static bool
inst_result_equals(struct qinst *a, struct qinst *b)
{
if (a->op != b->op ||
qir_depends_on_flags(a) ||
qir_depends_on_flags(b)) {
return false;
}
for (int i = 0; i < qir_get_op_nsrc(a->op); i++) {
if (!qir_reg_equals(a->src[i], b->src[i]) ||
src_file_varies_on_reread(a->src[i]) ||
src_file_varies_on_reread(b->src[i])) {
return false;
}
}
return true;
}
static struct qinst *
vc4_find_cse(struct vc4_compile *c, struct hash_table *ht,
struct qinst *inst, uint32_t sf_count,
uint32_t r4_count)
{
if (inst->dst.file != QFILE_TEMP ||
inst->op == QOP_MOV ||
qir_get_op_nsrc(inst->op) > 4) {
return NULL;
}
struct inst_key key;
memset(&key, 0, sizeof(key));
key.op = inst->op;
memcpy(key.src, inst->src,
qir_get_op_nsrc(inst->op) * sizeof(key.src[0]));
if (qir_depends_on_flags(inst))
key.implicit_arg_update_count = sf_count;
if (qir_reads_r4(inst))
key.implicit_arg_update_count = r4_count;
uint32_t hash = _mesa_hash_data(&key, sizeof(key));
struct hash_entry *entry =
_mesa_hash_table_search_pre_hashed(ht, hash, &key);
if (entry) {
if (debug) {
fprintf(stderr, "CSE found match:\n");
fprintf(stderr, " Original inst: ");
qir_dump_inst(c, entry->data);
fprintf(stderr, "\n");
fprintf(stderr, " Our inst: ");
qir_dump_inst(c, inst);
fprintf(stderr, "\n");
}
return entry->data;
}
struct inst_key *alloc_key = ralloc(ht, struct inst_key);
if (!alloc_key)
return NULL;
memcpy(alloc_key, &key, sizeof(*alloc_key));
_mesa_hash_table_insert_pre_hashed(ht, hash, alloc_key, inst);
if (debug) {
fprintf(stderr, "Added to CSE HT: ");
qir_dump_inst(c, inst);
fprintf(stderr, "\n");
}
return NULL;
}
static bool
qir_opt_peephole_sf_block(struct vc4_compile *c, struct qblock *block)
{
bool progress = false;
/* We don't have liveness dataflow analysis for flags, but we also
* never generate a use of flags across control flow, so just treat
* them as unused at block exit.
*/
bool sf_live = false;
struct qinst *last_sf = NULL;
/* Walk the block from bottom to top, tracking if the SF is used, and
* removing unused or repeated ones.
*/
qir_for_each_inst_rev(inst, block) {
if (inst->sf) {
if (!sf_live) {
/* Our instruction's SF isn't read, so drop it.
*/
dump_from(c, inst, "dead SF");
inst->sf = false;
dump_to(c, inst);
progress = true;
} else if (last_sf &&
inst_result_equals(last_sf, inst)) {
/* The last_sf sets up same value as inst, so
* just drop the later one.
*/
dump_from(c, last_sf, "repeated SF");
last_sf->sf = false;
dump_to(c, last_sf);
progress = true;
last_sf = inst;
} else {
last_sf = inst;
}
sf_live = false;
}
if (last_sf) {
if (inst_srcs_updated(last_sf, inst))
last_sf = NULL;
}
if (qir_depends_on_flags(inst))
sf_live = true;
}
return progress;
}
bool
qir_opt_peephole_sf(struct vc4_compile *c)
{
bool progress = false;
bool sf_live = false;
struct qinst *last_sf = NULL;
/* Walk the block from bottom to top, tracking if the SF is used, and
* removing unused or repeated ones.
*/
list_for_each_entry_rev(struct qinst, inst, &c->instructions, link) {
if (inst->sf) {
if (!sf_live) {
/* Our instruction's SF isn't read, so drop it.
*/
dump_from(c, inst, "dead SF");
inst->sf = false;
dump_to(c, inst);
progress = true;
} else if (last_sf &&
inst_result_equals(last_sf, inst)) {
/* The last_sf sets up same value as inst, so
* just drop the later one.
*/
dump_from(c, last_sf, "repeated SF");
last_sf->sf = false;
dump_to(c, last_sf);
progress = true;
last_sf = inst;
} else {
last_sf = inst;
}
sf_live = false;
}
if (last_sf) {
if (inst_srcs_updated(last_sf, inst))
last_sf = NULL;
}
if (qir_depends_on_flags(inst))
sf_live = true;
}
return progress;
}
bool
qir_opt_dead_code(struct vc4_compile *c)
{
bool progress = false;
bool *used = calloc(c->num_temps, sizeof(bool));
bool sf_used = false;
/* Whether we're eliminating texture setup currently. */
bool dce_tex = false;
struct simple_node *node, *t;
for (node = c->instructions.prev, t = node->prev;
&c->instructions != node;
node = t, t = t->prev) {
struct qinst *inst = (struct qinst *)node;
if (inst->dst.file == QFILE_TEMP &&
!used[inst->dst.index] &&
!inst->sf &&
(!qir_has_side_effects(c, inst) ||
inst->op == QOP_TEX_RESULT) &&
!has_nonremovable_reads(c, inst)) {
if (inst->op == QOP_TEX_RESULT) {
dce_tex = true;
c->num_texture_samples--;
}
for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
if (inst->src[i].file != QFILE_VPM)
continue;
uint32_t attr = inst->src[i].index / 4;
uint32_t offset = (inst->src[i].index % 4) * 4;
if (c->vattr_sizes[attr] == offset + 4) {
c->num_inputs--;
c->vattr_sizes[attr] -= 4;
}
}
dce(c, inst);
progress = true;
continue;
}
if (qir_depends_on_flags(inst))
sf_used = true;
if (inst->sf) {
if (!sf_used) {
if (debug) {
fprintf(stderr, "Removing SF on: ");
qir_dump_inst(c, inst);
fprintf(stderr, "\n");
}
inst->sf = false;
progress = true;
}
sf_used = false;
}
if (inst->op == QOP_TEX_RESULT)
dce_tex = false;
if (dce_tex && (inst->op == QOP_TEX_S ||
inst->op == QOP_TEX_T ||
inst->op == QOP_TEX_R ||
inst->op == QOP_TEX_B ||
inst->op == QOP_TEX_DIRECT)) {
dce(c, inst);
progress = true;
continue;
}
for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
if (inst->src[i].file == QFILE_TEMP)
used[inst->src[i].index] = true;
}
}
free(used);
return progress;
}
bool
qir_opt_dead_code(struct vc4_compile *c)
{
bool progress = false;
bool *used = calloc(c->num_temps, sizeof(bool));
bool sf_used = false;
/* Whether we're eliminating texture setup currently. */
bool dce_tex = false;
struct simple_node *node, *t;
for (node = c->instructions.prev, t = node->prev;
&c->instructions != node;
node = t, t = t->prev) {
struct qinst *inst = (struct qinst *)node;
if (inst->dst.file == QFILE_TEMP &&
!used[inst->dst.index] &&
(!qir_has_side_effects(inst) ||
inst->op == QOP_TEX_RESULT)) {
if (inst->op == QOP_TEX_RESULT) {
dce_tex = true;
c->num_texture_samples--;
}
dce(c, inst);
progress = true;
continue;
}
if (qir_depends_on_flags(inst))
sf_used = true;
if (inst->op == QOP_SF) {
if (!sf_used) {
dce(c, inst);
progress = true;
continue;
}
sf_used = false;
}
if (inst->op == QOP_TEX_RESULT)
dce_tex = false;
if (dce_tex && (inst->op == QOP_TEX_S ||
inst->op == QOP_TEX_T ||
inst->op == QOP_TEX_R ||
inst->op == QOP_TEX_B)) {
dce(c, inst);
progress = true;
continue;
}
for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
if (inst->src[i].file == QFILE_TEMP)
used[inst->src[i].index] = true;
}
}
free(used);
return progress;
}
