static void
ir3_instr_depth(struct ir3_instruction *instr, unsigned boost, bool falsedep)
{
struct ir3_instruction *src;
/* don't mark falsedep's as used, but otherwise process them normally: */
if (!falsedep)
instr->flags &= ~IR3_INSTR_UNUSED;
if (ir3_instr_check_mark(instr))
return;
instr->depth = 0;
foreach_ssa_src_n(src, i, instr) {
unsigned sd;
/* visit child to compute it's depth: */
ir3_instr_depth(src, boost, __is_false_dep(instr, i));
/* for array writes, no need to delay on previous write: */
if (i == 0)
continue;
sd = ir3_delayslots(src, instr, i) + src->depth;
sd += boost;
instr->depth = MAX2(instr->depth, sd);
}
void ir3_block_depth(struct ir3_block *block)
{
unsigned i;
block->head = NULL;
ir3_clear_mark(block->shader);
for (i = 0; i < block->noutputs; i++)
if (block->outputs[i])
ir3_instr_depth(block->outputs[i]);
/* mark un-used instructions: */
for (i = 0; i < block->shader->instrs_count; i++) {
struct ir3_instruction *instr = block->shader->instrs[i];
/* just consider instructions within this block: */
if (instr->block != block)
continue;
if (!ir3_instr_check_mark(instr))
instr->depth = DEPTH_UNUSED;
}
/* cleanup unused inputs: */
for (i = 0; i < block->ninputs; i++) {
struct ir3_instruction *in = block->inputs[i];
if (in && (in->depth == DEPTH_UNUSED))
block->inputs[i] = NULL;
}
}
static void
ir3_instr_depth(struct ir3_instruction *instr)
{
struct ir3_instruction *src;
/* if we've already visited this instruction, bail now: */
if (ir3_instr_check_mark(instr))
return;
instr->depth = 0;
foreach_ssa_src_n(src, i, instr) {
unsigned sd;
/* visit child to compute it's depth: */
ir3_instr_depth(src);
/* for array writes, no need to delay on previous write: */
if (i == 0)
continue;
sd = ir3_delayslots(src, instr, i) + src->depth;
instr->depth = MAX2(instr->depth, sd);
}
void ir3_block_depth(struct ir3_block *block)
{
unsigned i;
block->head = NULL;
ir3_clear_mark(block->shader);
for (i = 0; i < block->noutputs; i++)
if (block->outputs[i])
ir3_instr_depth(block->outputs[i]);
/* at this point, any unvisited input is unused: */
for (i = 0; i < block->ninputs; i++) {
struct ir3_instruction *in = block->inputs[i];
if (in && !ir3_instr_check_mark(in))
block->inputs[i] = NULL;
}
}
static void ir3_instr_depth(struct ir3_instruction *instr)
{
unsigned i;
/* if we've already visited this instruction, bail now: */
if (ir3_instr_check_mark(instr))
return;
instr->depth = 0;
for (i = 1; i < instr->regs_count; i++) {
struct ir3_register *src = instr->regs[i];
if (src->flags & IR3_REG_SSA) {
unsigned sd;
/* visit child to compute it's depth: */
ir3_instr_depth(src->instr);
sd = ir3_delayslots(src->instr, instr, i-1) +
src->instr->depth;
instr->depth = MAX2(instr->depth, sd);
}
}
/* meta-instructions don't add cycles, other than PHI.. which
* might translate to a real instruction..
*
* well, not entirely true, fan-in/out, etc might need to need
* to generate some extra mov's in edge cases, etc.. probably
* we might want to do depth calculation considering the worst
* case for these??
*/
if (!is_meta(instr))
instr->depth++;
insert_by_depth(instr);
}
