Always allocate saves

This sometimes allocates saves too eagerly, but it removes a branch
in release saves that executes on every thread failing which seems
slightly better.
This commit is contained in:
Maxime Coste 2024-03-15 09:18:42 +11:00
parent 83f12fc8e9
commit cde5f5a258

View File

@ -313,8 +313,10 @@ private:
template<bool copy> template<bool copy>
int16_t new_saves(Iterator* pos, uint32_t valid_mask) int16_t new_saves(Iterator* pos, uint32_t valid_mask)
{ {
if constexpr (mode & RegexMode::NoSaves)
return -1;
kak_assert(not copy or pos != nullptr); kak_assert(not copy or pos != nullptr);
const auto count = m_program.save_count;
if (m_first_free >= 0) if (m_first_free >= 0)
{ {
const int16_t res = m_first_free; const int16_t res = m_first_free;
@ -327,6 +329,7 @@ private:
return res; return res;
} }
const auto count = m_program.save_count;
auto* new_pos = reinterpret_cast<Iterator*>(operator new (count * sizeof(Iterator))); auto* new_pos = reinterpret_cast<Iterator*>(operator new (count * sizeof(Iterator)));
for (size_t i = 0; i < count; ++i) for (size_t i = 0; i < count; ++i)
new (new_pos+i) Iterator{copy ? pos[i] : Iterator{}}; new (new_pos+i) Iterator{copy ? pos[i] : Iterator{}};
@ -336,8 +339,9 @@ private:
void release_saves(int16_t index) void release_saves(int16_t index)
{ {
if (index < 0) if constexpr (mode & RegexMode::NoSaves)
return; return;
auto& saves = m_saves[index]; auto& saves = m_saves[index];
if (saves.refcount == 1) if (saves.refcount == 1)
{ {
@ -393,6 +397,7 @@ private:
(config.flags & RegexExecFlags::NotInitialNull and pos == config.begin)) (config.flags & RegexExecFlags::NotInitialNull and pos == config.begin))
return failed(); return failed();
if (m_captures >= 0)
release_saves(m_captures); release_saves(m_captures);
m_captures = thread.saves; m_captures = thread.saves;
m_found_match = true; m_found_match = true;
@ -419,7 +424,6 @@ private:
if (auto target = inst.param.split.target; if (auto target = inst.param.split.target;
instructions[target].last_step != current_step) instructions[target].last_step != current_step)
{ {
if (thread.saves >= 0)
++m_saves[thread.saves].refcount; ++m_saves[thread.saves].refcount;
if (not inst.param.split.prioritize_parent) if (not inst.param.split.prioritize_parent)
std::swap(thread.inst, target); std::swap(thread.inst, target);
@ -429,9 +433,7 @@ private:
case CompiledRegex::Save: case CompiledRegex::Save:
if constexpr (mode & RegexMode::NoSaves) if constexpr (mode & RegexMode::NoSaves)
break; break;
if (thread.saves < 0) if (auto& saves = m_saves[thread.saves]; saves.refcount > 1)
thread.saves = new_saves<false>(nullptr, 0);
else if (auto& saves = m_saves[thread.saves]; saves.refcount > 1)
{ {
--saves.refcount; --saves.refcount;
thread.saves = new_saves<true>(saves.pos, saves.valid_mask); thread.saves = new_saves<true>(saves.pos, saves.valid_mask);
@ -471,12 +473,13 @@ private:
bool exec_program(Iterator pos, const ExecConfig& config, auto&& idle_func) bool exec_program(Iterator pos, const ExecConfig& config, auto&& idle_func)
{ {
kak_assert(m_threads.current_is_empty() and m_threads.next_is_empty()); kak_assert(m_threads.current_is_empty() and m_threads.next_is_empty());
if (m_captures >= 0)
release_saves(m_captures); release_saves(m_captures);
m_captures = -1; m_captures = -1;
m_threads.ensure_initial_capacity(); m_threads.ensure_initial_capacity();
const int16_t first_inst = forward ? 0 : m_program.first_backward_inst; const int16_t first_inst = forward ? 0 : m_program.first_backward_inst;
m_threads.push_current({first_inst, -1}); m_threads.push_current({first_inst, new_saves<false>(nullptr, 0)});
const auto& start_desc = forward ? m_program.forward_start_desc : m_program.backward_start_desc; const auto& start_desc = forward ? m_program.forward_start_desc : m_program.backward_start_desc;
@ -518,7 +521,7 @@ private:
{ {
if (start_desc and m_threads.next_is_empty()) if (start_desc and m_threads.next_is_empty())
to_next_start(pos, config, *start_desc); to_next_start(pos, config, *start_desc);
m_threads.push_next({first_inst, -1}); m_threads.push_next({first_inst, new_saves<false>(nullptr, 0)});
} }
m_threads.swap_next(); m_threads.swap_next();
} }