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;
}
static void
replace_with_mov(struct vc4_compile *c, struct qinst *inst, struct qreg arg)
{
dump_from(c, inst);
if (qir_is_mul(inst))
inst->op = QOP_MMOV;
else if (qir_is_float_input(inst))
inst->op = QOP_FMOV;
else
inst->op = QOP_MOV;
inst->src[0] = arg;
inst->src[1] = c->undef;
dump_to(c, inst);
}
static bool
constant_fold(struct vc4_compile *c, struct qinst *inst)
{
int nsrc = qir_get_op_nsrc(inst->op);
uint32_t ui[nsrc];
for (int i = 0; i < nsrc; i++) {
struct qreg reg = inst->src[i];
if (reg.file == QFILE_UNIF &&
c->uniform_contents[reg.index] == QUNIFORM_CONSTANT) {
ui[i] = c->uniform_data[reg.index];
} else if (reg.file == QFILE_SMALL_IMM) {
ui[i] = reg.index;
} else {
return false;
}
}
uint32_t result = 0;
switch (inst->op) {
case QOP_SHR:
result = ui[0] >> ui[1];
break;
default:
return false;
}
dump_from(c, inst);
inst->src[0] = qir_uniform_ui(c, result);
for (int i = 1; i < nsrc; i++)
inst->src[i] = c->undef;
inst->op = QOP_MOV;
dump_to(c, inst);
return true;
}
bool
qir_opt_algebraic(struct vc4_compile *c)
{
bool progress = false;
qir_for_each_inst_inorder(inst, c) {
switch (inst->op) {
case QOP_FMIN:
if (is_1f(c, inst->src[1]) &&
inst->src[0].pack >= QPU_UNPACK_8D_REP &&
inst->src[0].pack <= QPU_UNPACK_8D) {
replace_with_mov(c, inst, inst->src[0]);
progress = true;
}
break;
case QOP_FMAX:
if (is_zero(c, inst->src[1]) &&
inst->src[0].pack >= QPU_UNPACK_8D_REP &&
inst->src[0].pack <= QPU_UNPACK_8D) {
replace_with_mov(c, inst, inst->src[0]);
progress = true;
}
break;
case QOP_FSUB:
case QOP_SUB:
if (is_zero(c, inst->src[1])) {
replace_with_mov(c, inst, inst->src[0]);
progress = true;
}
break;
case QOP_ADD:
if (replace_x_0_with_x(c, inst, 0) ||
replace_x_0_with_x(c, inst, 1)) {
progress = true;
break;
}
break;
case QOP_FADD:
if (replace_x_0_with_x(c, inst, 0) ||
replace_x_0_with_x(c, inst, 1)) {
progress = true;
break;
}
/* FADD(a, FSUB(0, b)) -> FSUB(a, b) */
if (inst->src[1].file == QFILE_TEMP &&
c->defs[inst->src[1].index] &&
c->defs[inst->src[1].index]->op == QOP_FSUB) {
struct qinst *fsub = c->defs[inst->src[1].index];
if (is_zero(c, fsub->src[0])) {
dump_from(c, inst);
inst->op = QOP_FSUB;
inst->src[1] = fsub->src[1];
progress = true;
dump_to(c, inst);
break;
}
}
/* FADD(FSUB(0, b), a) -> FSUB(a, b) */
if (inst->src[0].file == QFILE_TEMP &&
c->defs[inst->src[0].index] &&
c->defs[inst->src[0].index]->op == QOP_FSUB) {
struct qinst *fsub = c->defs[inst->src[0].index];
if (is_zero(c, fsub->src[0])) {
dump_from(c, inst);
inst->op = QOP_FSUB;
inst->src[0] = inst->src[1];
inst->src[1] = fsub->src[1];
dump_to(c, inst);
progress = true;
break;
}
}
break;
case QOP_FMUL:
if (!inst->dst.pack &&
(replace_x_0_with_0(c, inst, 0) ||
replace_x_0_with_0(c, inst, 1) ||
fmul_replace_one(c, inst, 0) ||
fmul_replace_one(c, inst, 1))) {
progress = true;
break;
}
break;
case QOP_MUL24:
if (!inst->dst.pack &&
(replace_x_0_with_0(c, inst, 0) ||
replace_x_0_with_0(c, inst, 1))) {
progress = true;
break;
}
break;
case QOP_AND:
if (replace_x_0_with_0(c, inst, 0) ||
replace_x_0_with_0(c, inst, 1)) {
progress = true;
break;
}
if (is_constant_value(c, inst->src[0], ~0)) {
replace_with_mov(c, inst, inst->src[1]);
progress = true;
break;
}
if (is_constant_value(c, inst->src[1], ~0)) {
replace_with_mov(c, inst, inst->src[0]);
progress = true;
break;
}
break;
case QOP_OR:
if (replace_x_0_with_x(c, inst, 0) ||
replace_x_0_with_x(c, inst, 1)) {
progress = true;
break;
}
break;
case QOP_RCP:
if (is_1f(c, inst->src[0])) {
replace_with_mov(c, inst, inst->src[0]);
progress = true;
break;
}
break;
default:
break;
}
}
return progress;
}
