namespace aco {
-struct sched_ctx {
+enum MoveResult {
+ move_success,
+ move_fail_ssa,
+ move_fail_rar,
+ move_fail_pressure,
+};
+
+struct MoveState {
+ RegisterDemand max_registers;
+
+ Block *block;
+ Instruction *current;
+ RegisterDemand *register_demand;
+ bool improved_rar;
+
std::vector<bool> depends_on;
+ /* Two are needed because, for downwards VMEM scheduling, one needs to
+ * exclude the instructions in the clause, since new instructions in the
+ * clause are not moved past any other instructions in the clause. */
std::vector<bool> RAR_dependencies;
- /* For downwards VMEM scheduling, same as RAR_dependencies but excludes the
- * instructions in the clause, since new instructions in the clause are not
- * moved past any other instructions in the clause. */
- std::vector<bool> new_RAR_dependencies;
+ std::vector<bool> RAR_dependencies_clause;
+
+ int source_idx;
+ int insert_idx, insert_idx_clause;
+ RegisterDemand total_demand, total_demand_clause;
+
+ /* for moving instructions before the current instruction to after it */
+ void downwards_init(int current_idx, bool improved_rar, bool may_form_clauses);
+ MoveResult downwards_move(bool clause);
+ void downwards_skip();
+
+ /* for moving instructions after the first use of the current instruction upwards */
+ void upwards_init(int source_idx, bool improved_rar);
+ bool upwards_check_deps();
+ void upwards_set_insert_idx(int before);
+ MoveResult upwards_move();
+ void upwards_skip();
+
+private:
+ void downwards_advance_helper();
+};
- RegisterDemand max_registers;
+struct sched_ctx {
int16_t num_waves;
int16_t last_SMEM_stall;
int last_SMEM_dep_idx;
+ MoveState mv;
};
/* This scheduler is a simple bottom-up pass based on ideas from
*/
template <typename T>
-void move_element(T& list, size_t idx, size_t before) {
+void move_element(T begin_it, size_t idx, size_t before) {
if (idx < before) {
- auto begin = std::next(list.begin(), idx);
- auto end = std::next(list.begin(), before);
+ auto begin = std::next(begin_it, idx);
+ auto end = std::next(begin_it, before);
std::rotate(begin, begin + 1, end);
} else if (idx > before) {
- auto begin = std::next(list.begin(), before);
- auto end = std::next(list.begin(), idx + 1);
+ auto begin = std::next(begin_it, before);
+ auto end = std::next(begin_it, idx + 1);
std::rotate(begin, end - 1, end);
}
}
return temp_registers;
}
+void MoveState::downwards_advance_helper()
+{
+ source_idx--;
+ total_demand.update(register_demand[source_idx]);
+}
+
+void MoveState::downwards_init(int current_idx, bool improved_rar_, bool may_form_clauses)
+{
+ improved_rar = improved_rar_;
+ source_idx = current_idx;
+
+ insert_idx = current_idx + 1;
+ insert_idx_clause = current_idx;
+
+ total_demand = total_demand_clause = register_demand[current_idx];
+
+ std::fill(depends_on.begin(), depends_on.end(), false);
+ if (improved_rar) {
+ std::fill(RAR_dependencies.begin(), RAR_dependencies.end(), false);
+ if (may_form_clauses)
+ std::fill(RAR_dependencies_clause.begin(), RAR_dependencies_clause.end(), false);
+ }
+
+ for (const Operand& op : current->operands) {
+ if (op.isTemp()) {
+ depends_on[op.tempId()] = true;
+ if (improved_rar && op.isFirstKill())
+ RAR_dependencies[op.tempId()] = true;
+ }
+ }
+
+ /* update total_demand/source_idx */
+ downwards_advance_helper();
+}
+
+MoveResult MoveState::downwards_move(bool clause)
+{
+ aco_ptr<Instruction>& instr = block->instructions[source_idx];
+
+ for (const Definition& def : instr->definitions)
+ if (def.isTemp() && depends_on[def.tempId()])
+ return move_fail_ssa;
+
+ /* check if one of candidate's operands is killed by depending instruction */
+ std::vector<bool>& RAR_deps = improved_rar ? (clause ? RAR_dependencies_clause : RAR_dependencies) : depends_on;
+ for (const Operand& op : instr->operands) {
+ if (op.isTemp() && RAR_deps[op.tempId()]) {
+ // FIXME: account for difference in register pressure
+ return move_fail_rar;
+ }
+ }
+
+ if (clause) {
+ for (const Operand& op : instr->operands) {
+ if (op.isTemp()) {
+ depends_on[op.tempId()] = true;
+ if (op.isFirstKill())
+ RAR_dependencies[op.tempId()] = true;
+ }
+ }
+ }
+
+ int dest_insert_idx = clause ? insert_idx_clause : insert_idx;
+ RegisterDemand register_pressure = clause ? total_demand_clause : total_demand;
+
+ const RegisterDemand candidate_diff = getLiveChanges(instr);
+ const RegisterDemand temp = getTempRegisters(instr);
+ if (RegisterDemand(register_pressure - candidate_diff).exceeds(max_registers))
+ return move_fail_pressure;
+ const RegisterDemand temp2 = getTempRegisters(block->instructions[dest_insert_idx - 1]);
+ const RegisterDemand new_demand = register_demand[dest_insert_idx - 1] - temp2 + temp;
+ if (new_demand.exceeds(max_registers))
+ return move_fail_pressure;
+
+ /* move the candidate below the memory load */
+ move_element(block->instructions.begin(), source_idx, dest_insert_idx);
+
+ /* update register pressure */
+ move_element(register_demand, source_idx, dest_insert_idx);
+ for (int i = source_idx; i < dest_insert_idx - 1; i++)
+ register_demand[i] -= candidate_diff;
+ register_demand[dest_insert_idx - 1] = new_demand;
+ total_demand_clause -= candidate_diff;
+ insert_idx_clause--;
+ if (!clause) {
+ total_demand -= candidate_diff;
+ insert_idx--;
+ }
+
+ downwards_advance_helper();
+ return move_success;
+}
+
+void MoveState::downwards_skip()
+{
+ aco_ptr<Instruction>& instr = block->instructions[source_idx];
+
+ for (const Operand& op : instr->operands) {
+ if (op.isTemp()) {
+ depends_on[op.tempId()] = true;
+ if (improved_rar && op.isFirstKill()) {
+ RAR_dependencies[op.tempId()] = true;
+ RAR_dependencies_clause[op.tempId()] = true;
+ }
+ }
+ }
+ total_demand_clause.update(register_demand[source_idx]);
+
+ downwards_advance_helper();
+}
+
+void MoveState::upwards_init(int source_idx_, bool improved_rar_)
+{
+ source_idx = source_idx_;
+ improved_rar = improved_rar_;
+
+ insert_idx = -1;
+
+ std::fill(depends_on.begin(), depends_on.end(), false);
+ std::fill(RAR_dependencies.begin(), RAR_dependencies.end(), false);
+
+ for (const Definition& def : current->definitions) {
+ if (def.isTemp())
+ depends_on[def.tempId()] = true;
+ }
+}
+
+bool MoveState::upwards_check_deps()
+{
+ aco_ptr<Instruction>& instr = block->instructions[source_idx];
+ for (const Operand& op : instr->operands) {
+ if (op.isTemp() && depends_on[op.tempId()])
+ return false;
+ }
+ return true;
+}
+
+void MoveState::upwards_set_insert_idx(int before)
+{
+ insert_idx = before;
+ total_demand = register_demand[before - 1];
+}
+
+MoveResult MoveState::upwards_move()
+{
+ assert(insert_idx >= 0);
+
+ aco_ptr<Instruction>& instr = block->instructions[source_idx];
+ for (const Operand& op : instr->operands) {
+ if (op.isTemp() && depends_on[op.tempId()])
+ return move_fail_ssa;
+ }
+
+ /* check if candidate uses/kills an operand which is used by a dependency */
+ for (const Operand& op : instr->operands) {
+ if (op.isTemp() && (!improved_rar || op.isFirstKill()) && RAR_dependencies[op.tempId()])
+ return move_fail_rar;
+ }
+
+ /* check if register pressure is low enough: the diff is negative if register pressure is decreased */
+ const RegisterDemand candidate_diff = getLiveChanges(instr);
+ const RegisterDemand temp = getTempRegisters(instr);
+ if (RegisterDemand(total_demand + candidate_diff).exceeds(max_registers))
+ return move_fail_pressure;
+ const RegisterDemand temp2 = getTempRegisters(block->instructions[insert_idx - 1]);
+ const RegisterDemand new_demand = register_demand[insert_idx - 1] - temp2 + candidate_diff + temp;
+ if (new_demand.exceeds(max_registers))
+ return move_fail_pressure;
+
+ /* move the candidate above the insert_idx */
+ move_element(block->instructions.begin(), source_idx, insert_idx);
+
+ /* update register pressure */
+ move_element(register_demand, source_idx, insert_idx);
+ for (int i = insert_idx + 1; i <= source_idx; i++)
+ register_demand[i] += candidate_diff;
+ register_demand[insert_idx] = new_demand;
+ total_demand += candidate_diff;
+
+ insert_idx++;
+
+ total_demand.update(register_demand[source_idx]);
+ source_idx++;
+
+ return move_success;
+}
+
+void MoveState::upwards_skip()
+{
+ if (insert_idx >= 0) {
+ aco_ptr<Instruction>& instr = block->instructions[source_idx];
+ for (const Definition& def : instr->definitions) {
+ if (def.isTemp())
+ depends_on[def.tempId()] = true;
+ }
+ for (const Operand& op : instr->operands) {
+ if (op.isTemp())
+ RAR_dependencies[op.tempId()] = true;
+ }
+ total_demand.update(register_demand[source_idx]);
+ }
+
+ source_idx++;
+}
+
static bool is_spill_reload(aco_ptr<Instruction>& instr)
{
return instr->opcode == aco_opcode::p_spill || instr->opcode == aco_opcode::p_reload;
if (current->opcode == aco_opcode::s_memtime || current->opcode == aco_opcode::s_memrealtime)
return;
- /* create the initial set of values which current depends on */
- std::fill(ctx.depends_on.begin(), ctx.depends_on.end(), false);
- for (const Operand& op : current->operands) {
- if (op.isTemp())
- ctx.depends_on[op.tempId()] = true;
- }
-
- /* maintain how many registers remain free when moving instructions */
- RegisterDemand register_pressure = register_demand[idx];
-
/* first, check if we have instructions before current to move down */
- int insert_idx = idx + 1;
int moving_interaction = barrier_none;
bool moving_spill = false;
+ ctx.mv.downwards_init(idx, false, false);
+
for (int candidate_idx = idx - 1; k < max_moves && candidate_idx > (int) idx - window_size; candidate_idx--) {
assert(candidate_idx >= 0);
+ assert(candidate_idx == ctx.mv.source_idx);
aco_ptr<Instruction>& candidate = block->instructions[candidate_idx];
bool can_reorder_candidate = can_reorder(candidate.get());
break;
if (candidate->isVMEM())
break;
- register_pressure.update(register_demand[candidate_idx]);
/* if current depends on candidate, add additional dependencies and continue */
- bool can_move_down = true;
bool writes_exec = false;
for (const Definition& def : candidate->definitions) {
- if (def.isTemp() && ctx.depends_on[def.tempId()])
- can_move_down = false;
if (def.isFixed() && def.physReg() == exec)
writes_exec = true;
}
if (writes_exec)
break;
+ bool can_move_down = true;
if (moving_spill && is_spill_reload(candidate))
can_move_down = false;
if ((moving_interaction & barrier_shared) && candidate->format == Format::DS)
moving_interaction |= get_barrier_interaction(candidate.get());
moving_spill |= is_spill_reload(candidate);
if (!can_move_down) {
- for (const Operand& op : candidate->operands) {
- if (op.isTemp())
- ctx.depends_on[op.tempId()] = true;
- }
+ ctx.mv.downwards_skip();
can_reorder_cur &= can_reorder_candidate;
continue;
}
- bool register_pressure_unknown = false;
- /* check if one of candidate's operands is killed by depending instruction */
- for (const Operand& op : candidate->operands) {
- if (op.isTemp() && ctx.depends_on[op.tempId()]) {
- // FIXME: account for difference in register pressure
- register_pressure_unknown = true;
- }
- }
- if (register_pressure_unknown) {
- for (const Operand& op : candidate->operands) {
- if (op.isTemp())
- ctx.depends_on[op.tempId()] = true;
- }
+ MoveResult res = ctx.mv.downwards_move(false);
+ if (res == move_fail_ssa || res == move_fail_rar) {
+ ctx.mv.downwards_skip();
can_reorder_cur &= can_reorder_candidate;
continue;
- }
-
- /* check if register pressure is low enough: the diff is negative if register pressure is increased */
- const RegisterDemand candidate_diff = getLiveChanges(candidate);
- const RegisterDemand tempDemand = getTempRegisters(candidate);
- if (RegisterDemand(register_pressure - candidate_diff).exceeds(ctx.max_registers))
- break;
- const RegisterDemand tempDemand2 = getTempRegisters(block->instructions[insert_idx - 1]);
- const RegisterDemand new_demand = register_demand[insert_idx - 1] - tempDemand2 + tempDemand;
- if (new_demand.exceeds(ctx.max_registers))
+ } else if (res == move_fail_pressure) {
break;
- // TODO: we might want to look further to find a sequence of instructions to move down which doesn't exceed reg pressure
-
- /* move the candidate below the memory load */
- move_element(block->instructions, candidate_idx, insert_idx);
-
- /* update register pressure */
- move_element(register_demand, candidate_idx, insert_idx);
- for (int i = candidate_idx; i < insert_idx - 1; i++) {
- register_demand[i] -= candidate_diff;
}
- register_demand[insert_idx - 1] = new_demand;
- register_pressure -= candidate_diff;
if (candidate_idx < ctx.last_SMEM_dep_idx)
ctx.last_SMEM_stall++;
- insert_idx--;
k++;
}
- /* create the initial set of values which depend on current */
- std::fill(ctx.depends_on.begin(), ctx.depends_on.end(), false);
- std::fill(ctx.RAR_dependencies.begin(), ctx.RAR_dependencies.end(), false);
- for (const Definition& def : current->definitions) {
- if (def.isTemp())
- ctx.depends_on[def.tempId()] = true;
- }
-
/* find the first instruction depending on current or find another MEM */
- insert_idx = idx + 1;
moving_interaction = barrier_none;
moving_spill = false;
can_reorder_cur = true;
+ ctx.mv.upwards_init(idx + 1, false);
+
bool found_dependency = false;
/* second, check if we have instructions after current to move up */
for (int candidate_idx = idx + 1; k < max_moves && candidate_idx < (int) idx + window_size; candidate_idx++) {
+ assert(candidate_idx == ctx.mv.source_idx);
assert(candidate_idx < (int) block->instructions.size());
aco_ptr<Instruction>& candidate = block->instructions[candidate_idx];
bool can_reorder_candidate = can_reorder(candidate.get());
break;
/* check if candidate depends on current */
- bool is_dependency = std::any_of(candidate->operands.begin(), candidate->operands.end(),
- [&ctx](const Operand& op) { return op.isTemp() && ctx.depends_on[op.tempId()];});
+ bool is_dependency = !found_dependency && !ctx.mv.upwards_check_deps();
/* no need to steal from following VMEM instructions */
if (is_dependency && candidate->isVMEM())
break;
moving_interaction |= get_barrier_interaction(candidate.get());
moving_spill |= is_spill_reload(candidate);
if (is_dependency) {
- for (const Definition& def : candidate->definitions) {
- if (def.isTemp())
- ctx.depends_on[def.tempId()] = true;
- }
- for (const Operand& op : candidate->operands) {
- if (op.isTemp())
- ctx.RAR_dependencies[op.tempId()] = true;
- }
if (!found_dependency) {
- insert_idx = candidate_idx;
+ ctx.mv.upwards_set_insert_idx(candidate_idx);
found_dependency = true;
- /* init register pressure */
- register_pressure = register_demand[insert_idx - 1];
}
}
break;
if (!found_dependency) {
+ ctx.mv.upwards_skip();
k++;
continue;
}
- /* update register pressure */
- register_pressure.update(register_demand[candidate_idx - 1]);
-
if (is_dependency) {
+ ctx.mv.upwards_skip();
can_reorder_cur &= can_reorder_candidate;
continue;
}
- assert(insert_idx != idx);
-
- // TODO: correctly calculate register pressure for this case
- bool register_pressure_unknown = false;
- /* check if candidate uses/kills an operand which is used by a dependency */
- for (const Operand& op : candidate->operands) {
- if (op.isTemp() && ctx.RAR_dependencies[op.tempId()])
- register_pressure_unknown = true;
- }
- if (register_pressure_unknown) {
- if (candidate->isVMEM())
+
+ MoveResult res = ctx.mv.upwards_move();
+ if (res == move_fail_ssa || res == move_fail_rar) {
+ /* no need to steal from following VMEM instructions */
+ if (res == move_fail_ssa && candidate->isVMEM())
break;
- for (const Definition& def : candidate->definitions) {
- if (def.isTemp())
- ctx.RAR_dependencies[def.tempId()] = true;
- }
- for (const Operand& op : candidate->operands) {
- if (op.isTemp())
- ctx.RAR_dependencies[op.tempId()] = true;
- }
+ ctx.mv.upwards_skip();
can_reorder_cur &= can_reorder_candidate;
continue;
- }
-
- /* check if register pressure is low enough: the diff is negative if register pressure is decreased */
- const RegisterDemand candidate_diff = getLiveChanges(candidate);
- const RegisterDemand temp = getTempRegisters(candidate);
- if (RegisterDemand(register_pressure + candidate_diff).exceeds(ctx.max_registers))
+ } else if (res == move_fail_pressure) {
break;
- const RegisterDemand temp2 = getTempRegisters(block->instructions[insert_idx - 1]);
- const RegisterDemand new_demand = register_demand[insert_idx - 1] - temp2 + candidate_diff + temp;
- if (new_demand.exceeds(ctx.max_registers))
- break;
-
- /* move the candidate above the insert_idx */
- move_element(block->instructions, candidate_idx, insert_idx);
-
- /* update register pressure */
- move_element(register_demand, candidate_idx, insert_idx);
- for (int i = insert_idx + 1; i <= candidate_idx; i++) {
- register_demand[i] += candidate_diff;
}
- register_demand[insert_idx] = new_demand;
- register_pressure += candidate_diff;
- insert_idx++;
k++;
}
- ctx.last_SMEM_dep_idx = found_dependency ? insert_idx : 0;
+ ctx.last_SMEM_dep_idx = found_dependency ? ctx.mv.insert_idx : 0;
ctx.last_SMEM_stall = 10 - ctx.num_waves - k;
}
bool can_reorder_vmem = true;
bool can_reorder_smem = true;
- /* create the initial set of values which current depends on */
- std::fill(ctx.depends_on.begin(), ctx.depends_on.end(), false);
- std::fill(ctx.RAR_dependencies.begin(), ctx.RAR_dependencies.end(), false);
- std::fill(ctx.new_RAR_dependencies.begin(), ctx.new_RAR_dependencies.end(), false);
- for (const Operand& op : current->operands) {
- if (op.isTemp()) {
- ctx.depends_on[op.tempId()] = true;
- if (op.isFirstKill())
- ctx.RAR_dependencies[op.tempId()] = true;
- }
- }
-
- /* maintain how many registers remain free when moving instructions */
- RegisterDemand register_pressure_indep = register_demand[idx];
- RegisterDemand register_pressure_clause = register_demand[idx];
-
/* first, check if we have instructions before current to move down */
- int indep_insert_idx = idx + 1;
- int clause_insert_idx = idx;
int moving_interaction = barrier_none;
bool moving_spill = false;
+ ctx.mv.downwards_init(idx, true, true);
+
for (int candidate_idx = idx - 1; k < max_moves && candidate_idx > (int) idx - window_size; candidate_idx--) {
+ assert(candidate_idx == ctx.mv.source_idx);
assert(candidate_idx >= 0);
aco_ptr<Instruction>& candidate = block->instructions[candidate_idx];
bool can_reorder_candidate = can_reorder(candidate.get());
bool can_stall_prev_smem = idx <= ctx.last_SMEM_dep_idx && candidate_idx < ctx.last_SMEM_dep_idx;
if (can_stall_prev_smem && ctx.last_SMEM_stall >= 0)
break;
- register_pressure_indep.update(register_demand[candidate_idx]);
bool part_of_clause = false;
if (current->isVMEM() == candidate->isVMEM()) {
if (current->isVMEM())
same_resource = candidate->operands[0].tempId() == current->operands[0].tempId();
bool can_reorder = can_reorder_vmem || can_reorder_candidate;
- int grab_dist = clause_insert_idx - candidate_idx;
+ int grab_dist = ctx.mv.insert_idx_clause - candidate_idx;
/* We can't easily tell how much this will decrease the def-to-use
* distances, so just use how far it will be moved as a heuristic. */
part_of_clause = can_reorder && same_resource && grab_dist < clause_max_grab_dist;
bool can_move_down = !is_vmem || part_of_clause;
bool writes_exec = false;
for (const Definition& def : candidate->definitions) {
- if (def.isTemp() && ctx.depends_on[def.tempId()])
- can_move_down = false;
if (def.isFixed() && def.physReg() == exec)
writes_exec = true;
}
moving_interaction |= get_barrier_interaction(candidate.get());
moving_spill |= is_spill_reload(candidate);
if (!can_move_down) {
- for (const Operand& op : candidate->operands) {
- if (op.isTemp()) {
- ctx.depends_on[op.tempId()] = true;
- if (op.isFirstKill()) {
- ctx.RAR_dependencies[op.tempId()] = true;
- ctx.new_RAR_dependencies[op.tempId()] = true;
- }
- }
- }
- register_pressure_clause.update(register_demand[candidate_idx]);
+ ctx.mv.downwards_skip();
can_reorder_smem &= candidate->format != Format::SMEM || can_reorder_candidate;
can_reorder_vmem &= !is_vmem || can_reorder_candidate;
continue;
}
- if (part_of_clause) {
- for (const Operand& op : candidate->operands) {
- if (op.isTemp()) {
- ctx.depends_on[op.tempId()] = true;
- if (op.isFirstKill())
- ctx.RAR_dependencies[op.tempId()] = true;
- }
- }
- }
-
- bool register_pressure_unknown = false;
- std::vector<bool>& RAR_deps = part_of_clause ? ctx.new_RAR_dependencies : ctx.RAR_dependencies;
- /* check if one of candidate's operands is killed by depending instruction */
- for (const Operand& op : candidate->operands) {
- if (op.isTemp() && RAR_deps[op.tempId()]) {
- // FIXME: account for difference in register pressure
- register_pressure_unknown = true;
- }
- }
- if (register_pressure_unknown) {
- for (const Operand& op : candidate->operands) {
- if (op.isTemp()) {
- ctx.depends_on[op.tempId()] = true;
- if (op.isFirstKill()) {
- ctx.RAR_dependencies[op.tempId()] = true;
- ctx.new_RAR_dependencies[op.tempId()] = true;
- }
- }
- }
- register_pressure_clause.update(register_demand[candidate_idx]);
+ MoveResult res = ctx.mv.downwards_move(part_of_clause);
+ if (res == move_fail_ssa || res == move_fail_rar) {
+ ctx.mv.downwards_skip();
can_reorder_smem &= candidate->format != Format::SMEM || can_reorder_candidate;
can_reorder_vmem &= !is_vmem || can_reorder_candidate;
continue;
- }
-
- int insert_idx = part_of_clause ? clause_insert_idx : indep_insert_idx;
- RegisterDemand register_pressure = part_of_clause ? register_pressure_clause : register_pressure_indep;
-
- /* check if register pressure is low enough: the diff is negative if register pressure is increased */
- const RegisterDemand candidate_diff = getLiveChanges(candidate);
- const RegisterDemand temp = getTempRegisters(candidate);;
- if (RegisterDemand(register_pressure - candidate_diff).exceeds(ctx.max_registers))
- break;
- const RegisterDemand temp2 = getTempRegisters(block->instructions[insert_idx - 1]);
- const RegisterDemand new_demand = register_demand[insert_idx - 1] - temp2 + temp;
- if (new_demand.exceeds(ctx.max_registers))
+ } else if (res == move_fail_pressure) {
break;
- // TODO: we might want to look further to find a sequence of instructions to move down which doesn't exceed reg pressure
-
- /* move the candidate below the memory load */
- move_element(block->instructions, candidate_idx, insert_idx);
-
- /* update register pressure */
- move_element(register_demand, candidate_idx, insert_idx);
- for (int i = candidate_idx; i < insert_idx - 1; i++) {
- register_demand[i] -= candidate_diff;
- }
- register_demand[insert_idx - 1] = new_demand;
- register_pressure_clause -= candidate_diff;
- clause_insert_idx--;
- if (!part_of_clause) {
- register_pressure_indep -= candidate_diff;
- indep_insert_idx--;
}
k++;
if (candidate_idx < ctx.last_SMEM_dep_idx)
ctx.last_SMEM_stall++;
}
- /* create the initial set of values which depend on current */
- std::fill(ctx.depends_on.begin(), ctx.depends_on.end(), false);
- std::fill(ctx.RAR_dependencies.begin(), ctx.RAR_dependencies.end(), false);
- for (const Definition& def : current->definitions) {
- if (def.isTemp())
- ctx.depends_on[def.tempId()] = true;
- }
-
/* find the first instruction depending on current or find another VMEM */
- RegisterDemand register_pressure;
- int insert_idx = idx;
moving_interaction = barrier_none;
moving_spill = false;
// TODO: differentiate between loads and stores (load-load can always reorder)
can_reorder_vmem = true;
can_reorder_smem = true;
+ ctx.mv.upwards_init(idx + 1, true);
+
bool found_dependency = false;
/* second, check if we have instructions after current to move up */
for (int candidate_idx = idx + 1; k < max_moves && candidate_idx < (int) idx + window_size; candidate_idx++) {
+ assert(candidate_idx == ctx.mv.source_idx);
assert(candidate_idx < (int) block->instructions.size());
aco_ptr<Instruction>& candidate = block->instructions[candidate_idx];
bool can_reorder_candidate = can_reorder(candidate.get());
is_dependency = !can_reorder_smem && !can_reorder_candidate;
if (is_vmem)
is_dependency = !can_reorder_vmem && !can_reorder_candidate;
- for (const Operand& op : candidate->operands) {
- if (op.isTemp() && ctx.depends_on[op.tempId()]) {
- is_dependency = true;
- break;
- }
- }
+ is_dependency |= !found_dependency && !ctx.mv.upwards_check_deps();
if (moving_spill && is_spill_reload(candidate))
is_dependency = true;
if ((moving_interaction & barrier_shared) && candidate->format == Format::DS)
moving_interaction |= get_barrier_interaction(candidate.get());
moving_spill |= is_spill_reload(candidate);
if (is_dependency) {
- for (const Definition& def : candidate->definitions) {
- if (def.isTemp())
- ctx.depends_on[def.tempId()] = true;
- }
- for (const Operand& op : candidate->operands) {
- if (op.isTemp())
- ctx.RAR_dependencies[op.tempId()] = true;
- }
/* update flag whether we can reorder other memory instructions */
can_reorder_smem &= candidate->format != Format::SMEM || can_reorder_candidate;
can_reorder_vmem &= !is_vmem || can_reorder_candidate;
if (!found_dependency) {
- insert_idx = candidate_idx;
+ ctx.mv.upwards_set_insert_idx(candidate_idx);
found_dependency = true;
- /* init register pressure */
- register_pressure = register_demand[insert_idx - 1];
- continue;
}
} else if (is_vmem) {
/* don't move up dependencies of other VMEM instructions */
for (const Definition& def : candidate->definitions) {
if (def.isTemp())
- ctx.depends_on[def.tempId()] = true;
+ ctx.mv.depends_on[def.tempId()] = true;
}
}
- /* update register pressure */
- register_pressure.update(register_demand[candidate_idx - 1]);
-
- if (is_dependency || !found_dependency)
+ if (is_dependency || !found_dependency) {
+ ctx.mv.upwards_skip();
continue;
- assert(insert_idx != idx);
-
- bool register_pressure_unknown = false;
- /* check if candidate uses/kills an operand which is used by a dependency */
- for (const Operand& op : candidate->operands) {
- if (op.isTemp() && op.isFirstKill() && ctx.RAR_dependencies[op.tempId()])
- register_pressure_unknown = true;
}
- if (register_pressure_unknown) {
- for (const Definition& def : candidate->definitions) {
- if (def.isTemp())
- ctx.depends_on[def.tempId()] = true;
- }
- for (const Operand& op : candidate->operands) {
- if (op.isTemp())
- ctx.RAR_dependencies[op.tempId()] = true;
- }
+
+ MoveResult res = ctx.mv.upwards_move();
+ if (res == move_fail_ssa || res == move_fail_rar) {
+ ctx.mv.upwards_skip();
can_reorder_smem &= candidate->format != Format::SMEM || can_reorder_candidate;
can_reorder_vmem &= !is_vmem || can_reorder_candidate;
continue;
- }
-
- /* check if register pressure is low enough: the diff is negative if register pressure is decreased */
- const RegisterDemand candidate_diff = getLiveChanges(candidate);
- const RegisterDemand temp = getTempRegisters(candidate);
- if (RegisterDemand(register_pressure + candidate_diff).exceeds(ctx.max_registers))
- break;
- const RegisterDemand temp2 = getTempRegisters(block->instructions[insert_idx - 1]);
- const RegisterDemand new_demand = register_demand[insert_idx - 1] - temp2 + candidate_diff + temp;
- if (new_demand.exceeds(ctx.max_registers))
+ } else if (res == move_fail_pressure) {
break;
-
- /* move the candidate above the insert_idx */
- move_element(block->instructions, candidate_idx, insert_idx);
-
- /* update register pressure */
- move_element(register_demand, candidate_idx, insert_idx);
- for (int i = insert_idx + 1; i <= candidate_idx; i++) {
- register_demand[i] += candidate_diff;
}
- register_demand[insert_idx] = new_demand;
- register_pressure += candidate_diff;
- insert_idx++;
k++;
}
}
int max_moves = POS_EXP_MAX_MOVES;
int16_t k = 0;
- /* create the initial set of values which current depends on */
- std::fill(ctx.depends_on.begin(), ctx.depends_on.end(), false);
- std::fill(ctx.RAR_dependencies.begin(), ctx.RAR_dependencies.end(), false);
- for (const Operand& op : current->operands) {
- if (op.isTemp()) {
- ctx.depends_on[op.tempId()] = true;
- if (op.isFirstKill())
- ctx.RAR_dependencies[op.tempId()] = true;
- }
- }
-
- /* maintain how many registers remain free when moving instructions */
- RegisterDemand register_pressure = register_demand[idx];
+ ctx.mv.downwards_init(idx, true, false);
/* first, check if we have instructions before current to move down */
- int insert_idx = idx + 1;
int moving_interaction = barrier_none;
bool moving_spill = false;
if (!can_move_instr(candidate, current, moving_interaction))
break;
- register_pressure.update(register_demand[candidate_idx]);
-
/* if current depends on candidate, add additional dependencies and continue */
- bool can_move_down = true;
bool writes_exec = false;
for (unsigned i = 0; i < candidate->definitions.size(); i++) {
- if (candidate->definitions[i].isTemp() && ctx.depends_on[candidate->definitions[i].tempId()])
- can_move_down = false;
if (candidate->definitions[i].isFixed() && candidate->definitions[i].physReg() == exec)
writes_exec = true;
}
if (writes_exec)
break;
+ bool can_move_down = true;
if (moving_spill && is_spill_reload(candidate))
can_move_down = false;
if ((moving_interaction & barrier_shared) && candidate->format == Format::DS)
moving_interaction |= get_barrier_interaction(candidate.get());
moving_spill |= is_spill_reload(candidate);
if (!can_move_down) {
- for (const Operand& op : candidate->operands) {
- if (op.isTemp()) {
- ctx.depends_on[op.tempId()] = true;
- if (op.isFirstKill())
- ctx.RAR_dependencies[op.tempId()] = true;
- }
- }
+ ctx.mv.downwards_skip();
continue;
}
- bool register_pressure_unknown = false;
- /* check if one of candidate's operands is killed by depending instruction */
- for (const Operand& op : candidate->operands) {
- if (op.isTemp() && ctx.RAR_dependencies[op.tempId()]) {
- // FIXME: account for difference in register pressure
- register_pressure_unknown = true;
- }
- }
- if (register_pressure_unknown) {
- for (const Operand& op : candidate->operands) {
- if (op.isTemp()) {
- ctx.depends_on[op.tempId()] = true;
- if (op.isFirstKill())
- ctx.RAR_dependencies[op.tempId()] = true;
- }
- }
+ MoveResult res = ctx.mv.downwards_move(false);
+ if (res == move_fail_ssa || res == move_fail_rar) {
+ ctx.mv.downwards_skip();
continue;
- }
-
- /* check if register pressure is low enough: the diff is negative if register pressure is increased */
- const RegisterDemand candidate_diff = getLiveChanges(candidate);
- const RegisterDemand temp = getTempRegisters(candidate);;
- if (RegisterDemand(register_pressure - candidate_diff).exceeds(ctx.max_registers))
- break;
- const RegisterDemand temp2 = getTempRegisters(block->instructions[insert_idx - 1]);
- const RegisterDemand new_demand = register_demand[insert_idx - 1] - temp2 + temp;
- if (new_demand.exceeds(ctx.max_registers))
+ } else if (res == move_fail_pressure) {
break;
- // TODO: we might want to look further to find a sequence of instructions to move down which doesn't exceed reg pressure
-
- /* move the candidate below the export */
- move_element(block->instructions, candidate_idx, insert_idx);
-
- /* update register pressure */
- move_element(register_demand, candidate_idx, insert_idx);
- for (int i = candidate_idx; i < insert_idx - 1; i++) {
- register_demand[i] -= candidate_diff;
}
- register_demand[insert_idx - 1] = new_demand;
- register_pressure -= candidate_diff;
- insert_idx--;
k++;
}
}
{
ctx.last_SMEM_dep_idx = 0;
ctx.last_SMEM_stall = INT16_MIN;
+ ctx.mv.block = block;
+ ctx.mv.register_demand = live_vars.register_demand[block->index].data();
/* go through all instructions and find memory loads */
for (unsigned idx = 0; idx < block->instructions.size(); idx++) {
if (current->definitions.empty())
continue;
- if (current->isVMEM() || current->isFlatOrGlobal())
+ if (current->isVMEM() || current->isFlatOrGlobal()) {
+ ctx.mv.current = current;
schedule_VMEM(ctx, block, live_vars.register_demand[block->index], current, idx);
- if (current->format == Format::SMEM)
+ }
+
+ if (current->format == Format::SMEM) {
+ ctx.mv.current = current;
schedule_SMEM(ctx, block, live_vars.register_demand[block->index], current, idx);
+ }
}
if ((program->stage & hw_vs) && block->index == program->blocks.size() - 1) {
if (current->format == Format::EXP) {
unsigned target = static_cast<Export_instruction*>(current)->dest;
- if (target >= V_008DFC_SQ_EXP_POS && target < V_008DFC_SQ_EXP_PARAM)
+ if (target >= V_008DFC_SQ_EXP_POS && target < V_008DFC_SQ_EXP_PARAM) {
+ ctx.mv.current = current;
schedule_position_export(ctx, block, live_vars.register_demand[block->index], current, idx);
+ }
}
}
}
void schedule_program(Program *program, live& live_vars)
{
sched_ctx ctx;
- ctx.depends_on.resize(program->peekAllocationId());
- ctx.RAR_dependencies.resize(program->peekAllocationId());
- ctx.new_RAR_dependencies.resize(program->peekAllocationId());
+ ctx.mv.depends_on.resize(program->peekAllocationId());
+ ctx.mv.RAR_dependencies.resize(program->peekAllocationId());
+ ctx.mv.RAR_dependencies_clause.resize(program->peekAllocationId());
/* Allowing the scheduler to reduce the number of waves to as low as 5
* improves performance of Thrones of Britannia significantly and doesn't
* seem to hurt anything else. */
ctx.num_waves = std::max<uint16_t>(ctx.num_waves, program->min_waves);
assert(ctx.num_waves > 0 && ctx.num_waves <= program->num_waves);
- ctx.max_registers = { int16_t(get_addr_vgpr_from_waves(program, ctx.num_waves) - 2),
- int16_t(get_addr_sgpr_from_waves(program, ctx.num_waves))};
+ ctx.mv.max_registers = { int16_t(get_addr_vgpr_from_waves(program, ctx.num_waves) - 2),
+ int16_t(get_addr_sgpr_from_waves(program, ctx.num_waves))};
for (Block& block : program->blocks)
schedule_block(ctx, program, &block, live_vars);
projects / mesa.git / blobdiff