 1; i <= 2; i++) {
if (!instr->src[i].src.is_ssa)
continue;
nir_instr *parent = instr->src[i].src.ssa->parent_instr;
if (parent->type != nir_instr_type_ssa_undef)
continue;
/* We can't just use nir_alu_src_copy, because we need the def/use
* updated.
*/
nir_instr_rewrite_src(&instr->instr, &instr->src[0].src,
instr->src[i == 1 ? 2 : 1].src);
nir_alu_src_copy(&instr->src[0], &instr->src[i == 1 ? 2 : 1],
ralloc_parent(instr));
nir_src empty_src;
memset(&empty_src, 0, sizeof(empty_src));
nir_instr_rewrite_src(&instr->instr, &instr->src[1].src, empty_src);
nir_instr_rewrite_src(&instr->instr, &instr->src[2].src, empty_src);
instr->op = nir_op_imov;
return true;
}
return false;
}
static void
lower_reduction(nir_alu_instr *instr, nir_op chan_op, nir_op merge_op,
void *mem_ctx)
{
unsigned num_components = nir_op_infos[instr->op].input_sizes[0];
nir_ssa_def *last = NULL;
for (unsigned i = 0; i < num_components; i++) {
nir_alu_instr *chan = nir_alu_instr_create(mem_ctx, chan_op);
nir_alu_ssa_dest_init(chan, 1);
nir_alu_src_copy(&chan->src[0], &instr->src[0], mem_ctx);
chan->src[0].swizzle[0] = chan->src[0].swizzle[i];
if (nir_op_infos[chan_op].num_inputs > 1) {
assert(nir_op_infos[chan_op].num_inputs == 2);
nir_alu_src_copy(&chan->src[1], &instr->src[1], mem_ctx);
chan->src[1].swizzle[0] = chan->src[1].swizzle[i];
}
nir_instr_insert_before(&instr->instr, &chan->instr);
if (i == 0) {
last = &chan->dest.dest.ssa;
} else {
nir_alu_instr *merge = nir_alu_instr_create(mem_ctx, merge_op);
nir_alu_ssa_dest_init(merge, 1);
merge->dest.write_mask = 1;
merge->src[0].src = nir_src_for_ssa(last);
merge->src[1].src = nir_src_for_ssa(&chan->dest.dest.ssa);
nir_instr_insert_before(&instr->instr, &merge->instr);
last = &merge->dest.dest.ssa;
}
}
assert(instr->dest.write_mask == 1);
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(last),
mem_ctx);
nir_instr_remove(&instr->instr);
}
static void
lower_alu_instr_scalar(nir_alu_instr *instr, void *mem_ctx)
{
unsigned num_src = nir_op_infos[instr->op].num_inputs;
unsigned i, chan;
assert(instr->dest.dest.is_ssa);
assert(instr->dest.write_mask != 0);
#define LOWER_REDUCTION(name, chan, merge) \
case name##2: \
case name##3: \
case name##4: \
lower_reduction(instr, chan, merge, mem_ctx); \
break;
switch (instr->op) {
case nir_op_vec4:
case nir_op_vec3:
case nir_op_vec2:
/* We don't need to scalarize these ops, they're the ones generated to
* group up outputs into a value that can be SSAed.
*/
return;
LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd);
LOWER_REDUCTION(nir_op_ball_fequal, nir_op_feq, nir_op_iand);
LOWER_REDUCTION(nir_op_ball_iequal, nir_op_ieq, nir_op_iand);
LOWER_REDUCTION(nir_op_bany_fnequal, nir_op_fne, nir_op_ior);
LOWER_REDUCTION(nir_op_bany_inequal, nir_op_ine, nir_op_ior);
LOWER_REDUCTION(nir_op_fall_equal, nir_op_seq, nir_op_fand);
LOWER_REDUCTION(nir_op_fany_nequal, nir_op_sne, nir_op_for);
LOWER_REDUCTION(nir_op_ball, nir_op_imov, nir_op_iand);
LOWER_REDUCTION(nir_op_bany, nir_op_imov, nir_op_ior);
LOWER_REDUCTION(nir_op_fall, nir_op_fmov, nir_op_fand);
LOWER_REDUCTION(nir_op_fany, nir_op_fmov, nir_op_for);
default:
break;
}
if (instr->dest.dest.ssa.num_components == 1)
return;
unsigned num_components = instr->dest.dest.ssa.num_components;
static const nir_op nir_op_map[] = {nir_op_vec2, nir_op_vec3, nir_op_vec4};
nir_alu_instr *vec_instr =
nir_alu_instr_create(mem_ctx, nir_op_map[num_components - 2]);
nir_alu_ssa_dest_init(vec_instr, num_components);
for (chan = 0; chan < 4; chan++) {
if (!(instr->dest.write_mask & (1 << chan)))
continue;
nir_alu_instr *lower = nir_alu_instr_create(mem_ctx, instr->op);
for (i = 0; i < num_src; i++) {
/* We only handle same-size-as-dest (input_sizes[] == 0) or scalar
* args (input_sizes[] == 1).
*/
assert(nir_op_infos[instr->op].input_sizes[i] < 2);
unsigned src_chan = (nir_op_infos[instr->op].input_sizes[i] == 1 ?
0 : chan);
nir_alu_src_copy(&lower->src[i], &instr->src[i], mem_ctx);
for (int j = 0; j < 4; j++)
lower->src[i].swizzle[j] = instr->src[i].swizzle[src_chan];
}
nir_alu_ssa_dest_init(lower, 1);
lower->dest.saturate = instr->dest.saturate;
vec_instr->src[chan].src = nir_src_for_ssa(&lower->dest.dest.ssa);
nir_instr_insert_before(&instr->instr, &lower->instr);
}
nir_instr_insert_before(&instr->instr, &vec_instr->instr);
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
nir_src_for_ssa(&vec_instr->dest.dest.ssa),
mem_ctx);
nir_instr_remove(&instr->instr);
}
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");
}
}
