#include "sid.h"
#include <map>
+#include <set>
#include <stack>
/*
std::stack<Block*> loop_header;
std::vector<std::map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_start;
std::vector<std::map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_end;
- std::vector<std::pair<RegClass, std::set<uint32_t>>> interferences;
+ std::vector<std::pair<RegClass, std::unordered_set<uint32_t>>> interferences;
std::vector<std::vector<uint32_t>> affinities;
std::vector<bool> is_reloaded;
std::map<Temp, remat_info> remat;
spill_ctx(const RegisterDemand target_pressure, Program* program,
std::vector<std::vector<RegisterDemand>> register_demand)
: target_pressure(target_pressure), program(program),
- register_demand(register_demand), renames(program->blocks.size()),
+ register_demand(std::move(register_demand)), renames(program->blocks.size()),
spills_entry(program->blocks.size()), spills_exit(program->blocks.size()),
processed(program->blocks.size(), false), wave_size(program->wave_size) {}
}
}
+ void add_interference(uint32_t first, uint32_t second)
+ {
+ if (interferences[first].first.type() != interferences[second].first.type())
+ return;
+
+ bool inserted = interferences[first].second.insert(second).second;
+ if (inserted)
+ interferences[second].second.insert(first);
+ }
+
uint32_t allocate_spill_id(RegClass rc)
{
- interferences.emplace_back(rc, std::set<uint32_t>());
+ interferences.emplace_back(rc, std::unordered_set<uint32_t>());
is_reloaded.push_back(false);
return next_spill_id++;
}
}
-void compute_global_next_uses(spill_ctx& ctx, std::vector<std::set<Temp>>& live_out)
+void compute_global_next_uses(spill_ctx& ctx)
{
ctx.next_use_distances_start.resize(ctx.program->blocks.size());
ctx.next_use_distances_end.resize(ctx.program->blocks.size());
}
unsigned loop_end = i;
+ /* keep live-through spilled */
+ for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block_idx - 1]) {
+ if (pair.second.first < loop_end)
+ continue;
+
+ Temp to_spill = pair.first;
+ auto it = ctx.spills_exit[block_idx - 1].find(to_spill);
+ if (it == ctx.spills_exit[block_idx - 1].end())
+ continue;
+
+ ctx.spills_entry[block_idx][to_spill] = it->second;
+ spilled_registers += to_spill;
+ }
+
/* select live-through vgpr variables */
while (new_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr) {
unsigned distance = 0;
assert(idx != 0 && "loop without phis: TODO");
idx--;
RegisterDemand reg_pressure = ctx.register_demand[block_idx][idx] - spilled_registers;
+ /* Consider register pressure from linear predecessors. This can affect
+ * reg_pressure if the branch instructions define sgprs. */
+ for (unsigned pred : block->linear_preds) {
+ reg_pressure.sgpr = std::max<int16_t>(
+ reg_pressure.sgpr, ctx.register_demand[pred].back().sgpr - spilled_registers.sgpr);
+ }
+
while (reg_pressure.sgpr > ctx.target_pressure.sgpr) {
unsigned distance = 0;
Temp to_spill;
for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
if (pair.first.type() == RegType::sgpr &&
ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
- ctx.next_use_distances_start[block_idx][pair.first].second > block_idx) {
+ ctx.next_use_distances_start[block_idx][pair.first].first != block_idx) {
ctx.spills_entry[block_idx].insert(pair);
spilled_registers.sgpr += pair.first.size();
}
for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
if (pair.first.type() == RegType::vgpr &&
ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
- ctx.next_use_distances_start[block_idx][pair.first].second > block_idx) {
+ ctx.next_use_distances_start[block_idx][pair.first].first != block_idx) {
ctx.spills_entry[block_idx].insert(pair);
spilled_registers.vgpr += pair.first.size();
}
}
}
} else {
+ for (unsigned i = 0; i < idx; i++) {
+ aco_ptr<Instruction>& instr = block->instructions[i];
+ assert(is_phi(instr));
+ /* Killed phi definitions increase pressure in the predecessor but not
+ * the block they're in. Since the loops below are both to control
+ * pressure of the start of this block and the ends of it's
+ * predecessors, we need to count killed unspilled phi definitions here. */
+ if (instr->definitions[0].isKill() &&
+ !ctx.spills_entry[block_idx].count(instr->definitions[0].getTemp()))
+ reg_pressure += instr->definitions[0].getTemp();
+ }
idx--;
}
reg_pressure += ctx.register_demand[block_idx][idx] - spilled_registers;
+ /* Consider register pressure from linear predecessors. This can affect
+ * reg_pressure if the branch instructions define sgprs. */
+ for (unsigned pred : block->linear_preds) {
+ reg_pressure.sgpr = std::max<int16_t>(
+ reg_pressure.sgpr, ctx.register_demand[pred].back().sgpr - spilled_registers.sgpr);
+ }
+
while (reg_pressure.sgpr > ctx.target_pressure.sgpr) {
assert(!partial_spills.empty());
std::set<Temp>::iterator it = partial_spills.begin();
- Temp to_spill = *it;
- unsigned distance = ctx.next_use_distances_start[block_idx][*it].second;
+ Temp to_spill = Temp();
+ unsigned distance = 0;
while (it != partial_spills.end()) {
assert(ctx.spills_entry[block_idx].find(*it) == ctx.spills_entry[block_idx].end());
assert(!partial_spills.empty());
std::set<Temp>::iterator it = partial_spills.begin();
- Temp to_spill = *it;
- unsigned distance = ctx.next_use_distances_start[block_idx][*it].second;
+ Temp to_spill = Temp();
+ unsigned distance = 0;
while (it != partial_spills.end()) {
assert(ctx.spills_entry[block_idx].find(*it) == ctx.spills_entry[block_idx].end());
RegisterDemand get_demand_before(spill_ctx& ctx, unsigned block_idx, unsigned idx)
{
if (idx == 0) {
- RegisterDemand demand_before = ctx.register_demand[block_idx][idx];
+ RegisterDemand demand = ctx.register_demand[block_idx][idx];
aco_ptr<Instruction>& instr = ctx.program->blocks[block_idx].instructions[idx];
- for (const Definition& def : instr->definitions)
- demand_before -= def.getTemp();
- for (const Operand& op : instr->operands) {
- if (op.isFirstKill())
- demand_before += op.getTemp();
- }
- return demand_before;
+ aco_ptr<Instruction> instr_before(nullptr);
+ return get_demand_before(demand, instr, instr_before);
} else {
return ctx.register_demand[block_idx][idx - 1];
}
for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
if (var == pair.first)
continue;
- ctx.interferences[def_spill_id].second.emplace(pair.second);
- ctx.interferences[pair.second].second.emplace(def_spill_id);
+ ctx.add_interference(def_spill_id, pair.second);
}
/* check if variable is already spilled at predecessor */
for (std::pair<Temp, uint32_t> exit_spill : ctx.spills_exit[pred_idx]) {
if (exit_spill.first == pair.first)
continue;
- ctx.interferences[exit_spill.second].second.emplace(pair.second);
- ctx.interferences[pair.second].second.emplace(exit_spill.second);
+ ctx.add_interference(exit_spill.second, pair.second);
}
/* variable is in register at predecessor and has to be spilled */
uint32_t spill_id = ctx.allocate_spill_id(to_spill.regClass());
/* add interferences with currently spilled variables */
- for (std::pair<Temp, uint32_t> pair : current_spills) {
- ctx.interferences[spill_id].second.emplace(pair.second);
- ctx.interferences[pair.second].second.emplace(spill_id);
- }
- for (std::pair<Temp, std::pair<Temp, uint32_t>> pair : reloads) {
- ctx.interferences[spill_id].second.emplace(pair.second.second);
- ctx.interferences[pair.second.second].second.emplace(spill_id);
- }
+ for (std::pair<Temp, uint32_t> pair : current_spills)
+ ctx.add_interference(spill_id, pair.second);
+ for (std::pair<Temp, std::pair<Temp, uint32_t>> pair : reloads)
+ ctx.add_interference(spill_id, pair.second.second);
current_spills[to_spill] = spill_id;
spilled_registers += to_spill;
RegisterDemand spilled_registers = init_live_in_vars(ctx, block, block_idx);
/* add interferences for spilled variables */
- for (std::pair<Temp, uint32_t> x : ctx.spills_entry[block_idx]) {
- for (std::pair<Temp, uint32_t> y : ctx.spills_entry[block_idx])
- if (x.second != y.second)
- ctx.interferences[x.second].second.emplace(y.second);
+ for (auto it = ctx.spills_entry[block_idx].begin(); it != ctx.spills_entry[block_idx].end(); ++it) {
+ for (auto it2 = std::next(it); it2 != ctx.spills_entry[block_idx].end(); ++it2)
+ ctx.add_interference(it->second, it2->second);
}
bool is_loop_header = block->loop_nest_depth && ctx.loop_header.top()->index == block_idx;
Operand(rsrc_conf));
}
-void assign_spill_slots(spill_ctx& ctx, unsigned spills_to_vgpr) {
- std::map<uint32_t, uint32_t> sgpr_slot;
- std::map<uint32_t, uint32_t> vgpr_slot;
- std::vector<bool> is_assigned(ctx.interferences.size());
+void add_interferences(spill_ctx& ctx, std::vector<bool>& is_assigned,
+ std::vector<uint32_t>& slots, std::vector<bool>& slots_used,
+ unsigned id)
+{
+ for (unsigned other : ctx.interferences[id].second) {
+ if (!is_assigned[other])
+ continue;
- /* first, handle affinities: just merge all interferences into both spill ids */
- for (std::vector<uint32_t>& vec : ctx.affinities) {
- for (unsigned i = 0; i < vec.size(); i++) {
- for (unsigned j = i + 1; j < vec.size(); j++) {
- assert(vec[i] != vec[j]);
- for (uint32_t id : ctx.interferences[vec[i]].second)
- ctx.interferences[id].second.insert(vec[j]);
- for (uint32_t id : ctx.interferences[vec[j]].second)
- ctx.interferences[id].second.insert(vec[i]);
- ctx.interferences[vec[i]].second.insert(ctx.interferences[vec[j]].second.begin(), ctx.interferences[vec[j]].second.end());
- ctx.interferences[vec[j]].second.insert(ctx.interferences[vec[i]].second.begin(), ctx.interferences[vec[i]].second.end());
+ RegClass other_rc = ctx.interferences[other].first;
+ unsigned slot = slots[other];
+ std::fill(slots_used.begin() + slot, slots_used.begin() + slot + other_rc.size(), true);
+ }
+}
- bool reloaded = ctx.is_reloaded[vec[i]] || ctx.is_reloaded[vec[j]];
- ctx.is_reloaded[vec[i]] = reloaded;
- ctx.is_reloaded[vec[j]] = reloaded;
+unsigned find_available_slot(std::vector<bool>& used, unsigned wave_size,
+ unsigned size, bool is_sgpr, unsigned *num_slots)
+{
+ unsigned wave_size_minus_one = wave_size - 1;
+ unsigned slot = 0;
+
+ while (true) {
+ bool available = true;
+ for (unsigned i = 0; i < size; i++) {
+ if (slot + i < used.size() && used[slot + i]) {
+ available = false;
+ break;
}
}
+ if (!available) {
+ slot++;
+ continue;
+ }
+
+ if (is_sgpr && ((slot & wave_size_minus_one) > wave_size - size)) {
+ slot = align(slot, wave_size);
+ continue;
+ }
+
+ std::fill(used.begin(), used.end(), false);
+
+ if (slot + size > used.size())
+ used.resize(slot + size);
+
+ return slot;
}
- for (ASSERTED uint32_t i = 0; i < ctx.interferences.size(); i++)
- for (ASSERTED uint32_t id : ctx.interferences[i].second)
- assert(i != id);
+}
- /* for each spill slot, assign as many spill ids as possible */
- std::vector<std::set<uint32_t>> spill_slot_interferences;
- unsigned slot_idx = 0;
- bool done = false;
-
- /* assign sgpr spill slots */
- while (!done) {
- done = true;
- for (unsigned id = 0; id < ctx.interferences.size(); id++) {
- if (is_assigned[id] || !ctx.is_reloaded[id])
- continue;
- if (ctx.interferences[id].first.type() != RegType::sgpr)
- continue;
+void assign_spill_slots_helper(spill_ctx& ctx, RegType type,
+ std::vector<bool>& is_assigned,
+ std::vector<uint32_t>& slots,
+ unsigned *num_slots)
+{
+ std::vector<bool> slots_used(*num_slots);
- /* check interferences */
- bool interferes = false;
- for (unsigned i = slot_idx; i < slot_idx + ctx.interferences[id].first.size(); i++) {
- if (i == spill_slot_interferences.size())
- spill_slot_interferences.emplace_back(std::set<uint32_t>());
- if (spill_slot_interferences[i].find(id) != spill_slot_interferences[i].end() || i / ctx.wave_size != slot_idx / ctx.wave_size) {
- interferes = true;
- break;
- }
- }
- if (interferes) {
- done = false;
+ /* assign slots for ids with affinities first */
+ for (std::vector<uint32_t>& vec : ctx.affinities) {
+ if (ctx.interferences[vec[0]].first.type() != type)
+ continue;
+
+ for (unsigned id : vec) {
+ if (!ctx.is_reloaded[id])
continue;
- }
- /* we found a spill id which can be assigned to current spill slot */
- sgpr_slot[id] = slot_idx;
- is_assigned[id] = true;
- for (unsigned i = slot_idx; i < slot_idx + ctx.interferences[id].first.size(); i++)
- spill_slot_interferences[i].insert(ctx.interferences[id].second.begin(), ctx.interferences[id].second.end());
-
- /* add all affinities: there are no additional interferences */
- for (std::vector<uint32_t>& vec : ctx.affinities) {
- bool found_affinity = false;
- for (uint32_t entry : vec) {
- if (entry == id) {
- found_affinity = true;
- break;
- }
- }
- if (!found_affinity)
- continue;
- for (uint32_t entry : vec) {
- sgpr_slot[entry] = slot_idx;
- is_assigned[entry] = true;
- }
+ add_interferences(ctx, is_assigned, slots, slots_used, id);
+ }
+
+ unsigned slot = find_available_slot(slots_used, ctx.wave_size,
+ ctx.interferences[vec[0]].first.size(),
+ type == RegType::sgpr, num_slots);
+
+ for (unsigned id : vec) {
+ assert(!is_assigned[id]);
+
+ if (ctx.is_reloaded[id]) {
+ slots[id] = slot;
+ is_assigned[id] = true;
}
}
- slot_idx++;
}
- unsigned sgpr_spill_slots = spill_slot_interferences.size();
- spill_slot_interferences.clear();
- slot_idx = 0;
- done = false;
+ /* assign slots for ids without affinities */
+ for (unsigned id = 0; id < ctx.interferences.size(); id++) {
+ if (is_assigned[id] || !ctx.is_reloaded[id] || ctx.interferences[id].first.type() != type)
+ continue;
- /* assign vgpr spill slots */
- while (!done) {
- done = true;
- for (unsigned id = 0; id < ctx.interferences.size(); id++) {
- if (is_assigned[id] || !ctx.is_reloaded[id])
- continue;
- if (ctx.interferences[id].first.type() != RegType::vgpr)
- continue;
+ add_interferences(ctx, is_assigned, slots, slots_used, id);
- /* check interferences */
- bool interferes = false;
- for (unsigned i = slot_idx; i < slot_idx + ctx.interferences[id].first.size(); i++) {
- if (i == spill_slot_interferences.size())
- spill_slot_interferences.emplace_back(std::set<uint32_t>());
- /* check for interference and ensure that vector regs are stored next to each other */
- if (spill_slot_interferences[i].find(id) != spill_slot_interferences[i].end()) {
- interferes = true;
- break;
- }
- }
- if (interferes) {
- done = false;
- continue;
- }
+ unsigned slot = find_available_slot(slots_used, ctx.wave_size,
+ ctx.interferences[id].first.size(),
+ type == RegType::sgpr, num_slots);
- /* we found a spill id which can be assigned to current spill slot */
- vgpr_slot[id] = slot_idx;
- is_assigned[id] = true;
- for (unsigned i = slot_idx; i < slot_idx + ctx.interferences[id].first.size(); i++)
- spill_slot_interferences[i].insert(ctx.interferences[id].second.begin(), ctx.interferences[id].second.end());
-
- /* add all affinities: there are no additional interferences */
- for (std::vector<uint32_t>& vec : ctx.affinities) {
- bool found_affinity = false;
- for (uint32_t entry : vec) {
- if (entry == id) {
- found_affinity = true;
- break;
- }
- }
- if (!found_affinity)
- continue;
- for (uint32_t entry : vec) {
- vgpr_slot[entry] = slot_idx;
- is_assigned[entry] = true;
- }
+ slots[id] = slot;
+ is_assigned[id] = true;
+ }
+
+ *num_slots = slots_used.size();
+}
+
+void assign_spill_slots(spill_ctx& ctx, unsigned spills_to_vgpr) {
+ std::vector<uint32_t> slots(ctx.interferences.size());
+ std::vector<bool> is_assigned(ctx.interferences.size());
+
+ /* first, handle affinities: just merge all interferences into both spill ids */
+ for (std::vector<uint32_t>& vec : ctx.affinities) {
+ for (unsigned i = 0; i < vec.size(); i++) {
+ for (unsigned j = i + 1; j < vec.size(); j++) {
+ assert(vec[i] != vec[j]);
+ bool reloaded = ctx.is_reloaded[vec[i]] || ctx.is_reloaded[vec[j]];
+ ctx.is_reloaded[vec[i]] = reloaded;
+ ctx.is_reloaded[vec[j]] = reloaded;
}
}
- slot_idx++;
}
+ for (ASSERTED uint32_t i = 0; i < ctx.interferences.size(); i++)
+ for (ASSERTED uint32_t id : ctx.interferences[i].second)
+ assert(i != id);
- unsigned vgpr_spill_slots = spill_slot_interferences.size();
+ /* for each spill slot, assign as many spill ids as possible */
+ unsigned sgpr_spill_slots = 0, vgpr_spill_slots = 0;
+ assign_spill_slots_helper(ctx, RegType::sgpr, is_assigned, slots, &sgpr_spill_slots);
+ assign_spill_slots_helper(ctx, RegType::vgpr, is_assigned, slots, &vgpr_spill_slots);
for (unsigned id = 0; id < is_assigned.size(); id++)
assert(is_assigned[id] || !ctx.is_reloaded[id]);
continue;
assert(ctx.is_reloaded[vec[i]] == ctx.is_reloaded[vec[j]]);
assert(ctx.interferences[vec[i]].first.type() == ctx.interferences[vec[j]].first.type());
- if (ctx.interferences[vec[i]].first.type() == RegType::sgpr)
- assert(sgpr_slot[vec[i]] == sgpr_slot[vec[j]]);
- else
- assert(vgpr_slot[vec[i]] == vgpr_slot[vec[j]]);
+ assert(slots[vec[i]] == slots[vec[j]]);
}
}
}
bool can_destroy = true;
for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[block.linear_preds[0]]) {
- if (sgpr_slot.find(pair.second) != sgpr_slot.end() &&
- sgpr_slot[pair.second] / ctx.wave_size == i) {
+ if (ctx.interferences[pair.second].first.type() == RegType::sgpr &&
+ slots[pair.second] / ctx.wave_size == i) {
can_destroy = false;
break;
}
if (!ctx.is_reloaded[spill_id]) {
/* never reloaded, so don't spill */
- } else if (vgpr_slot.find(spill_id) != vgpr_slot.end()) {
+ } else if (!is_assigned[spill_id]) {
+ unreachable("No spill slot assigned for spill id");
+ } else if (ctx.interferences[spill_id].first.type() == RegType::vgpr) {
/* spill vgpr */
ctx.program->config->spilled_vgprs += (*it)->operands[0].size();
- uint32_t spill_slot = vgpr_slot[spill_id];
+ uint32_t spill_slot = slots[spill_id];
bool add_offset_to_sgpr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size + vgpr_spill_slots * 4 > 4096;
unsigned base_offset = add_offset_to_sgpr ? 0 : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
for (unsigned i = 0; i < temp.size(); i++)
split->definitions[i] = bld.def(v1);
bld.insert(split);
- for (unsigned i = 0; i < temp.size(); i++)
- bld.mubuf(opcode, scratch_rsrc, Operand(), scratch_offset, split->definitions[i].getTemp(), offset + i * 4, false);
+ for (unsigned i = 0; i < temp.size(); i++) {
+ Instruction *instr = bld.mubuf(opcode, scratch_rsrc, Operand(v1), scratch_offset, split->definitions[i].getTemp(), offset + i * 4, false, true);
+ static_cast<MUBUF_instruction *>(instr)->sync = memory_sync_info(storage_vgpr_spill, semantic_private);
+ }
} else {
- bld.mubuf(opcode, scratch_rsrc, Operand(), scratch_offset, temp, offset, false);
+ Instruction *instr = bld.mubuf(opcode, scratch_rsrc, Operand(v1), scratch_offset, temp, offset, false, true);
+ static_cast<MUBUF_instruction *>(instr)->sync = memory_sync_info(storage_vgpr_spill, semantic_private);
}
- } else if (sgpr_slot.find(spill_id) != sgpr_slot.end()) {
+ } else {
ctx.program->config->spilled_sgprs += (*it)->operands[0].size();
- uint32_t spill_slot = sgpr_slot[spill_id];
+ uint32_t spill_slot = slots[spill_id];
/* check if the linear vgpr already exists */
if (vgpr_spill_temps[spill_slot / ctx.wave_size] == Temp()) {
spill->operands[1] = Operand(spill_slot % ctx.wave_size);
spill->operands[2] = (*it)->operands[0];
instructions.emplace_back(aco_ptr<Instruction>(spill));
- } else {
- unreachable("No spill slot assigned for spill id");
}
} else if ((*it)->opcode == aco_opcode::p_reload) {
uint32_t spill_id = (*it)->operands[0].constantValue();
assert(ctx.is_reloaded[spill_id]);
- if (vgpr_slot.find(spill_id) != vgpr_slot.end()) {
+ if (!is_assigned[spill_id]) {
+ unreachable("No spill slot assigned for spill id");
+ } else if (ctx.interferences[spill_id].first.type() == RegType::vgpr) {
/* reload vgpr */
- uint32_t spill_slot = vgpr_slot[spill_id];
+ uint32_t spill_slot = slots[spill_id];
bool add_offset_to_sgpr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size + vgpr_spill_slots * 4 > 4096;
unsigned base_offset = add_offset_to_sgpr ? 0 : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
for (unsigned i = 0; i < def.size(); i++) {
Temp tmp = bld.tmp(v1);
vec->operands[i] = Operand(tmp);
- bld.mubuf(opcode, Definition(tmp), scratch_rsrc, Operand(), scratch_offset, offset + i * 4, false);
+ Instruction *instr = bld.mubuf(opcode, Definition(tmp), scratch_rsrc, Operand(v1), scratch_offset, offset + i * 4, false, true);
+ static_cast<MUBUF_instruction *>(instr)->sync = memory_sync_info(storage_vgpr_spill, semantic_private);
}
bld.insert(vec);
} else {
- bld.mubuf(opcode, def, scratch_rsrc, Operand(), scratch_offset, offset, false);
+ Instruction *instr = bld.mubuf(opcode, def, scratch_rsrc, Operand(v1), scratch_offset, offset, false, true);
+ static_cast<MUBUF_instruction *>(instr)->sync = memory_sync_info(storage_vgpr_spill, semantic_private);
}
- } else if (sgpr_slot.find(spill_id) != sgpr_slot.end()) {
- uint32_t spill_slot = sgpr_slot[spill_id];
+ } else {
+ uint32_t spill_slot = slots[spill_id];
reload_in_loop[spill_slot / ctx.wave_size] = block.loop_nest_depth > 0;
/* check if the linear vgpr already exists */
reload->operands[1] = Operand(spill_slot % ctx.wave_size);
reload->definitions[0] = (*it)->definitions[0];
instructions.emplace_back(aco_ptr<Instruction>(reload));
- } else {
- unreachable("No spill slot assigned for spill id");
}
} else if (!ctx.remat_used.count(it->get()) || ctx.remat_used[it->get()]) {
instructions.emplace_back(std::move(*it));
/* initialize ctx */
spill_ctx ctx(register_target, program, live_vars.register_demand);
- compute_global_next_uses(ctx, live_vars.live_out);
+ compute_global_next_uses(ctx);
get_rematerialize_info(ctx);
/* create spills and reloads */