shader: prove inactive resources across masked traversal merges

This commit is contained in:
nmzik
2026-10-13 08:37:44 +02:00
parent 259a40abfc
commit 9ade7e3582
5 changed files with 287 additions and 55 deletions
@@ -585,6 +585,19 @@ bool SrtWalker::Arg(const Inst& inst, size_t index, uint64_t& result) {
}
bool SrtWalker::EvaluatePhi(const Inst& inst, uint64_t& result) {
if (const auto control = m_program.control_phi_predecessors.find(&inst);
control != m_program.control_phi_predecessors.end()) {
bool found = false;
for (size_t index = 0; index < inst.NumArgs(); ++index) {
if (m_reachable_blocks != nullptr &&
!m_reachable_blocks->at(control->second[index])) continue;
uint64_t incoming = 0;
if (!Arg(inst, index, incoming) || (found && incoming != result)) return false;
result = incoming;
found = true;
}
return found;
}
const auto value = ResolveInvariantPhi(m_program, Value(const_cast<Inst*>(&inst)));
return !value.IsEmpty() && EvaluateWide(value, result);
}
@@ -1019,7 +1032,13 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
}
return false;
case ValueOpcode::ULessThanEqual32:
if (binary()) {
// A zero lower bound accepts every U32, including a varying lane value.
if (!Arg(inst, 0, a)) return false;
if (a == 0u) {
result = 1u;
return true;
}
if (Arg(inst, 1, b)) {
result = static_cast<uint32_t>(a) <= static_cast<uint32_t>(b);
return true;
}
@@ -1043,9 +1062,22 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
result = 0u;
return true;
}
if (!Arg(inst, 1, b) || (b != 0u && !left)) return false;
result = b != 0u;
return true;
if (Arg(inst, 1, b) && (b == 0u || left)) {
result = b != 0u;
return true;
}
if (!left) {
// Only lanes satisfying the left EXEC factor can make the conjunction
// true. Use a separate memo so this assumption cannot escape its RHS.
SrtWalker active(m_program, m_clean_evaluator != nullptr
? CleanRuntime(m_runtime) : m_runtime, nullptr, inst.Arg(0).Resolve());
active.m_reachable_blocks = m_reachable_blocks;
if (active.EvaluateWide(inst.Arg(1), b) && b == 0u) {
result = 0u;
return true;
}
}
return false;
}
case ValueOpcode::LogicalOr: {
const bool left = Arg(inst, 0, a);
@@ -1124,35 +1156,48 @@ bool SrtWalker::RefreshFlatBuffer(std::vector<uint32_t>& flat) {
}
return true;
}
auto& visited = m_program.visited_blocks;
auto& reachable = m_program.reachable_blocks;
auto& pending = m_program.pending_blocks;
visited.assign(m_program.control_flow.size(), 1u);
m_reachable_blocks = &visited;
if (m_clean_evaluator != nullptr) m_clean_evaluator->m_reachable_blocks = &visited;
for (;;) {
reachable.assign(m_program.control_flow.size(), 0u);
pending.clear();
pending.push_back(0u);
while (!pending.empty()) {
const auto index = pending.back();
pending.pop_back();
if (reachable.at(index)) continue;
reachable[index] = 1u;
const auto& block = m_program.control_flow[index];
uint32_t condition = 0;
auto& predicate = m_clean_evaluator != nullptr ? *m_clean_evaluator : *this;
if (!block.condition.IsEmpty() &&
(m_runtime.read_specialization_memory != nullptr || !m_program.capture_specialization_reads) &&
predicate.Evaluate(block.condition, condition)) {
pending.push_back(block.successors[condition != 0u ? 0u : 1u]);
} else {
pending.insert(pending.end(), block.successors.begin(), block.successors.end());
}
}
if (reachable == visited) break;
visited.swap(reachable);
if (m_program.control_phi_predecessors.empty()) break;
// Only unknown values need another evaluation: successful Phi facts remain
// true as their possible predecessors shrink. Keep uniform-read memos stable.
}
flat.assign(m_program.srt_reads.size(), 0u);
active.assign(m_program.descriptor_sources.size(), 1u);
for (const auto& block: m_program.control_flow) {
for (const auto& block: m_program.control_flow)
for (const auto source: block.sources) active.at(source) = 0u;
}
auto& visited = m_program.visited_blocks;
auto& pending = m_program.pending_blocks;
visited.assign(m_program.control_flow.size(), 0u);
pending.clear();
pending.push_back(0u);
while (!pending.empty()) {
const auto index = pending.back();
pending.pop_back();
if (visited.at(index)) continue;
visited[index] = 1u;
for (uint32_t index = 0; index < visited.size(); ++index) {
if (!visited[index]) continue;
const auto& block = m_program.control_flow[index];
for (const auto source: block.sources) active[source] = 1u;
for (const auto slot: block.srt_reads) {
for (const auto slot: block.srt_reads)
if (!refresh(slot)) return false;
}
uint32_t condition = 0;
auto& predicate = m_clean_evaluator != nullptr ? *m_clean_evaluator : *this;
if (!block.condition.IsEmpty() &&
(m_runtime.read_specialization_memory != nullptr || !m_program.capture_specialization_reads) &&
predicate.Evaluate(block.condition, condition)) {
pending.push_back(block.successors[condition != 0u ? 0u : 1u]);
} else {
pending.insert(pending.end(), block.successors.begin(), block.successors.end());
}
}
return true;
}