static bool
is_zero(struct vc4_compile *c, struct qreg reg)
{
reg = qir_follow_movs(c, reg);
return is_constant_value(c, reg, 0);
}
static bool
is_1f(struct vc4_compile *c, struct qreg reg)
{
reg = qir_follow_movs(c, reg);
return is_constant_value(c, reg, fui(1.0));
}
bool
qir_opt_small_immediates(struct vc4_compile *c)
{
bool progress = false;
qir_for_each_inst_inorder(inst, c) {
/* The small immediate value sits in the raddr B field, so we
* can't have 2 small immediates in one instruction (unless
* they're the same value, but that should be optimized away
* elsewhere).
*/
bool uses_small_imm = false;
for (int i = 0; i < qir_get_nsrc(inst); i++) {
if (inst->src[i].file == QFILE_SMALL_IMM)
uses_small_imm = true;
}
if (uses_small_imm)
continue;
/* Don't propagate small immediates into the top-end bounds
* checking for indirect UBO loads. The kernel doesn't parse
* small immediates and rejects the shader in this case. UBO
* loads are much more expensive than the uniform load, and
* indirect UBO regions are usually much larger than a small
* immediate, so it's not worth updating the kernel to allow
* optimizing it.
*/
if (inst->op == QOP_MIN_NOIMM)
continue;
for (int i = 0; i < qir_get_nsrc(inst); i++) {
struct qreg src = qir_follow_movs(c, inst->src[i]);
if (src.file != QFILE_UNIF ||
src.pack ||
c->uniform_contents[src.index] !=
QUNIFORM_CONSTANT) {
continue;
}
if (qir_is_tex(inst) &&
i == qir_get_tex_uniform_src(inst)) {
/* No turning the implicit uniform read into
* an immediate.
*/
continue;
}
uint32_t imm = c->uniform_data[src.index];
uint32_t small_imm = qpu_encode_small_immediate(imm);
if (small_imm == ~0)
continue;
if (debug) {
fprintf(stderr, "opt_small_immediate() from: ");
qir_dump_inst(c, inst);
fprintf(stderr, "\n");
}
inst->src[i].file = QFILE_SMALL_IMM;
inst->src[i].index = imm;
if (debug) {
fprintf(stderr, "to: ");
qir_dump_inst(c, inst);
fprintf(stderr, "\n");
}
progress = true;
break;
}
}
