Regex: store saves in a copy on write structure

This commit is contained in:
Maxime Coste 2017-10-03 10:54:43 +08:00
parent 11b9c996ea
commit 589cde67f0

View File

@ -68,10 +68,35 @@ struct ThreadedRegexVM
ThreadedRegexVM(const CompiledRegex& program) ThreadedRegexVM(const CompiledRegex& program)
: m_program{program} { kak_assert(m_program); } : m_program{program} { kak_assert(m_program); }
struct Saves
{
int refcount;
Vector<Iterator> pos;
};
Saves* clone_saves(Saves* saves)
{
Saves* res = nullptr;
if (not m_free_saves.empty())
{
res = m_free_saves.back();
m_free_saves.pop_back();
}
else
{
m_saves.push_back(std::make_unique<Saves>());
res = m_saves.back().get();
}
res->refcount = 1;
res->pos = saves->pos;
return res;
}
struct Thread struct Thread
{ {
const char* inst; const char* inst;
Vector<Iterator> saves = {}; Saves* saves;
}; };
enum class StepResult { Consumed, Matched, Failed }; enum class StepResult { Consumed, Matched, Failed };
@ -97,33 +122,35 @@ struct ThreadedRegexVM
case CompiledRegex::AnyChar: case CompiledRegex::AnyChar:
return StepResult::Consumed; return StepResult::Consumed;
case CompiledRegex::Jump: case CompiledRegex::Jump:
{ thread.inst = prog_start + *reinterpret_cast<const CompiledRegex::Offset*>(thread.inst);
auto inst = prog_start + *reinterpret_cast<const CompiledRegex::Offset*>(thread.inst);
// if instruction is already going to be executed by another thread, drop this thread
if (std::find_if(m_threads.begin(), m_threads.end(),
[inst](const Thread& t) { return t.inst == inst; }) != m_threads.end())
return StepResult::Failed;
thread.inst = inst;
break; break;
}
case CompiledRegex::Split_PrioritizeParent: case CompiledRegex::Split_PrioritizeParent:
{ {
auto new_thread_inst = prog_start + *reinterpret_cast<const CompiledRegex::Offset*>(thread.inst); auto parent = thread.inst + sizeof(CompiledRegex::Offset);
thread.inst += sizeof(CompiledRegex::Offset); auto child = prog_start + *reinterpret_cast<const CompiledRegex::Offset*>(thread.inst);
add_thread(thread_index+1, new_thread_inst, thread.saves); thread.inst = parent;
++thread.saves->refcount;
m_threads.insert(m_threads.begin() + thread_index + 1, {child, thread.saves});
break; break;
} }
case CompiledRegex::Split_PrioritizeChild: case CompiledRegex::Split_PrioritizeChild:
{ {
auto new_thread_inst = thread.inst + sizeof(CompiledRegex::Offset); auto parent = thread.inst + sizeof(CompiledRegex::Offset);
thread.inst = prog_start + *reinterpret_cast<const CompiledRegex::Offset*>(thread.inst); auto child = prog_start + *reinterpret_cast<const CompiledRegex::Offset*>(thread.inst);
add_thread(thread_index+1, new_thread_inst, thread.saves); thread.inst = child;
++thread.saves->refcount;
m_threads.insert(m_threads.begin() + thread_index + 1, {parent, thread.saves});
break; break;
} }
case CompiledRegex::Save: case CompiledRegex::Save:
{ {
const char index = *thread.inst++; const char index = *thread.inst++;
thread.saves[index] = m_pos.base(); if (thread.saves->refcount > 1)
{
--thread.saves->refcount;
thread.saves = clone_saves(thread.saves);
}
thread.saves->pos[index] = m_pos.base();
break; break;
} }
case CompiledRegex::Matcher: case CompiledRegex::Matcher:
@ -194,8 +221,8 @@ struct ThreadedRegexVM
bool found_match = false; bool found_match = false;
m_threads.clear(); m_threads.clear();
const auto start_offset = (flags & RegexExecFlags::Search) ? 0 : CompiledRegex::search_prefix_size; const auto start_offset = (flags & RegexExecFlags::Search) ? 0 : CompiledRegex::search_prefix_size;
add_thread(0, m_program.bytecode.data() + start_offset, m_saves.push_back(std::make_unique<Saves>(Saves{1, Vector<Iterator>(m_program.save_count, Iterator{})}));
Vector<Iterator>(m_program.save_count, Iterator{})); m_threads.push_back({m_program.bytecode.data() + start_offset, m_saves.back().get()});
m_begin = begin; m_begin = begin;
m_end = end; m_end = end;
@ -204,9 +231,14 @@ struct ThreadedRegexVM
if (flags & RegexExecFlags::NotInitialNull and m_begin == m_end) if (flags & RegexExecFlags::NotInitialNull and m_begin == m_end)
return false; return false;
auto release_saves = [this](Saves* saves) {
if (--saves->refcount == 0)
m_free_saves.push_back(saves);
};
for (m_pos = Utf8It{m_begin, m_begin, m_end}; m_pos != m_end; ++m_pos) for (m_pos = Utf8It{m_begin, m_begin, m_end}; m_pos != m_end; ++m_pos)
{ {
for (int i = 0; i < m_threads.size(); ) for (int i = 0; i < m_threads.size(); ++i)
{ {
const auto res = step(i); const auto res = step(i);
if (res == StepResult::Matched) if (res == StepResult::Matched)
@ -214,11 +246,12 @@ struct ThreadedRegexVM
if (not (flags & RegexExecFlags::Search) or // We are not at end, this is not a full match if (not (flags & RegexExecFlags::Search) or // We are not at end, this is not a full match
(flags & RegexExecFlags::NotInitialNull and m_pos == m_begin)) (flags & RegexExecFlags::NotInitialNull and m_pos == m_begin))
{ {
m_threads.erase(m_threads.begin() + i); m_threads[i].inst = nullptr;
release_saves(m_threads[i].saves);
continue; continue;
} }
m_captures = std::move(m_threads[i].saves); m_captures = std::move(m_threads[i].saves->pos);
if (flags & RegexExecFlags::AnyMatch) if (flags & RegexExecFlags::AnyMatch)
return true; return true;
@ -226,17 +259,25 @@ struct ThreadedRegexVM
m_threads.resize(i); // remove this and lower priority threads m_threads.resize(i); // remove this and lower priority threads
} }
else if (res == StepResult::Failed) else if (res == StepResult::Failed)
m_threads.erase(m_threads.begin() + i); {
m_threads[i].inst = nullptr;
release_saves(m_threads[i].saves);
}
else else
{ {
auto it = m_threads.begin() + i; auto it = m_threads.begin() + i;
if (std::find_if(m_threads.begin(), it, [inst = it->inst](auto& t) if (std::find_if(m_threads.begin(), it, [inst = it->inst](auto& t)
{ return t.inst == inst; }) != it) { return t.inst == inst; }) != it)
m_threads.erase(it); {
else m_threads[i].inst = nullptr;
++i; release_saves(m_threads[i].saves);
} }
} }
}
// Remove dead threads
m_threads.erase(std::remove_if(m_threads.begin(), m_threads.end(),
[](auto& t) { return t.inst == nullptr; }),
m_threads.end());
// we should never have more than one thread on the same instruction // we should never have more than one thread on the same instruction
kak_assert(m_threads.size() <= m_program.bytecode.size()); kak_assert(m_threads.size() <= m_program.bytecode.size());
if (m_threads.empty()) if (m_threads.empty())
@ -250,21 +291,13 @@ struct ThreadedRegexVM
{ {
if (step(i) == StepResult::Matched) if (step(i) == StepResult::Matched)
{ {
m_captures = std::move(m_threads[i].saves); m_captures = std::move(m_threads[i].saves->pos);
return true; return true;
} }
} }
return false; return false;
} }
void add_thread(int index, const char* inst, Vector<Iterator> saves)
{
if (std::find_if(m_threads.begin(), m_threads.end(),
[inst](const Thread& t) { return t.inst == inst; }) == m_threads.end())
m_threads.insert(m_threads.begin() + index, {inst, std::move(saves)});
kak_assert(m_threads.size() < m_program.bytecode.size());
}
bool is_line_start() const bool is_line_start() const
{ {
return (m_pos == m_begin and not (m_flags & RegexExecFlags::NotBeginOfLine)) or return (m_pos == m_begin and not (m_flags & RegexExecFlags::NotBeginOfLine)) or
@ -294,6 +327,9 @@ struct ThreadedRegexVM
Utf8It m_pos; Utf8It m_pos;
RegexExecFlags m_flags; RegexExecFlags m_flags;
Vector<std::unique_ptr<Saves>> m_saves;
Vector<Saves*> m_free_saves;
Vector<Iterator> m_captures; Vector<Iterator> m_captures;
}; };